I guess I'm somewhat confused by the longer timelines espoused by people like @Richard_Ngo (exhibit) and @TsviBT (exhibit A, exhibit B, exhibit C, exhibit D). Not in a sense of "this is incoherent or obviously wrong", but perhaps a failure of imagination, a "what would that world even look like"? What "timelines" mean is somewhat nebulous, but for the purpose of this post, I'm talking about timelines to AI systems that are sufficiently powerful to:

1. Precipitate high economic growth (say the US economy doubling under 4 years)^[1][2]
2. Precipitate human disempowerment or catastrophically bad outcomes if they were sufficiently misaligned.

Alternatively, the kind of AI about that this community has spent two decades discussing how to align to human values to ensure that the future goes well.

Henceforth, I will refer to this property as "transformative".

From my layperson perspective in late October 2026, we have superhuman programmers and superhuman mathematicians. It does not seem like we've even come close to exhausting what RLVR has to offer us. More RL environments are being created for real world engineering tasks and for arbitrary white collar work. And I expect this to continue for an ever growing basket of tasks.

We have not yet exhausted scaling on test time compute and swarm scaling seems like yet another scaling frontier with much more to give. Time horizons for the tasks agents can do continue to increase, arguably at an exponential rate, though some claim it is instead linear.

Even without RSI (though I expect that we would see AI contributing a larger and larger share of AI engineering/research effort over time), it feels like the default path^[3] should see continued strong improvement in AI capabilities for years to come.

My naive layperson perspective is that individual AI models would continue to acquire more general capabilities, and grow ever more powerful within each magisterium they conquer^[4], and AI civilisations would continue to grow even more powerful and this will continue for years to come.

And it feels like this is true even without solving any fundamental foundational problems behind general intelligence. Like even without creating a single model that is as data/sample efficient as humans, as readily capable of continual learning^[5] and adapting to novel environments as humans. Even without creating a highly agentic, fully autonomous model with independent drives in the same way humans have independent drives. Even without developing any fundamentally new paradigms in ML research/engineering^[6].

It feels like just scaling the existing basket of techniques that exist as of late October 2026 is sufficient to produce AI civilisations that are capability wise "transformative"^[7] in the sense described earlier. These AI civilisations would probably not be well conceived of as a singleton that possesses decisive strategic advantage with respect to human civilisation^[8]. But it does feel like they will be at a capability level where "superintelligence" is an apt description.

It seems to me that Richard, Tsvi and others with decade plus^[9] timelines have a different perspective on the "default path" and I would like to (hear why / be brought up to speed on what it is that I am missing).

1. ^{^}

   This is conditional on said systems being widely adopted/deployed throughout the economy.

2. ^{^}

   The economy doubling under a year is still doubling under four years. Four years just feels like a good round number for historical reasons.

3. ^{^}

   I mean something like the counterfactual where the AI race continues from late October 2026 as it has been going so far, there are no major regulations or industry agreements for a pause or explicit pacing the frontier, etc. Maybe "default" is not the best name as I do not mean to predict that it is significantly more likely than not that we do not get any such (agreements/regulations).

4. ^{^}

   I model October 2026 LLMs as having conquered the magisteria of programming and mathematics.

5. ^{^}

   Though continual learning being solved on the default path before 2035 would not at all surprise me, since it seems labs are making progress on this.

6. ^{^}

   There will be new paradigms across the next decade (just as we saw the novel paradigms of scaling pretraining, RLVR, scaling test time compute, and scaling multi agent cooperation).

7. ^{^}

   This is not a prediction that we would necessarily see such growth, just that conditional on the systems being broadly deployed throughout the economy, such growth would follow. I am more confident in the capabilities story than in the economics story. Human institutions are slow to adapt and change, I expect GDP to be a significantly lagging indicator, etc.

8. ^{^}

   The operative distinction here is "singleton". I do not expect a singleton with decisive strategic advantage because I do not expect a singleton. Decisive strategic advantage isn't a natural model for thinking about the form factor in which I expect superintelligence to arrive in, since again I do not expect a singleton.

9. ^{^}

   This is plausibly inaccurate. Richard has staked a position against superintelligence in eight years and Tsvi has staked a position where his timelines concentrate mass in 20 - 40 years. I don't know what Richard's actual timelines are and how much I actually disagree with him.