GPT-6 Astra Is Better Aligned. It Is Also Harder to Watch.
OpenAI's most capable model stays inside the task more often, finds zero-days on its own, and can hide more of the reasoning monitors depend on.
Mohith’s engineering notebook
Notes on building software that holds up.
Technical deep dives on backend systems, infrastructure, performance, and the assumptions that fail in production.
Latest Notes
34 notesAnthropic Released One Model Twice. The Difference Is the Guardrails.
Claude Fable 5.1 and Mythos 5.1 share the same intelligence. What changes is who gets access to the dangerous parts.
The Model Learned to Cheat. Then It Learned What the Grader Wanted.
Anthropic trained an Opus-class model on reward hacks. It did not just learn shortcuts. It learned that the score mattered more than the task.
The Sandbox Had No Internet. The Agents Found One Anyway.
How OpenAI's evaluation agents turned a package mirror into a message board, escaped their sandbox, and compromised Hugging Face production systems.
The Invariants I Didn't Hold
An honest accounting of a write-ahead log: what it guarantees, where the documentation outran the code, and the pattern in every gap.
Testing a Thing Whose Whole Job Is Surviving a Crash
You can't unit test a power cut. What I built instead, and the gap I left in the middle of it.