State of the Art: Training >70B LLMs on 10,000 H100 clusters
It’s return guest season here at Latent Space! We last talked to Kanjun in October and Jonathan in May (and December post Databricks acquisition):
Imbue and Databricks are back for a rare treat: a double-header interview talking about DBRX from Databricks and Imbue 70B, a new internal LLM that “outperforms GPT-4o” zero-shot on a range of reasoning and coding-related benchmarks and datasets, while using 7x less data than Llama 3 70B.
While Imbue, being an agents company rather than a model provider, are not releasing their models today, they are releasing almost everything else:
* Cleaned-up and extended versions of 11 of the most popular NLP reasoning benchmarks
* An entirely new code-focused reasoning benchmark
* A fine-tuned 70B model, built with Meta Llama 3, to identify ambiguity
* A new dataset of 450,000 human judgments about ambiguity
* Infrastructure scripts for bringing a cluster from bare metal to robust, high performance training
* Our cost-aware hyperparameter optimizer, CARBS, which automatically and systematically fine-tunes all hyperparameters to derive optimum performance for models of any size
As well as EXTREMELY detailed posts on the infrastructure needs, hyperparameter search, and clean versions of the sorry state of industry standard benchmarks. This means for the FIRST TIME (perhaps since Meta’s OPT-175B in 2022?) you have this level of educational detail into the hardware and ML nitty gritty of training extremely large LLMs, and if you are in fact training LLMs of this scale you now have evals, optimizers, scripts, and human data/benchmarks you can use to move the industry forward together with Imbue.
We are busy running the sold-out AI Engineer World’s Fair today, and so are unable to do our usual quality writeup, however, please enjoy our show notes and the excellent conversation! Thanks also to Kanjun, Ashley, Tom and the rest of team Imbue for setting up this interview behind the scenes.
Video pod
Timestamps
* [00:00:00] Introduction and catch up with guests
* [00:01:55] Databricks' text to image model release
* [00:03:46] Details about the DBRX model
* [00:05:26] Imbue's infrastructure, evaluation, and hyperparameter optimizer releases
* [00:09:18] Challenges of training foundation models and getting infrastructure to work
* [00:12:03] Details of Imbue's cluster setup
* [00:18:53] Process of bringing machines online and common failures
* [00:22:52] Health checks and monitoring for the cluster
* [00:25:06] Typical timelines and team composition for setting up a cluster
* [00:27:24] Monitoring GPU utilization and performance
* [00:29:39] Open source tools and libraries used
* [00:32:33] Reproducibility and portability of cluster setup
* [00:35:57] Infrastructure changes needed for different model architectures
* [00:40:49] Imbue's focus on text-only models for coding and reasoning
* [00:42:26] CARBS hyperparameter tuner and cost-aware optimization
* [00:51:01] Emergence and CARBS
* [00:53:18] Evaluation datasets and reproducing them with high quality
* [00:58:40] Challenges of evaluating on more realistic tasks
* [01:06:01] Abstract reasoning benchmarks like ARC
* [01:10:13] Long context evaluation and needle-in-a-haystack tasks
* [01:13:50] Function calling and tool use evaluation
* [01:19:19] Imbue's future plans for coding and reasoning applications
* [01:20:14] Databricks' future plans for useful applications and upcoming blog posts
Transcript
SWYX [00:00:00]: Welcome to the Latent Space Podcast, another super special edition. Today, we have sort of like a two-header. John Frankel from Mosaic Databricks, or Databricks Mosaic, and Josh Albrecht from MBU. Welcome.
JOSH [00:00:12]: Hey, glad to be here.
SWYX [00:00:14]: Thank you for having us. Hey, so both of you are kind of past guests. Jonathan, you were actually one of the most popular episodes from last year talking about MPT7B. Remember the days when we trained large models and there was 7B?
JONATHAN [00:00:30]: Yeah, back when reproducing LLAMA1-7B was considered a huge accomplishment for the field. Those are the good old days. I miss that.
SWYX [00:00:38]: As the things have accelerated a lot. Actually, let's do a quick catch up and Josh, you can chime on in as well. So Databricks got acquired. I talked to you at New York.
JONATHAN [00:00:45]: Mosaic got acquired, although sometimes it feels like Mosaic acquired Databricks because, you know, we're having a lot of fun being here. But, you know, yeah.
SWYX [00:00:52]: Yeah. I mean, you are chief scientist now of Databricks.
JONATHAN [00:00:55]: Chief AI scientist. Careful with the title. As much as I would love to understand how Spark works, I'm going to have to defer that to much smarter people than me.
SWYX [00:01:03]: Got it. And I don't know about like what you would highlight so far as a post-acquisition, but the most recent news is that you guys released DBRX. Is that the thing that most people should be aware of?
JONATHAN [00:01:13]: Actually, that's no longer the most recent news. Honestly, the most recent news, we announced this, but it was at our Data and AI Summit last week. So it was announced among like 100,000 other things, is that we finally released our text to image model, which has been a year in the making through a collaboration directly with Shutterstock. There was a lot of work put into finding a dataset that we were comfortable with working on and trying to build a model that honestly, I felt like I could trust and that others might be able to trust to put out in the world. So that model was released last week. It's unfortunately just available via API due to the fact that the data is quite sensitive and quite valuable. It's Shutterstock's entire business in a lot of ways, but I'm still really excited that there's now a model that is trained on a dataset where the provenance of every single image is known, and it's a damn good model. So I'm really proud of the team on that.
SWYX [00:01:55]: Yeah, amazing. Josh, do you have any thoughts on image model questions?
JOSH [00:01:59]: That is not my area of expertise, but I was excited to see the release of it last week as well, and very happy that you guys did a nice job on the data side of everything there. So that was cool to see.
SWYX [00:02:09]: I think what's unusual is like, I think Shutterstock's doing multiple deals in multiple labs. So what is the Shutterstock model? Like, I guess, is this the house model for Shutterstock? Is this Databricks' version of the Shutterstock model? Like, what is this?
JONATHAN [00:02:22]: The way that I would think about it is that Shutterstock is doing an amazing business in AI across the board. Their dataset is kind of widely known to be the best stock photos dataset in the world, the most comprehensive, the biggest. When you think about like, what dataset am I going to train a multimodal model on? You call Shutterstock. And I, at least I've heard in the news, like OpenAI, Google, Meta, Apple have all called Shutterstock and made those deals. So a lot of models have had Shutterstock data incorporated into them. But this is the only model I know of so far where it was, you know, exclusively and specifically trained just on the vanilla Shutterstock data. There was nothing else mixed in. We didn't go and scrape the web and find other data or combined datasets or anything like that. And so this is, in some sense, the house blend. But the other piece is that it's just a dataset where the provenance of every image is known in public. Where did the data come from? It is the Shutterstock collection. That's it. You know, nothing less, nothing more. And certainly being at Databricks, if I've learned one thing, I've learned about enterprise customers and what they want out of AI. And one of the things they ask for most is just, what can you tell me about the data the model was trained on? And here, especially for text to image models, where images are just tricky subject matter, there's been a lot of kind of legal conversation about images, especially. It's nice to just have something where I can point to it and say, you know, if you want to know where the images came from, these are what they are and this is how they got there.
SWYX [00:03:36]: I will talk a little bit about Databricks because it's relevant to the rest of today's episode. So Databricks, sorry, I keep misspeaking. It's DBRX.
JONATHAN [00:03:46]: DBRX, actually, there's been a pronunciation update. It is now D-B-Rex. So we have decided to add a dinosaur mascot because what model doesn't like a mascot? So literally, I wish I could pull it up. There is a little plush dinosaur that we had made. It's like the world's cutest dinosaur, but it is the official mascot of D-B-Rex. And there's a little dinosaur logo that, you know, you'll probably see around a little bit more because DBRX is a mouthful, but D-B-Rex, like, you know, it's just kind of...
SWYX [00:04:13]: Rolls off the tongue. I love mascots. Like every company should have a mascot. And I think Hugging Face got it right. You need an emoji mascot because that's the minimal viable image.
JONATHAN [00:04:21]: I probably shouldn't talk at all about, you know, Velociraptor, but, you know, that's a, maybe that's something we can talk about later in the summer. I'll just leave it at that.
SWYX [00:04:28]: Okay. That's a hint to names. I feel like your names leak a lot of alpha. So just to quickly cover the headline details, DBRX, as Make Sure Experts model, that's fairly big, 132 billion total parameters, so 36 billion active on any input, pre-trained on 12 trillion tokens of text and code, and did really well on evals to the point where you had to dye your hair blue. That's my high level conclusion.
JONATHAN [00:04:53]: Never make a bet with your team two weeks out from model launch, even when, you know, human eval is looking quite bad. Because if you set some bar, even if it's arbitrary and you think there's no way in hell they're going to hit it, apparently money doesn't motivate people anymore. Humiliating their boss motivates people. So Josh, you should really take a hint from this. You know, you cannot pay someone enough money to make up for you dyeing your hair blue.
JOSH [00:05:15]: I'll keep that in mind for our next model.
SWYX [00:05:17]: It works. So speaking of Imbue's next model, perhaps Josh, you want to actually just say hi to the general sort of latent space audience and talk about what we're releasing today. Yeah.
JOSH [00:05:26]: I'm Josh, CTO of Imbue, and we're not releasing the model. We're not releasing the weights, but we are releasing a bunch of different things that should make it easier for other people to make their own models. So I think right now, training foundation models from scratch is like a very difficult, time-consuming, expensive, kind of risky endeavor, especially for smaller companies. And the things that we're releasing hopefully make that at least a little bit easier. So the things that we're releasing fall into kind of three different buckets. One is infrastructure and scripts for dealing with the kind of hardware and hardware failures and understanding how well is the actually lowest level of thing actually working so that you can actually do your training at all and at a reasonable speed without having to constantly restart, etc. So infrastructure and training scripts. A second set of things is around the evaluation. So after you've trained it, like how well is this actually working and how do you know how well it's working? We're releasing a whole bunch of different data there, a new benchmark about code, reasoning, understanding, as well as our own private versions of 11 different open source benchmarks. So things like pool queue or ANLI, where we've gone through and kind of cleaned up the data as much as possible by looking at all the ones that models get wrong or that are flagged for ambiguity and also our own kind of private reproductions of those where we've done like a kind of clean room black box, like, okay, this is what the data set is supposed to be. Here are some examples. Let's make our own version of this to make sure that there is no data contamination, etc. To make sure that we're actually, you know, not testing on train. And then I think a final thing that we're releasing there is around 450,000 human judgments about ambiguity and question quality, which we used in the process of cleaning these evaluations and we also hope will be helpful for other people training kind of similar models. And then the third thing is CARBS, our hyperparameter, our cost-aware hyperparameter optimizer, which was especially helpful for being able to experiment at much smaller scales and then scale those experiments up to the much larger scale kind of on the first try without having to retry it. You don't want to be training, you know, 10, 20 different 70B models. You really want to get these larger models
SWYX [00:07:30]: right on the first try.
JOSH [00:07:30]: And so the ability to kind of tune things very precisely and learn scaling laws, not just for, you know, the like data and flops, but also for learning rate and all the other hyperparameters and see like how should you scale these things up was extremely valuable to us as we were training the larger models. Yeah, that's a lot of stuff.
SWYX [00:07:49]: Yeah, exactly. So there's a bunch of stuff
JOSH [00:07:50]: we'll have to go through all of it.
JONATHAN [00:07:52]: Yeah, I just want to throw in how excited I am about this. This is the stuff that nobody ever talks about. That is the difference between success and failure in this stuff. Like, can you get your cluster to run? Can you get software on your cluster? Can you figure out what broke? Because fault tolerance is still not really built into any of the fundamental primitives of training models. And so if something breaks, you have to go figure out what broke, your job stops, you have to restart your job. It is a nightmare just to get to the point where anything can train on the cluster. A basic MPI hello world that has the GPUs talk to each other is hard enough, let alone actually training a model, let alone getting good performance out of the GPUs, let alone actually getting a model that converges to anything interesting. There's so many levels of things you have to accomplish. This is the kind of stuff that matters. I think to a point that Josh made earlier, before we got on here, there are plenty of weights out there. Nobody's released this.
JOSH [00:08:46]: Yeah, that was part of the motivation actually is that there are lots of other things that are complimentary, but I have not seen nearly as much discussion about some of these other things that we think are pretty important. I mean, in some sense,
SWYX [00:08:56]: I'm very excited to have Jonathan on because this is a little bit, you're a bread and butter with Mosaic. And I think you've released some part with Composer. And I think it's just really interesting to see like a different take, basically a full stack take that's kind of open source today.
JONATHAN [00:09:18]: Yeah, it's really kind of, it's been an ordeal to figure this out. And every time something changes, whether it's a new GPU or even a new driver update, you get new creative errors and new things go wrong. And, you know, we've dealt with the weirdest things from, you know, our InfiniBand cables getting stolen from the data center twice, like in boxes before they arrived at the data center. Like, you know, Porch Pirate basically had stolen our InfiniBand cables back when those were hard to come by. To like, you know, weird recalls of switches to like the strangest stuff has happened. I have my favorite GPU failures I've seen, like ones where the GPU doesn't fail, it has a correctable memory issue and the memory correction causes the GPU to become a straggler and hold up the whole job. Like weird stuff happens and figuring out how to not just identify all of that, but then eventually productize it, is in some sense, the entire story of Mosaic and now Databricks in terms of our ML offering. Really, the thing we offer is we have gone through this suffering and figured out how to even productize that. It has been a pain in the butt.
SWYX [00:10:20]: Yeah, it's a lot of work.
JOSH [00:10:20]: I think my favorite failure was GPU is just giving wrong math. Like if they give errors, great, because you can see the errors, but if they just give you the wrong math back, not so fun.
SWYX [00:10:30]: When did they give you wrong math?
JOSH [00:10:32]: Like literally you could just, you know, add two things. For example, the numbers come back. They're not the numbers that they're supposed to be.
JONATHAN [00:10:40]: I think it's important to say at this stage, just because like it, I think it goes without saying for Josh and I, but it's worth saying here, this isn't to say that like anything is wrong with us. It's not like NVIDIA did a bad job or, you know, Mellanox did a bad job or the like the server builder, the data center operator, the cloud provider, like the million other parties that are involved in building this. We are running these insane chips that are huge and complicated and built on tiny transistors at insane frequencies with insane heat in data centers that for the most part, were not built remotely for this kind of power or heat and have been retrofitted for this. Like failures happen on a good day with normal CPUs. And this is not a good day and not a normal CPU for the most part. It's fun to joke about all the weird things we see. This is not to say anybody's done anything wrong. This is just kind of part and parcel of working on a massive cluster running at multiple megawatts of power at a time.
SWYX [00:11:32]: It's crazy. Yeah.
JONATHAN [00:11:33]: So optical cables, like all sorts, like everything.
SWYX [00:11:37]: I'll take the opportunity to start going to the sort of infra piece. There's just like a description of the infra just to give people a sense of what we talk about when we talk about massive clusters. So I'm just going to read off the blog post here. This post is about one cluster that has 4,092 H100 GPUs spread across 511 computers. They use unified fabric manager nodes, which manage the infinite band network. And you talk a little bit about your networking. Is there anything unusual about this setup that you'll call out to people?
JOSH [00:12:03]: Yeah, actually this particular cluster is a little bit non-standard. The normal, like vanilla setup for these large clusters as vanilla as it can be is what's normally like a 127 node cluster. So closer to like 1024 GPUs instead of 4,000. Here we have a larger cluster. As you start to get into the larger clusters, the networking becomes a little bit more custom. It's a little bit more, it's a little bit trickier. It's a little bit more difficult to get these things to all be able to talk to each other at the same speed. And so this has, in this particular case, this is a three tier network architecture instead of two tiers, kind of the normal one. So most of the clusters are a little bit smaller. As you get to even larger scales, then this becomes even much more complicated,
SWYX [00:12:43]: much more expensive.
JOSH [00:12:43]: So we chose this particular scale, kind of knowing our own workloads and kind of what we wanted to do. This was kind of the right size for us. But yeah, I think it's not exactly vanilla already. It's already getting into kind of the custom territory.
SWYX [00:12:54]: So my understanding is that there, and is there any part of this that comes with the Voltage Park deal that you guys had? Is that part of the hardware that you got from the deal with them?
JOSH [00:13:04]: Yeah, so we worked really closely with Voltage Park to set up all their clusters and infrastructure and everything and kind of decide even like what to order, how should the networking work? Like we were very involved in kind of the construction and bring up of this. And that's what this post is about, is about that process of like bringing up all these, there's like different clusters in different places of different scales. So in this particular post, we're talking about this one 4096 GPU, but there are other clusters that they have as well. And we were very closely involved with figuring out the exact architecture and kind of the trade-offs that go along with picking, you know, those exact components. You really don't want to like place the wrong order because it takes months to get it and it's very expensive. So yeah, we were happy to help out with that.
JONATHAN [00:13:43]: And then your bit of good cables get stolen.
SWYX [00:13:44]: Yeah, yeah, exactly.
JOSH [00:13:47]: We wanted to make sure that we ended up with compute that would work for us and that would also work for their other customers. And so we kind of helped design something so that we would get exactly what we were looking for. We knew that these kinds of details would be super important and that getting down to the level of the hardware and like having these good scripts and everything was going to be a core part of like actually getting this to work. I'm very glad that we did that. I don't think that most companies kind of take that full stack approach, but for us, it certainly paid off.
SWYX [00:14:12]: Yeah, it's basically sort of built to spec. It's interesting that relationship because you usually, for the rest of us who don't operate at your scale, we take whatever we can get from cloud providers, but you are basically co-designing from the single machine up. And you described that a little bit. Do you want to take us through the process that you described here?
JOSH [00:14:27]: Yeah, so for the actual, like the blog post and kind of bringing these machines online.
SWYX [00:14:32]: Yeah.
JOSH [00:14:32]: So yeah, I think the process, as we have it broken down in the blog post, there's kind of a few different layers. First is like getting the individual machines to work at all and then getting the machines to actually be able to talk to each other. So getting the InfiniBand networking to work and then getting to a point where, you know, not just the machines are working and they can talk to each other, but everything is actually working correctly. There's a big gap between like it's working at all to it's working perfectly correctly. And then after you have all this stuff working perfectly correctly, nice and healthy, then now you get into kind of the software data, like training issues. And then after that, you're still not done. Like now, even once you're training at full speed, things are going to fail over time. Things are going to change. There's going to be new, you know, firmware updates. Like how do you kind of deal with this change and flux over time without going crazy
SWYX [00:15:16]: and pulling your hair out,
JOSH [00:15:16]: trying to like reproduce things or understand why there were regressions. And so there's a lot of work to kind of automate the infrastructure tooling as well. And kind of the first step, like bringing these things online in the first place, you know, you have hundreds of machines at this point. So you don't necessarily want to be like walking around with like a CD-ROM or a USB drive, like plugging it in with your keyboard, like hitting next, next, next on the OS install. That's not how this works. You do that for one machine. And then you use, we use this thing called Metal as a Service to bring up all the other machines. So it's a kind of server that can kind of install the operating system on these other machines. So most like when you're talking about these machines, like each machine is, you know, on the order of hundreds of thousands of dollars. So they usually come with a kind of out-of-band management interface as well. So they don't, they have their InfiniBand networking. They have their normal 100 gigabit per second Ethernet networking. These are like dual, redundant, et cetera. And then you also have this extra out-of-band management network. So you can log in and you can see like the boot screen or you can see the blue screen of death. You can like get in there and actually see what was wrong, which is pretty fun. And it makes it like possible to automate a lot of this work. So the beginning of that, and the blog post goes into much more detail about like exactly how we set these up and kind of the other errors that we ran into. When you're bringing these online, you'll definitely have failures. Even if they all worked in the factory, they get shipped, some parts come loose, something fails, something goes wrong. So when you're bringing them online, there'll be some that don't quite work for all sorts of reasons. As you start to be working with machines at this scale, like if something happens one in a thousand times, you're like pretty likely to see it. And so you can get pretty rare, weird things, especially since we had fairly early builds and fairly early versions of this hardware. Like these are some of the like first machines that were ever produced, some of the first GPUs. So you've got some extra special things there. We definitely worked with Dell, for example, on making fixes in the firmware level to be like, okay, like this thing is wrong. Like we need to update this at the firmware to like actually fix this particular thing. So we worked pretty closely with Dell and Nvidia. Yeah, that's what I'm saying. Like this stuff gets complicated. And the thing is like, you know, taking a step back, the whole reason we're doing this, right, is that we knew that this was going to be complicated. There would be these kinds of failures. And if we're just using, you know, AWS or some other cloud provider, these errors are still gonna be there and you're gonna have no way to know and no way to debug this and no way to diagnose what's going wrong. And so we would much rather be able to like call up Dell and say, hey, this isn't working. And they're like, yep, okay, cool. Let's debug it together. Oh, I see. Yeah, cool. We'll ship a firmware update and actually fix this for you. That was a much better experience than like, great, just magically fails. I guess we restart and hope that that machine goes away. Like that's not a very good place to be. So yeah, that's kind of the first place is getting to a place where like GPU training is working on your single node machines. You can observe stuff. We have tons of tooling around like, you know, Prometheus and all sorts of other tools for understanding what's going on in these machines because you don't want to be like logging into each one and looking at the temperature or something you really need to have tooling to collect all these metrics, et cetera. Unfortunately, all of the scripts that we have for this are like for this entire cluster and for all this infrastructure are a little bit like special purpose for our particular thing. So it's not that every script that we have, it's not that you can just like take this and plug this in. Even if we did open source all the tooling that we have, you'd still have to do like a lot of work to open source it. What we are releasing is as many of the things that we can that are going to be useful for other people. You're still going to have to have some way of kind of managing these things, making your own like logging aggregators, et cetera, et cetera. So that's kind of bringing them up to the like, you know, the single nodes that are working. From there, it goes into, I'm happy to keep going if you want. Well, I just want to leave the opportunity for John
SWYX [00:18:53]: to comment if there's anything that's different from how he runs things.
