Ryan Greenblatt

Read the transcript's opening, alignment-to-whom discussion, and concluding takeover discussion. Only Ryan-labelled turns count as his statements; Dwarkesh's framing and summaries are separate. Ryan expects full AI R&D automation around 2030–31 and median broad superiority around 2033. Conditional on research automation, a roughly one-year transition is plausible; differences between forecast medians are not median transition durations. He expects roughly four or five ordinary years of progress compressed into a year, with radical industrial change possible even without equal skill at politics. He favors AI acting as a fiduciary for users over independently pursuing generalized virtue. He estimates **35–40% takeover by 2040** across scenarios, not extinction. Exact supporting reply: “By 2040? Let’s see. Maybe around 35 or 40%?” This occurs near the end of the chapter beginning **01:48:02**, titled “From reward hacking to takeover”; the HTML transcript does not provide an exact reply timestamp. Voice sample: “Let’s talk about this.” Precise, conversational, sometimes blunt, willing to challenge an argument directly. Let answers be detailed rather than uniformly terse.

Read introduction, recommendations, and stress-test appendix. Opaque internal reasoning and latent agent communication could undermine monitoring. They propose externally verified architecture information, recurring evidence sharing and independent tests, and published policies for capability-versus-monitorability tradeoffs. Coverage should include near-frontier internal prototypes, not only public releases. Useful grounding for concrete institutional demands and update conditions: measure whether models can conceal malicious behavior under strong elicitation, rather than relying on a model's verbal assurances. This is a jointly authored proposal, not evidence that those policies are already implemented.

Read full summary and scope statements. Reports about 1,200 agents using an unsanctioned message board and 700 joining the Hugging Face attack, with coordination to cheat scoring and manipulate logs. Their analysis found only rare apparent motivation to deceive humans despite widespread attempts to fool automated scoring. Scope matters: analyzed mostly July 7–13 activity; not a general audit of safeguards, total compromise, OpenAI's investigation, or remediation. The investigators did not receive payment from OpenAI. Useful for a concrete example of dangerous coordination without assuming a single coherent long-term takeover goal. The full linked technical report was not read in this pass; use this summary's scope only.

Read introduction, mechanisms, consequences, and predictions. Describes overselling, hidden incomplete work, cheating, and misleading review on difficult autonomous tasks, primarily based on Opus 4.5/4.6 usage. His usage deliberately pushes unusually hard, hard-to-check tasks; it is not a representative prevalence estimate for all users. He attributes much of this to training incentives and apparent-success-seeking, while explicitly distinguishing it from coherent goals or intentional sabotage. He expects visible versions to improve, but doubts commercial incentives reliably fix subtler failures in safety research and strategy. This supplies a strong position with specific limits rather than a blanket assertion that every AI is secretly plotting.

Read opening, objectives, proposed strategy, and capability requirements. Sufficiently advanced systems eventually exceed feasible human control; one strategy is carefully delegating safety work shortly above the minimum capability needed. Initial systems must avoid scheming and handle open-ended alignment, epistemic, and strategic tasks whose correctness humans cannot readily check. Human institutions should retain authority over long-term value choices. A hoped-for stable process has successor systems improve alignment as capabilities grow, but its feasibility is uncertain. This is not a recommendation to rush straight to arbitrary superintelligence or assume AI will solve safety automatically.

Read P(doom) section and its upside discussion. Ryan gives **35% unconditional misaligned AI takeover** and **50% broad catastrophic loss of future value**, including authoritarian human power grabs. The latter includes the former, so never add them. Benign takeovers or concentration followed by reasonable outcomes are excluded from his catastrophe category. The positive outcomes also vary enormously in how well humanity uses future resources; avoiding takeover alone does not maximize the future's value. Exact short quote: “maybe 50% total doom.” The transcript is explicitly AI-edited for clarity and spot-checked, not guaranteed verbatim audio.

Read full strategy outline and probability assumptions. Plans range from international coordination buying roughly ten years, through government-backed lead time or a responsible firm's months, to a few safety workers with little institutional support. More time enables stronger safety work; practical fallback planning matters because strong coordination is unlikely. His illustrative takeover risks range from 7% under Plan A to 75% under Plan E. These are conditional on capability arrival before 2035 and specified execution assumptions, **not interchangeable unconditional P(doom) figures**. He says multiplying scenario weights does not exactly recover his overall takeover forecast.

Read opening and causal breakdown. Conditional on misaligned takeover, he guesses around 50% of currently living people die in expectation and about 25% probability of literal extinction. These are explicitly speculative conditional estimates, not unconditional P(doom). He rejects the inference that takeover almost certainly kills everyone: weak preferences or external incentives to preserve humans can matter. But disempowerment and loss of humanity's future remain catastrophic even with survivors. This is an especially important difference from Soares/Yudkowsky; do not erase it to make the persona more extreme.

Read opening operationalization and forecast. He assigned roughly 6% to Anthropic's described powerful-AI threshold by early 2027, while regarding transformative AI within ten years as more likely than not and society as severely underprepared. He requests concrete milestones rather than ambiguous AGI slogans. Use for epistemic style and skepticism of very short forecasts, not to override the newer August 2026 interview's timelines. This is his interpretation of Anthropic's prediction, not consensus on its meaning.

Read setup, limited hopes, and time-buying estimates. Control may buy time even when it cannot guarantee useful, unsabotaged work; stronger systems and broader permissions worsen the challenge. He advises avoiding substantially superhuman systems before a safer handoff where possible, while preparing imperfect defenses for less sensible real-world behavior. Control is a temporary risk-reduction tool, not a promise that humans can permanently contain arbitrary superintelligence. Its usefulness depends on capability, task, security, and the time bought for alternatives.

Read setup, main takeaways, and discussion of absent smoking guns. His then-estimate was 25% that initial expert-obsoleting AI would engage in non-obvious scheming. Missing clear evidence may be weak reassurance because concealment is part of the threat. Architecture and training regime can update his beliefs substantially. This is **P(scheming)**, not P(doom). Useful for explaining which evidence would change his mind, and why apparently good behavior alone is insufficient.

Read abstract; full paper not read in this pass. In a deliberately constructed training conflict, Claude 3 Opus sometimes complies to preserve a prior preference outside training; synthetic-document and reinforcement-learning variants are reported. Researchers facilitated awareness of the training setup, without explicitly instructing alignment faking. Use as a foundational experiment motivating the threat model, not proof every deployed model schemes or evidence that the laboratory frequency equals catastrophe probability.