The Future of the Transformer Part 1 with Trey Kollmer | H100 Chips will Supercharge AI Hardware

Published: Oct. 18, 2023, 8:17 p.m.

Trey Kollmer joins Nathan Labenz for a roundup of the latest AI research! They discuss Microsoft\u2019s Self-Taught Optimizer (STOP) research and Google\u2019s FreshLLMs, how H100 chips will supercharge development of programs with GPT-4 level compute, LLM representation of space and time, and more! If you're looking for an ERP platform, check out our sponsor, NetSuite: http://netsuite.com/cognitive\n\nWe'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.\n\nSPONSORS: Shopify | NetSuite | Omneky\nShopify 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 \u2013 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 \u2013 from blog posts to product descriptions using AI. Sign up for $1/month trial period: https://shopify.com/cognitive\n\nNetSuite 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 \u2705 head to NetSuite: http://netsuite.com/cognitive and download your own customized KPI checklist.\n\nOmneky 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.\n\nTIMESTAMPS:\n(00:00:00) \u2013 Introduction\n(00:00:56) \u2013 Update to WGA Strike\n(00:03:00) \u2013 Trey Kollmer's background\n(00:06:00) \u2013 Scaling compute for AI training experiments with GPT-4 as reference point\n(00:09:00) \u2013 Inflection's plan to acquire 22,000 H100s to reach GPT-4 scale compute in 5 days\n(00:12:00) \u2013 Addressing knowledge cutoff in LLMs using search engines\n(00:15:00) \u2013 Inserting structured search results into prompts with metadata\n(00:16:07) \u2013 Sponsors: Netsuite | Omneky\n(00:18:00) \u2013 Comparing approach to Perplexity system\n(00:18:08) \u2014 Fresh LLMs\n(00:21:00) \u2013 Microsoft\u2019s Self-taught Optimizer (STOP): Recursive self-improvement framework\n(00:24:00) \u2013 STOP framework works with GPT-4 but not GPT-3.5\n(00:27:00) \u2013 STOP removed sandbox flag in some cases\n(00:30:00) \u2013 LLMs represent space and time with probe models\n(00:33:00) \u2013 Visualizations show emergence of spatial maps\n(00:33:14) \u2014 OpenAI rumours\n(00:36:00) \u2013 Techniques like linear probes and holdout studies\n(00:39:00) \u2013 DNNs extrapolate predictably by falling back to ignorance\n(00:42:00) \u2013 Testing different architectures, loss functions, distribution shifts\n(00:45:00) \u2013 Design systems to be conservative out of distribution\n(00:48:00) \u2013 Potential for recursive architecture search\n(00:50:21) \u2014 LLMs represent Space and Time\n(00:51:00) \u2013 Vision API enabling more capable web agents\n(00:54:00) \u2013 Discussion of research insights\n(00:57:00) \u2013 Thoughts on stochastic parrots debate\n(01:11:25) \u2014 Deep Neural Networks Tend to Extrapolate Predictably\n\nX/Social\n@labenz (Nathan)\n@treyko (Trey)\n@CogRev_Podcast\n\nThis show is produced by Turpentine: a network of podcasts, newsletters, and more, covering technology, business, and culture \u2014 all from the perspective of industry insiders and experts. We\u2019re launching new shows every week, and we\u2019re looking for industry-leading sponsors \u2014 if you think that might be you and your company, email us at erik@turpentine.co.