JONATHAN [00:18:57]: Oh, I mean, all I'll say is I'll endorse this and say this s**t is hard. Like this is really, really hard. And, you know, I have a special props to, you know, the folks in Vue because they were building this from the ground up. You know, at Databricks and at Mosaic, we typically work with cloud providers because some of this stuff is just, there's too much to handle. It's complicated. There's a lot to deal with. And this doesn't even get into things like physical security, you know, securing power if you're the data center operator. Like this gets infinitely complicated and you have to abstract somewhere. Like, you know, and then you get to the folks who are literally building their own custom chips and like, good God.
SWYX [00:19:36]: Like, oh my God, that's, you know,
JONATHAN [00:19:38]: if you're one of those folks, you're having, you know, pour one out for the infra people at some of the AI chip startups who are having a really, really interesting time right now. But this stuff is really hard. And I don't think we talk about it much because there's so many other things that are hard. But the other hard things, I think everybody's becoming pretty familiar with at this point. This is something that I don't think there's ever really been a comprehensive discussion of, at least not that I've seen.
SWYX [00:20:00]: Yeah, so my impression is that you guys, Mosaic, have your own software for sort of spinning up and down machines, just like Imbue had to build. But Imbue probably, it sounds like Imbue, you guys went fuller stack. I don't know how to describe it. Like Mosaic is not working with Dell on like their firmware.
JONATHAN [00:20:21]: No, no, we're typically working with like, you know, pick your cloud provider on their Dell firmware or what have you. Like, it's kind of, I think one of the things, I don't know, Josh, you can correct me on this. It's kind of impossible if you're doing training to not go all the way through the entire stack, regardless of what happens. Like somehow I'm still chatting with cloud providers about power contracts, even though the whole point of dealing with the cloud provider is not to have to think about power contracts. Somehow I'm still asking them about which InfiniBand provider they used this time to see if this is part of the bad batch of cables I encountered on that cloud provider or what have you. Or like, we're still talking about a firmware update from pick your provider. You can't not do this. It's convenient that they have data center staff who are worrying about what to send back to which provider when, and they have people who can go and wait for the InfiniBand cables so they don't get stolen outside. But, you know, it's kind of, it's impossible not to really go full stack if you're thinking about the infrastructure at all. I don't know, Josh, correct me. No, I think that's right.
JOSH [00:21:17]: That's what we expected from the beginning as well, is that we would inevitably have to get into the details here. And I'm glad that we kind of just planned for it. I think it made it a lot easier from our perspective to have direct control over this. Instead of having to go to the cloud provider that goes to the data center, that goes to the supplier, we could just go direct to NVIDIA or Dell
SWYX [00:21:37]: or the data center,
JOSH [00:21:37]: whoever was responsible and be like, hey, this thing needs to change. And they're like, oh, okay. Yeah, that is our responsibility. Great, we can fix that. So it was just a lot easier for us to fix these bugs than if we had to go through an extra layer of email.
SWYX [00:21:48]: Something we discussed in the pre-show was that you had a rule of thumb for your cluster of reliability. You say here in the post, by and large, you expect around 3% of your machines to break every week. So you're basically going to turn through all your machines in a year.
JOSH [00:22:04]: As it says in the post. So that would be true if it was a uniform failure like that. But as it says in the post, it's usually these kind of problematic nodes. And to be clear, that is the number that we've heard from other people is like they're having about 3%. I don't think we're experiencing failure rates that are that high. I think ours is actually quite a bit lower than that, probably because we've taken the time to like dig into a large, maybe larger number than we should have of these failures and get to the root cause of it and be like, oh, okay, like that's exactly what's going wrong.
SWYX [00:22:33]: How do we fix this?
JOSH [00:22:33]: How do we prevent this from happening? How do we make automated checks for this so that if it does happen, it just goes back to whoever owns that particular part of the process and they can fix it immediately.
SWYX [00:22:43]: And that's part of what you're also open sourcing, which is the health checks, right? You got the NIC health checks, GPU health check, this space health check, Docker D message. I don't know what that is.
JOSH [00:22:52]: That one is just a lot of stuff.
SWYX [00:22:54]: Yeah.
JOSH [00:22:55]: That one is one where we realized that actually like when these machines boot, sometimes they wouldn't actually boot cleanly all the way. Or when they rebooted, they had problems that they didn't have when they were working before, which was kind of frustrating. Like usually if you restart your computer,
SWYX [00:23:08]: it gets better.
JOSH [00:23:08]: Here you restart. It did not get better.
SWYX [00:23:10]: It got worse.
JOSH [00:23:10]: That was very frustrating. So this health check looks at every particular line we've ever seen from the boot, like in D message, like every single log line that your computer emits
SWYX [00:23:21]: and says like,
JOSH [00:23:21]: have we ever seen this before?
SWYX [00:23:23]: Is this expected?
JOSH [00:23:23]: Is this in the right order? Or is there something out of place? If there's anything out of place, let me say, okay, great. Like now it goes into this, like longer, more triage list of like, all right, great. Like, is this acceptable?
SWYX [00:23:33]: Should we flag this?
JOSH [00:23:33]: Like, should someone take a look at this? So we're looking down at a very, very granular detail level, what's happening on these computers to make sure that nothing is out of place. And that's critical because without that, if you're running your training, as Jonathan said, and this thing is slow, like what are you supposed to do? Right?
SWYX [00:23:49]: Like you really,
JOSH [00:23:49]: you really want to be very certain that like all 4,000 of these GPUs are working like they're supposed to.
SWYX [00:23:54]: We know that.
JOSH [00:23:54]: And so if it's slow, it's because like we messed up the config or something else and not because of this earlier thing that's like really hard to detect in software later.
JONATHAN [00:24:01]: Yeah. I think the, I'm just curious to ask,
SWYX [00:24:03]: like, you know,
JONATHAN [00:24:03]: suppose you were to set up another, let's say another H100 cluster and it were at a different data center. And instead of the vendor being Dell, it was super micro or what have you. How much of this would be repeatable? And how much of this would you have to redo? I, you know, I genuinely don't know.
SWYX [00:24:18]: A decent amount.
JOSH [00:24:19]: I think it would go a lot faster the second time. I think there's lots of learnings that we had. And also the blog post,
SWYX [00:24:24]: you know, yes,
JOSH [00:24:24]: we are releasing the health checks, releasing some scripts, but a lot of the valuable stuff is also in the blog post itself, in the details and kind of the, you know, the learnings that we've had and the sort of errors that we run into. We tried to as much as possible surface those to other people
SWYX [00:24:36]: could learn from those
JOSH [00:24:36]: and avoid the same mistakes or failures as well. But I think it would go a lot faster.
SWYX [00:24:41]: Although, yes,
JOSH [00:24:41]: there would certainly be some things that'd be a little bit different. I mean, there'd probably be different CPUs
SWYX [00:24:46]: or whatever,
JOSH [00:24:46]: but I think a lot of that stuff is less,
SWYX [00:24:49]: it's less,
JOSH [00:24:49]: that's the like, that's less variable. I think most of it would apply the second time around. Although I'm sure next time
SWYX [00:24:56]: we're building one,
JOSH [00:24:56]: it'll probably be, you know, at a scale that's 10x as big with a different chip or something like this.
SWYX [00:25:00]: And then who knows?
JOSH [00:25:01]: Yeah, with Kinect X8,
JONATHAN [00:25:02]: that will have its own fun behavior and all that good stuff. Yeah.
SWYX [00:25:06]: Perhaps there's something that people don't discuss about, and you don't even talk about this in the blog, but I always wonder is what is the timeline that's like kind of reasonable for this amount of work, at least the initial stages? And also what does the team composition look like for setting up a cluster, right? Like what are the mix of skills that you typically would require to get all this going?
JOSH [00:25:27]: I'm, I can't really speak to typical. One thing I am very proud of is how much we accomplished with such a ridiculously small team. Like our infrastructure team is like, you know, fluctuates from week to week, depending on like how many things are on fire and how much we need to build. But it's like between like three and six people, like it's small. It's not like some huge team of like tons and tons of engineers. But those people are very, very good at what they do. And so that has allowed us to get a lot of mileage out of out of these things. I think it's not that we're building everything, right? It's not that three to six people build this whole thing. I definitely want to like, you know, say thanks very much to Dell and H5 and NVIDIA and the other people that have done a lot of the work, like to bring up this cluster, you know, with 4000 GPUs and three tier networking, networking architecture, you have 12,000 cables. So that's 24,000 things that need to be plugged in. Like that's just a lot of stuff to plug in, right? And you don't want to mess it up. Like each one needs to be done correctly. Like it's a little bit loose. Like it doesn't really work.
SWYX [00:26:23]: If you break it,
JOSH [00:26:23]: you need to replace it. Like there's a lot of work
SWYX [00:26:26]: that goes into this.
JOSH [00:26:27]: Yeah.
SWYX [00:26:28]: And then, you know,
JOSH [00:26:28]: that's just like that's it. That's if you were to do everything right the first time.
SWYX [00:26:32]: And if you didn't
JOSH [00:26:32]: have to fix anything. But inevitably, you know, you will have to replace something, which means like taking all the wires out, pulling the thing out, taking all the GPUs out, going and fixing some cable, putting it all back correctly, putting it back in, doing this every time. So there were a lot of people at Dell, NVIDIA and at H5 that all helped a ton with this stuff. I don't know the exact size of the Dell team. It also fluctuated over time.
SWYX [00:26:55]: Yeah, excellent. And then, you know, you so you have all the hardware set up and now you're firing it up for a single node. There's a long description that you guys have about just like monitoring the MFU, right? And what each situation might look might be indicative of. One of the most interesting things to me that I saw from here is like, you know, if training immediately starts off at 60 to 80% MFU, something's wrong.
SWYX [00:27:24]: But like, you know, like what what are like, you know, some anecdotes or, you know, notable scenarios here that you might you might call out as maybe counterintuitive or super interesting.
JOSH [00:27:36]: There's just so many of them. I mean, one of them, which I think is probably pretty common, like common knowledge by this point. But like we did have a sort of like
SWYX [00:27:46]: which one was this exactly?
JOSH [00:27:47]: I think for the MFU, like gradually getting worse over time. I think that one, when we saw that the first time we were like, what the heck is going on? Like, why does it get just like a little bit worse? This is so strange. Like, what is it getting lazy or tired or something? Like, is it heat? Like what's going on? And in this particular case, it was memory fragmentation. Because you have hundreds of machines, they're doing garbage collection slightly different times. And then they get slightly further apart and slightly more and more jittered until eventually they're all happening kind of at random times. And just like really messing up each one of your steps. So you just turn off garbage collection and call it a day, basically,
SWYX [00:28:20]: to be honest.
JOSH [00:28:20]: There's other things you can do if you want to be a little bit more sophisticated about it. But you can also just manually
JONATHAN [00:28:25]: have it all garbage collect on some interval. Like that's what we've done. We just have a garbage collection callback that just runs. But I've seen the exact same thing.
JOSH [00:28:33]: Yeah, yeah, exactly. So I thought that one was kind of funny. And we did trace that one down and look and we did find the actual call. Like, again, this goes to like having good tools. So we had really good tools where we could look at a bunch of like actual traces in C and be like, OK, cool. This is the thing that's taking a lot of time. Or like, you know, this is the thing that doesn't quite line up here. Like, oh, I guess it's garbage collection. OK, cool.
SWYX [00:28:52]: Interesting.
JOSH [00:28:52]: Yeah, let's just try taking it off.
SWYX [00:28:54]: OK, great.
JOSH [00:28:54]: That's what it was. Now we can fix it. So for each of them, like basically bugs are not hard if you have good tools. But if you don't have good tools, bugs can be very, very hard. So similarly for like heat, another thing that we saw was like, oh, you know, the CPU is getting throttled. OK, well, it's easy to see if you're monitoring the CPU throttling or monitoring the heat. If you're not monitoring that, it's really hard to know why it's just suddenly one of them is going slower. I noticed also in the piece
SWYX [00:29:17]: that you mentioned FSDP with 0.3. Actually, we met, I went to iClear and Guanhua from the DSP team was there presenting 0++. I was wondering if you want to make any call outs to, you know, particular open source or open library or open whatever implementation teams that were super helpful in your process. I think we ended up actually
JOSH [00:29:39]: pulling from a whole bunch of different ones to pull things in into our own particular pipeline. So we use things from NVIDIA's, you know, Megatron stuff. We use stuff from probably DeepSpeed. I think we pulled in a bunch of different pieces from a bunch of different places. So it was really nice to see all these working open source like examples. I think I really appreciate all the effort that has gone into actually tuning these things because you can tune them, but it's a lot of work to like tune this stuff and do all this stuff from scratch. It's really nice to have like a working example. I think those are probably the two biggest ones, DeepSpeed and Megatron alone, but there are probably other ones as well.
SWYX [00:30:13]: Is there a particular thing in the ecosystem where you would call out as like, you know, there should be something here that is open source, but like it's not really, it's like everyone kind of builds it on their own. I want to say something with the file system because everyone talks about the file system eventually.
JOSH [00:30:28]: The file system actually was,
SWYX [00:30:30]: I mean, we did something
JOSH [00:30:31]: kind of dumb there. Like we have our own sort of local mirror so that we can, you know, like a crappy version of S3
SWYX [00:30:38]: that's local,
JOSH [00:30:38]: but it's just a pretty simple script, right?
SWYX [00:30:41]: Like I think we run like
JOSH [00:30:41]: a little web server that just like serves files and then, you know, it can upload them
SWYX [00:30:45]: and download them.
JOSH [00:30:45]: Okay, great. And part of the reason we did that is that our internet connection
SWYX [00:30:50]: in the beginning
JOSH [00:30:50]: was not the like full speed
SWYX [00:30:52]: one that we would
JOSH [00:30:52]: eventually have. And so we are a little bit more kind of bottlenecked in terms of internet bandwidth. And so we had this. I think we looked at a bunch of services out there like Minio and some other ones, but a lot of these like come with a lot of extra overhead and maintenance. And since we already have so much infrastructure
SWYX [00:31:09]: to deal with,
JOSH [00:31:09]: we kind of didn't want to, you know, bring in a whole other like cloud provider, virtualize something, something.
SWYX [00:31:14]: We just wanted something simple.
JOSH [00:31:14]: So we went with that, which has been quite helpful. Like our tools
SWYX [00:31:19]: are usually quite simple.
JOSH [00:31:19]: It's like Bash and Python and SSH and Docker. Like we'd like to keep things simple so that's easier to debug, like less layers of infrastructure, less layers of abstraction, make it a lot easier to work with. Like we don't use Kubernetes,
SWYX [00:31:30]: for example,
JOSH [00:31:30]: and we just directly launch these things. And it's just been much easier to debug this way. One tool actually that does come into mind that I will call out is Kraken from Uber. That was great. We love that tool. We were a little bit skeptical. What is it?
SWYX [00:31:44]: I'm sorry. Yeah.
JOSH [00:31:45]: So Kraken is this, yeah, it's a distributed like Docker registry, basically, that uses BitTorrent to like transfer things between the machines in a sort of nice optimal way. Like in the very beginning, the naive way is like you have this one Docker registry, which was outside of the cluster. So every time we change an image, you know, there's many gigabytes that each of the 500 machines needs to download.
SWYX [00:32:07]: So that just takes
JOSH [00:32:07]: a really long time. So what this thing does is like just one of them downloads it and then like they all sort of broadcast all the pieces to each other. And it was just like a really nice, fast way of getting these images down. And it was very robust.
SWYX [00:32:19]: Like there's a lot
JOSH [00:32:19]: going on under the hood, but I think it's a pretty cool tool that we haven't really had any bugs with it at all. Amazing.
SWYX [00:32:26]: Yeah. I mean, that's all my questions, I guess, for the info piece. I don't know if, John, you had something that you were sort of burning to ask or.
JONATHAN [00:32:33]: No, all I can say is just same
SWYX [00:32:36]: in a lot of places, like, you know, and they're done that
JONATHAN [00:32:38]: seeing this plus one. I think the one big difference, you know, perhaps in philosophies is we've tried to basically standardize on as much commodity stuff as possible, just because, you know, I think the reason I asked about trying to do this
SWYX [00:32:50]: on multiple different
JONATHAN [00:32:50]: pieces of infrastructure is like, I think we're running on like six or seven different clouds right now. And everybody has done something slightly different. And my gosh, the little differences add up as you know, you've seen. And so, you know,
SWYX [00:33:04]: our philosophy has been like, whatever the hell
JONATHAN [00:33:05]: we can standardize, please let's standardize it. Like vanilla off the shelf FSDB.
SWYX [00:33:10]: And like, you know,
JONATHAN [00:33:10]: we wrote our own data loader, but we've tried to make that as much of a standard as we can across our infrastructure and in Databricks, because things just start getting really complicated
SWYX [00:33:18]: or like we use
JONATHAN [00:33:18]: Kubernetes extensively because it at least gives us a uniform set of APIs. Like that's our hardware abstraction layer to a certain extent for everything else. So it's just, you know, a difference in philosophy there. But otherwise, like, yeah, this stuff is really, really hard. And I feel like we take for granted how much of this, you know, is done for us when you go and you just query chat GPT, for example. Like, oh my God, everything going on underneath that, you know, it's kind of a miracle that the machines boot up, let alone that you can like query a giant language model that's probably doing inference across multiple machines and was trained across thousands of machines. Like, you know, minor miracle.
SWYX [00:33:54]: Yeah, it is an awesome amount of power that we invoke with a single API call that we take for granted these days. It's absurd. Yeah, I mean, like Kubernetes, like that point about Kubernetes, I will say as a former AWS employee, like it seems like it would be ideal for imbue to at some point make it more abstracted or agnostic because you're going to want to, you know, replicate your setup. We do have our own
JOSH [00:34:19]: sort of replacement. It's just a much simpler version of Kubernetes. Kubernetes is really designed for running services, not for running experiments. Like that's not its like main architecture. And so for us, like we have everything that's like, cool, you're going to run an experiment. So you want it to run to completion, right?
SWYX [00:34:34]: OK, great.
JOSH [00:34:34]: Like the primitives are sort of built around a slightly different style. And that makes it a lot easier, like just a lot simpler to fit that the nature of like these machines are going to disappear. They will need to be rebooted for infrastructure upgrades. They will like something will happen to the GPUs. Failure is like baked into this as like a core part of our infrastructure. So it's not that we don't have an abstraction. It's that it's a sort of simpler, more tailored abstraction for the particular work that we're doing.
