potential title:
You see AI and Technology Everywhere but in the Productivity Statistics
(OK, the title would actually be in the form of ‘AI Will Transform the Economy in 2026’ or ‘AI Will Not…’, but we don’t want to start with the Bottom Line or whatever the rationalism idiom is.
The automation of short tasks is not transformative. New tools automate short tasks all the time, it’s generally not a big deal. The automation of long tasks is a big deal. In general, on average, the longer the task that a technology automates, the bigger of a deal it is. So, when LLMs automate short tasks, the effect isn’t really seen in the productivity numbers. That’s what we’ve seen so far. Our latest productivity stats are from Q3 2025, during which SOTA time horizon was mostly, roughly three and a half hours. We’re now at 14 and a half, and we’ll be at 60 hours long before the year is over.
Now, even three and a half hours doesn’t sound like an insignificant period of time. So, it’s interesting that productivity would not increase noticeably with time horizons of that length. There are a lot of points one could make in response to this. One obvious one is that the majority of knowledge workers were not even using AI, let alone SOTA AI. Some programmers were, and some AI enthusiasts or workers at companies whose leadership were taking AI seriously. Another is that impactful work takes time, and is generally bottlenecked by long subtasks (that is, it’s not merely a series of shorter subtasks, which would be automatable). So, shorter time horizon AI can speed up some of the less important subtasks, and it can be a useful tool for completing the longer parts. But a cool new tool isn’t revolutionary. And automating, say, 10% of the total task time while remaining bottlenecked by the rest has at most a uh 10% speedup? Plug it into the Amdahl’s Law formula, I guess. Not much, anyway. And it’s a 50% time horizon, so it’s not like the AI can reliably do tasks of 3.5 hours. The 80% time horizon of these models were only about 30 minutes.
So, as time horizons extend, each additional minute is more valuable than the last (going from 0 minutes → 1 minute unlocks 1 minute tasks which aren’t very useful, going from 100 minutes → 101 minutes unlocks 101 minute tasks which are quite useful). What this means is that with each doubling of time horizon, model utility more than doubles. Now, reality is a little messy, but to speak roughly: when model time horizon goes from t to 2t utility doubles, so if utility is boosting gdp by 1.01 now, it will be 1.02 after a double, then 1.04, then 1.08, then 1.16, then 1.32, then 1.64. Except… even faster because utility doesn’t merely double. Also adoption accelerates when the models are proved more effective (obviously), and adoption becomes less important as models become increasingly autonomous. Autonomous agents don’t need to be adopted by non-AI companies, merely brought online and directed by the AI company itself.
There is a limitation here, though. Compute is finite. There will be a lot more compute in late 2026 than there was in late 2025, but still not enough to service the whole economy? So, things won’t really look like workers getting replaced at massive scale, they’ll look like… I mean, listen. The obvious use for AI will be for accelerating AI research. The logic isn’t complicated. Think about what the economy looks like in a couple years. You have AI which scales way beyond human cognitive ability in terms of both quality and speed. Whoever controls this will be wealthy beyond any comparison with today’s economy or today’s largest companies. Trying to capture some slice of today’s economy by providing your models to the population is not the play. The play is trying to maximise your capture of the future economy. You do that by turning your GPUs inward. Datacentres of AI research LLM agents managing training runs, conducting experiments, designing RL environments, etc. A few high paying customers who can make the best use of the models, sure, and a few charity cases, but 90% of compute directed at AI research and training. Nothing else would make sense. OK, so what does that look like? What’s the result? What follows from this?
We shouldn’t think about this at the economy-scale. Not yet. Although come to think of it OpenAI only has like 1,000,000 h100 equivilants for inference and ChatGPT has like 100m DAU. That’s 100 DAU per user. Let’s say each user uses the site for an average of 20 mins a day. That’s about 1 GPU per presently-active user. Let’s say AI companies have 10,000,000 h100e for inference at sep 2026, that’s 10,000,000 active users, that’s not a small fraction of knowledge workers. The most productive knowledge workers could be pretty well-serviced. If we assume something like a 5x uplift, and apply that to the top 10% of knowledge workers in terms of pay/status, I wonder what overall knowledge worker productivity uplift would result. Anyway, this is an interesting line-of-thought but probably not ultimately going to be productive. But does counter the narrative of a complete inward compute move being necessarily the future we’re getting. I suspect we will get it anyway, because it makes by far the most sense, but because labs can service a lot of workers with a relatively small portion of their overall compute we can imagine maybe they’d have some PR or fundingraising incentive to continue to service the populous at large scale. Also, older hardware might become a bit less useful for serving their highest-powered models… but would still be useful for automated experimentation… so that’s not ultimately a compelling point against inwardness.
So, regardless of whether AI labs keep servicing the public in late 2026, they’ll certainly have a lot of compute dedicated dedicated to LLM AI researchers and to training and to experiments. It will become crystalised in the consciousnesses of these orgs that the scale of the economy resultant of proximate AI systems is simply so much larger than the present economy that focusing on the present economy doesn’t make sense (unless required for funding/PR as an instrument that the AI lab can maximise its expected share of the post-ASI economy, i.e. of our cosmic endowment…).
So, we have AI which… let’s do it like this. Take current uplift as 1.25x (roughly what Anth has found internally IIRC). Roughly speaking, that’s Opus 4.6 providing a 1.25x uplift which is a model released on Feb 5 with a 50% time horizon of 14 hours 30 minutes. My best-estimate of doubling times as of January was 2.86 months (see @jimfund.com thread on Bsky), and this figure is decreasing due to the facts that the figure tends to infinity, investment has increased massively into AI, and that post-Opus 4.5 AI is beginning to (increasingly) contribute to the rate of AI progress. So, for the short-term let’s take the figure of 2.86 months (how to decompose this in terms of how much it’s constrained by the compute requirement of training new models vs how much that can be overcome by increased intelligence vs how much the doubling time is just constrained by human researcher inefficiencies still a jim-open question.
Anyway…
uplift as of Feb 5: 1.25x
time horizon doubling time: 2.86 months
utility scales with time horizon: superlinearly (increasingly superly)
So, Mayday we should see uplift of >1.5x
End of July we should see >2x
Mid October we should see >3x
As massively more compute comes online, and researchers lock in for the final stretch, and as doubling difficult decreases, and as productivity increases over 100%, over 200%, doubling times drop rapidly.