AI 2026

Are we in a AI bubble of over hype? At some point in the future, superintelligent AI is almost certain to become the most important technology that has ever existed - but it doesn’t necessarily have to happen in the next few years.

One way or another, 2026 should tell us the reliability of a lot of the leading AI scenarios out there right now:

• if you take the AI 2027 scenario seriously, 2026 is when we first see the arrival of superhuman coders, helping to speed up the process of AI research itself

• at the current rate of progress on the METR benchmark of increasing time horizon length, by the end of 2026 we should be seeing agents capable of completing day long tasks in complexity. That’s more enough to write a short story, a new compelling blog or put together a report.

• Since GPT-4, we've only had minimal scaling of pre-training compute, with the focus more on post-training and test-time computer. In 2026, major new facilities such as OpenAI's Stargate or xAI's Colossus 2 start to come online - potentially unlocking a return to the previous trend of faster compute growth.

• 2025 was to be the ‘year of agents’ according to most of the labs - and to be fair to them, we have really seen significant progress, particularly in assistants for coding and desk research. What will 2026 be the year of? If you listen to enough interviews with them, there’s starting to be a repeated theme: 2026 is the year when AI starts making discoveries on its own.

If you combine all these predictions, if scaling on its own will be genuinely enough to create a recursive intelligence explosion - then in 2026, we should start to see signs of real creativity. That doesn't have to mean fundamentally reshaping physics: it could just be getting increasingly capable at writing a surprisingly insightful blog post, or reliably creating a moving short story.

By contrast, the biggest current argument against their capabilities is that their lack of a robust world model makes it hard for them to come up with complex new ideas, and this is a fundamental bottleneck that simply scaling up compute won’t be enough to overcome.

By the end of 2026, we will have the compute, the predicted time horizon and all the underlying tools - and we should have a better idea of which path we're on.