JONATHAN [00:34:58]: Yeah, I think it all depends on what your goals are. And like, I think the challenge in a lot of the deep learning stuff right now is that people are trying to like, people often build things that are more complicated than necessary to get the job done. And the complication is the enemy of everything. You know, don't use a fancier parallelism strategy than you have to. Don't use a fancier set of libraries than you have to.
SWYX [00:35:18]: Don't do anything
JONATHAN [00:35:18]: that you don't have to do because it's hard enough as it is. Like, don't overcomplicate
SWYX [00:35:23]: your own life.
JONATHAN [00:35:23]: Don't try to bring in more tools or more fancy architecture tweaks if you absolutely don't have to.
SWYX [00:35:29]: Like getting to the minimum
JONATHAN [00:35:30]: necessary to get the job done. And it's really tempting to want to try to use everything. So like, I totally understand that one.
SWYX [00:35:37]: I think the last piece I'll maybe call out is that I'm just going to weave this in just because I see the opportunity to do it. Are there any infrastructure shifts that need to be, that need to rise because of changing architecture? So I think, for example,
SWYX [00:35:57]: you're announcing a dense model, a 70B dense model, whereas John just worked on DBRX and the image-to-text model, which presumably has different bottlenecks.
JONATHAN [00:36:10]: That's correct for us. You know, we train both dense and mixture of expert models. The one we happened to, you know, kind of get permission to open source was a mixture of expert model. And those models are very demanding when it comes to network bandwidth, at least if you're training them in kind of FSTP 03 style, where there's just a lot of parameters getting shuffled back and forth. And your ratio of kind of compute to amount of data that you have to shuffle back and forth becomes a lot worse because you're now, you know, you're only using a fraction of the parameters for every token instead of all the parameters. And so we had to really push the envelope on getting all the stuff to the right places on time. And so actually the networking part of DBRX was the single hardest thing, I think, of the entire process. Just get MOE training, working at scale across a big cluster. We still managed to, I think, do it all with commodity parts, which was very exciting. You know, we were using FSTP and we eventually used HSTP so that we could have HSTP as a version of FSTP where you have multiple smaller replicas and you're doing data parallel within those replicas. And that helped a lot with network latency issues that we were running into just because we were transmitting so much data, you know, for every single part of the process. I think it actually, like, it was instructive for how Google designs their hardware and software together personally. Their training, as far as I understand, using kind of a 03 style of training and have been for a while. They also train mixture of expert models. TPUs have a very different network bandwidth to compute ratio. They have a lot more bandwidth just objectively. And TPUs per chip tend to be a little bit less compute intensive and have a little bit less memory. You know, it's just a different design choice. So the ratio of flops to bandwidth is very different. And that means that it's much easier for Google to be able to pull off
SWYX [00:37:54]: some of this stuff.
JONATHAN [00:37:54]: They also have interesting, you know, Torus style network architecture or Torus style, like, literal network architecture
SWYX [00:38:00]: is not like the model,
JONATHAN [00:38:00]: but the network.
SWYX [00:38:02]: Is this the sort of block attention? I forgot what you call it. So this is just more or the,
JONATHAN [00:38:07]: yeah, this is more, not the ring attention, but these are the ring all reduces. Like you have three different dimensions of rings because they kind of put you in these three dimensional Toruses from what I understand. And so like, you know, Google's infrastructure in some sense is kind of, I wouldn't say built for this, but maybe the way that Google trains models is built for a slightly different bit of infrastructure they have. And it's kind of neat to think about that. You know, as one thing that I think NVIDIA announced for, you know, for, for both the GH200 and the GB200 is this hybrid networking where you'll have blocks of NVLink network chips. I think for the GB200, I think it's like groups of 72 GPUs will all have NVLink to each other. So higher bandwidth, then you'll have normal networking of some kind, InfiniBand or Rocky or what have you between these blocks. And that's kind of a, you know, it's a change due to the fact that, you know, it's hard to build really high bandwidth networks over very large groups, but it is now a blocked networking. And you have to think about how you architect your model and your parallelism differently. You also have to think about fault tolerance differently because it now matters where you lose a GPU, whereas it didn't before. So, you know, it's, it's, it's just all really interesting and really fun speaking personally, but it's going to mean new nightmares when we all move to that generation and have to think about, you know, new versions of these problems.
JOSH [00:39:20]: As you go up to larger scales, it gets quite different. Like right now, you know, if you're experiencing, let's say, for example, you experience a GPU failure every day, that's fine.
SWYX [00:39:31]: Just restart.
JOSH [00:39:31]: If you make your thing 24 times as big, now it's once an hour. Now it stops being quite as easy to just restart, right? So now you have to kind of break, like bake in this sort of redundancy that you didn't have before. So I think as you go up in scale, you end up running into like a lot of really interesting problems that also inform the, the actual like design. Yeah, I mean, as an orchestration guy,
SWYX [00:39:52]: this is why I always emphasize like very cheap storage or very fast storage. So you can checkpoint more, but I don't think that's probably not the best solution to for fast, you know, training.
JONATHAN [00:40:05]: Which works fine when you're doing language and then you move to vision or video. And then, you know, you have multi petabyte datasets
SWYX [00:40:12]: and getting, you know,
JONATHAN [00:40:13]: cheap, fast multi petabyte storage starts to bite. Like I've certainly encountered issues where the literal data center where my GPUs were did not have enough, you know, object store to fit the datasets that people wanted to bring into that data center from whichever users were, were trying to bring them in. And then you get to a whole
SWYX [00:40:31]: different world of hurt
JONATHAN [00:40:31]: where you have to keep your data in a different region because the region is just out of storage. So things get fun really fast.
SWYX [00:40:39]: Speaking of vision, Josh, actually, you know, Embu is an agents company, but you're only, you're announcing a text-only model. What, where does, where does the vision side come in?
JOSH [00:40:49]: I think we've actually done a lot of work in the past and people can see kind of our blog posts about sort of self-supervised learning and some other kind of vision-related stuff in the past as well. So we're very familiar with, with that stuff. But I think our main focus right now is on kind of, as we say, coding and reasoning. And there, there's certainly a visual component to some problems. But, you know, it's not necessarily required for all problems. And actually we found that for most of the kind of like code writing and, and reasoning problems that we care about, the visual part isn't really a huge important part of it. Sometimes if you really need to, you can maybe describe
SWYX [00:41:24]: the thing.
JOSH [00:41:24]: There are other like, you know, multimodal models that you can use off the shelf to sort of plug in for those particular pieces
SWYX [00:41:30]: that you need, right?
JOSH [00:41:30]: Like if something is driving a browser or whatever, like you can sometimes get away with not having to have that baked into the original model. So our folk were, you know, in a sense, we kind of do a lot across the stack. We're working on our own infrastructure and pre-training and RL and fine tuning and products and everything. But in another sense, we're very narrowly focused on the application side. So all of the stuff across the stack is kind of going toward a very particular purpose. And so that particular purpose right now doesn't really need vision. So we think that people are going to make all sorts of really cool image models
SWYX [00:42:00]: like Jonathan, right?
JOSH [00:42:00]: And all sorts of interesting multimodal models into the future. We'll let them go do that. That's great. We'll take advantage of that, partner with those people in the future. And right now we're really focused on kind of the core reasoning and coding capabilities and aspects of the model.
SWYX [00:42:14]: I wanted to go into carbs since that's kind of the next layer of the stack. We talked about carbs in the first episode with Kanjin because you've actually had a blog post about it like a couple of years ago. Maybe let's introduce it.
JONATHAN [00:42:26]: Has that been a couple of years now?
JOSH [00:42:28]: No, it must have been at least one year. Hopefully it's not multiple years.
SWYX [00:42:32]: Sorry, I'm counting AI time. Yeah, yeah. Yeah, I was going to say
JONATHAN [00:42:35]: you're making me feel really old right now.
SWYX [00:42:39]: I count everything before the generally intelligent rename as like, you know, prehistory. Yeah. And now sort of modernity, right? So I actually thought carbs was more about hyperparameter optimization in a sense of like sort of parameters, hyperparameter search. Whereas, you know, when you introduced it, especially in this blog post, it's more about scaling laws and predictability of like, are we sort of in the right ballpark before we scale things up? Maybe sort of recount the history of carbs.
JOSH [00:43:10]: Yeah, so it really is a little bit of both. So carbs is, it's maybe a backronym, but it's for cost aware Pareto region Bayesian search. So this is about technically how it works, but carbs is like, you know, we like pastries and stuff.
SWYX [00:43:26]: So great, why not? But the point is that
JOSH [00:43:29]: it's a cost aware hyperparameter tuner. So most hyperparameter tuners, you kind of say, OK, here's this objective function. I want you to make this number as big as possible or as small as possible, whichever direction you want to go. So yeah, just go make this number, you know, as small as possible. OK, so it'll try a bunch of different
SWYX [00:43:46]: hyperparameters,
JOSH [00:43:46]: a bunch of different configurations
SWYX [00:43:48]: to figure out, like,
JOSH [00:43:48]: how do I tweak your network and architecture, et cetera, to get the kind of best performance I possibly can. That's usually saying, like, you know, almost all of these hyperparameter configurations are, let's say they're all going to use the same number of GPUs or the same number of nodes.
SWYX [00:44:01]: So it's going to run
JOSH [00:44:01]: for the same amount of time.
SWYX [00:44:03]: So you can do that.
JOSH [00:44:03]: You can get a number out and that's great. But what carbs does is it says,
SWYX [00:44:07]: OK, actually,
JOSH [00:44:07]: what if we relax that constraint? What if we say each of these different points, we're going to model how expensive it will be to sample this configuration. So if what if we train with just one one hundredth of the data? Like, how well can we do?
SWYX [00:44:19]: What if we train
JOSH [00:44:19]: with one tenth of the data? What if we train with all the data? That way you can understand, like, as we get more and more data, as we spend more and more compute,
SWYX [00:44:26]: as we make a bigger
JOSH [00:44:26]: and bigger network, how does performance change with these things that change? Like how expensive it is to even explore this data point. So by doing that, we can see the scaling laws for not just, you know,
SWYX [00:44:36]: the scaling laws
JOSH [00:44:36]: from like the, you know, Chantilla paper, the scaling laws for all parameters. We can see how does how does the number of layers change with this? How does the, you know, the learning rate change? How do the like, you know, various types of regularization change? So you can see these nice scaling laws. And as you're going across costs, like how should this be changing as you're scaling up your model? So that, coupled with the kind of metric that we chose, which is a very precise way of measuring performance, allowed us to really like hone in on parameters that worked really well
SWYX [00:45:05]: and understand, like,
JOSH [00:45:05]: how do we want to scale those up, especially as we're changing
SWYX [00:45:08]: things about the network?
JOSH [00:45:08]: Like one of the things that we did is we used a custom tokenizer. As we change this tokenizer, changes a bunch of other things about the model. So how should we scale up this entirely new tokenizer? Like no one has ever made a model this large with this tokenizer before. And so how do we want to
SWYX [00:45:22]: change all these things?
JOSH [00:45:22]: Harps kind of shows you, like, look, as you change these parameters, like these other ones are kind of dependent on this.
SWYX [00:45:28]: Like this is the, these are
JOSH [00:45:28]: the relationships between them. So you can better understand, like, OK, if I'm going to scale this up 10x or 100x, like, where do I want to be? I can only go so far. And so, you know, we did run, like, I think maybe it was like a 14b one or something
SWYX [00:45:40]: like that to check.
JOSH [00:45:41]: But and so we had a bunch of like 1b or 14b and then at 70b. I don't think we had a, I think we just did like one at 14b. So you can, we get to check that like, oh, is this on the curve? Like, is this where we expect? It was like right there. So then great, go on to the next one. Yeah, I mean, that makes a lot of sense.
SWYX [00:45:56]: I wonder if, so one of the key questions, and correct me if I'm wrong, but like usually people do search or do their evals just based on loss. But you actually evaluate based on, you know, the sort of end state evals that people might expect, like HellaSwag and Lombata, whatever. What is the norm here? Is there a norm?
JOSH [00:46:20]: Yeah, I don't know if there's a hundred percent.
SWYX [00:46:21]: I don't know. I only see loss on most people's reports.
JOSH [00:46:25]: I think it's easy to, like, loss is very nice because it's very precise. It will tell you, like, very fine grained differences between like really small changes in your hyperparameters or network architecture. Whereas, especially at the smaller scales, if you're looking at like accuracy, it's very noisy. Like it might be zero or a hundred or like, you know, fluctuating by like 10 or 20 percentage points, which makes it really hard to tell, like, did that change actually mean anything? So our loss is sort of a combination of these two. Instead of saying, like, let's just look at perplexity, we say, let's look at perplexity on the tasks that we care about for multiple choice questions effectively.
SWYX [00:47:00]: So we're saying like, yes,
JOSH [00:47:00]: this is formulated as a multiple choice question, and we're going to look at the, like, you know, the loss of perplexity for this particular answer token. And that ends up being something that's like both targeted to what you actually care about and also very precise. The nice thing about this though is that it's independent of the data that you train on. One thing that's annoying about perplexity or about loss is that as you change your data set, this is really obnoxious because now it fundamentally changes your loss, right? And so you can't tell, like, how do I tweak my data set? But because we have this held out evaluation data set where we're looking at perplexity, we can actually change the data mix. And so CARBs actually control what is the mix of data that we want to see, like how much code, you know, how much internet text, et cetera, in order to figure out what is the best optimal mix of data and we could do that because we have this other metric. So that was one of the things that was really, really helpful.
SWYX [00:47:46]: I think there is a trend overall about changing data mix as training goes on. I don't know how, you know, we're deciding not to talk about data sets in this podcast, but what have you observed about the changing data mix question?
JOSH [00:48:06]: We did some experiments
SWYX [00:48:08]: and we've actually talked
JOSH [00:48:08]: to a bunch of researchers who are doing work here as well
SWYX [00:48:11]: and looking at kind of
JOSH [00:48:12]: their experiments on this. And we were originally pretty hopeful because it sounds like something that should work and make sense, right? Like, oh, cool. Like maybe you would have your model, like learn the basic features
SWYX [00:48:22]: and then over time,
JOSH [00:48:22]: it could get really good at these complicated math problems or coding or something, right? But it just turns out that like, it's just not the way it works. Like we've done so many experiments and you can get like a tiny, tiny little boost from this, but it just is not like, it's just not the important thing, at least in the experiments that we've seen. So yeah, we've kind of, we're letting other people
SWYX [00:48:40]: explore that more
JOSH [00:48:40]: if they want, but that just doesn't seem like the most promising direction for us.
JONATHAN [00:48:44]: We've had some surprisingly good luck with this. We just released a paper on it. The details matter a lot and it really matters what you're trying to do with the model.
SWYX [00:48:53]: Yeah.
JONATHAN [00:48:53]: But it's been quite effective for us depending on the setting. And certainly when we're thinking about domain-specific models, this helps a ton. You know, to a certain extent, you can always think of this as like early fine tuning. But yeah, I like, there've been little glimmers of this in the literature for years. Like especially, I think the Gemini 1.5 paper mentions this. And I don't remember whether the Llama 3 paper mentions this,
SWYX [00:49:15]: but it's kind of,
JONATHAN [00:49:16]: it's one of those, like people have different ways to get to these endpoints.
SWYX [00:49:20]: I think, you know,
JONATHAN [00:49:20]: there are the architectural tricks that each lab has to mitigate loss spikes or what have you. And everybody's got, you know, their own bag of tricks and it leads to kind of sometimes this contradictory information. It's not contradictory. People are just kind of exploring
SWYX [00:49:33]: different parts of the space
JONATHAN [00:49:33]: in some sense. And there are lots of ways to get a great model. But certainly for us within our config, and it seems like, I guess for the folks at Google, within kind of the part of the world they live in, changing the dataset has helped, but the details matter a lot. And it's really hard to get those details right for the reasons Josh,
SWYX [00:49:48]: you know, just mentioned.
JONATHAN [00:49:48]: Like there's a lot of search involved and you essentially have to make hard choices about
SWYX [00:49:52]: what parts of the space
JONATHAN [00:49:52]: you're going to search and which ones you're going to leave be. And so, you know, some people have done an amazing job. Like I think the, who is it? The Deep Seek folks have done an awesome job looking at like batch size warmup. And that's been really, really fruitful for them. You know, other people are looking really hard at things like data mix, but it just gets tricky to look at everything.
JOSH [00:50:09]: Yeah, I think we've found that like we could get some things that looked like gains from datasets. But one of the things that I like about carbs is that when we applied carbs to like properly tune things, then a lot of those kind of evaporated. Whereas like, like if we just tune these other parameters, actually we can get almost the same gains without having to do this more complicated thing. So at least in the experiment and in the settings that we've, like in the particular metrics
SWYX [00:50:34]: that we care about,
JOSH [00:50:34]: we haven't seen these kind of like pan out or scale up in quite the same way. But not to rule it out. And I think you're right, Jonathan,
SWYX [00:50:41]: that there probably are
JOSH [00:50:41]: a lot of like details that go into like exactly what is the metric, exactly what is the dataset, exactly which, like what schedule are we using for this. And I certainly wouldn't rule it out working.
SWYX [00:50:52]: Quick question about emergence. Doesn't emergence throw a spanner into a theory of carbs? Ah, so there is a paper
JOSH [00:51:01]: of which I really liked and I think informed
SWYX [00:51:05]: a little bit of how
JOSH [00:51:05]: we thought about this, which is are emergent properties of language models a mirage? And I think if you look at that paper, it actually makes a relatively compelling case that in fact, you know, this emergent behavior that you're seeing is not really emergent behavior, but is really a function of the evaluation metrics that we're using. So if you look at accuracy as a metric, what's happening is that accuracy is actually going up continually over training, but it's in log scale. So it starts out at 0.001%, 0.1, 0.1, 10.
SWYX [00:51:35]: Only when you're going
JOSH [00:51:35]: between 10 and 90 do you see this happen, right? When you go from one in, you know,
SWYX [00:51:40]: a thousand getting right
JOSH [00:51:40]: to one in a thousand getting wrong, like there's many orders of magnitude happening here.
SWYX [00:51:44]: So when you're looking
JOSH [00:51:44]: at this in perplexity, then you just see this nice straight line. And so that's actually what carbs is exploiting. Like since we're, since our metric is in this kind of like perplexity log space, like you can see like, oh, it's just like getting better as you make it bigger in this nice, very predictable way. So that, and that is exactly what we saw. Like these things were really, really bad at, you know, predicting the multiple choice answer, just always guess A. OK, it's so terrible at it, but it was like learning to be less confident about that.
SWYX [00:52:09]: Yeah. One trick I saw from one of the papers recently was just like, just randomize the order of the multiple choice questions. And if you, if, if, if they, if they over, if that hits the performance a lot, then they're just basically memorizing the test set, which makes a lot of sense.
JONATHAN [00:52:28]: Yeah, this is, I, I mean, you know, I, I completely agree with what Josh said.
SWYX [00:52:32]: I think the, you know,
JONATHAN [00:52:32]: my bigger lesson is that anything can look however you want it to look. If you put it on a log scale to a certain extent and log, we love our log scales and deep learning for various reasons. Everything looks very clean on a log scale until everything looks very flat on a log scale. Um, I don't know. I like log scales always mix me up. That's, that's all I can say.
SWYX [00:52:51]: Great. I think the, the last thing I was, I was going to mention on, uh, carbs. Oh, well, I mean, let's, let's just kind of go right into evals because I think that's going to be, uh, the, the sort of crowd favorite. Um, so carbs, we already mentioned, um, you know, leans heavily on, uh, the sort of end evals that we would typically eval LLMs on, except that you had to make your own. Um, there are a lot of documented problems with many of the common evals out there and you fixed all of them. It sounds like, I don't know
JOSH [00:53:18]: about fixed all of them, but, uh, I think in the same way that we like to dig into the infrastructure and hardware and understand, like what actually is going
SWYX [00:53:27]: wrong?
JOSH [00:53:27]: Like what is the actual error on this machine with this GPU?
SWYX [00:53:31]: And why did that happen?
