GPT-6 Astra Fixed Another AI by Editing 0.016% of Weights
Om avsnittet
MIT researchers let GPT-6 Astra edit a smaller model's weights directly, and it broke the model's endless reasoning loop by changing just 0.016% of them, with no retraining. If AI can be patched like software, the multimillion-dollar economics of improving models are in question.
• DeepValues app, a research team in your pocket: intrinsic value and margin of safety for any stock, reports down to the footnotes
• Is the market overvalued? The Buffett Indicator app
Android (Buffett Indicator & Shiller PE)
Chapters
(0:00) Weight-Space Programming Overview
(4:36) Neural Networks as Software Artifacts
(5:35) Microscopic Edits, Massive Gains
(6:42) Repairing Capabilities vs Creating
(9:05) Nvidia's Compute Paradox
(10:49) Tail Risk: Neoclouds and Debt
(13:07) Apple's Advantage in Edge AI
(14:27) Tech Giants' Strategic Positioning
(15:26) Nvidia's Compute Paradox
(17:12) Model Supply Chain Security
(19:14) Verdict and What to Watch
(19:39) Get the apps
DeepValues toolkit
• Run the numbers on any stock yourself, free
• Free Margin of Safety Weekly, every Saturday (guru moves, market temperature, the week's research; no account needed)
• Open-source research on 300+ companies
• Deep Value Intelligence, the multi-agent AI analyst
• Developer API and MCP server for your AI agents
Deep Values explains AI from first principles: the research papers, models, chips and companies shaping the AI era, and what they mean for builders and investors. Understand AI from first principles.
Podcast (Deep Values Frontier): Spotify · Apple Podcasts
The two hosts' voices are AI-generated from the sources discussed. Educational content only; not investment advice.
#AI #ArtificialIntelligence #LLM
Deep Values Frontier: AI Research Explained med DeepValues finns tillgänglig på flera plattformar. Informationen på denna sida kommer från offentliga podd-flöden.