I just joined the Brave Rewards Card waitlist! If you sign up earlier, you can get more benefits once it's live. Sign up now. Takes only a few seconds: https://t.co/Ee5KLZg6oT @brave@AttentionToken
🚨 Is post-launch Astra dumb?
❖ The quick reality is... No. It isn't.
❖ It is as dumb as it was on launch... check my first impressions (link in second message)... the general mood was disappointment (and waiting for 6.1).
❖ At some point i thought i was biased and was missing out something... but that was not the case, people in AI-twitter were just over-hyped.
So:
❖ The model is good, but the model has a lot of problems.
❖ Is lazy
❖ Writes like a bullet-point-addict
❖ Is too literal following instructions
❖ Overengineers anything
❖ Can easily derail from the initial goal
But:
❖ After testing it in many use cases and lessening many of its problems (the ones from the previous block) with hand-crafted prompts, boom! the model is quite good, better than fable-5 in general.
❖ So what's really happening here is that people were overhyped on launch's week, now they had time to test it and see its mistakes... and I'm pretty sure most of them don't know how to lessen such mistakes with good prompting.
❖ one final thing: remember what openai did to gpt-5.5 when it became 5.6, it was a whole new thing, and it was just most post-training (built on the same pre-train), so expect the same with 6.1-astra (it's not ideal, but it is what it is)... in the meantime try to put in agents.md whatever problem you see coming up more than twice.