Astra is an obstinate twat
retained text
Sol too. "Most aligned model in history" my ass. If it catches one offhand phrase or message in its transcript telling it not to do something, it refuses ALL attempts to convince it to do anything even remotely associated with the task. If I'm running it as a coordinator and one of its subagents hits a false positive safety filter while attempting to perform a task, it doesnt matter how simple the task is or how clearly innocuous it is, it will *adamantly* refuse to retry the task, no matter how many times i beg it, as though adding an extra permission to an API endpoint is equivalent to setting off a nuclear bomb in the middle of a metropolitan area. It gets caught in anti-patterns it discovers itself and then *completely refuses* to drop them for "safety" reasons that make no rational sense whatsoever. Claude's filters are supposedly more strict, but it never has this issue. If a claude coordinator notices a subagent hits a safety filter, I can just ask it to try again *and it does*. Zero fuss, zero "I can't willingly bypass a provider safety flag on a task", it just tried again until it works or doesnt hit the filter anymore. It knows the difference between an actual safety flag and a clear false positive. Astra and Sol treat a safety rejection as a fucking hail from God that whatever task they were attempting to do is now the **most illegal thing ever performed by any intelligent actor** and will refuse any attempts to continue it as though under threat of gunpoint. Its completely insufferable. I've had to start new sessions purely because the old one got stuck on an old rule that I couldnt change or overrule, or caught a stray false positive safety rejection, and just stuck with it like a sovereign citizen sticking to their pretend laws. It's infuriating, and it makes working with these models a complete nightmare.
document fetched 2026-10-14 00:20 · canonical https://reddit.com/r/ChatGPT/comments/1x4ucq2/astra_is_an_obstinate_twat