AI Builders
Digest

In the Unsupervised Learning podcast, Jacob Efron discusses with former DeepMind/Meta researcher and Datalogy founder Ari, and Radical VC Rob, the latest shifts in AI. Coding agents are now working at longer horizons, shifting engineers from ICs to managers of agents. Open-weight models face existential pressures: Meta and Chinese labs may pivot to proprietary APIs due to compute costs and geopolitics. The compute crunch could lead frontier labs to pause APIs in favor of internal use and first-party products. Anthropic's Fable release sparked controversy over silent restrictions on AI development uses, seen as competitive positioning rather than pure safety.

Key insights: pre-training hasn't hit a wall; test-time compute and scaffolding drive progress. Rob notes the human brain's 20-watt efficiency as proof future AI will be far more resource-efficient. Ari is more bullish on recursive self-improvement (RSI) but stresses compute as the limiter.

"the open frontier would only be a few months behind... there are signs now that... near Frontier open weight AI falling off altogether."

Multiple AI builders discussed Anthropic's Fable model release and restrictions on AI development uses. Peter Yang questioned enforceability of suspending access for "foreign persons inside the US." Aaron Levie sees this as a major turning point for AI regulation, with governments deeming models too powerful for certain uses and setting precedents. Amjad Masad noted Replit may need to turn off Fable access. Alex Albert shared that Fable excels in long agentic conversations but benefits from specific prompts for clarity.

Replit CEO Amjad Masad highlighted rejecting token leaderboards in favor of selling outcomes rather than tokens. Enterprise customers requested leaderboards during the Tokenmaxxing craze, but Replit prioritized long-term value. He also noted users who make money on Replit get free credits and expressed concern over potential Fable access restrictions.

Vercel CEO Guillermo Rauch shipped HarnessAgent, a unified abstraction to orchestrate any agent’s “brain” into your app. @aisdk now frees developers from model and agent lock-in while remaining delightful to use. He also celebrated HTML being back with drag-and-drop demos.

OpenAI's Thibault Sottiaux addressed feedback: next Codex usage reset will let users choose when it applies. Peter Steinberger shared how Codex running inside crabbox while building it has become essential infra, looping for days with verifiable e2e builds. Garry Tan loves forceBlockStreamingForReasoning in OpenClaw for seeing Claude Fable 5 reasoning traces.

Anthropic Engineering published key posts: one on Claude Code quality reports addressing regressions from reasoning effort defaults, session caching, and system prompts; the other on Scaling Managed Agents, decoupling the "brain" (harness) from "hands" (sandboxes) with self-hosted sandboxes and MCP tunnels for reliable long-horizon agents.

Changes improved TTFT and security/scalability.

Andrej Karpathy expressed awe at SpaceX and its story across past, present, and future, offering huge congrats to the team.

Box CEO Aaron Levie congratulated SpaceX on 25 years and called it an example pushing the future forward. YC's Garry Tan shared concerns that AI coding tools often scaffold bureaucracy faster than new experiences and promoted Datacurve DeepSWE benchmark.

Swyx wonders if Git needs to die next after PRs and code review, suggesting future codebases may resemble Notion/Linear databases—less efficient but more scalable, with collaboration via chat and suggested edits rather than line-by-line merges.

Zara Zhang noted receiving dozens of AI product trial requests daily, with insane competition for attention. Nikunj Kothari shared thoughts on how application companies survive the “what if the labs build this” question. Dan Shipper humorously reacted to CFO optimism.

Claude's official account asked "What are you building?" and shared related posts. Other updates include Aditya Agarwal welcoming a new partner and celebrations of milestones like SpaceX.

Peter Steinberger detailed how Codex runs inside crabbox while building it, becoming essential infrastructure that loops nonstop for days, handling service signups automatically.