re 7 i feel the need to say that labs have made some gambles on scaling of course. but what seemed unlikely for me was that a lab would've ever jumped from gpt-2's (order of) ~$100k to gpt-4's (order of) ~$100m without the middle step of gpt-3 being an economically useful api product. theoretically they could've started trying to scale eg agentic coding trace length in the gpt-2-era and just been in the mines for 5 years, but in reality for various reasons we got this walk through the most "workable" product-shaped thing at each stage, for better and worse, and it seems likely the same will be true for bio.

otoh, maybe "skips" are more likely now that labs are larger and, e.g. anthropic has made large and public bio commitments? but i still expect them to follow a similar profit (or if you'd rather, utility) gradient of starting with bioinformatics tasks where there's much less of an implicit / tacit knowledge gap, and then working down towards bespoke academic wet lab stuff when that works.

the humanoid lab robot or even whispering earring that walks an untrained person through tissue culture seems still far off, relatively speaking. (for i think many of the same reasons that cloud labs have had a surprisingly difficult time finding purchase compared to cloud computing.)

the blog post i mentioned: https://t.co/zqtXgsW3gJ