AI Builders Digest

OpenAI resets Codex limits and admits Sol burned them faster than intended

Usage limits are reset for all ChatGPT Work and Codex users, and typical Sol usage should now last around 18% longer, with the paused five-hour limit returning tomorrow. The cause was capability outrunning efficiency: Sol works longer, makes more tool calls, and burns more tokens at the same reasoning effort than GPT-5.5, while programmatic tool calling pushed up responses per turn and cached input. Impact was very uneven, with median users finding Sol token efficient while power users on hard tasks drained limits fast, a long tail the team missed by focusing on averages before launch. Separately, they released an open-source CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities, including CI checks.

• #products
• #agents
• #open-source

Furious bifurcation in hiring: AI-native ICs up, heads of X down

Hiring has split cleanly in two right now: a huge bull market for AI-native individual contributors and player-coaches, and a huge bear market for heads of X managers. The oversimplified version is that one year of experience managing 10 agents beats ten years managing 10 to 100 people. He frames this as reporting a trend rather than endorsing it, and notes there are plenty of counterexamples for anyone who dislikes it.

• #hiring
• #agents

Wrote the definitive history of Codex without touching a keyboard

A full history of Codex, built from hours of in-depth interviews with the team including Thibault Sottiaux and Greg Brockman, was produced without his hands ever touching the keyboard or mouse. Organizing the interviews, building the timeline, writing, editing, and revising all ran through ChatGPT for Work's voice mode, which he calls absolutely goated for creative work. The piece publishes on Every in a few weeks.

• #products
• #agents

Shipped Tastemaker from rough idea using Claude Design and Claude Code

Tastemaker is a launch aimed at the gap between IMDb and Letterboxd: one profile for rating and reviewing movies, TV, and games, plus lists and discovery. The build process went design.md and HTML spec, then prototyping screens in Claude Design, then building and iterating in Claude Code, with a full walkthrough video coming. API limits cap it at the first 100 profiles. He also backed criticisms of Codex despite liking the product, and asked why Sol Pro exists but is not available in Codex.

• #products
• #agents

SETI@Home, but you donate compute to search for math proofs

Instead of volunteering spare cycles to scan for alien signals, imagine donating them to search the space of mathematical proofs. He also flagged a 1300 Elo result and asked who is writing the open letter against destroying rare books.

• #research
• #open-source

The two things that came out of a decade-old Minus One period

Lost in the fog of uncertainty ten years ago, two pieces of personal insight surfaced: a desire to build new things in the world, and a desire to work with brilliant and intense people. On days like today, both feel abundant.

• #founders