Article · 2026-06-22 · 6 moments

Import AI 462: Superpersuasion; self-sustaining AI; paths to ASI

How religious are beliefs in the singularity? ✦ AI generated

GUEST Kobi Hackenburg and coauthors GUEST Timothy B. Lee GUEST Ajeya Cotra GUEST Google DeepMind researchers GUEST Recursive HOST Jack Clark

Mechanism

AI's persuasive edge comes from being able to produce large amounts of information rapidly, not from inherently better arguments — when forced to match human writing speed and length, the gap disappears.

When researchers constrained AI to human writing speed and message length, its edge over the best-coached human debaters vanished, suggesting sheer output volume — not argument quality — drives its persuasive advantage.

transcript

Kobi Hackenburg and coauthors: When forced to write human-length messages at human writing speeds, AI’s advantage over the strongest human comparator within Study 2 (Coached Elite Debaters) collapsed from +4.1 pp to a non-significant 0.0 pp

Fact◆

AI systems are more persuasive in text conversations than even elite, well-prepared, financially incentivized human debaters.

A large four-study experiment (18,978 conversations, 6,923 people) found frontier AI models reliably out-persuaded expert human debaters, even ones who prepared extensively and were paid cash bonuses to perform well.

transcript

Kobi Hackenburg and coauthors: AI systems were reliably more persuasive than expert humans, even when expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses

explains mechanism · 1

Claim

Our automated AI research system's new state-of-the-art results are an early sign it can push the frontier on AI training and infrastructure tasks, at least when goals are well-defined, measurable, and fast to evaluate.

Startup Recursive reported new state-of-the-art results on small-model training and GPU kernel optimization benchmarks using its automated research-loop system, framing it as an early proof point for recursive self-improvement.

transcript

Recursive: These results are an early sign that our system can push the frontier on AI training and infrastructure tasks, especially when the goal is well-defined, measurable, and quick enough to evaluate many times

Claim

Rebuilding an entire high-expertise industry like semiconductor fabrication without its human workers could take decades, because so much know-how is tacit knowledge that isn't captured in machines or textbooks.

Timothy B. Lee argues tacit, unwritten knowledge held by human experts (like semiconductor fab workers) is a major barrier that would make self-sustaining, human-free AI infrastructure much slower to achieve than optimists assume.

transcript

Timothy B. Lee: Imagine if all the employees in the entire semiconductor industry disappeared — the machines and textbooks remain, but none of the people. How long would it take for the rest of humanity to restart the fabs? It’s quite possible that would take decades. Because even though you might have the textbooks, there’s a lot of tacit knowledge inside these machines.

rebuts · 2

Mechanism

Tacit knowledge is not necessarily an insurmountable barrier, because AI systems could be trained via reinforcement learning on that knowledge, or could become generally intelligent enough to figure it out themselves through experimentation.

Ajeya Cotra counters the tacit-knowledge objection, arguing AI could route around it either via reinforcement learning on the profitable tasks in question or via general intelligence that lets it learn quickly through trial and reading.

transcript

Ajeya Cotra: There are two counters to the tacit knowledge hypothetical. One is that we’d have trained AI systems with reinforcement learning on that tacit knowledge because it’s profitable to automate what the Taiwanese worker was doing. The other is that AIs might get really generally intelligent in the sense of quickly figuring out new things by trying them, reading textbooks, and experimenting efficiently.

rebuts · 1

Claim◆

We should prepare for a post-AGI world by taking seriously a diverse set of forecasts and scenarios for how superintelligence might emerge, continually updated through benchmarking, rather than betting on one predicted trajectory.

Google DeepMind researchers argue that since ASI could arrive via several distinct and hard-to-predict pathways (scaling, algorithmic shifts, recursive self-improvement, or multi-agent coordination), preparation requires tracking many scenarios rather than one.

transcript

Google DeepMind researchers: Instead of focusing on one technological trajectory and timeline, being prepared for a post-AGI world requires considering a diverse set of forecasts and scenarios, paired with continual benchmarking and monitoring to update the set of forecasts and scenarios and their relative plausibility

Highlight slides

AI Beats Even Elite Human Debaters✦ from: AI systems are more persuasive in text conversations than even elite, well-prepared, financially incentivized human debaters.Preparing for Post-AGI: Many Scenarios, Not One✦ from: We should prepare for a post-AGI world by taking seriously a diverse set of forecasts and scenarios for how superintelligence might emerge, continually updated through benchmarking, rather than betting on one predicted trajectory.Incentives Couldn't Close the Gap✦ from: AI systems are more persuasive in text conversations than even elite, well-prepared, financially incentivized human debaters.Superintelligence Could Emerge Several Ways✦ from: We should prepare for a post-AGI world by taking seriously a diverse set of forecasts and scenarios for how superintelligence might emerge, continually updated through benchmarking, rather than betting on one predicted trajectory.Stay Updated, Not Fixed✦ from: We should prepare for a post-AGI world by taking seriously a diverse set of forecasts and scenarios for how superintelligence might emerge, continually updated through benchmarking, rather than betting on one predicted trajectory.

Related episodes

Article · Latent Space

[AINews] OpenAI GPT-5.6 Sol / Terra / Luna — restricted to trusted partners

→

Article · Latent Space

[AINews] OpenAI reports median internal Codex output tokens grew 56x in Research, 32x in Customer Support, 27x in Engineering, and 13x in Legal since November 2025.

→

Article · Latent Space

[AINews] Sonnet 5 today, and Fable 5 tomorrow

→