AI Builders Digest

Sutton: LLMs are a breakthrough, but only 20% of intelligence

Rich Sutton and his co-founder Khurram Javed argue that today's LLMs are a dead end because they stop learning after deployment. Sutton calls synthetic data 'a big mistake' — the real world is infinitely complex, and no simulation can capture it. His lab, Oak, is betting on continual deep learning with per-weight step sizes and a 'generate and test' approach to grow new abstractions over time. He estimates a truly intelligent system can run on 2,000 watts today, scaling to 20 watts in a decade.

• #agents
• #research
• #open-source

Levie: Better open weights enable vertical model plays

Box CEO Aaron Levie argues that as open-weight models improve and post-training infrastructure matures, companies with proprietary data can now train their own specialized models without competing with frontier labs on research. While general-purpose models will still lead on breadth, he predicts far more domain-specific models emerging across every vertical.

• #open-source
• #funding
• #products

Rauch: Your next design system is Markdown

Vercel CEO Guillermo Rauch announces that DESIGN.md — a Markdown-based design system — is helping solve AI slop and scale design taste across large organizations. He also notes that AI Gateway now supports per-user budgets for coding tokens, addressing the governance problem of letting developers 'go wild' with API keys.

• #products
• #open-source

Shipper: Anthropomorphizing AI is fine — when it helps us understand it

Dan Shipper argues that treating AI like a person is useful for understanding, predicting, and using it better, but becomes harmful when it sows panic or makes false claims about consciousness. He says 'pragmatists stay winning' in the age of AI.

• #policy
• #agents

Nan Yu joins OpenAI to work on Codex and ChatGPT

Former Linear head of product Nan Yu announces she is joining OpenAI's product staff, working on Codex and ChatGPT. She calls a separate post 'concerning' but does not elaborate.

• #products
• #funding

Meta AI director: PMs must understand the model frontier better than researchers

Madhu Guru, Sr Director of AI at Meta, advises product managers to deeply understand what models in each size category can and cannot do today, build workarounds for failures, and anticipate what models will be capable of in 2-3 months. He calls this 'a core part of the PM job now.'

• #products
• #agents

Garry Tan: Circleback beats Granola, new GBrain evals released

Y Combinator CEO Garry Tan calls Circleback 'so much better than Granola' for note-taking, citing Granola's lack of multiple-person disambiguation. He also announces new GBrain evals claiming SOTA retrieval-for-AI-agent memory without an LLM in the loop, plus evals for saving memory from agent transcripts.

• #agents
• #open-source
• #products

Peter Yang: Trust is the biggest barrier and driver of personal agent adoption

Peter Yang argues that trust will determine whether people adopt personal AI agents. The rest of his posts are promotional or retweets without substantive commentary.

• #agents
• #products

OpenAI's Thibault Sottiaux asks what holds people back from Codex

An OpenAI engineer working on Codex and ChatGPT asks users who have considered but not tried Codex what the single barrier is.

• #products

Replit CEO shares iOS interaction tips

Amjad Masad shares a link to good iOS interaction tips without further commentary.

• #products

VC posts link without context

Nikunj Kothari shared a link with no substantive content or commentary.