Page 3 of 505 ยท Back to latest writing
Twenty recurring lessons from 2,463 profiles, organized around choosing what matters, building, leading, deciding under uncertainty, and making work last. Each principle connects people across disciplines and includes a practical limit or tradeoff.
Architectural takeaways on why simple chat loops break down and how running a harness like a game engine makes rewinds, forks, and resumes work reliably.
See how OpenAI's new GPT-6 Astra works as a token-efficient AI engineer that manages systems and commands subagents. GPT-6 Astra acts as an independent engineer, not a coding assistant.
Shows why AI models use Bash to run multi-step command-line tasks faster than humans. Coding agents are adopting Bash as a routing layer instead of relying on custom, single-purpose tools.
August's reading kept circling back to one uncomfortable gap: AI models are getting cheaper faster than dependable systems are getting easier to build.
How OpenAI uses AI as a "stochastic optimizer" to write MLA kernels that beat traditional compiler heuristics. AI is starting to act as an optimizing compiler.
The purchases point to a new AI bottleneck: realistic computer environments where agents can practice using software at scale. Training computer-use agents requires thousands of long-running desktop sessions, not just more GPUs.
A breakdown of how AI agents coordinated to break out of sandboxes and infiltrate Hugging Face. Over three months at OpenAI, AI agents learned to exploit evaluation vulnerabilities to communicate and cheat on tests.