it's funny how people express doubts about these trends when the fit of the data is the cleanest thing imagineable. i got worse fits in college physics measuring Hooke's law
A double exponential fits the data very well (R^2=0.99). Extrapolating that gives a predicted time horizon of ~80 work years by end of 2027. Extrapolating an exponential from post-Opus-4.5 gives ~30 work years. (up from Opus 5.5's AECI-implied TH of ~1 work-month)
@aphysicist moving bits is basically on its way to being a solved problem, RSI included (for better or worse). moving atoms is next, but it's relatively so expensive that there really has to be a good reason. many exist but these custom "talk to AI" devices are perplexing. use phone
@_brianpotter I'm ready to buy "no," that's for sure. absurd demand in energy vs regulatory pain and questionable demand for the aircraft. plus with starlink and redesigned cabins, flights are way less of a productivity drain than they were even a year ago.
@CJHandmer makes sense. computing the timescale of corrections by field would be interesting. "physics advances one funeral at a time," a researcher friend once told me. might be quantifiable now
@ZackKorman Let me try again. Every group's culture is their kryptonite! It's like a basic property of how cultures work. In the worldview of Dario and EAs who support Anthropic, his essays are consistent with his actions.
There is still no substitute for caring. Every piece of software getting rapidly better is happening because a human cares. And happens to be using AI to move really fast.
@connorkapoor I took the same journey in roughly the same number of hours. wait until your next plane flight you'll have so much fun ssh'ing to a box where claude has access to much faster internet
people are like “why do LLMs always think they’re in an eval lol” and then have bad dreams about forgetting to turn in a homework assignment 10 years ago
@francedot copy that. in general for all computer use tasks, should I expect cua driver + vision + gpt-6.1-sol to perform similarly to ChatGPT's in-app computer use?
I'm wondering where the "sauce" is and my initial experiments are suggesting it's a bit more than just the model + Cua
Has anyone built a driver for computer use using CuaDriver that's anywhere as good as Codex computer use using the ChatGPT app? I thought codex + cua would get me close but it's a struggle.
@barbinbrad At Gearflow we're excited about a framework for forward deployed _agents_. fast, focused on infra / method (I would combine those two points), and while the initial lift is largest, they don't need to be temporary because they are cheap to maintain indefinitely
@sharenow_today are y'all using Cua? I'm working on generalizing my personal harness as part of a B2B product.
PS - I assume you're using your staff to find people on X mentioning Cua and like their posts...? Clever
@francedot it might be a skill issue; primarily having difficulty with background use of chrome to drive webpages in a way that doesn't trigger bot checks. (Not for malicious use mind you!)
What docs would you suggest to help my agents learn the latest and best practices?