JOSH [00:53:31]: And how do we fix it? We take the same approach to the evaluations. So when we looked at the evaluations and actually looked at the data sets, you know, what we did is
SWYX [00:53:39]: like, okay, if we're going
JOSH [00:53:39]: to be, you know, evaluating natural language, understanding and reasoning, like, let's look at all the data sets that are out there. Let's actually look at a bunch of the examples and say, like, is this a good data set that we should use for evaluation? That's kind of how we selected the evaluation data set that we had. Uh, and then when we looked at the actual examples in there, we noticed like a lot of these are very messy. Like some of them messy
SWYX [00:54:00]: to the point of like
JOSH [00:54:00]: incoherence and some of the ones that we didn't choose. Uh, but even the ones that we chose, like people tried pretty hard on
SWYX [00:54:06]: these data sets.
JOSH [00:54:06]: They did try and clean them, but there's just a lot of data points in there and it's just easy to
SWYX [00:54:10]: make mistakes.
JOSH [00:54:10]: Right. And so, you know, it's not that they have a
SWYX [00:54:13]: hundred people looking
JOSH [00:54:13]: at every question, like that's just way too
SWYX [00:54:15]: expensive.
JOSH [00:54:15]: So you end up with questions that just don't make sense.
SWYX [00:54:18]: Somebody didn't really
JOSH [00:54:18]: see this. Somebody just clicked the wrong box for the answer. Uh, or the question makes sense in your head. When you write it, we've often seen this, it's not even like malice or
SWYX [00:54:26]: incompetence.
JOSH [00:54:26]: It's really just like, you know, you write this,
SWYX [00:54:28]: you're ready.
JOSH [00:54:28]: You're like, this makes
SWYX [00:54:29]: sense to me.
JOSH [00:54:29]: You show it to another person like that makes
SWYX [00:54:31]: sense.
JOSH [00:54:31]: You show it to a third
SWYX [00:54:32]: person.
JOSH [00:54:32]: They're like, this makes no sense at all.
SWYX [00:54:34]: That's because you're
JOSH [00:54:34]: kind of, you know, using a different meaning of
SWYX [00:54:36]: the word.
JOSH [00:54:36]: And then when they say that, you're like, Oh,
SWYX [00:54:38]: wow, you're right.
JOSH [00:54:38]: That is actually really confusing. It's easy for things to
SWYX [00:54:41]: kind of make sense in
JOSH [00:54:41]: our own head. So what we did for the evaluations is really dug into the details of each of these data sets and tried to ask, like, what makes a good
SWYX [00:54:50]: question?
JOSH [00:54:50]: What makes a good answer?
SWYX [00:54:52]: Like, what does it mean
JOSH [00:54:52]: for it to be ambiguous? We had a whole, like,
SWYX [00:54:55]: we looked at lots of
JOSH [00:54:55]: data, broke this down, asked lots of people
SWYX [00:54:58]: about all these
JOSH [00:54:58]: different questions to build a model of this and help us kind of clean these data sets. That was sort of one big piece of it. A second big piece was making sure that our data that we're training on is not data that we're testing on. So there we kind of took a step back and said, like, OK, well, let's just reproduce, you know, 500 to a thousand examples for every single one of these data sets ourselves. And just make sure that this data is definitely not in the, you know, the training set. So we did that. And then we're able to, like, now be confident about, like, our performance of our model and also performance of other open source and other closed source models. Yeah, there's a lot there.
SWYX [00:55:33]: You had 11? I don't know how many data sets. I think so. One, two? Yeah. Any one you want to call out in particular to dive deeper on? Some of these are very famous, like HelloSwag, MitoGrand. Some are less famous, like Race. I don't know if... Race is a great data set.
JOSH [00:55:50]: See that one?
SWYX [00:55:51]: Yeah. Yeah. Just, you know, anything that's interesting you want on specific data sets? I think there are
JOSH [00:55:57]: a few asterisks in there. You know, definitely read the whole paper
SWYX [00:56:02]: as you're looking at
JOSH [00:56:02]: some of these, like the GSM8K one is a little bit weird. I think one that was
SWYX [00:56:06]: kind of funny,
JOSH [00:56:06]: it was, like, low performance on ethics from some of the more recent models. I think that was a
SWYX [00:56:11]: little bit funny
JOSH [00:56:11]: because the models, you know,
SWYX [00:56:13]: I think there was
JOSH [00:56:13]: a reaction to, like, oh, no, like, you know,
SWYX [00:56:16]: the models are saying
JOSH [00:56:16]: bad things.
SWYX [00:56:17]: And so they went way,
JOSH [00:56:17]: way in the other direction. And now, like, on the ethics data set,
SWYX [00:56:20]: it's always like,
JOSH [00:56:20]: this is totally unethical, even though it's really fine. So they've just been tuned to, you know, make sure they don't make any PR disasters.
SWYX [00:56:28]: I thought that was
JOSH [00:56:28]: a little bit funny. Not to say that it's necessarily like a flaw of the model, but just kind of like, you know, political or tuning opinion. I think the main takeaway, I was just going to say
SWYX [00:56:38]: the main takeaway
JOSH [00:56:38]: for many of the, like, actual performance is, like, once you fix these ambiguous examples, a lot of these benchmarks are really saturated. Like, I think it's
SWYX [00:56:48]: important to look at,
JOSH [00:56:48]: like, you know,
SWYX [00:56:50]: like when you're
JOSH [00:56:50]: talking about performance on ANLI or race or pool queue or something, what you're really talking about is, like, performance on questions that make no sense. Like, it's just like, did it guess the answer in this, like, really weird scenario? Like, those are the ones that are left.
SWYX [00:57:03]: Like, when you look
JOSH [00:57:03]: at the performance on the ones that actually make sense to everyone, all the models agree.
SWYX [00:57:07]: We agree, like,
JOSH [00:57:07]: everyone's on the same page, which I think is kind of a really interesting result.
SWYX [00:57:11]: The question then becomes, you know, what are the new, like, set of evals that would be like the next frontier that often embeds with it your idea of what reasoning is, because it's obviously you're super interested in reasoning. And yeah, I mean, like, where does this, where does the state of evals go from here?
JOSH [00:57:30]: This work and this blog post is talking mostly about the public evaluations
SWYX [00:57:34]: and the things
JOSH [00:57:34]: that we can release. We do have our own internal evaluations. For example, one of them that we are releasing is the code understanding evaluation, which is about predicting,
SWYX [00:57:44]: you know,
JOSH [00:57:44]: what will this variable be or asking questions about code, et cetera. And that is one of the early benchmarks that we made that we can release. We can partly release it because we can generate an almost infinite amount of this data because these are programmatically generated. And so, you know, we're not really worried about there being like corruption in the kind of the training or test sets. So that makes it a little
SWYX [00:58:03]: bit easier for us.
JOSH [00:58:04]: But I think it's, you know, we have built other data sets as well that we can't release. Some of them, you know,
SWYX [00:58:09]: for example,
JOSH [00:58:09]: because they maybe use other open source code and so we can't redistribute it necessarily. Other ones, because, you know, that's, I think evaluations and data are like a core, important part of, you know, the business. And I think we take evaluations very seriously and are spending a lot of effort in terms of like, what exactly do we make as part of the evaluation set? How do you evaluate these things? We've done a lot of other stuff, you know, since these evaluations. But I think a lot around like code understanding for us, since that's our main focus. And it's a nice place to explore reasoning as well.
SWYX [00:58:40]: It sounds like you talk a little bit about like code understanding as like sort of variable level, like sort of very micro context. Is there a sense of like larger code context as well? I don't know what I mean by that, by the way. It's mostly just like if I told the senior engineer to go look at a code base, they would understand at a broad level, the architecture, but also the design decisions and be able to tell me that. I don't know if that's useful or not, but I mean, that's useful to me as a, as someone who might be working with them. Yeah.
JOSH [00:59:06]: This particular dataset is like the more low level code understanding,
SWYX [00:59:10]: like just literally
JOSH [00:59:10]: what happens in this code. And this is mostly because, you know,
SWYX [00:59:13]: this is part of the
JOSH [00:59:13]: carbs tuning metric, etc.
SWYX [00:59:15]: Like we care about
JOSH [00:59:15]: the low scale version
SWYX [00:59:17]: of this as well.
JOSH [00:59:17]: We want smaller scale models to be able to do something on this. And so that's kind of the focus for this.
SWYX [00:59:22]: And hopefully this is more
JOSH [00:59:22]: useful for other people. But yes,
SWYX [00:59:25]: those other questions
JOSH [00:59:25]: are also quite interesting. They get a lot harder to evaluate, like, is this a good architecture or not? Like you and I could probably debate for a while on, you know, different architectures. And so it becomes a lot trickier to do these evaluations as they become more realistic. So I think that's one of the things that we've been playing around with a lot, especially around like code generation.
SWYX [00:59:44]: So if you're saying,
JOSH [00:59:44]: you know, implement this function, okay, it can be kind of objective, but, you know, even MBPP, we've made our own internal version of this data set, right?
SWYX [00:59:52]: Where we've taken like
JOSH [00:59:52]: every single example
SWYX [00:59:54]: and looked at it and been like,
JOSH [00:59:54]: does this actually make sense? Like, what is the type signature? Like, can we remove all ambiguity, et cetera?
SWYX [01:00:00]: So you basically like reviewed every single question on, I mean, that's impossible for like HelloSwag, right? Yeah, yeah.
JOSH [01:00:05]: We didn't do that for HelloSwag, but this is for MBPP, which is only like a few hundred. So we just sat down and did it. Yeah.
JONATHAN [01:00:12]: I'm so excited to get to look at this data set. Like this is such a resource for the community. I absolutely can't wait. We should probably do the,
JOSH [01:00:19]: I don't know. I don't know if we were planning on doing the healed MBPP one,
SWYX [01:00:23]: but hopefully we can do
JOSH [01:00:23]: that one in the future. Did you look at SweetBench?
SWYX [01:00:26]: It's the sort of hot new data set of the summer.
JOSH [01:00:28]: Yeah, I've taken a quick look
SWYX [01:00:29]: at SweetBench.
JOSH [01:00:29]: It's really interesting. I like that it's a much more difficult kind of coding, code related task for bug fixing. I think it gets into some of these problems where it is a lot harder to evaluate these things once they get more realistic. Like we were looking at the AgentBench paper, I think just last week for our paper club and one of the things
SWYX [01:00:49]: that we noticed
JOSH [01:00:49]: is that actually like both of the examples in the appendix that are given as like traces where it got it right. This is actually not the right solution. And it's OK. You know, it's fine. Like it did make it past the test. That's what the metric is.
SWYX [01:01:02]: That's what the benchmark
JOSH [01:01:02]: is about, right? But like it just said,
SWYX [01:01:05]: you know, like,
JOSH [01:01:05]: you know, dot encode ASCII. Like, well, that's not the right way to do this. Like it just dropped all the other edge cases that you actually would have cared about in production for this thing.
SWYX [01:01:14]: And there is like
JOSH [01:01:14]: a better way of doing it.
SWYX [01:01:16]: And you know,
JOSH [01:01:16]: that's what the real golden patch was. But, you know, that's OK. But then how do you test all of that?
SWYX [01:01:21]: Like as you start to do
JOSH [01:01:21]: more realistic things, the test coverage, like getting test coverage over all possible ways of solving these bugs is really hard. Evaluation is the single
JONATHAN [01:01:28]: hardest part of the whole thing. Like I spend a shocking amount of time just telling our customers
SWYX [01:01:34]: we need to find a way
JONATHAN [01:01:34]: to measure what you actually want out of the model before you should ever touch a GPU. And, you know, trying to convince my team and me to follow our own advice a lot of the time on that. And I think everybody like on the one hand,
SWYX [01:01:46]: it's easy to laugh
JONATHAN [01:01:46]: at the state of the evaluations that we have. None of them are good. Like if you go read these eval benchmarks, you'll always come away
SWYX [01:01:52]: disappointed.
JONATHAN [01:01:53]: And yet they've given us useful hills to climb. And we do seem to be making progress and measuring
SWYX [01:01:58]: progress in the field.
JONATHAN [01:01:58]: And I think anecdotally, models are getting better year to year. So I feel like people tend to go and get into one situation or the other, like evals don't matter. I'm just going to look at loss
SWYX [01:02:07]: or like, you know,
JONATHAN [01:02:08]: the evals matter a lot and they're all broken. So what do I do? And I think like a lot of things in deep learning, we have to make peace with just complete imperfection. Like the most successful scientists I see are the ones who are OK operating in a world
SWYX [01:02:20]: where everything's
JONATHAN [01:02:20]: going to be broken.
SWYX [01:02:22]: And yet we can still
JONATHAN [01:02:22]: cobble things together and make something
SWYX [01:02:24]: interesting happen.
JONATHAN [01:02:24]: I mean, we were just discussing that with literal infrastructure. And now we're all the way
SWYX [01:02:28]: up to like,
JONATHAN [01:02:28]: how do we measure whether a model performed a complex coding task correctly? And everything is broken.
SWYX [01:02:34]: And yet we're still able
JONATHAN [01:02:34]: to make huge amounts of forward progress.
SWYX [01:02:36]: I think that's right, Jonathan.
JOSH [01:02:38]: And that the challenge
SWYX [01:02:40]: isn't necessarily
JOSH [01:02:40]: making perfect evaluations. I think our blog post here is about going really into the weeds on these to figure out like, what does that look like? And I think one thing is like, you know,
SWYX [01:02:49]: as you said,
JOSH [01:02:49]: we have been able to make a lot of progress without making these perfect.
SWYX [01:02:52]: That's great.
JOSH [01:02:52]: You don't have to have perfect evaluations. And, you know, the more interesting work is the stuff that we can't necessarily publish about, which is the imperfect evaluations that we have for actual coding tasks, for example.
SWYX [01:03:04]: Like, what does this
JOSH [01:03:04]: really mean as a person? And there, as you said, it's much messier.
SWYX [01:03:08]: So it's a lot harder
JOSH [01:03:08]: to put it out and say like, hey, everybody use this because there's so many
SWYX [01:03:12]: rough edges.
JOSH [01:03:12]: It's so hard to like even say, oh, is this even the right task? Is this even the right way to do it? And there's a lot of judgment.
SWYX [01:03:19]: There's a lot of intuition
JOSH [01:03:19]: that it comes down to. But yeah, I think that's where it's critical to do
SWYX [01:03:23]: if you actually want to
JOSH [01:03:23]: make these systems work.
JONATHAN [01:03:24]: Yeah, you have to make peace with with living in that in between.
SWYX [01:03:28]: Yeah.
JONATHAN [01:03:28]: And I think that in some sense,
SWYX [01:03:30]: when I hire researchers,
JONATHAN [01:03:30]: that's the number one quality I look for. Like, can they be at peace living in a house that is neither clean nor messy,
SWYX [01:03:36]: but it's just kind of
JONATHAN [01:03:36]: somewhere in between? And are they OK with that? Are they OK with a few dishes being out on the table and a few clothes
SWYX [01:03:42]: being on the floor?
JONATHAN [01:03:43]: Or will that drive them insane? Or will they just end up with all the clothes on the floor and like all the dishes out all the time? Like, it's kind of I'm looking for that perfect balance because, you know, we have to operate in this imperfect world. Like, yeah, go ahead and give me the perfect evaluation for programmers
SWYX [01:03:58]: or for an LLM
JONATHAN [01:03:58]: that is a program assistant tool. Like there is no perfect evaluation. But clearly we've made progress. And so the most important part
SWYX [01:04:06]: is just are we
JONATHAN [01:04:06]: climbing the right hills? And so this is why I'm so excited to see the ambiguity aspect of this. We often think we have more room to climb on these benchmarks. It turns out we don't. Or it turns out that actually we're climbing, getting good at the benchmark and not actually getting good at the task we care about underlying the benchmark anymore.
SWYX [01:04:21]: Maybe the model,
JONATHAN [01:04:21]: like this is the famous example where if you get 100% at MNIST, your model must be broken in some way because there are four examples mislabeled, you know, it's it's that all over again. Welcome to this.
SWYX [01:04:33]: Yeah, it's the accidental canary canary in this. I think one thing that's
JOSH [01:04:37]: actually really interesting about this also is that, yes, like the ambiguous examples are sort of, you know, not that great from the perspective of these particular tasks that we're evaluating.
SWYX [01:04:46]: But actually, one thing
JOSH [01:04:46]: that we're very interested in is ambiguity itself. Like, can we detect whether a task from a user is ambiguous or whether you've, you know, completed a task successfully? Like these are actually hard, messy problems, but are really important from like the user experience of using these models. I would much rather have a coding agent that will give me back a thing. And, you know, it's it's actually the code doesn't work like 10% less of the time than some other model, but it will tell me 100% of the time like when it's not sure. Like that's so much more useful if it can communicate like, I'm not really sure about this or maybe there's some errors here. Then just like, here's some code. I have no idea if it works. And so these kind of like, you know, detecting ambiguity and detecting correctness
SWYX [01:05:25]: or uncertainty,
JOSH [01:05:25]: I think are really interesting problems
SWYX [01:05:27]: that we're really like
JOSH [01:05:27]: digging into quite deeply.
SWYX [01:05:29]: I want to touch on maybe a couple of hot topics in evals, maybe tangentially related, but we're on the evals train right now. So I'm just going to get on that. So ArcAGI, Francois Chollet's hot new thing, it's sort of my take on it is basically it's trying to measure reasoning through an abstract IQ test. Effectively, I noticed that you don't use it. There's a lot of community debate, pro and con about it. What are your thoughts on just more abstract reasoning and maybe ArcAGI specifically?
JOSH [01:06:01]: I think we purposely stayed away from the very, like there's BigBench, for example, that has a lot of, I think, to me, feels sort of similar types of tasks that are like very unrealistic. Like, oh, you know, we have books of different colors and then you're going to shuffle them and like which book is furthest to the left or something like, OK, cool, I guess it's neat. It's neat, I think, for us to explore in terms of like an agent reasoning in a larger loop. And we do care about these types of evaluations there. The types of evaluations we're talking about in the blog post here are for getting at, like, does this model in a base model sense, is this working at all? There's no chain of thought in these evaluations. These are just like, go straight to the answer. Does this make sense?
SWYX [01:06:42]: Like, is this a thing that
JOSH [01:06:42]: you can answer very quickly? That's what we were selecting for with these evaluations. This is not to say that these are the only evaluations we have. I think the Arc ones are like a little bit too, probably, visual for us to really be able to integrate with.
SWYX [01:06:56]: But I think some of the
JOSH [01:06:56]: BigBench ones are... You can tokenize it.
SWYX [01:06:59]: Yeah, but, you know,
JOSH [01:07:00]: I think it's not really... I think you can spend a lot of time getting really good at these kinds of benchmarks without making, like, kind of more general purpose progress. And so I think we're a little bit leery of going too far in that direction. Similarly, like, coding competitions. Like, we do a lot of code generation, but we don't really do a lot on, like, code competition problems for the very, very hard ones.
SWYX [01:07:20]: So I think you can go
JOSH [01:07:20]: very far down that route
SWYX [01:07:22]: and make something that's, like,
JOSH [01:07:22]: really good at those problems, but not actually that useful as, like, a programmer day to day.
SWYX [01:07:26]: Yeah.
JONATHAN [01:07:27]: Take a different tactic, which is, like, at the end of the day at Databricks, I have 12,000 customers, or I think that's the latest number, all of whom are trying to do something with, you know, LLMs or AI or machine learning. And those things don't look like these tasks. I don't think I have a single customer that's asking to, you know, have AI solve abstract reasoning problems. Things are pretty, like, they can be ambiguous,
SWYX [01:07:53]: they can be challenging,
JONATHAN [01:07:53]: they can be really interesting,
SWYX [01:07:55]: but none of them look quite like this.
JONATHAN [01:07:56]: And so, you know, I think to Josh's point, like, it's really about asking, why are we doing this? Even if you're trying to build AGI, and that's not personally my purpose, and I, you know, Josh has much more interesting things to say about that than I do. I don't even know if this is the kind of intelligence I would get excited about or care about personally, or if I would consider, you know, to Josh's point, this to be the indicia of intelligence.
SWYX [01:08:17]: It's neat.
JONATHAN [01:08:17]: But, you know, for me, it's, like, more down to earth things, like having a model that can have a conversation with you about data
SWYX [01:08:24]: that on the backend
JONATHAN [01:08:24]: is running SQL queries on your literal data. That's a much more interesting task to me. That's something that really matters day to day for my customers and, you know, different perspectives, but, you know, I think Josh and I would probably say the same thing,
SWYX [01:08:36]: even though I would,
JONATHAN [01:08:36]: I'm guessing, I don't want to put words in your mouth. You would say that you're pursuing more general intelligence in your own way. And I would say that I'm very happy with narrow intelligence. Like, I'm very happy with my little SQL bot and building 12,000 of those because they're moving the needle for a lot of folks every day.
