Are Emergent Behaviors in LLMs an Illusion? with Sanmi Koyejo - #671
Today we’re joined by Sanmi Koyejo, assistant professor at Stanford University, to continue our NeurIPS 2024 series. In our conversation, Sanmi discusses his two recent award-winning papers. First, we dive into his paper, “Are Emergent Abilities of Large Language Models a Mirage?”. We discuss the different ways LLMs are evaluated and the excitement surrounding their“emergent abilities” such as the ability to perform arithmetic Sanmi describes how evaluating model performance using nonlinear metrics can lead to the illusion that the model is rapidly gaining new capabilities, whereas linear metrics show smooth improvement as expected, casting doubt on the significance of emergence. We continue on to his next paper, “DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models,” discussing the methodology it describes for evaluating concerns such as the toxicity, privacy, fairness, and robustness of LLMs.
The complete show notes for this episode can be found at twimlai.com/go/671.
198 - SUAVE Explained with Phil Daian & Andrew Miller
Phil Daian is a crypto-economic researcher! Phil is the lead author behind the landmark paper, “Flash Boys 2.0,” which introduced and defined the MEV problem in the Ethereum landscape, over 4 years ago. He is the cofounder of FlashBots, which is a research and dev organization with the mission of mitigating the negative externalities of MEV.
Andrew Miller is an Associate Professor at the University of Illinois, SGX Bull, and visiting researcher at Flashbots, and big believer in cypherpunk values through crypto-economic mechanisms.
------ ✨ DEBRIEF | Ryan & David unpacking the episode: https://www.bankless.com/debrief-suave
Have you ever heard of Flashbots? Flashbots is the silent protector of Ethereum, a shield against the forces of centralization that silently sneak into our cryptosystems and turn them into TradFi. It’s one of the unsung heroes of crypto.
The guests we have on today are from Flashbots, Phil Daian, and Andrew Miller. They believe the biggest centralization threats in crypto still lie ahead. And they’ve got a new platform that’s being built to defend against them. It’s called “Suave”, and we’ve been hearing a lot about it recently - “Solving MEV” - “Decentralizing block building” - but going into this episode, we weren’t sure exactly what it was. We found out during this episode.
----- 🏹 Airdrop Hunter is HERE, join your first HUNT today https://bankless.cc/JoinYourFirstHUNT
------ BANKLESS SPONSOR TOOLS:
🐙KRAKEN | MOST-TRUSTED CRYPTO EXCHANGE https://k.xyz/bankless-pod-q2
🦊METAMASK PORTFOLIO | MANAGE YOUR WEB3 EVERYTHING https://bankless.cc/MetaMask
⚖️ARBITRUM | SCALING ETHEREUM https://bankless.cc/Arbitrum
🔗CELO | CEL2 COMING SOON https://bankless.cc/Celo
👾GMX | V2 IS NOW LIVE https://bankless.cc/GMX
💲 USDV | NATIVE OMNICHAIN STABLECOIN https://bankless.cc/usdv
------ TIMESTAMPS
0:00 Intro 8:20 History of Flashbots & MEV 15:35 MEV Boost 19:25 Andrew’s Journey 23:20 What’s at Stake? 25:00 Solving MEV 33:44 What’s Suave? 41:13 Intents 45:25 Suavechain? 54:53 Decentralized Block Building 1:02:27 Privacy of Suave 1:08:02 Suave’s Importance & Goal 1:12:40 L2s 1:15:18 Alt-L1s 1:18:00 MEV Naiveness 1:21:31 Benefits of Suave 1:30:36 Suave Roadmap 1:32:23 Why the World Needs Suave 1:38:25 Closing & Disclaimers ------ RESOURCES
How to get Started Building on Suave https://collective.flashbots.net/t/suave-wiki/2018
Learn about MEV Crypto’s Existential Threat https://www.youtube.com/watch?v=rOVz7dOrGyY Flashbots Saves Crypto https://www.youtube.com/watch?v=xex9WVmVfbg
Learn about Blockchain Supply Chains Ethereum’s Hidden Power Structures https://www.youtube.com/watch?v=rQxuPBy8tJg
Learn about Moloch Slaying Moloch https://www.youtube.com/watch?v=903tHM4RA9k Why we can’t have nice things? https://www.youtube.com/watch?v=YeQ2_ep0Cg4
------ Not financial or tax advice. See our investment disclosures here: https://www.bankless.com/disclosures
Cryptography Will Revolutionize AI Data Privacy with Daniel Kang
In this episode, Nathan sits down with Daniel Kang, Assistant Professor of Computer Science at the University of Illinois. Kang has done pioneering work bringing zero knowledge cryptographic proofs to AI. In this episode, they chat about the cryptographic theory behind Daniel's work, how cryptography allows us to balance the tradeoff between privacy and authenticity, and how cryptography usage is needed in a world where LLMs are increasingly embedded into our daily lives. If you're looking for an ERP platform, check out our sponsor, NetSuite: http://netsuite.com/cognitive
We're hiring across the board at Turpentine and for Erik's personal team on other projects he's incubating. He's hiring a Chief of Staff, EA, Head of Special Projects, Investment Associate, and more. For a list of JDs, check out: eriktorenberg.com.
