Tomorrow isn't in the training data

AI can now predict the future nearly as well as the best human forecasters

The Crystal Ball (1902) by John William Waterhouse (1849–1917)

There are so many opaque AI benchmarks that it can sometimes be hard to see whether AI capabilities are still improving. Today Scott Alexander discussed a specific type of AI capability that’s actually quite easy to assess.

He discusses AI superforecasters. Superforecasters are traditionally people who are exceptionally good at predicting future events. These can include, for example, wars, stock market trends, or even elections. They don’t do this by looking into a crystal ball, but by making well-informed judgments based on expertise, background knowledge, probability models, and calculations.

AI superforecasters are frontier models that perform the same work, using deep search and analysis in structured software environments specifically tailored to support them, which are called “scaffolds.” Only they’re doing it much faster and more cost-effectively. At the most recent Metaculus Cup, a well-known competition for forecasters, humans took first and second place, but the AI start-up Preseen’s AI superforecaster took third place. If development continues as the trend suggests, AI superforecasters will leave even the best human counterparts behind in about half a year. Incidentally, the Metaculus forecasters themselves estimate a 95% probability that this will happen before 2030.

Of course, the commercial potential is immense. According to Alexander, Preseen itself claims that its model turned just $35 into a whopping $2 million on the Kalshi forecasting market in only seven months. Another company called FutureSearch even claims to significantly outperform the stock market. I’m always quite skeptical of such claims, because with a bit of luck and survivorship bias, even utter nonsense can look convincing at first glance. I find the experiment Alexander conducted more convincing: He had the FutureSearch model calculate the probability of a future event for $8, after which it spat out a result and a list of the 212 sources used in just 5 minutes. He also gave the same task to the Preseen model and a human superforecaster. All three arrived at nearly the same result. Not bad for a couple of “stochastic parrots”, right?

The “stochastic parrot” phrase — that is, the notion that LLMs essentially just remix what’s already in their training data and then regurgitate it — is hard to reconcile with results like these. Tomorrow isn’t in the training data. Of course, it’s possible that certain basic assumptions about probabilities and causal relationships were part of the training data. But analyzing and predicting new events on that basis — and outperforming even the best humans in the process — is precisely not “parroting.” We generally call that “judgment.” And the benchmark in this case isn’t just some test that someone created and for which the AI can potentially be prepared, but reality itself.

So how much longer will it take for the capabilities of AI superforecasters to eclipse those of all humans? As mentioned, human superforecasters believe it won’t be long. Skeptics of superhuman AI capabilities see it differently. Alexander cites an article on the topic by the computer scientists Arvind Narayanan and Sayash Kapoor, authors of the book AI Snake Oil, in which they argue that human capabilities in some areas represent the limit of what AIs can achieve:

Specifically, we propose two such areas: forecasting and persuasion. We predict that AI will not be able to meaningfully outperform trained humans (particularly teams of humans and especially if augmented with simple automated tools) at forecasting geopolitical events (such as elections). We make the same prediction for the task of persuading people to act against their own self-interest.

Just two weeks ago, we reported that frontier AI models can now out-persuade even expert human persuaders. That may not yet meet the strict criteria set by Narayanan and Kapoor, but you don’t need to be a superforecaster to doubt that it won’t be long before that changes, too. In any case, it’s a testable prediction by the two skeptics that’s currently being seriously shaken on both fronts.

Alexander thinks that superforecasting could become a kind of “opinion layer” of AI. That would enable far more people to make the best-informed decisions even in cases where no one would shell out for a human superforecaster.

But I would actually go one step further. In the online resources for If Anyone Builds It, Everyone Dies, Eliezer Yudkowsky and Nate Soares explain the danger of creating something that is better at predicting the world and steering it than humans are. Superforecasting is prediction and basically already constitutes half of that. And if superprediction and superpersuasion are already drawing nearer, how far away is superhuman research and the resulting threat of an intelligence explosion through smarter and smarter AIs building their own successors?

We’re not there yet, and we still have a chance to hit the brakes. But every new improvement increases AI’s ability to devise and implement plans — including plans nobody asked for.

---

The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.