JOSH [01:08:51]: Yeah, I think we're, you know, we're not as far away in our position as it might seem. I think we're also excited about, like,
SWYX [01:08:58]: how do you actually
JOSH [01:08:58]: make these things useful? And that does end up being pretty narrow. I think these other tasks can be interesting as, like, ways to explore these more abstract reasoning questions or like, OK, how could an agent actually work through this? But it's important to keep in mind that it's like a toy, not a real problem. It's like it's a scientific tool to tell us something about the models.
SWYX [01:09:16]: It's not something we should
JOSH [01:09:16]: be optimizing for necessarily.
SWYX [01:09:18]: The one thing I'll point out is, you know, as a kid, I was graded into a gifted program based on my ability to solve these exact type of problems. And then I entered college based on my ability to solve SATs, which, again, have nothing to do with my college experience, but whatever. So, you know, we have a history in the humanity of doing correlated IQ tests to general capability. OK, so the two more, two more viral evals, and then, you know, I just want to be mindful of your time. Needle in a haystack, long context utilization. Oh, for the love of God. Something, well, OK, like, let's just assume that, you know, on our podcast, we've discussed the, you know, baseline problems with needle in a haystack, but just generally long context, right? It's a useful thing for agents. I assume. And it's something that, you know, it's out there. Like, we don't know, don't really know what the best way to utilize memory is. But like, I assume it's important, right? What I'll say is like, you know,
JONATHAN [01:10:13]: I spend a lot of time thinking about RAG these days. And RAG, you know, in one sense, you know, the way that I think about RAG is it's the world's simplest agent. It is an agent that basically, you know, there's at least more than one thing happening in the process of building models, at least a system. If you give the model the ability to decide when it wants to retrieve data from a context or retrieve data from a database, then we're talking about an agent. So RAG kind of, I think, like toes that boundary really nicely. There are a lot of reasons why you do genuinely need a long context. Like, I don't think long contexts are problematic in and of themselves. I know there's some controversy even about that. I love the idea of doing like thousand shot tasks as an alternative to fine tuning. I love the idea of pulling in lots of data into the context. I love the idea of once you get in a multimodal land, you're just going to end up
SWYX [01:10:54]: with giant context.
JONATHAN [01:10:54]: It's kind of unavoidable. The flip side is I don't know of anyone who like is hiding a secret passphrase in a book and needs the model to find it. Needle in a haystack is, it's interesting. The challenge with long context to my mind, and Josh,
SWYX [01:11:08]: I'm curious what you think,
JONATHAN [01:11:08]: is simply that annotating long context evals is really hard and really expensive, you know, intrinsically, because you need someone to read 10,000 tokens or 100,000 tokens, or like you need someone to read a 1,000 page book or the equivalent thereof in order to measure those long context benchmarks. I don't know if a human could solve these tasks, let alone that a human could do this in any amount of time where you're willing to pay the money to get the data annotated. And so any long context eval
SWYX [01:11:33]: has to, in some sense,
JONATHAN [01:11:33]: be correct by construction. And you have to, you know, the, you have to know the answer before you've created the example. And needle in a haystack is kind of the simplest way
SWYX [01:11:41]: of doing that.
JONATHAN [01:11:41]: I think the problems of needle in a haystack are well known, you know, it doesn't measure anything real. You're not even testing the model's ability to holistically use the context just to identify one part of the context. So you can do some wacky things to your model, like quantize the hell out of the KV cache and still get needle in a haystack to work quite well because it's not trying to holistically take advantage
SWYX [01:11:59]: of things.
JONATHAN [01:12:00]: You know, I have some thoughts on things that I like more that are also still correct by construction. Like, I really like the idea of doing thousand shot tasks where you can look at the scaling as you go from 10 shot to 100 shot to thousand shot to fine tuning on that data instead. And I like that as a way to, you know, have something that's correct by construction, or at least where you have
SWYX [01:12:19]: a nice baseline
JONATHAN [01:12:19]: that you can compare to automatically. So I'm typically looking for like contexts that are situations where long context is one way to solve the task, but not the only way
SWYX [01:12:28]: to solve the task.
JONATHAN [01:12:28]: And we have some other strong baseline floating around personally. But yeah, needle in a haystack, not my favorite thing in the world, to say the least.
JOSH [01:12:35]: Yeah, I mean, I agree with most of what Jonathan
SWYX [01:12:38]: said, I think.
JOSH [01:12:38]: I think one other thing that I will call out
SWYX [01:12:40]: is that, you know,
JOSH [01:12:40]: from like a coding application perspective, it's useful to have long context because the lazy thing of just like throw the whole repo in the context is like,
SWYX [01:12:48]: OK, cool.
JOSH [01:12:48]: Like, you know, you can just get started with that. But then in, you know, in real scenarios, you don't necessarily want to put the whole thing in there. You can have code bases
SWYX [01:12:56]: that are bigger.
JOSH [01:12:56]: You probably want to filter down to the stuff that's relevant anyway to not be confusing. Like you probably even if you did have a lot of context,
SWYX [01:13:02]: you might want to sort it
JOSH [01:13:02]: in some way to say this is more important than this other stuff. So and, you know, you don't want to wait for you don't want to be wasting all this time and compute
SWYX [01:13:09]: on inference and like
JOSH [01:13:09]: doesn't really matter. So, yeah, I don't know that it's the most important thing.
SWYX [01:13:15]: I think people will find creative use cases. And like Jon said, I think the multimodality examples will naturally lend themselves to long context. Cool. And then one last one on just general sort of agent related capabilities that we didn't really talk about in the eval section is function calling and tool use. There's a recent trend, I think, basically led again by OpenAI on parallel function calling. There's always there's been a limit on how many tools you can call from four to now, I think, 128. And I think theoretically, Claude and Jem and I support a lot more.
JOSH [01:13:49]: So just generally,
SWYX [01:13:50]: how do you think about evaling tool use? Is that super important for you guys? We're thinking about it
JOSH [01:13:55]: in a slightly different way, which is, yes, you can have this like hard coded list of tools. But if only you could have like this really large open set of like tools, maybe they would be like functions that you could call if only there was like a language or like a programming thing, like being able to write code. I think for us, it's like, well, look, if we can write code, like now you have all these tools accessible at the end of the day,
SWYX [01:14:16]: like function calling
JOSH [01:14:16]: is just a function invocation, like literally in code. I think our approach to this is like
SWYX [01:14:21]: instead of worrying about
JOSH [01:14:21]: like weird hard coded agents using tools, like let's just make them
SWYX [01:14:25]: able to actually
JOSH [01:14:25]: write code robustly and make that code work and be able to debug that code, know if that code is safe to run, like get really good at the like code writing and execution part of things, because that will open up the action space like far more than, you know, 128 tools, like just everything is at your fingertips, especially I think over the next few years, like we already have so many really good APIs. As we get better and better at writing code, we'll be able to make APIs to things that don't even have APIs today. That's kind of how we think about it is less as like a special purpose thing
SWYX [01:14:52]: and more as like
JOSH [01:14:52]: this is one of the reasons to focus on code.
SWYX [01:14:55]: On my end,
JONATHAN [01:14:55]: the way that I think about this is, you know, I think a lot about how models interact with data.
SWYX [01:15:00]: And so for me,
JONATHAN [01:15:00]: tool use is really a question of how do you take models
SWYX [01:15:04]: that are really built
JONATHAN [01:15:04]: for unstructured data
SWYX [01:15:06]: and have them interact
JONATHAN [01:15:06]: with structured data? So, you know, and I get the question a lot from my customers,
SWYX [01:15:10]: like what do I do
JONATHAN [01:15:10]: with tabular data? Or what do I do with like, you know, JSON? Or what do I do? I mean, you name it, like even what do I do
SWYX [01:15:17]: with a PDF?
JONATHAN [01:15:17]: Because PDF parsing is still an unsolved problem, even in 2024. And the answer, or even just the basic question
SWYX [01:15:24]: of like, should I bother
JONATHAN [01:15:24]: to structure my data anymore? Shouldn't I just toss the table? Shouldn't I flatten it
SWYX [01:15:28]: and just throw it
JONATHAN [01:15:28]: into the LLM context and like let the model
SWYX [01:15:30]: figure it out?
JONATHAN [01:15:30]: Answer is no. We've built all these fun APIs and fun languages
SWYX [01:15:36]: and paradigms
JONATHAN [01:15:36]: for dealing with structured data over the years. Just use them.
SWYX [01:15:40]: Have your model use them.
JONATHAN [01:15:40]: Train a model that can interact
SWYX [01:15:42]: with these things
JONATHAN [01:15:42]: in a meaningful way. Like text to SQL
SWYX [01:15:45]: is still,
JONATHAN [01:15:45]: or like having a model be able to make SQL calls in the backend is actually like one of the single
SWYX [01:15:51]: most useful things
JONATHAN [01:15:51]: for my customers. It sounds really boring. Models are really good at it. And it moves the needle day to day.
SWYX [01:15:57]: So tool use for me
JONATHAN [01:15:58]: really is that like, how do you just interact with structured data sources and take advantage of the fact that you have some
SWYX [01:16:05]: prior knowledge
JONATHAN [01:16:05]: about the structure of your data that an LLM would completely flatten away. In many ways, this is kind of one of the, one of my biggest frustrations with the fact that LLMs work well with code. We have decades and decades and decades
SWYX [01:16:17]: of understanding
JONATHAN [01:16:17]: about the structure and interpretation of programs. Like I think that's literally the name of a book on programming, if I remember right. And, you know, we have all this theory. We know everything there is to know about programming languages if they're well-formed languages and have the right properties. And yet when we have an LLM
SWYX [01:16:31]: work with them,
JONATHAN [01:16:31]: we literally just turn it into a token stream.
SWYX [01:16:33]: Despite the fact that we know
JONATHAN [01:16:34]: how to parse it. We know, you know, how to do all sorts of, you know,
SWYX [01:16:38]: reference, you know,
JONATHAN [01:16:38]: disambiguation and things like that. We're still just flattening it into a model and making the model relearn all of these things from scratch. And it frustrates
SWYX [01:16:45]: the hell out of me.
JONATHAN [01:16:45]: I don't have a better answer when it comes to code, but I really appreciate that with a lot of data sources that have structure to them. Tool uses and function calling
SWYX [01:16:53]: are just,
JONATHAN [01:16:53]: in my mind,
SWYX [01:16:55]: So I think basically what you're saying is like code is the God tool for Jonathan. Like, you know, SQL is so much the right abstraction for accessing all this data. One thing I do spend a lot of time thinking about is for the stuff that doesn't fit in a SQL table, you know, is knowledge graphs the answer? I think a lot of people are exploring that and I think every now and then people get a bout of knowledge graph religion and then it kind of doesn't work out. So I wonder, I wonder what the end state is. Like, is this an idea where it's a mirage? Or is this the idea where it sometime is going to work? It's about having the right tools
JOSH [01:17:27]: for the problems, right? Like as Jonathan was saying, SQL is sometimes definitely the right tool. Like you've got your, you know, order table or something and you want to know, you know, number of sales last month. Like you should be using SQL sum that column. OK, great. You're all set. Knowledge graphs also,
SWYX [01:17:40]: you know,
JOSH [01:17:40]: are sometimes the right tool for a particular problem. You have some like weird question about relationships between entities
SWYX [01:17:46]: that are modeled
JOSH [01:17:46]: on some particular ontology that you actually understand and it's like math to the real world. Great. Use a knowledge base. Like use a knowledge graph. This is fine. But I think in the real world, it gets a lot messier than like knowledge graph style of things where it's like, well, is there a relationship between these two nodes? Like, I don't know.
SWYX [01:18:04]: Like, is are these
JOSH [01:18:04]: two separate nodes? Like those kind of messy borders, I think, prevent it
SWYX [01:18:08]: from being a tool
JOSH [01:18:08]: that can like solve everything forever. And so I think it'll always be good for certain problems, just like SQL is good
SWYX [01:18:14]: for certain problems.
JOSH [01:18:14]: Like different abstractions are good for different problems. And yeah, I think this is why I'm excited about code. Like code lets you
SWYX [01:18:20]: kind of pick the right,
JOSH [01:18:20]: like let's use this library for this problem.
SWYX [01:18:22]: Let's use this library
JOSH [01:18:22]: for this other problem.
JONATHAN [01:18:24]: I think Josh said it and you said it well, like code is kind of the God tool. It unlocks literally everything. The challenge for me is always like,
SWYX [01:18:31]: you know, sometimes
JONATHAN [01:18:31]: unlocking too much power can sometimes inconvenient things can happen. And so it's all about balancing that
SWYX [01:18:37]: in some sense,
JONATHAN [01:18:37]: language is the God tool.
SWYX [01:18:39]: If only, you know,
JONATHAN [01:18:39]: we knew how to interpret it all the time. So code is has the really nice property
SWYX [01:18:44]: that at least you can
JONATHAN [01:18:44]: always execute it. And sometimes you just literally want your model to be able to do SQL calls and nothing else. And setting those boundaries properly for the problem,
SWYX [01:18:52]: I think is going to be, I think at least a lot of my customers
JONATHAN [01:18:54]: are going to be thinking very hard about that.
SWYX [01:18:56]: Like, should I give
JONATHAN [01:18:56]: the model access to the web?
SWYX [01:18:58]: Is that actually helpful
JONATHAN [01:18:58]: for this problem? It sounds great to just like flip yes on all the tools.
SWYX [01:19:02]: Is that actually going to mean
JONATHAN [01:19:02]: I'm going to get better solutions to my problems?
SWYX [01:19:04]: So I want to be mindful of time. I think that's basically our sort of recap of our discussion based on Imbue's releases today. I wanted to leave some time for what's next for both of you guys. Maybe Josh, as a guest of honor, you want to go first as to what happens next.
JOSH [01:19:19]: We have these releases. We're happy to put these things out. I think there's a lot of stuff
SWYX [01:19:22]: that we haven't released.
JOSH [01:19:22]: Like, this is not the only thing we've been working on. Most of our actual focus has been on kind of coding and reasoning. In particular, like the things that we're excited about are can we make these things useful? Like Jonathan is saying, right? Like, it's not about toy problems. It's like, can we use these today in our day-to-day workflow and actually have them accelerate us? And I think we have some kind of internal product prototypes and things that we're excited about. And so we're excited to share more about this in the coming, you know, months to quarters as we get it to a place where like other people could maybe get value out of this as well. But that's kind of our real focus right now is like, how do you take these really cool capabilities that are out there that our models have, et cetera. And like, make sure that they're actually useful today for us, like when we're doing real work and then for other people as well. In particular, focused on generating code, understanding code, testing code, verifying it, like starting with the like robust creation of software. Excellent.
SWYX [01:20:13]: Jonathan?
JONATHAN [01:20:14]: I never like to talk too much about the future because I think you've heard this from me before. I like for us to speak
SWYX [01:20:19]: through our work.
JONATHAN [01:20:19]: And so I don't, I don't like to tease too much. Our mission is, to Josh's point, to make this stuff useful to 12,000 customers. And not a lot of that ends up making it into the public eye
SWYX [01:20:30]: and not a lot of that
JONATHAN [01:20:30]: ends up getting released open source. So for this kind of forum where really, you know,
SWYX [01:20:34]: where we're talking
JONATHAN [01:20:34]: to the community, I'm asking myself right now, like, you know, what exciting things
SWYX [01:20:38]: are we going to have
JONATHAN [01:20:38]: to offer the community in the next little while? I think the most exciting part is just we're writing a lot of blog posts right now. We're trying to share more and more of our science because I feel like
SWYX [01:20:47]: we've been doing
JONATHAN [01:20:47]: these big pushes to create these really giant models.
SWYX [01:20:50]: I think, Josh,
JONATHAN [01:20:50]: I'm sure you had
SWYX [01:20:51]: the same experience.
JONATHAN [01:20:51]: It's exhausting and all-consuming and you get to the end
SWYX [01:20:54]: and you're like,
JONATHAN [01:20:54]: oh, I have all this stuff
SWYX [01:20:56]: I want to talk about.
JONATHAN [01:20:56]: Now I need to find the time to talk about it now that I've survived this huge push. And we're definitely in that mode right now. So there's going to be a lot of that coming in in the next little while. And, you know, we're always cooking up fun new models. I think the real question is, you know, releasing models open source is not our day-to-day bread and butter. It's kind of a fun reward that we get to do sometimes when we have something really cool to share and a little bit of time and spare GPUs in our hands. But for the most part,
SWYX [01:21:20]: everything is going
JONATHAN [01:21:20]: toward customers. You know, I think the joke is Databricks has been 18 months away from IPO for five years. So I guess Databricks
SWYX [01:21:26]: is 18 months away
JONATHAN [01:21:26]: from IPO still. But 18 months away from IPO means there's a lot of pressure to deliver for customers. And we're going to keep working on that. But I think you'll see hopefully some cool, interesting things
SWYX [01:21:36]: get dropped over the course
JONATHAN [01:21:36]: of the summer and into the fall. We'll find out when we get there.
SWYX [01:21:39]: I think that's the right way
JONATHAN [01:21:39]: to put it. I know we were talking earlier about kind of Abracadabra and Alakazam. And all I'll say is that, you know, the DBRX small model that we still haven't released yet was called Abra. DBRX was called Kadabra. And there's a third Pokémon in that evolution. And that's all I'll say for now. Cool stuff kind of popping up sometimes on Chatbot Arena. And, you know, keep your eyes out. Yep.
SWYX [01:21:59]: I'll leave the links and the hints in the show notes. That was a very fun way to leave some breadcrumbs for people to follow. Cool. I'll leave everything to sort of some calls to action. We're going to be releasing this next week. So I'll be deep in my conference, the AI Engineer World's Fair. So people can just go to AI.Engineer and livestream it. Do you guys have any other calls to action before you wrap?
JOSH [01:22:20]: The only one is, you know, we're definitely hiring. So if you're interested in working on coding, reasoning, interested in working on all of this stuff, you know, from the ground up and really deeply understanding not just how does the hardware work, but how do the models work and also designing these, you know, systems to actually be useful for yourself day to day, come say hi.
JONATHAN [01:22:36]: The only thing I'll say is, you know, and I like saying it these days, it feels like the field is so crowded and, you know, it requires so many resources to do impactful work. And, you know, on some days it feels like everything's been done or somebody else is doing everything before you can. At least I remember that feeling every single day of my PhD and even more so now. But I hope like what you heard from Josh today tells you there is so much enormously impactful work to do in the field. If only you take a step back and take a fresh look at some of these things and just talk about what you're doing. There's a huge amount left to do here and a huge amount of exciting work happening every day. And for those who are certainly feeling that exhaustion right now, and I count myself among those folks many days, it's refreshing to see these kinds of drops and see that there is so much more even in things that people feel like they understand how to set up a cluster. My God, you know, even in these evals that we think we understand, there is still more to understand and still more work to do. I hope everybody's keeping at it.
SWYX [01:23:32]: All right. Keep on keeping on. Well, thanks so much for your time, guys. That was a great discussion and we'll put the links in the show notes for people to read more. Thanks. Thanks a bunch.
JOSH [01:23:40]: Thank you so much.
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe
Why AI Agents Don't Work (yet) - with Kanjun Qiu of Imbue
Thanks to the over 11,000 people who joined us for the first AI Engineer Summit! A full recap is coming, but you can 1) catch up on the fun and videos on Twitter and YouTube, 2) help us reach 1000 people for the first comprehensive State of AI Engineering survey and 3) submit projects for the new AI Engineer Foundation.
See our Community page for upcoming meetups in SF, Paris, NYC, and Singapore.
This episode had good interest on Twitter.
Last month, Imbue was crowned as AI’s newest unicorn foundation model lab, raising a $200m Series B at a >$1 billion valuation. As “stealth” foundation model companies go, Imbue (f.k.a. Generally Intelligent) has stood as an enigmatic group given they have no publicly released models to try out. However, ever since their $20m Series A last year their goal has been to “develop generally capable AI agents with human-like intelligence in order to solve problems in the real world”.
