DG's AI Timelines Late October 2026

On the default path, I expect transformatively/dangerously capable AI within a decade.

What I mean by "timelines"

"Timelines" is kind of nebulous, so for my purposes: timelines to AI systems sufficiently powerful to

1. precipitate high economic growth (say the US economy doubling in under four years), conditional on the systems being broadly deployed throughout the economy;
2. precipitate human disempowerment or catastrophically bad outcomes if they were sufficiently misaligned.

I.e. the kind of AI LW has spent two decades discussing how to align to human values. Let's call this property "transformative". Four years is a round number chosen for historical reasons; an economy doubling in under a year is still doubling in under four.

By "default path" I mean the AI race continues from late October 2026 as it has so far: no major regulations, no industry agreements to pause or explicitly pace the frontier. This is a counterfactual. I'm not predicting such agreements are unlikely.

Where we are now

I model October 2026 LLMs as having conquered the magisteria of programming and mathematics. AI proofs are still generally "slop", but I don't think eisegesis explains the superhuman feats in either domain.

October 2026 LLMs are still short of transformative. Broad deployment wouldn't intuitively double the US economy in four years, and they don't seem sufficient to disempower humanity. Even if they were badly misaligned, I expect 2026 Earth could withstand takeover attempts from them. Humans aren't asleep at the wheel.

Why I expect that to change within a decade

I think we're far from the practical limits of scaling:

• RLVR. We haven't come close to exhausting it. RL environments are being built for real world engineering tasks and arbitrary white collar work, and I expect this to continue for an ever-growing basket of tasks. We can just do a lot more RL on more magisteria to conquer them too?
• Test-time compute. Not exhausted.
• Population scaling. Even when we run out of efficient test-time compute scaling on a single model, there's a lot more we can do with swarms.
• Time horizons keep increasing, arguably at an exponential rate (some claim linear).
• New paradigms. The last decade gave us scaling pretraining, RLVR, test-time compute and multi-agent cooperation. I expect more paradigms of scaling we haven't thought of yet.
• AI doing AI research. I don't need RSI for this story, but I do expect AI to take a larger and larger share of AI engineering/research effort.

What I don't think is needed

Scaling the basket of techniques that exists as of late October 2026 feels sufficient, even without:

• a single model as data/sample efficient as humans;
• continual learning (though it being solved before 2035 wouldn't surprise me, labs are making progress);
• a highly agentic, fully autonomous model with independent drives the way humans have them;
• any fundamentally new paradigm of ML research/engineering.

Economically valuable tasks (e.g. software engineering, mathematics, research, etc.) are generally composite: they bundle many forms of thought, and current AIs are good at some and bad at others. My implicit belief is that the missing cognitive skills get subsumed by scaling, whether pretraining, RL post-training, test-time compute or population scaling. Intuition especially seems like one of the main things that improves with scale. The famed "big model smell".

Form Factor: AI Civilisations

I expect dangerously capable AI to arrive as AI civilisations/economies: a society of agents whose whole is strongly superhuman on the relevant basket of strategic tasks, while no single model is. "Superintelligence" seems apt for the whole.

Economies feel "sticky" to me. Gains from specialisation, benefits of trade, etc. Even if a superhuman AI civilisation sets out to design a single model strongly superhuman on strategically important tasks, smarter models could just be used to create new AI firms that participate in the AI economy. It's plausible AI civilisations converge on a single coherent, unified agent as they grow more capable, but I lack the intuition that intelligence amplification converges towards a singleton in the limit. Economies as a form factor seem like a stable attractor?

So decisive strategic advantage isn't a natural model for how I expect superintelligence to arrive. If anything, AI economies seem like a form factor that would precipitate human disempowerment sooner: humans could lose control before any single actor has decisive strategic advantage.

Economic Growth as a Lagging Indicator

I'm more confident in the capabilities story than the economics story. Domestic policy and the speed at which human institutions adapt make actual GDP growth a lagging indicator, so I expect transformatively capable AI before a profound economic transformation.

Same for robotics. AI would be dangerously capable well before it has a billion humanoid robots at its disposal. In a conflict, a misaligned AI civilisation would co-opt human civilisation's infrastructure and cooperate far more effectively than we do, avoiding a conventional kinetic contest. And when robotics does truly kick off, I'd expect growth faster than doubling every year.

Provenance

This post is a synthesis of my comments in the following shortform^[1]. Said synthesis was originally written by Claude (and then lightly edited by me) from the following prompt:

Read over this entire comment thread (including all footnotes) and then reply with a comprehensive write-up [in my (voice/style/writing register)] of my position as expressed/revealed in this thread.

The writeup should express my position on AI timelines on its own terms, not be framed as a response to Richard and TsviBT.

I want to cross post this on Discord.

1. ^{^}

   My comments in the shortform itself was written by me.