For those with access to Mythos, it does not appear to be some kind of step function change in vulnerability detection. Rather, it seems to show code agents are good at finding vulnerabilities and...
There's a lot of chatter about models getting worse after launch, during peak hours, on subscription plans vs. pay-per-use APIs.
I've started using Codex more consistently. It took me entirely too long to locate the plan usage page: https://chatgpt.com/codex/cloud/settings/analytics, roughly equivalent to Claude Code's...
It will be interesting to look back in retrospect if April 2026 was peak LLM hype. During and since the month, subsidies for inference have started to go away meaningfully, and people finally seem to...
No one can explain to me what reinforcement learning actually is
Switched to running a mix of claude and codex for the first time for day-to-day dev.
Opus 4.7's tendency to use acronyms drives me crazy, especially when I have no idea what it's referring to, even though I'm its collaborator.
TIL, "ennui": not just "nothing to do" but "nothing seems worth doing"
Why you need stateful agents. You can be ambiguous, concise, and still be successful because the agent has the surrounding context.
I really have enjoyed writing about building software in the past but lately I've been struggling to find a foothold. The prompts to an agent are now most of the software. There is trial and error...