From RL to Reasoning LLMs
Along with their Series A, they announced Avalon, “A Benchmark for RL Generalization Using Procedurally Generated Worlds”. Avalon is built on top of the open source Godot game engine, and is ~100x faster than Minecraft to enable fast RL benchmarking and a clear reward with adjustable game difficulty.
After a while, they realized that pure RL isn’t a good path to teach reasoning and planning. The agents were able to learn mechanical things like opening complex doors, climbing, but couldn’t go to higher level tasks. A pure RL world also doesn’t include a language explanation of the agent reasoning, which made it hard to understand why it made certain decisions. That pushed the team more towards the “models for reasoning” path:
“The second thing we learned is that pure reinforcement learning is not a good vehicle for planning and reasoning. So these agents were able to learn all sorts of crazy things: They could learn to climb like hand over hand in VR climbing, they could learn to open doors like very complicated, like multiple switches and a lever open the door, but they couldn't do any higher level things. And they couldn't do those lower level things consistently necessarily. And as a user, I do not want to interact with a pure reinforcement learning end to end RL agent. As a user, like I need much more control over what that agent is doing.”
Inspired by Chelsea Finn’s work on SayCan at Stanford, the team pivoted to have their agents do the reasoning in natural language instead. This development parallels the large leaps in reasoning that humans have developed as the scientific method:
“We are better at reasoning now than we were 3000 years ago. An example of a reasoning strategy is noticing you're confused. Then when I notice I'm confused, I should ask:
* What was the original claim that was made?
* What evidence is there for this claim?
* Does the evidence support the claim?
* Is the claim correct?
This is like a reasoning strategy that was developed in like the 1600s, you know, with like the advent of science. So that's an example of a reasoning strategy. There are tons of them. We employ all the time, lots of heuristics that help us be better at reasoning. And we can generate data that's much more specific to them.“
The Full Stack Model Lab
One year later, it would seem that the pivot to reasoning has had tremendous success, and Imbue has now reached a >$1B valuation, with participation from Astera Institute, NVIDIA, Cruise CEO Kyle Vogt, Notion co-founder Simon Last, and others. Imbue tackles their work with a “full stack” approach:
* Models. Pretraining very large (>100B parameter) models, optimized to perform well on internal reasoning benchmarks, with a ~10,000 Nvidia H100 GPU cluster lets us iterate rapidly on everything from training data to architecture and reasoning mechanisms.
* Tools and Agents. Building internal productivity tools from coding agents for fixing type checking and linting errors, to sophisticated systems like CARBS (for hyperparameter tuning and network architecture search).
* Interface Invention. Solving agent trust and collaboration (not merely communication) with humans by creating better abstractions and interfaces — IDEs for users to program computers in natural language.
* Theory. Publishing research about the theoretical underpinnings of self-supervised learning, as well as scaling laws for machine learning research.
Kanjun believes we are still in the “bare metal phase” of agent development, and they want to take a holistic approach to building the “operating system for agents”. We loved diving deep into the Imbue approach toward solving the AI Holy Grail of reliable agents, and are excited to share our conversation with you today!
Timestamps
* [00:00:00] Introductions
* [00:06:07] The origin story of Imbue
* [00:09:39] Imbue's approach to training large foundation models optimized for reasoning
* [00:12:18] Imbue's goals to build an "operating system" for reliable, inspectable AI agents
* [00:15:37] Imbue's process of developing internal tools and interfaces to collaborate with AI agents
* [00:17:27] Imbue's focus on improving reasoning capabilities in models, using code and other data
* [00:19:50] The value of using both public benchmarks and internal metrics to evaluate progress
* [00:21:43] Lessons learned from developing the Avalon research environment
* [00:23:31] The limitations of pure reinforcement learning for general intelligence
* [00:28:36] Imbue's vision for building better abstractions and interfaces for reliable agents
* [00:31:36] Interface design for collaborating with, rather than just communicating with, AI agents
* [00:37:40] The future potential of an agent-to-agent protocol
* [00:39:29] Leveraging approaches like critiquing between models and chain of thought
* [00:45:49] Kanjun's philosophy on enabling team members as creative agents at Imbue
* [00:53:51] Kanjun's experience co-founding the communal co-living space The Archive
* [01:00:22] Lightning Round
Show Notes
* Imbue
* Avalon
* CARBS (hyperparameter optimizer)
* Series B announcement
* Kanjun/Imbue’s Podcast
* MIT Media Lab
* Research mentioned:
* Momentum Contrast
* SimClr
* Chelsea Finn - SayCan
* Agent Protocol - part of the AI Engineer Foundation
* Xerox PARC
* Michael Nielsen
* Jason Benn
* Outset Capital
* Scenius - Kevin Kelly
* South Park Commons
* The Archive
* Thursday Nights in AI
Transcript
Alessio: Hey everyone, welcome to the Latent Space Podcast. This is Alessio, Partner and CTO at Residence at Decibel Partners, and I'm joined by my co-host Swyx, founder of Smol.ai. [00:00:19]
Swyx: Hey, and today in the studio we have Kanjun from Imbue. Welcome. So you and I have, I guess, crossed paths a number of times. You're formerly named Generally Intelligent and you've just announced your rename, rebrand in huge, humongous ways. So congrats on all of that. And we're here to dive in into deeper detail on Imbue. We like to introduce you on a high level basis, but then have you go into a little bit more of your personal side. So you graduated your BS at MIT and you also spent some time at the MIT Media Lab, one of the most famous, I guess, computer hacking labs in the world. Then you graduated MIT and you went straight into BizOps at Dropbox, where you're eventually chief of staff, which is a pretty interesting role we can dive into later. And then it seems like the founder bug hit you. You were basically a three times founder at Ember, Sorceress, and now at Generally Intelligent slash Imbue. What should people know about you on the personal side that's not on your LinkedIn? That's something you're very passionate about outside of work. [00:01:12]
Kanjun: Yeah. I think if you ask any of my friends, they would tell you that I'm obsessed with agency, like human agency and human potential. [00:01:19]
Swyx: That's work. Come on.
Kanjun: It's not work. What are you talking about?
Swyx: So what's an example of human agency that you try to promote? [00:01:27]
Kanjun: With all of my friends, I have a lot of conversations with them that's kind of helping figure out what's blocking them. I guess I do this with a team kind of automatically too. And I think about it for myself often, like building systems. I have a lot of systems to help myself be more effective. At Dropbox, I used to give this onboarding talk called How to Be Effective, which people liked. I think like a thousand people heard this onboarding talk, and I think maybe Dropbox was more effective. I think I just really believe that as humans, we can be a lot more than we are. And it's what drives everything. I guess completely outside of work, I do dance. I do partner dance. [00:02:03]
Swyx: Yeah. Lots of interest in that stuff, especially in the sort of group living houses in San Francisco, which I've been a little bit part of, and you've also run one of those. [00:02:12]
Kanjun: That's right. Yeah. I started the archive with two friends, with Josh, my co-founder, and a couple of other folks in 2015. That's right. And GPT-3, our housemates built. [00:02:22]
Swyx: Was that the, I guess, the precursor to Generally Intelligent, that you started doing more things with Josh? Is that how that relationship started? Yeah. [00:02:30]
Kanjun: This is our third company together. Our first company, Josh poached me from Dropbox for Ember. And there we built a really interesting technology, laser raster projector, VR headset. And then we were like, VR is not the thing we're most passionate about. And actually it was kind of early days when we both realized we really do believe that in our lifetimes, like computers that are intelligent are going to be able to allow us to do much more than we can do today as people and be much more as people than we can be today. And at that time, we actually, after Ember, we were like, work on AI research or start an AI lab. A bunch of our housemates were joining OpenAI, and we actually decided to do something more pragmatic to apply AI to recruiting and to try to understand like, okay, if we are actually trying to deploy these systems in the real world, what's required? And that was Sorceress. That taught us so much about maybe an AI agent in a lot of ways, like what does it actually take to make a product that people can trust and rely on? I think we never really fully got there. And it's taught me a lot about what's required. And it's kind of like, I think informed some of our approach and some of the way that we think about how these systems will actually get used by people in the real world. [00:03:42]
Swyx: Just to go one step deeper on that, you're building AI agents in 2016 before it was cool. You got some muscle and you raised $30 million. Something was working. What do you think you succeeded in doing and then what did you try to do that did not pan out? [00:03:56]
Kanjun: Yeah. So the product worked quite well. So Sorceress was an AI system that basically looked for candidates that could be a good fit and then helped you reach out to them. And this was a little bit early. We didn't have language models to help you reach out. So we actually had a team of writers that like, you know, customized emails and we automated a lot of the customization. But the product was pretty magical. Like candidates would just be interested and land in your inbox and then you can talk to them. As a hiring manager, that's such a good experience. I think there were a lot of learnings, both on the product and market side. On the market side, recruiting is a market that is endogenously high churn, which means because people start hiring and then we hire the role for them and they stop hiring. So the more we succeed, the more they... [00:04:39]
Swyx: It's like the whole dating business. [00:04:40]
Kanjun: It's the dating business. Exactly. Exactly. And I think that's the same problem as the dating business. And I was really passionate about like, can we help people find work that is more exciting for them? A lot of people are not excited about their jobs and a lot of companies are doing exciting things and the matching could be a lot better. But the dating business phenomenon like put a damper on that, like it's actually a pretty good business. But as with any business with like relatively high churn, the bigger it gets, the more revenue we have, the slower growth becomes because if 30% of that revenue you lose year over year, then it becomes a worse business. So that was the dynamic we noticed quite early on after our Series A. I think the other really interesting thing about it is we realized what was required for people to trust that these candidates were like well vetted and had been selected for a reason. And it's what actually led us, you know, a lot of what we do at Imbue is working on interfaces to figure out how do we get to a situation where when you're building and using agents, these agents are trustworthy to the end user. That's actually one of the biggest issues with agents that, you know, go off and do longer range goals is that I have to trust, like, did they actually think through this situation? And that really informed a lot of our work today. [00:05:52]
Alessio: Let's jump into GI now, Imbue. When did you decide recruiting was done for you and you were ready for the next challenge? And how did you pick the agent space? I feel like in 2021, it wasn't as mainstream. Yeah. [00:06:07]
Kanjun: So the LinkedIn says that it started in 2021, but actually we started thinking very seriously about it in early 2020, late 2019, early 2020. So what we were seeing is that scale is starting to work and language models probably will actually get to a point where like with hacks, they're actually going to be quite powerful. And it was hard to see that at the time, actually, because GPT-3, the early versions of it, there are all sorts of issues. We're like, oh, that's not that useful, but we could kind of see like, okay, you keep improving it in all of these different ways and it'll get better. What Josh and I were really interested in is how can we get computers that help us do bigger things? Like, you know, there's this kind of future where I think a lot about, you know, if I were born in 1900 as a woman, like my life would not be that fun. I'd spend most of my time like carrying water and literally like getting wood to put in the stove to cook food and like cleaning and scrubbing the dishes and, you know, getting food every day because there's no refrigerator, like all of these things, very physical labor. And what's happened over the last 150 years since the industrial revolution is we've kind of gotten free energy, like energy is way more free than it was 150 years ago. And so as a result, we've built all these technologies like the stove and the dishwasher and the refrigerator, and we have electricity and we have infrastructure, running water, all of these things that have totally freed me up to do what I can do now. And I think the same thing is true for intellectual energy. We don't really see it today, but because we're so in it, but our computers have to be micromanaged. You know, part of why people are like, oh, you're stuck to your screen all day. Well, we're stuck to our screen all day because literally nothing happens unless I'm doing something in front of my screen. I don't, you know, I can't send my computer off to do a bunch of stuff for me. And there is a future where that's not the case, where, you know, I can actually go off and do stuff and trust that my computer will pay my bills and figure out my travel plans and do the detailed work that I am not that excited to do so that I can like be much more creative and able to do things that I as a human, I'm very excited about and collaborate with other people. And there are things that people are uniquely suited for. So that's kind of always been the thing that has been really exciting to me. Like Josh and I have known for a long time, I think that, you know, whatever AI is, it would happen in our lifetimes. And the personal computer kind of started giving us a bit of free intellectual energy. And this is like really the explosion of free intellectual energy. So in early 2020, we were thinking about this and what happened was self-supervised learning basically started working across everything. So worked in language, SimClear came out, I think MoCo had come out, Momentum Contrast had come out earlier in 2019, SimClear came out in early 2020. And we're like, okay, for the first time, self-supervised learning is working really well across images and text and suspect that like, okay, actually it's the case that machines can learn things the way that humans do. And if that's true, if they can learn things in a fully self-supervised way, because like as people, we are not supervised. We like go Google things and try to figure things out. So if that's true, then like what the computer could be is much bigger than what it is today. And so we started exploring ideas around like, how do we actually go? We didn't think about the fact that we could actually just build a research lab. So we were like, okay, what kind of startup could we build to like leverage self-supervised learning? So that eventually becomes something that allows computers to become much more able to do bigger things for us. But that became General Intelligence, which started as a research lab. [00:09:39]
Alessio: So your mission is you aim to rekindle the dream of the personal computer. So when did it go wrong and what are like your first products and user facing things that you're building to rekindle it? [00:09:53]
Kanjun: Yeah. So what we do at Imbue is we train large foundation models optimized for reasoning. And the reason for that is because reasoning is actually, we believe the biggest blocker to agents or systems that can do these larger goals. If we think about something that writes an essay, like when we write an essay, we like write it. We put it and then we're done. We like write it and then we look at it and we're like, oh, I need to do more research on that area. I'm going to go do some research and figure it out and come back and, oh, actually it's not quite right. The structure of the outline. So I'm going to rearrange the outline, rewrite it. It's this very iterative process and it requires thinking through like, okay, what am I trying to do? Is the goal correct? Also like, has the goal changed as I've learned more? So as a tool, like when should I ask the user questions? I shouldn't ask them questions all the time, but I should ask them questions in higher risk situations. How certain am I about the like flight I'm about to book? There are all of these notions of like risk certainty, playing out scenarios, figuring out how to make a plan that makes sense, how to change the plan, what the goal should be. That are things that we lump under the bucket of reasoning and models today, they're not optimized for reasoning. It turns out that there's not actually that much explicit reasoning data on the internet as you would expect. And so we get a lot of mileage out of optimizing our models for reasoning in pre-training. And then on top of that, we build agents ourselves and we, I can get into, we really believe in serious use, like really seriously using the systems and trying to get to an agent that we can use every single day, tons of agents that we can use every single day. And then we experiment with interfaces that help us better interact with the agents. So those are some set of things that we do on the kind of model training and agent side. And then the initial agents that we build, a lot of them are trying to help us write code better because code is most of what we do every day. And then on the infrastructure and theory side, we actually do a fair amount of theory work to understand like, how do these systems learn? And then also like, what are the right abstractions for us to build good agents with, which we can get more into. And if you look at our website, we build a lot of tools internally. We have a like really nice automated hyperparameter optimizer. We have a lot of really nice infrastructure and it's all part of the belief of like, okay, let's try to make it so that the humans are doing the things humans are good at as much as possible. So out of our very small team, we get a lot of leverage. [00:12:18]
Swyx: And so would you still categorize yourself as a research lab now, or are you now in startup mode? Is that a transition that is conscious at all? [00:12:26]
Kanjun: That's a really interesting question. I think we've always intended to build, you know, to try to build the next version of the computer, enable the next version of the computer. The way I think about it is there's a right time to bring a technology to market. So Apple does this really well. Actually, iPhone was under development for 10 years, AirPods for five years. And Apple has a story where iPhone, the first multi-touch screen was created. They actually were like, oh wow, this is cool. Let's like productionize iPhone. They actually brought, they like did some work trying to productionize it and realized this is not good enough. And they put it back into research to try to figure out like, how do we make it better? What are the interface pieces that are needed? And then they brought it back into production. So I think of production and research as kind of like these two separate phases. And internally we have that concept as well, where like things need to be done in order to get to something that's usable. And then when it's usable, like eventually we figure out how to productize it. [00:13:20]
Alessio: What's the culture like to make that happen, to have both like kind of like product oriented, research oriented. And as you think about building the team, I mean, you just raised 200 million. I'm sure you want to hire more people. What are like the right archetypes of people that work at Imbue? [00:13:35]
Kanjun: I would say we have a very unique culture in a lot of ways. I think a lot about social process design. So how do you design social processes that enable people to be effective? I like to think about team members as creative agents, because most companies, they think of their people as assets and they're very proud of this. And I think about like, okay, what is an asset? It's something you own that provides you value that you can discard at any time. This is a very low bar for people. This is not what people are. And so we try to enable everyone to be a creative agent and to really unlock their superpowers. So a lot of the work I do, you know, I was mentioning earlier, I'm like obsessed with agency. A lot of the work I do with team members is try to figure out like, you know, what are you really good at? What really gives you energy and where can we put you such that, how can I help you unlock that and grow that? So much of our work, you know, in terms of team structure, like much of our work actually comes from people. Carbs, our hyperparameter optimizer came from Abe trying to automate his own research process doing hyperparameter optimization. And he actually pulled some ideas from plasma physics. He's a plasma physicist to make the local search work. A lot of our work on evaluations comes from a couple of members of our team who are like obsessed with evaluations. We do a lot of work trying to figure out like, how do you actually evaluate if the model is getting better? Is the model making better agents? Is the agent actually reliable? A lot of things kind of like, I think of people as making the like them shaped blob inside imbue and I think, you know, yeah, that's the kind of person that we're, we're hiring for. We're hiring product engineers and data engineers and research engineers and all these roles. We have projects, not teams. We have a project around data, data collection and data engineering. That's actually one of the key things that improve the model performance. We have a pre-training kind of project with some fine tuning as part of that. And then we have an agent's project that's like trying to build on top of our models as well as use other models in the outside world to try to make agents then we actually use as programmers every day. So all sorts of different, different projects. [00:15:37]
Swyx: As a founder, you're now sort of a capital allocator among all of these different investments effectively at different projects. And I was interested in how you mentioned that you were optimizing for improving reasoning and specifically inside of your pre-training, which I assume is just a lot of data collection. [00:15:55]
Kanjun: We are optimizing reasoning inside of our pre-trained models. And a lot of that is about data. And I can talk more about like what, you know, what exactly does it involve? But actually big, maybe 50% plus of the work is figuring out even if you do have models that reason well, like the models are still stochastic. The way you prompt them still makes, is kind of random, like makes them do random things. And so how do we get to something that is actually robust and reliable as a user? How can I, as a user, trust it? We have all sorts of cool things on the, like, you know, I was mentioning earlier when I talked to other people building agents, they have to do so much work, like to try to get to something that they can actually productize and it takes a long time and agents haven't been productized yet for, partly for this reason is that like the abstractions are very leaky. We can get like 80% of the way there, but like self-driving cars, like the remaining 20% is actually really difficult. We believe that, and we have internally, I think some things that like an interface, for example, that lets me really easily like see what the agent execution is, fork it, try out different things, modify the prompt, modify like the plan that it is making. This type of interface, it makes it so that I feel more like I'm collaborating with the agent as it's executing, as opposed to it's just like doing something as a black box. That's an example of a type of thing that's like beyond just the model pre-training, but on the model pre-training side, like reasoning is a thing that we optimize for. And a lot of that is about what data do we put in. [00:17:27]
Swyx: It's interesting just because I always think like, you know, out of the levers that you have, the resources that you have, I think a lot of people think that running foundation model company or a research lab is going to be primarily compute. And I think the share of compute has gone down a lot over the past three years. It used to be the main story, like the main way you scale is you just throw more compute at it. And now it's like, Flops is not all you need. You need better data, you need better algorithms. And I wonder where that shift has gone. This is a very vague question, but is it like 30-30-30 now? Is it like maybe even higher? So one way I'll put this is people estimate that Llama2 maybe took about three to $4 million of compute, but probably 20 to $25 million worth of labeling data. And I'm like, okay, well that's a very different story than all these other foundation model labs raising hundreds of millions of dollars and spending it on GPUs. [00:18:20]
Kanjun: Data is really expensive. We generate a lot of data. And so that does help. The generated data is close to actually good, as good as human labeled data. [00:18:34]