CALL FOR FEEDBACK:
To borrow from a meme… we're in the podcast arena trying stuff. Some will work. Some won't. But we're always learning.
http://bit.ly/TCRFeedback
Fill out the above form to let us know how we can continue delivering great content to you or sending the feedback on your mind to tcr@turpentine.co.
TIMESTAMPS:
(00:00) Episode Preview
(00:01:04) Nathan's Introduction
(00:07:06) Motivation for bringing zero-knowledge proofs to AI
(00:10:27) Verifying model execution without revealing model details
(00:12:42) Verifying medical AI services haven't been tampered with
(00:15:09) Sponsors: Netsuite | Omneky
(00:18:54) Cryptographic hashes for commitments
(00:22:42) Assumptions underlying cryptographic hashes
(00:25:20) Adding entropy through salting
(00:26:24) Z case snarks and the proving process
(00:31:00) Using lookup tables for nonlinearities
(00:33:35) Floating point vs fixed point calculations
(00:34:08) Quantizing models for efficiency
(00:35:55) Using polynomials to represent arbitrary computations
(00:37:26) What are finite fields?
(00:41:23) Toxic waste for cryptographic secrecy
(00:45:51) Computational costs
(00:47:39) The experience of using a cryptography application to verify model output
(00:49:05) Verification key doesn't reveal model weights
(00:56:36) What using crypto infrastructure in AI enables and challenges to its implementation
(01:01:26) Potential for 10-100x cost reductions
(01:04:51) Authenticating images with attested cameras
(01:11:56) How cryptography in AI could impact daily life
(01:14:25) On-device credential verification
(01:15:50) Potential for regulation of hardware authentication
(01:18:52) Upcoming work to reduce proof costs
LINKS:
Daniel's website
SPONSORS: Shopify | NetSuite | Omneky
Shopify is the global commerce platform that helps you sell at every stage of your business. Shopify powers 10% of ALL eCommerce in the US. And Shopify's the global force behind Allbirds, Rothy's, and Brooklinen, and 1,000,000s of other entrepreneurs across 175 countries.From their all-in-one e-commerce platform, to their in-person POS system – wherever and whatever you're selling, Shopify's got you covered. With free Shopify Magic, sell more with less effort by whipping up captivating content that converts – from blog posts to product descriptions using AI. Sign up for $1/month trial period: https://shopify.com/cognitive
NetSuite has 25 years of providing financial software for all your business needs. More than 36,000 businesses have already upgraded to NetSuite by Oracle, gaining visibility and control over their financials, inventory, HR, eCommerce, and more. If you're looking for an ERP platform ✅ head to NetSuite: http://netsuite.com/cognitive and download your own customized KPI checklist.
Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off.
Music Credit: GoogleLM
A Historic Disruption To The World's Wheat Supply
When it comes to commodities, Russia's invasion of Ukraine has had a notable impact on the price of oil and natural gas. It's also a huge deal for wheat, and food more broadly. Ukraine is a massive player in the global wheat market, and the planting season is basically right now. What's more, Russia is also a big wheat seller, and Belarus is a big supplier of fertilizer. On this episode of Odd Lots, Tracy Alloway and Joe Weisenthal speak with Scott Irwin, an agriculture economist at the University of Illinois, about what he calls the biggest disruption he's seen in his career.
See omnystudio.com/listener for privacy information.
How Chinese Buying Is Causing a Boom in Agricultural Commodities
There are lots of hot areas in the market, which everybody knows. Stocks are obviously hot, as are industrial commodities like copper. Agricultural commodities are surging as well. If you look at a chart of corn or soy or even oats, they've been on a tear. One big factor: Chinese demand, in part driven by a desire to stock up on supplies. Meanwhile, China is launching agricultural futures of its own, including a new contract on hogs. On this episode, we speak about what's going on right now in agricultural commodities with Scott Irwin, an economist at the University of Illinois, who helps us break it all down.
See omnystudio.com/listener for privacy information.
Metric Elicitation and Robust Distributed Learning with Sanmi Koyejo - #352
The unfortunate reality is that many of the most commonly used machine learning metrics don't account for the complex trade-offs that come with real-world decision making. This is one of the challenges that Sanmi Koyejo, assistant professor at the University of Illinois, has dedicated his research to address. Sanmi applies his background in cognitive science, probabilistic modeling, and Bayesian inference to pursue his research which focuses broadly on “adaptive and robust machine learning.”