Post

This is a 5% stupid question but bear with me. I thought that large closed models were really cool before Mythos came out, and then I was blown away like how it was when GPT 3.5 shipped for the general public. What are the chances of a similar “holy cow?!” moment for a future model deployment?

Yea, my guess is that we're in a similar phase to '24-'25, where models score higher on benchmarks but that doesn't translate into additional utility for most users. I'm sure there will be edge cases, maybe I'm underestimating computer use, but for swe i haven't felt a shift this year

give it a week

I think the recent improvements in vision on the frontier models have been very useful, we could really use some focus along the more secondary axes like that.

I don't care if it gets any better than sonnet 5 tbqh (aside from the odd situation where I do need a fable team to beat the shit out of some bug hunting) But also if fable is too expensive to use I can also use my own brain just a lot more slowly

I was under the impression the pretraining base is where you should look for improvements in flexibility and genuine novel approaches? Feels like Fable is a world apart from Opus in this sense. And by the sounds of it Astra from OpenAI should be a big step up.

That’s interesting. I find myself doing pretty intensive long design sessions before being able to see things clearly enough to break things into a straightforward task graph. I do find sol being more useful/mature than alternatives during this phase? Luna can crush individual tasks.

Defining “we” is the trick.