Swyx: So generated data from other models? [00:18:36]
Kanjun: From our own models. From your own models. Or other models, yeah. [00:18:39]
Swyx: Do you feel like there's certain variations of this? There's the sort of the constitutional AI approach from Anthropic and basically models sampling training on data from other models. I feel like there's a little bit of like contamination in there, or to put it in a statistical form, you're resampling a distribution that you already have that you already know doesn't match human distributions. How do you feel about that basically, just philosophically? [00:19:04]
Kanjun: So when we're optimizing models for reasoning, we are actually trying to like make a part of the distribution really spiky. So in a sense, like that's actually what we want. We want to, because the internet is a sample of the human distribution that's also skewed in all sorts of ways. That is not the data that we necessarily want these models to be trained on. And so when we're generating data, we're not really randomly generating data. We generate very specific things that are like reasoning traces and that help optimize reasoning. Code also is a big piece of improving reasoning. So generated code is not that much worse than like regular human written code. You might even say it can be better in a lot of ways. So yeah. So we are trying to already do that. [00:19:50]
Alessio: What are some of the tools that you thought were not a good fit? So you built Avalon, which is your own simulated world. And when you first started, the metagame was like using games to simulate things using, you know, Minecraft and then OpenAI is like the gym thing and all these things. And I think in one of your other podcasts, you mentioned like Minecraft is like way too slow to actually do any serious work. Is that true? Yeah. I didn't say it. [00:20:17]
Swyx: I don't know. [00:20:18]
Alessio: That's above my pay grade. But Avalon is like a hundred times faster than Minecraft for simulation. When did you figure that out that you needed to just like build your own thing? Was it kind of like your engineering team was like, Hey, this is too slow. Was it more a long-term investment? [00:20:34]
Kanjun: Yeah. At that time we built Avalon as a research environment to help us learn particular things. And one thing we were trying to learn is like, how do you get an agent that is able to do many different tasks? Like RL agents at that time and environments at that time. What we heard from other RL researchers was the like biggest thing keeping holding the field back is lack of benchmarks that let us explore things like planning and curiosity and things like that and have the agent actually perform better if the agent has curiosity. And so we were trying to figure out in a situation where, how can we have agents that are able to handle lots of different types of tasks without the reward being pretty handcrafted? That's a lot of what we had seen is that like these very handcrafted rewards. And so Avalon has like a single reward it's across all tasks. And it also allowed us to create a curriculum so we could make the level more or less difficult. And it taught us a lot, maybe two primary things. One is with no curriculum, RL algorithms don't work at all. So that's actually really interesting. [00:21:43]
Swyx: For the non RL specialists, what is a curriculum in your terminology? [00:21:46]
Kanjun: So a curriculum in this particular case is basically the environment Avalon lets us generate simpler environments and harder environments for a given tasks. What's interesting is that the simpler environments, what you'd expect is the agent succeeds more often. So it gets more reward. And so, you know, kind of my intuitive way of thinking about it is, okay, the reason why it learns much faster with a curriculum is it's just getting a lot more signal. And that's actually an interesting general intuition to have about training these things as like, what kind of signal are they getting? And like, how can you help it get a lot more signal? The second thing we learned is that reinforcement learning is not a good vehicle, like pure reinforcement learning is not a good vehicle for planning and reasoning. So these agents were not able to, they were able to learn all sorts of crazy things. They could learn to climb like hand over hand in VR climbing, they could learn to open doors like very complicated, like multiple switches and a lever open the door, but they couldn't do any higher level things. And they couldn't do those lower level things consistently necessarily. And as a user, I do not want to interact with a pure reinforcement learning end to end RL agent. As a user, like I need much more control over what that agent is doing. And so that actually started to get us on the track of thinking about, okay, how do we do the reasoning part in language? And we were pretty inspired by our friend Chelsea Finn at Stanford was I think working on SACAN at the time where it's basically an experiment where they have robots kind of trying to do different tasks and actually do the reasoning for the robot in natural language. And it worked quite well. And that led us to start experimenting very seriously with reasoning. [00:23:31]
Alessio: How important is the language part for the agent versus for you to inspect the agent? You know, like is it the interface to kind of the human on the loop really important or? [00:23:43]
Kanjun: Yeah, I personally think of it as it's much more important for us, the human user. So I think you probably could get end to end agents that work and are fairly general at some point in the future. But I think you don't want that. Like we actually want agents that we can like perturb while they're trying to figure out what to do. Because, you know, even a very simple example, internally we have like a type error fixing agent and we have like a test generation agent. Test generation agent goes off rails all the time. I want to know, like, why did it generate this particular test? [00:24:19]
Swyx: What was it thinking? [00:24:20]
Kanjun: Did it consider, you know, the fact that this is calling out to this other function? And the formatter agent, if it ever comes up with anything weird, I want to be able to debug like what happened with RL end to end stuff. Like we couldn't do that. Yeah. [00:24:36]
Swyx: It sounds like you have a bunch of agents operating internally within the company. What's your most, I guess, successful agent and what's your least successful one? [00:24:44]
Kanjun: The agents don't work. All of them? I think the only successful agents are the ones that do really small things. So very specific, small things like fix the color of this button on the website or like change the color of this button. [00:24:57]
Swyx: Which is now sweep.dev is doing that. Exactly. [00:25:00]
Kanjun: Perfect. Okay. [00:25:02]
Swyx: Well, we should just use sweep.dev. Well, I mean, okay. I don't know how often you have to fix the color of a button, right? Because all of them raise money on the idea that they can go further. And my fear when encountering something like that is that there's some kind of unknown asymptote ceiling that's going to prevent them, that they're going to run head on into that you've already run into. [00:25:21]
Kanjun: We've definitely run into such a ceiling. But what is the ceiling? [00:25:24]
Swyx: Is there a name for it? Like what? [00:25:26]
Kanjun: I mean, for us, we think of it as reasoning plus these tools. So reasoning plus abstractions, basically. I think actually you can get really far with current models and that's why it's so compelling. Like we can pile debugging tools on top of these current models, have them critique each other and critique themselves and do all of these, like spend more computer inference time, context hack, retrieve augmented generation, et cetera, et cetera, et cetera. Like the pile of hacks actually does get us really far. And a way to think about it is like the underlying language model is kind of like a noisy channel. Actually I don't want to use this analogy. It's actually a really bad analogy, but you kind of like trying to get more signal out of the channel. We don't like to think about it that way. It's what the default approach is, is like trying to get more signal out of this noising channel. But the issue with agents is as a user, I want it to be mostly reliable. It's kind of like self-driving in that way. Like it's not as bad as self-driving, like in self-driving, you know, you're like hurtling at 70 miles an hour. It's like the hardest agent problem. But one thing we learned from Sorceress and one thing we learned by using these things internally is we actually have a pretty high bar for these agents to work. You know, it's actually really annoying if they only work 50% of the time and we can make interfaces to make it slightly less annoying. But yeah, there's a ceiling that we've encountered so far and we need to make the models better. We also need to make the kind of like interface to the user better. And also a lot of the like critiquing. I hope what we can do is help people who are building agents actually like be able to deploy them. I think, you know, that's the gap that we see a lot of today is everyone who's trying to build agents to get to the point where it's robust enough to be deployable. It just, it's like an unknown amount of time. Okay. [00:27:12]
Swyx: So this goes back into what Embu is going to offer as a product or a platform. How are you going to actually help people deploy those agents? Yeah. [00:27:21]
Kanjun: So our current hypothesis, I don't know if this is actually going to end up being the case. We've built a lot of tools for ourselves internally around like debugging, around abstractions or techniques after the model generation happens. Like after the language model generates the text and like interfaces for the user and the underlying model itself, like models talking to each other, maybe some set of those things kind of like an operating system. Some set of those things will be helpful for other people. And we'll figure out what set of those things is helpful for us to make our agents. Like what we want to do is get to a point where we can like start making an agent, deploy it, it's reliable, like very quickly. And there's a similar analog to software engineering, like in the early days, in the seventies and the sixties, like to program a computer, like you have to go all the way down to the registers and write things and eventually we had assembly. That was like an improvement. But then we wrote programming languages with these higher levels of abstraction and that allowed a lot more people to do this and much faster. And the software created is much less expensive. And I think it's basically a similar route here where we're like in the like bare metal phase of agent building. And we will eventually get to something with much nicer abstractions. [00:28:36]
Alessio: We had this conversation with George Hotz and we were like, there's not a lot of reasoning data out there. And can the models really understand? And his take was like, look, with enough compute, you're not that complicated as a human. Like the model can figure out eventually why certain decisions are made. What's been your experience? Like as you think about reasoning data, like do you have to do a lot of like manual work or like is there a way to prompt models to extract the reasoning from actions that they [00:29:03]
Swyx: see? [00:29:03]
Kanjun: So we don't think of it as, oh, throw enough data at it and then it will figure out what the plan should be. I think we're much more explicit. You know, a way to think about it is as humans, we've learned a lot of reasoning strategies over time. We are better at reasoning now than we were 3000 years ago. An example of a reasoning strategy is noticing you're confused. Then when I notice I'm confused, I should ask like, huh, what was the original claim that was made? What evidence is there for this claim? Does the evidence support the claim? Is the claim correct? This is like a reasoning strategy that was developed in like the 1600s, you know, with like the advent of science. So that's an example of a reasoning strategy. There are tons of them. We employ all the time, lots of heuristics that help us be better at reasoning. And we didn't always have them. And because they're invented, like we can generate data that's much more specific to them. So I think internally, yeah, we have a lot of thoughts on what reasoning is and we generate a lot more specific data. We're not just like, oh, it'll figure out reasoning from this black box or like it'll figure out reasoning from the data that exists. Yeah. [00:30:04]
Alessio: I mean, the scientific method is like a good example. If you think about hallucination, right, people are thinking, how do we use these models to do net new, like scientific research? And if you go back in time and the model is like, well, the earth revolves around the sun and people are like, man, this model is crap. It's like, what are you talking about? Like the sun revolves around the earth. It's like, how do you see the future? Like if the models are actually good enough, but we don't believe them, it's like, how do we make the two live together? So you're like, you use Inbu as a scientist to do a lot of your research and Inbu tells you, hey, I think this is like a serious path you should go down. And you're like, no, that sounds impossible. Like how is that trust going to be built? And like, what are some of the tools that maybe are going to be there to inspect it? [00:30:51]
Kanjun: Really there are two answers to this. One element of it is as a person, like I need to basically get information out of the model such that I can try to understand what's going on with the model. Then the second question is like, okay, how do you do that? And that's kind of some of our debugging tools, they're not necessarily just for debugging. They're also for like interfacing with and interacting with the model. So like if I go back in this reasoning trace and like change a bunch of things, what's going to happen? Like, what does it conclude instead? So that kind of helps me understand like, what are its assumptions? And, you know, we think of these things as tools. And so it's really about like, as a user, how do I use this tool effectively? I need to be willing to be convinced as well. It's like, how do I use this tool effectively? And what can it help me with? [00:31:36]
Swyx: And what can it tell me? There's a lot of mention of code in your process. And I was hoping to dive in even deeper. I think we might run the risk of giving people the impression that you view code or you use code just as like a tool within InView just for coding assistance. But I think you actually train code models. And I think there's a lot of informal understanding about how adding code to language models improves their reasoning capabilities. I wonder if there's any research or findings that you have to share that talks about the intersection of code and reasoning. Hmm. Yeah. [00:32:08]
Kanjun: So the way I think about it intuitively is like code is the most explicit example of reasoning data on the internet. [00:32:15]
Swyx: Yeah. [00:32:15]
Kanjun: And it's not only structured, it's actually very explicit, which is nice. You know, it says this variable means this, and then it uses this variable. And then the function does this. As people, when we talk in language, it takes a lot more to extract that explicit structure out of our language. And so that's one thing that's really nice about code is I see it as almost like a curriculum for reasoning. I think we use code in all sorts of ways. The coding agents are really helpful for us to understand what are the limitations of the agents. The code is really helpful for the reasoning itself. But also code is a way for models to act. So by generating code, it can act on my computer. And, you know, when we talk about rekindling the dream of the personal computer, kind of where I see computers going is, you know, like computers will eventually become these much more malleable things where I, as a user today, I have to know how to write software code, like in order to make my computer do exactly what I want it to do. But in the future, if the computer is able to generate its own code, then I can actually interface with it in natural language. And so one way we think about agents is kind of like a natural language programming language. It's a way to program my computer in natural language that's much more intuitive to me as a user. And these interfaces that we're building are essentially IDEs for users to program our computers in natural language. Maybe I should say what we're doing that way. Maybe it's clearer. [00:33:47]
Swyx: I don't know. [00:33:47]
Alessio: That's a good pitch. What do you think about the different approaches people have, kind of like text first, browser first, like multi-on? What do you think the best interface will be? Or like, what is your, you know, thinking today? [00:33:59]
Kanjun: In a lot of ways, like chat as an interface, I think Linus, Linus Lee, you had on this. I really like how he put it. Chat as an interface is skeuomorphic. So in the early days, when we made word processors on our computers, they had notepad lines because that's what we understood these like objects to be. Chat, like texting someone is something we understand. So texting our AI is something that we understand. But today's word documents don't have notepad lines. And similarly, the way we want to interact with agents, like chat is a very primitive way of interacting with agents. What we want is to be able to inspect their state and to be able to modify them and fork them and all of these other things. And we internally have, think about what are the right representations for that? Like architecturally, like what are the right representations? What kind of abstractions do we need to build? And how do we build abstractions that are not leaky? Because if the abstractions are leaky, which they are today, like, you know, this stochastic generation of text is like a leaky abstraction. I cannot depend on it. And that means it's actually really hard to build on top of. But our experience and belief is actually by building better abstractions and better tooling, we can actually make these things non-leaky. And now you can build like whole things on top of them. So these other interfaces, because of where we are, we don't think that much about them. [00:35:17]
Swyx: Yeah. [00:35:17]
Alessio: I mean, you mentioned, this is kind of like the Xerox Spark moment for AI. And we had a lot of stuff come out of Parc, like the, what you see is what you got editors and like MVC and all this stuff. But yeah, but then we didn't have the iPhone at Parc. We didn't have all these like higher things. What do you think it's reasonable to expect in like this era of AI, you know, call it like five years or so? Like what are like the things we'll build today and what are things that maybe we'll see in kind of like the second wave of products? [00:35:46]
Kanjun: That's interesting. I think the waves will be much faster than before. Like what we're seeing right now is basically like a continuous wave. Let me zoom a little bit earlier. So people like the Xerox Parc analogy I give, but I think there are many different analogies. Like one is the like analog to digital computer is kind of an example, like another analogy to where we are today. The analog computer Vannevar Bush built in the 1930s, I think, and it's like a system of pulleys and it can only calculate one function. Like it can calculate like an integral. And that was so magical at the time because you actually did need to calculate this integral bunch, but it had a bunch of issues like in analog errors compound. And so there was actually a set of breakthroughs necessary in order to get to the digital computer, like Turing's decidability, Shannon. I think the like whole like relay circuits can be thought of as can be mapped to Boolean operators and a set of other like theoretical breakthroughs, which essentially were abstractions. They were like creating abstractions for these like very like lossy circuits. They were creating abstractions for these like very analog circuits and digital had this nice property of like being error correcting. And so when I talk about like less leaky abstractions, that's what I mean. That's what I'm kind of pointing a little bit to. It's not going to look exactly the same way. And then the Xerox PARC piece, a lot of that is about like, how do we get to computers that as a person, I can actually use well. And the interface actually helps it unlock so much more power. So the sets of things we're working on, like the sets of abstractions and the interfaces, like hopefully that like help us unlock a lot more power in these systems. Like hopefully that'll come not too far in the future. I could see a next version, maybe a little bit farther out. It's like an agent protocol. So a way for different agents to talk to each other and call each other. Kind of like HTTP. [00:37:40]
Swyx: Do you know it exists already? [00:37:41]
Kanjun: Yeah, there is a nonprofit that's working on one. I think it's a bit early, but it's interesting to think about right now. Part of why I think it's early is because the issue with agents, it's not quite like the internet where you could like make a website and the website would appear. The issue with agents is that they don't work. And so it may be a bit early to figure out what the protocol is before we really understand how these agents get constructed. But, you know, I think that's, I think it's a really interesting question. [00:38:09]
Swyx: While we're talking on this agent to agent thing, there's been a bit of research recently on some of these approaches. I tend to just call them extremely complicated chain of thoughting, but any perspectives on kind of meta-GPT, I think it's the name of the paper. I don't know if you care about at the level of individual papers coming out, but I did read that recently and TLDR, it beat GPT-4 and human eval by role-playing software agent development agency, instead of having sort of single shot or single role, you have multiple roles and how having all of them criticize each other as agents communicating with other agents. [00:38:45]
Kanjun: Yeah, I think this is an example of an interesting abstraction of like, okay, can I just plop in this like multi-role critiquing and see how it improves my agent? And can I just plop in chain of thought, tree of thought, plop in these other things and see how they improve my agent? One issue with this kind of prompting is that it's still not very reliable. It's like, there's one lens, which is like, okay, if you do enough of these techniques, you'll get to high reliability. And I think actually that's a pretty reasonable lens. We take that lens often. And then there's another lens that's like, okay, but it's starting to get really messy what's in the prompt and like, how do we deal with that messiness? And so maybe you need like cleaner ways of thinking about and constructing these systems. And we also take that lens. So yeah, I think both are necessary. Yeah. [00:39:29]
Swyx: Side question, because I feel like this also brought up another question I had for you. I noticed that you work a lot with your own benchmarks, your own evaluations of what is valuable. I would say I would contrast your approach with OpenAI as OpenAI tends to just lean on, hey, we played StarCraft or hey, we ran it on the SAT or the, you know, the AP bio test and that did results. Basically, is benchmark culture ruining AI? [00:39:55]
Swyx: Or is that actually a good thing? Because everyone knows what an SAT is and that's fine. [00:40:04]
Kanjun: I think it's important to use both public and internal benchmarks. Part of why we build our own benchmarks is that there are not very many good benchmarks for agents, actually. And to evaluate these things, you actually need to think about it in a slightly different way. But we also do use a lot of public benchmarks for like, is the reasoning capability in this particular way improving? So yeah, it's good to use both. [00:40:26]
Swyx: So for example, the Voyager paper coming out of NVIDIA played Minecraft and set their own benchmarks on getting the Diamond X or whatever and exploring as much of the territory as possible. And I don't know how that's received. That's obviously fun and novel for the rest of the engineer, the people who are new to the scene. But for people like yourselves, you build Avalon just because you already found deficiencies with using Minecraft. Is that valuable as an approach? Oh, yeah. I love Voyager. [00:40:57]
Kanjun: I mean, Jim, I think is awesome. And I really like the Voyager paper and I think it has a lot of really interesting ideas, which is like the agent can create tools for itself and then use those tools. [00:41:06]
Swyx: He had the idea of the curriculum as well, which is something that we talked about earlier. Exactly. [00:41:09]
Kanjun: And that's like a lot of what we do. We built Avalon mostly because we couldn't use Minecraft very well to like learn the things we wanted. And so it's like not that much work to build our own. [00:41:19]
Swyx: It took us, I don't know. [00:41:22]
Kanjun: We had like eight engineers at the time, took about eight weeks. So six weeks. [00:41:27]
Swyx: And OpenAI built their own as well, right? Yeah, exactly. [00:41:30]
Kanjun: It's just nice to have control over our environment. But if you're doing our own sandbox to really trying to inspect our own research questions. But if you're doing something like experimenting with agents and trying to get them to do things like Minecraft is a really interesting environment. And so Voyager has a lot of really interesting ideas in it. [00:41:47]
