A few people have made the prediction that there’s inherent superexponentiality in time horizons. One way to define inherently superexponential time horizons is:

Even without substantial AI R&D automation, there is some reason to expect time horizons to grow faster than exponentially at some level of capabilities.

There are two common arguments that I’ve seen for this:

1. Under some definitions of time horizon, a superintelligence has an infinite time horizon. An infinite time horizon is not achievable in finite time if growth is exponential. Therefore, if a superintelligence is achievable in finite time, the rate at which the time horizon grows will have to be faster than exponential.
2. The difference between the difficulty of tasks is much higher for short tasks than it is for long tasks.

I. Superintelligence in finite time implies superexponentiality

I don’t think this argument is very strong. It somewhat begs the question. If it’s the case that superintelligence is unachievable without superexponential growth of time horizons, then is this mainly evidence for:

1. “superintelligence is not achievable on our current trajectory”, or
2. “time horizons are superexponential”?

My guess is that it’s mostly evidence that superintelligence just isn’t achievable on our current trajectory. Which is not to say that we won’t switch trajectories soon.

I also think many definitions of superintelligence also doesn’t necessarily imply an infinite time horizon. If we define the time horizon as the time it takes a human to complete some well-specified tasks that the AI has a 50% chance of completing, then it’s still possible that for a superintelligence, such tasks would be on the order of thousands of years, millions of years, or billions of years, but still not infinitely long. But that would mean that it will take a while to reach superintelligence given exponential growth of time horizons.

II. Difference in task difficulty lowers with task length

It’s possible that going from 1-hour tasks to 10-hour tasks involves a much bigger jump in “skill” compared to going from a 1-week task to a 10-week task, despite there being the same relative increase in task length.^[1]

I think this is possibly the case for humans doing research tasks, but there are a few factors that complicate generalizing from humans to AIs.

For instance, humans often do week-long work trials for jobs they will do for years. For some reason, there seems to be a high correlation on how successful humans are at weeklong tasks and tasks that take many months. So could we say the same for AIs? Will an AI that can do a work trial for a job also be able to do that job?

Well, obviously not. AIs are able to do lots of work trials at this point, but they can’t do nearly as many jobs. The correlation in performance between short-term tasks and long-term tasks seems to be much higher for humans than it is for AIs. This is also apparent from the fact that AIs are very capable at short IQ tests, but this has almost nothing to say about how economically useful they are. Whereas for humans, there seems to be a correlation between scores on such tests and career success.

Also, the tasks in the time horizon suite that take a day are much closer in appearance to work trial tasks than they are to real-world one-day tasks. This will probably also be true for weeklong tasks once METR adds those to the suite.

The intuitive argument

But I also do somewhat buy the intuitive argument that if there were an AI that could do a week-long task, then surely it wouldn’t be so hard to just chain many of those together and have them do a year-long task. If you had many AIs that could do a simple ML project, then maybe hundreds of thousands of them would be able to output papers of the quality of a senior ML researcher?

But I think this argument is slightly sneaky because it depends on succeeding at weeklong real-world tasks instead of weeklong time horizon suite tasks. So yes, it might be the case that once we reach weeklong tasks, we’ll have superexponential growth in the length of tasks AIs can do. But the measured time horizon on METR tasks, which are very different than real-world tasks, will at that point be much higher than one week — maybe it will be multiple months.^[2]

III. AI R&D might be more superexponential than other capabilities

I also find it it pretty plausible that the kind of work that happens inside AGI companies is uniquely suited to being split up into smaller tasks. My guess is that, because of the pace of progress, very few projects can survive inside an AGI company much more than a year, and most projects are probably less than half a year long (but I wouldn’t know). So it might be the case that capability growth is not superexponential in most of the world, but is superexponential in the place where it most matters—inside the AGI companies.^[3]

Conclusion

I find it pretty plausible that time horizons are inherently superexponential — that even without massive increases in AI R&D labor, they could increase faster than exponentially in the next five years. The main reason I think this is because once there exist AIs that can do weeks of real-world labor, it seems plausible that they will be able to do years, or even decades, of labor with slight modifications. I’m more sure of the superexponentiality of AI R&D capabilities than I am in the superexponentiality of other capabilities. But the skill profiles of AIs are different enough to human skill profiles that I’m pretty uncertain.

1. ^{^}

   There’s a similar view which says that there are some meta-skills, like long-term planning, that will be achieved soon and suddenly “unlock” long-horizon performance.

2. ^{^}

   This is related to a separate can of worms, which is that in the real world, people seem to very rarely perform months of software work without any outside feedback, and a lot of work seems to happen in teams. It’s unclear what the definition of task length will mean once we add those outside influences in — Steve Newman has a great post related to this question.

3. ^{^}

   I already held this opinion, but seeing Herbie Bradley’s comment on Steve Newman’s post prompted me to add this paragraph.