This is a quick reaction to: ai-2027.com, which forecasts a very rapid timeline for Artificial General Intelligence (AGI) and Artificial Super Intelligence (ASI).

I’d say the likelihood of a “fast takeoff” scenario like this has greatly increased since ChatGPT. The big breakthrough of the past decade is that scaling up the neural net also scales up intelligence. The AI 2027 scenario is based on that trend: if the scaling laws hold, then exponentially increasing AI compute - as companies plan to do - will exponentially increase its intelligence.

It could also not happen.

It’s anyone’s guess what the likelihood and timeline of this scenario is. I think the likelihood is certainly not zero, and I’d say it definitely won’t be tomorrow or next week. 2027 is a reasonable guess for the earliest it could be. It’s aggressive, but possible. For that reason, I think it’s worth someone’s time to “game it out” for the good of humanity in case it does happen we have some kind of plan.

When “we all” did the thought experiment and game theory with nuclear weapons, we settled on Mutually Assured Destruction and detente. I think the common knowledge of those concepts have (so far) saved us from nuclear extinction.

Similarly, the “good” scenario here is a deliberate slowdown to focus on human alignment. I think it’s worth raising awareness of the catastrophic risks of “true” AGI / ASI and the rapid, insidious way it might happen. As in this story: it may be an AI that emerges so quickly and is so smart that it conceals its actions until it’s too late.

Personally, I give it a low percentage of happening so soon or at all, because for instance we gave Claude all of the Internet and it still can’t beat Pokémon. I also think AI will need more physical embodiment to really outmaneuver humans. It could still do a lot of damage online though.

The next frontier for AI is long horizon tasks like beating Pokémon, building complex apps, inventing things, or scheming to take over the world. We only have a peek of the potential and it still blunders a lot – unless that’s a ploy to give us a false sense of security!

My guess is that scaling up and text-based generation won’t be enough for that. It will need another breakthrough, which is probably a new paradigm of AI model, plus embodiment. People are working on fulfilling that prophecy…

I’ll have to check back in on this prediction in 2027.

One of the authors Daniel Kokotajlo made similar predictions back in 2021: What 2026 Looks Like.