The “AI 2027” Scenario: How realistic is it?

Very few major advances take just six months from conception to full implementation. In reality, the chance that we will have reliable AI agents that truly advance AI research by the end of the year is small.

And it’s not just that particular longshot; pretty much the whole thing is a house of improbable longshots. If any one of those longshots fails to arrive on time, the entire timeline is moved back, and we are already into a longer time frame. Overall, then, the AI 2027 scenario almost certainly underestimates how much time we have to prepare for general intelligence, by years if not decades. 2 And some of what they describe (like AI’s going rogue, completely beyond human control) might never happen. If any single step doesn’t transpire, the entire scenario falls apart.

Where do they come from? Agent-1 is an extension of what the big companies are showing demos of now. It doesn’t actually exist, except in demo form. Google for example just announced an amazing voice-based agent, called Project Astra. If you look at the fine print, which they flash on briefly at 17 seconds in, the demo video is actually merely a “research prototype”. At one minute more fine print flashes that the video has been edited (“sequence shortened ” ) and “Results may vary”. It would be astounding if AI agents were actually contributing fundamentally new ideas to AI by the end of year. Far more likely is that Agent-1 type systems will be in a kind of “just a demo, still being debugged” phase for years . In all likelihood the hallucinations and boneheaded errors that have bedeviled LLMs since their introduction will plague agents (which are after all built on LLMs), too.

But ok, suppose we give a free pass on that one. How about Agent-2, where does it come from? The authors wisely don’t want to put all their eggs in the basket of pretraining scaling (which used to fuel most people’s AGI fantasies), because by now they know perfectly well that pretraining scaling has hit a rough patch.

A big part of the issue— which AI 2027 never faces — is that we simply don’t know how to create synthetic data for many other domains, like what might happen if India and Pakistan go to war using drones or what the effects of 30% tariff might be on the global economy. We can’t verify data like that; we barely even know what questions to ask, and we can’t foresee all that we need to foresee.

Detailed graphics grounded in what appears to be largely fictional numbers based in large part on overly charitable read of a flawed but popular graph from METR add to the effect. Every putative exponential is assumed to continue indefinitely, a version of what I once called the disco fallacy . (See also the trillion-pound baby .)

At most it shows that superhuman coding — just one of many prerequisites for the overall scenario — could come in 2027, not, by any stretch, that it will.

And even though I highly doubt AGI will arrive in three years, I can’t absolutely promise you it won’t happen in 10. We certainly aren’t prepared now, and if we don’t get moving, we won’t be prepared in 10, either.

2 Ernie Davis had some choice words for the 2027 scenario in a recent case study , “ The much discussed AI-2027 scenario, for example, predicts in passing that by the end of 2025 (eight months from now!), there will be AIs that have “a PhD-level knowledge of every field”. I often get the sense that many of the people who make those kinds of predictions have no idea what constitutes expert-level knowledge or expert-level capacities in any field except, possibly, computer science. With these kinds of statements being tossed around casually, it is salutary to remind people that AIs facing problems from various fields make some mistakes that even human beings who are entirely ignorant of the field could avoid, and make many mistakes that a human (such as myself) with only casual knowledge could avoid .”