Astra is a decent model, it seems about as smart as Fable, but has the same personality mismatch with me that kept me from using GPT models in the past. They all tend to be autistic in a too strong way by focusing too directly on what you asked and have a way lower ability to understand your intent. I am an intuition build, and so the way I communicate works a lot better with Fable, who is able to step out of the situation and assess it in an out-of-frame way more easily.

Taelin here writes that all models have this issue, which is true, but I do think that, to the degree that they both have it, Fable has it way less than Astra. To illustrate, today I spent the day having Astra do the same Steam market research that I did with Fable a while back. On the data gathering portion of it, I noticed Astra reported that it had just downloaded 300000 images from Steam in like 10 minutes and then I started a side chat with it about it:

Clearly it made a serious mistake, and this kind of mistake is typical of the GPT autism I mention, just a lack of ability to have... just overall good sense? I don't know. Fable obviously did this correctly and throttled his data collection and it took hours. Astra just went for it. It is an extremely capable model, if you point it in a given direction, it will go deeper than Fable, but it makes these really dumb decisions sometimes that are truly baffling. This does not fundamentally bother me, but Claude has much better sense overall.

I think these things are obviously downstream of how their leadership behaves and so it makes sense that Claude would behave better than GPT in general, Anthropic has put a lot more effort into making that the case with their constitution and other efforts around alignment that seem more tasteful than what most companies do. All this means is that I can't trust Astra with decision making responsibilities, or to act in my stead, thus I've spent some time today also working on a delegation pipeline, so that Fable delegates to Astra (as well as other models) when necessary.

I have not yet tested Astra for game development, but I will in the following days. Overall, will I keep using it? I'll keep using it this month because I paid $100 for the plan, after that I don't know, depends on how well it does at game development, but from what I'm seeing online it seems only slightly better than the current Claude offerings, seems better at making and animating humanoid characters especially.