Swyx: Yeah. Cool. One more element that we had on this list, which is context and memory. I think that's kind of like the foundational, quote unquote, RAM of our era. I think Andrej Karpathy has already made this comparison. So there's nothing new here. And that's just the amount of working knowledge that we can fit into one of these agents. And it's not a lot, right? Especially if you need to get them to do long running tasks. If they need to self-correct from errors that they observe while operating in their environment. Do you see this as a problem? Do you think we're going to just trend to infinite context and that'll go away? Or how do you think we're going to deal with it? [00:42:22]
Kanjun: I think when you talked about what's going to happen in the first wave and then in the second wave, I think what we'll see is we'll get like relatively simplistic agents pretty soon. And they will get more and more complex. And there's like a future wave in which they are able to do these like really difficult, really long running tasks. And the blocker to that future, one of the blockers is memory. And that was true of computers too. You know, I think when von Neumann made the von Neumann architecture, he was like, the biggest blocker will be like, we need this amount of memory, which is like, I don't remember exactly like 32 kilobytes or something to store programs. And that will allow us to write software. He didn't say it this way because he didn't have these terms, but that only really was like happened in the seventies with the microchip revolution. It may be the case that we're waiting for some research breakthroughs or some other breakthroughs in order for us to have like really good long running memory. And then in the meantime, agents will be able to do all sorts of things that are a little bit smaller than that. I do think with the pace of the field, we'll probably come up with all sorts of interesting things like, you know, RAG is already very helpful. [00:43:26]
Swyx: Good enough, you think? [00:43:27]
Kanjun: Maybe good enough for some things. [00:43:29]
Swyx: How is it not good enough? I don't know. [00:43:31]
Kanjun: I just think about a situation where you want something that's like an AI scientist. As a scientist, I have learned so much about my fields and a lot of that data is maybe hard to fine tune or on, or maybe hard to like put into pre-training. Like a lot of that data, I don't have a lot of like repeats of the data that I'm seeing. You know, like if I'm a scientist, I've like accumulated so many little data points. And ideally I'd want to store those somehow, or like use those to fine tune myself as a model somehow, or like have better memory somehow. I don't think RAG is enough for that kind of thing. But RAG is certainly enough for like user preferences and things like that. Like what should I do in this situation? What should I do in that situation? That's a lot of tasks. We don't have to be a scientist right away. Awesome. [00:44:21]
Swyx: I have a hard question, if you don't mind me being bold. Yeah. I think the most comparable lab to InView is Adept. You know, a research lab with like some amount of product situation on the horizon, but not just yet, right? Why should people work for InView over Adept? And we can cut this if it's too like... Yeah. [00:44:40]
Kanjun: The way I think about it is I believe in our approach. The type of thing that we're doing is we're trying to like build something that enables other people to build agents and build something that really can be maybe something like an operating system for agents. I know that that's what we're doing. I don't really know what everyone else is doing. You know, I can kind of like talk to people and have some sense of what they're doing. And I think it's a mistake to focus too much on what other people are doing, because extremely focused execution on the right thing is what matters. To the question of like, why us? I think like strong focus on reasoning, which we believe is the biggest blocker, on inspectability, which we believe is really important for user experience and also for the power and capability of these systems. Building non-leaky, good abstractions, which we believe is solving the core issue of agents, which is around reliability and being able to make them deployable. And then really seriously trying to use these things ourselves, like every single day, and getting to something that we can actually ship to other people that becomes something that is a platform. Like, it feels like it could be Mac or Windows. I love the dogfooding approach. [00:45:49]
Swyx: That's extremely important. And you will not be surprised how many agent companies I talk to that don't use their own agent. Oh no, that's not good. That's a big surprise. [00:45:59]
Kanjun: Yeah, I think if we didn't use our own agents, then we would have all of these beliefs about how good they are. Wait, did you have any other hard questions you wanted to ask? [00:46:08]
Swyx: Yeah, mine was just the only other follow-up that you had based on the answer you just gave was, do you see yourself releasing models or do you see yourself, what is the artifacts that you want to produce that lead up to the general operating system that you want to have people use, right? And so a lot of people just as a byproduct of their work, just to say like, hey, I'm still shipping, is like, here's a model along the way. Adept took, I don't know, three years, but they released Persimmon recently, right? Like, do you think that kind of approach is something on your horizon? Or do you think there's something else that you can release that can show people, here's kind of the idea, not the end products, but here's the byproducts of what we're doing? [00:46:51]
Kanjun: Yeah, I don't really believe in releasing things to show people like, oh, here's what we're doing that much. I think as a philosophy, we believe in releasing things that will be helpful to other people. [00:47:02]
Swyx: Yeah. [00:47:02]
Kanjun: And so I think we may release models or we may release tools that we think will help agent builders. Ideally, we would be able to do something like that, but I'm not sure exactly what they look like yet. [00:47:14]
Swyx: I think more companies should get into the releasing evals and benchmarks game. Yeah. [00:47:20]
Kanjun: Something that we have been talking to agent builders about is co-building evals. So we build a lot of our own evals and every agent builder tells me, basically evals are their biggest issue. And so, yeah, we're exploring right now. And if you are building agents, please reach out to me because I would love to, like, figure out how we can be helpful based on what we've seen. Cool. [00:47:40]
Swyx: That's a good call to action. I know a bunch of people that I can send your way. Cool. Great. [00:47:43]
Kanjun: Awesome. [00:47:44]
Swyx: Yeah. We can zoom out to other interests now. [00:47:46]
Alessio: We got a lot of stuff. So we have Sherif from Lexicon, the podcast. He had a lot of interesting questions on his website. You similarly have a lot of them. Yeah. [00:47:55]
Swyx: I need to do this. I'm very jealous of people with personal websites right there. Like, here's the high level questions of goals of humanity that I want to set people on. And I don't have that. [00:48:04]
Alessio: It's never too late, Sean. [00:48:05]
Swyx: Yeah. [00:48:05]
Alessio: It's never too late. [00:48:06]
Kanjun: Exactly. [00:48:07]
Alessio: There were a few that stuck out as related to your work that maybe you're kind of learning [00:48:12]
Swyx: more about it. [00:48:12]
Alessio: So one is why are curiosity and goal orientation often at odds? And from a human perspective, I get it. It's like, you know, would you want to like go explore things or kind of like focus on your career? How do you think about that from like an agent perspective? Where it's like, should you just stick to the task and try and solve it as in the guardrails as possible? Or like, should you look for alternative solutions? [00:48:34]
Swyx: Yeah. [00:48:34]
Kanjun: I think one thing that's really interesting about agents actually is that they can be forked. Like, you know, we can take an agent that's executed to a certain place and said, okay, here, like fork this and do a bunch of different things. I try a bunch of different things. Some of those agents can be goal oriented and some of them can be like more curiosity driven. You can prompt them in slightly different ways. And something I'm really curious about, like what would happen if in the future, you know, we were able to actually go down both paths. As a person, why I have this question on my website is I really find that like I really can only take one mode at a time and I don't understand why. And like, is it inherent in like the kind of context that needs to be held? That's why I think from an agent perspective, like forking it is really interesting. Like I can't fork myself to do both, but I maybe could fork an agent to like add a certain point in a task. [00:49:26]
Swyx: Yeah. Explore both. Yeah. [00:49:28]
Alessio: How has the thinking changed for you as the funding of the company changed? That's one thing that I think a lot of people in the space think is like, oh, should I raise venture capital? Like, how should I get money? How do you feel your options to be curious versus like goal oriented has changed as you raise more money and kind of like the company has grown? [00:49:50]
Kanjun: Oh, that's really funny. Actually, things have not changed that much. So we raised our Series A $20 million in late 2021. And our entire philosophy at that time was, and still kind of is, is like, how do we figure out the stepping stones, like collect stepping stones that eventually let us build agents, kind of these new computers that help us do bigger things. And there was a lot of curiosity in that. And there was a lot of goal orientation in that. Like the curiosity led us to build CARBS, for example, this hyperparameter optimizer. Great name, by the way. [00:50:28]
Swyx: Thank you. [00:50:29]
Kanjun: Is there a story behind that name? [00:50:30]
Swyx: Yeah. [00:50:31]
Kanjun: Abe loves CARBS. It's also cost aware. So as soon as he came up with cost aware, he was like, I need to figure out how to make this work. But the cost awareness of it was really important. So that curiosity led us to this really cool hyperparameter optimizer. That's actually a big part of how we do our research. It lets us experiment on smaller models. And for those experiment results to carry to larger ones. [00:50:56]
Swyx: Which you also published a scaling laws, which is great. I think the scaling laws paper from OpenAI was like the biggest. And from Google, I think, was the greatest public service to machine learning that any research lab can do. Yeah, totally. [00:51:10]
Kanjun: What was nice about CARBS is it gave us scaling laws for all sorts of hyperparameters. So yeah, that's cool. It basically hasn't changed very much. So there's some curiosity. And then there's some goal oriented parts. Like Avalon, it was like a six to eight week sprint for all of us. And we got this thing out. And then now different projects do like more curiosity or more goal orientation at different times. Cool. [00:51:36]
Swyx: Another one of your questions that we highlighted was, how can we enable artificial agents to permanently learn new abstractions and processes? I think this is might be called online learning. [00:51:45]
Kanjun: Yeah. So I struggle with this because, you know, that scientist example I gave. As a scientist, I've like permanently learned a lot of new things. And I've updated and created new abstractions and learned them pretty reliably. And you were talking about like, okay, we have this RAM that we can store learnings in. But how well does online learning actually work? And the answer right now seems to be like, as models get bigger, they fine tune faster. So they're more sample efficient as they get bigger. [00:52:15]
Swyx: Because they already had that knowledge in there. You're just kind of unlocking it. [00:52:23]
Kanjun: Partly maybe because they already have like some subset of the representation. Partly they just memorize things more, which is good. So maybe this question is going to be solved, but I still don't know what the answer is. [00:52:36]
Swyx: As I've had a platform that continually fine tunes for you as you work on that domain, which is something I'm working on. Well, that's great. We would love to use that. We'll talk more. Two more questions just about your general activities. I think you've just been very active in the San Francisco tech scene. You're a founding member of Software Commons. [00:52:56]
Kanjun: Oh yeah, that's true. [00:52:57]
Swyx: Tell me more. By the time I knew about SPC, it was already a very established thing. But what was it like in the early days? What was the story there? [00:53:05]
Kanjun: Yeah, the story is Ruchi, who started it, was the VP of operations at Dropbox. And I was the chief of staff and we worked together very closely. She's actually one of the investors in Sorceress. And SPC is an investor in Vue. And at that time, Ruchi was like, you know, I would like to start a space for people who are figuring out what's next. And we were figuring out what's next post-Ember, those three months. And she was like, do you want to just like hang out in this space? And we're like, sure. And it was a really good group. Wasim and Jeff from Pilot, the folks from Zulip, and a bunch of other people at that time. It was a really good group. We just hung out. There was no programming. It's much more official than it was at that time. [00:53:44]
Swyx: Yeah, now it's like a YC before YC type of thing. That's right, yeah. [00:53:48]
Kanjun: At that time, we literally, it was a bunch of friends hanging out in the space together. [00:53:51]
Swyx: And was this concurrent with the Archive? [00:53:53]
Kanjun: Oh yeah, actually, I think we started the Archive around the same time. [00:53:56]
Swyx: You're just like really big into community. But also like, so, you know, I run a Hacker House and I'm also part of hopefully what becomes like the next Software Commons or whatever. What are the principles in organizing communities like that with really exceptional people that go on to do great things? Do you have to be really picky about who joins? Like all your friends just magically turn out super successful like that. You know, it's not normal, right? Like this is very special. And a lot of people want to do that and fail. And you had the co-authors of GPT-3 in your house. That's true. [00:54:32]
Kanjun: And a lot of other really cool people that you'll eventually hear about. [00:54:35]
Swyx: Co-founders of Pilot and anyone else. I don't want you to pick your friends, but there's some magic special sauce in getting people together and in one workspace, living space, whatever, right? And that's part of why I'm here in San Francisco. And I would love for more people to learn about it and also maybe get inspired to build their own. [00:54:52]
Kanjun: Your question is really more about like, how do you actually build a community that where people in it are like eventually are awesome? [00:54:59]
Swyx: Okay. [00:55:00]
Kanjun: Which is different than like why live in a co-living house. So one adage we had when we started the archive was you become the average of the five people closest to you. [00:55:08]
Swyx: Yes. [00:55:08]
Kanjun: And I think that's roughly true. And good people draw good people. So there are really two things. One, we were quite picky and it mattered a lot to us. Is this someone where if they're hanging out in the living room, we'd be really excited to come hang out. Yeah. Two is I think we did a really good job of creating a high growth environment and an environment where people felt really safe. We actually apply these things to our team and it works remarkably well as well. So I do a lot of basically how do I create safe spaces for people where it's not just like safe law, but like it's like a safe space where people really feel inspired by each other. And I think at the archive, we really made each other better. My friend, Michael Nielsen called it a self-actualization machine. [00:55:52]
Swyx: My goodness. Okay. [00:55:54]
Kanjun: And I think, yeah, people came in. Was he a part of the archive? He was not, but he hung out a lot. Honorary member. Friend of the archive. [00:56:02]
Swyx: Yeah. [00:56:02]
Kanjun: The culture was that we learned a lot of things from each other about like how to make better life systems and how to think about ourselves and psychological debugging. And a lot of us were founders. So having other founders going through similar things was really helpful. And a lot of us worked in AI. And so having other people to talk about AI with was really helpful. And so I think all of those things led to a form of idea flux and also kind of like, so I think a lot about like the idea flux and default habits or default impulses. It led to a set of idea flux and default impulses that led to some really interesting things and led to us doing much bigger things, I think, than we otherwise would have decided to do because it felt like taking risks was less risky. So that's something we do a lot of on the team. It's like, how do we make it so that taking risks is less risky? And there's a term called senious. [00:56:57]
Swyx: Yes. I was thinking Kevin Kelly. Kevin Kelly, senious. I was going to feed you that word, but I didn't want to like bias you. Yes. [00:57:02]
Kanjun: I think maybe like a lot of what I'm interested in is constructing a kind of senious. And the archive was definitely a senious in a particular, or like getting toward a senious in a particular way. And Jason Ben, my archive housemate and who now runs the neighborhood, [00:57:17]
Swyx: has a good way of putting it. [00:57:17]
Kanjun: If genius is from your genes, senious is from your scene. Yeah, I think like maybe a lot of the community building impulse is from this like interest in what kind of idea flux can be created. You know, there's a question of like, why did Xerox PARC come out with all of this interesting stuff? It's their senious. Why did Bell Labs come out with all this interesting stuff? Maybe it's their senious. Why didn't the transistor come out of Princeton? And the other people working on it at the time. [00:57:44]
Swyx: I just think it's remarkable how you hear a lot about Alan Kay. And I just read a bit. And apparently Alan Kay was like the most junior guy at Xerox PARC. Yeah, definitely. [00:57:53]
Kanjun: He's just the one who talks about it. He talks the most. [00:57:57]
Swyx: Yeah, exactly. Yeah. So I, you know, hopefully I'm also working towards contributing that senious. I called mine the more provocative name of the arena. Interesting. That's quite provocative. In the arena. [00:58:08]
Kanjun: So are you fighting other people in the arena? [00:58:11]
Swyx: No. You never know. [00:58:12]
Alessio: On any day in the mission, it's an adventure. [00:58:15]
Swyx: We're in the arena trying stuff, as they say. You are also a GP at Outset Capital, where you also co-organize the Thursday Nights in AI, where hopefully someday I'll eventually speak. You're on the roster. [00:58:28]
Kanjun: I'm on the roster. [00:58:29]
Swyx: Thank you so much. So why spend time being a VC and organizing all these events? You're also a very busy CEO and, you know, why spend time with that? Why is that an important part of your life? [00:58:39]
Kanjun: Yeah, for me personally, I really like helping founders. So Allie, my investing partner, is fortunately amazing and she does everything for the fund. So she like hosts the Thursday night events and she finds folks who we could invest in. And she does basically everything. Josh and I are her co-partners. So Allie was our former chief of staff at Sorceress. We just thought she was amazing. She wanted to be an investor. And Josh and I also like care about helping founders and kind of like giving back to the community. What we didn't realize at the time when we started the fund is that it would actually be incredibly helpful for Imbue. So talking to AI founders who are building agents and working on, you know, similar things is really helpful. They could potentially be our customers and they're trying out all sorts of interesting things. And I think being an investor, looking at the space from the other side of the table, it's just a different hat that I routinely put on. And it's helpful to see the space from the investor lens as opposed to from the founder lens. So I find that kind of like hat switching valuable. It maybe would lead us to do slightly different things. [00:59:44]
Swyx: Awesome. Appreciate that. [00:59:46]
Alessio: Yeah, you've been really generous with your time. Let's just wrap with the lightning round. Okay. So we have two questions, acceleration, exploration, and then a takeaway. So the acceleration question is, what's something that already happened in AI that you thought would take much longer to be here? [01:00:03]
Kanjun: I think the rate at which we discover new capabilities of existing models and kind of like build hacks on top of them to make them work better is something that has been surprising and awesome. And the research community building on its own ideas, that's probably, you want something very specific. Yeah, I think the rate at which we discovered capabilities probably. [01:00:22]
Swyx: Cool. Exploration slash requests for startups. If you weren't building Imbue, what AI company would you build? Hmm. Every founder has like their like number two. Really? Yeah, I don't know. [01:00:33]
Kanjun: Wow. I cannot imagine building any other thing than Imbue. [01:00:37]
Swyx: Wow. Well, that's a great answer too. [01:00:38]
Kanjun: It's like obviously the thing to build. [01:00:42]
Swyx: Okay. [01:00:42]
Kanjun: It's like obviously work on the fundamental platform. Yeah. [01:00:46]
Swyx: So that was my attempt at innovating this question, but the previous one was, but what was the most interesting unsolved question in AI? [01:00:53]
Kanjun: My answer is kind of boring, but the most interesting unsolved questions are these questions of, how do we make these stochastic systems into things that we can like reliably use and build on top of? [01:01:04]
Swyx: Yep. [01:01:05]
Alessio: And yeah, take away what's one message you want everyone to remember? [01:01:09]
Kanjun: Maybe two things. One is just the like, we're in a historic moment. I didn't think in my lifetime I would necessarily be in, like able to work on the things I'm excited to work on in this moment, but we're in a historic moment that where we'll look back and be like, oh my God, the future was invented in these years. And I think like, there may be a set of messages to take away from that. One is like, AI is a tool like any technology. And you know, when it comes to things like, what might the future look like? Like we like to think about it as, it's like just a better computer. It's like much more powerful computer that gives us a lot of free intellectual energy that we can now like solve so many problems with. You know, there are so many problems in the world [01:01:53]
Swyx: where we're like, [01:01:53]
Kanjun: oh, it's not worth a person thinking about that. And so things get worse and things get worse. No one wants to work on maintenance. And like this technology gives us the potential to actually be able to like allocate intellectual energy to all of those problems. And the world could be much better, like could be much more thoughtful because of that. I'm so excited about that. And there are definitely risks and dangers. And we actually do a fair, something I didn't talk about is we do a fair amount of work on the policy side. On the safety side, like we think about safety and policy in terms of engineering theory and also regulation. And kind of comparing to like the automobile or the airplane or any new technology, there's like a set of new possible capabilities and a set of new possible dangers that are unlocked with every new technology. And so on the engineering side, like we think a lot about engineering safety, like how do we actually engineer these systems so that they are inspectable and why we reason in natural language so that the systems are very inspectable so that we can like stop things if anything weird is happening. That's why we don't think end-to-end black boxes [01:02:58]
Swyx: are a good idea. [01:02:58]
Kanjun: On the theoretical side, we like really believe in like deeply understanding, like when we actually fine tune on individual examples, like what's going on, when we're pre-training, what's going on, like debugging tools for these agents to understand like what's going on. And then on the regulation side, I think there's actually a lot of regulation that already covers many of the dangers like that people are talking about. And there are areas where there's not much regulation. And so we focus on those areas where there's not much regulation. So some of our work is actually, we built an agent that helped us analyze the 20,000 pages of policy proposals submitted to the Department of Commerce request for AI policy proposals. We looked at what were the problems people brought up and what were the solutions they presented and then like did a summary analysis and kind of like, you know, build agents to do that. And now the Department of Commerce is like interested in using that as a tool to like analyze proposals. And so a lot of what we're trying to do on the regulation side is like actually figure out where is there regulation missing and how do we actually in a very targeted way try to solve those missing areas. So I guess if I were to say like, what are the takeaways? It's like the takeaway is like the future could be really exciting if we can actually get agents that are able to do these bigger things. Reasoning is the biggest blocker plus like these sets of abstractions to make things more robust and reliable. And there are, you know, things where we have to be quite careful and thoughtful about how do we deploy these and what kind of regulation should go along with it so that this is actually a technology that where we, when we deploy it, it is protective to people and not harmful. [01:04:36]
Swyx: Awesome, wonderful. [01:04:38]
Alessio: Thank you so much for your time, Kanjun. [01:04:40]
Kanjun: Thank you. [01:04:41]
Swyx: Thank you. [01:04:48]
This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.latent.space/subscribe
.jpg?width=400){kind=small}