Watch Episode Here
Listen to Episode Here
Show Notes
In Part 2 of The AI Scouting Report, Nathan Labenz builds on Part 1: Fundamentals of AI (https://www.youtube.com/watch?v=0hvtiVQ_LqQ&t=3026s) and delves into recent trends and practical applications for AI. Nathan's aim is to impart the equivalent of a high school AP course understanding to listeners in 90 minutes. If you're looking for an ERP platform, check out our sponsor, NetSuite: http://netsuite.com/COGNITIVE
Get your questions answered in the podcast by emailing TCR@turpentine.co
The Cognitive Revolution is a part of the Turpentine podcast network. Learn more: Turpentine.co
TIMESTAMPS:
(00:00) Episode Preview and Recap of Pt 1
(02:35) New Custom Hardware
(07:51) Startups vs incumbents in the foundational model space
(09:55) History of fine tuning
(15:17) Sponsor: NetSuite
(22:08) Fine tuning, illustrated
(27:10) Mode collapse
(28:41) Better data, better models
(31:34) Chat UI - Assistants and beyond
(34:26) Erik and Nathan's favorite chatbots
(35:24) Claude 2’s advantages
(36:47) Character AI
(38:11) AI girlfriend story on LessWrong
(40:43) Replika's new romance app
(42:08) Trends in implementation
(42:50) Chain of thought prompting
(46:07) Memory & Retrieval with embeddings
(47:18) The trouble with hallucinations
(47:47) Vector store
(49:02) Self-teaching Tool Use and having language models call APIs
(53:01) Best AI-enabled search experience and Perplexity AI
(56:08) AI agents
(01:00:05) Incumbents adding AI vs AI-native companies
(01:06:00) Creating the best human plus AI bundle
(01:08:37) LLM as a planner and LLMs producing aspirin
(01:13:47) LLM as a lifelong learner and LLMs playing Minecraft
(01:18:36) LLMs reflecting on memories and past experiences to make decisions
(01:22:13) The great embedding: HD model-to-model communication
(01:26:40) Multi-modality
(01:28:50) Multi-modal outputs
(01:31:13) Gato, You The Real Agent
(01:31:46) Efficiency measures: Quantization, Distillation
(01:36:55) Efficiency measures: Mixture of Experts
(01:44:22) Sponsor: Omneky
TWITTER:
@labenz (Nathan)
@eriktorenberg (Erik)
SPONSORS:
NetSuite provides financial software for all your business needs. More than 36,000 companies have already upgraded to NetSuite, gaining visibility and control over their financials, inventory, HR, eCommerce, and more. If you're looking for an ERP platform, take advantage of a special financing offer and defer payments of a FULL NetSuite implementation for six months. ✅ NetSuite: http://netsuite.com/COGNITIVE
Thank you Omneky (www.omneky.com) for sponsoring The Cognitive Revolution. 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.
SPONSOR:
Music: GoogleLM
Music license:
8AI8UGU7SZ0SWPDB
Full Transcript
Transcript
Nathan Labenz: (0:00) Welcome back, weekend warriors, to the AI Scouting Report part 2. What's up, Erik?
Erik Torenberg: (0:07) Hey, Nathan. Stoked to get into part 2. Just to set ourselves up for where we left off: part 1 was all about setting up how we got to this AI moment. I've really come to believe more and more recently that given the existence of web-scale data and web-scale compute, it was really only a matter of time before somebody figured out an algorithm that was going to work. And the algorithm that we figured out first that seems to be generalizing to just about everything that we try to use it for is the transformer. We went through the architecture of the transformer and covered a lot of the jargon. And for me, the most interesting part was getting into what we can really say about what they do and don't know. Obviously, that led us to a discussion of grokking, which is this generalization and apparently very real, even if sometimes alien, conceptual understanding that is definitely more than just memorization. We can see that because we're able to, in some cases, even reverse engineer the algorithm that the transformer has learned and break it down and show that this really is a general solution to the problem.
So that brings us to this interesting moment where these frontier language models like GPT-4 know a lot of stuff, and it's unclear in many cases whether they know that because they've memorized statistical correlations and they're still operating in the stochastic paradigm, or whether they've truly grokked conceptually whatever the subject matter is and have really learned something in a robust way. So that debate rages on. And I think every month as we look at updating this, already since the publication of part 1, there have been some incremental grokking results, and people continue to really explore what they do and don't know—where are they guessing and where are they really coming to understand conceptually what we're asking them to do.
But today, part 2 is going to be much more practical, much more applied. Now that we've got these large-scale language models like GPT-4, what can we do with them, and how do we overcome the weaknesses that they have so that they can actually start to do really useful stuff for us? And that is where this recent trends part kicks off. How's that sound?
Nathan Labenz: (2:37) Sounds great. Let's get into it.
Erik Torenberg: (2:39) You will need to watch the video to get the full experience. Now without further ado, follow the link to our YouTube channel for my AI scouting report.
Read next
