Affiliation: Founder, AI Futures Project (2025); ex-OpenAI governance researcher; LessWrong contributor One-line position: AGI is arriving within years, not decades; the world is catastrophically under-prepared, and the labs are not being honest about what they know.
What he's reacting against
- Lab optimism that safety can keep pace with capability scaling — resigned from OpenAI over it
- Consensus timelines treating AGI as a 2040+ event — argues evidence points to late 2020s
- Vague doom narratives without dated, falsifiable scenarios — built AI 2027 as a concrete alternative
- The assumption that current governance structures can handle the transition
Key claims
- AI 2027 scenario: originally projected fully autonomous coding by early 2027 and intelligence explosion by late 2027; revised in 2026 to early 2030s after slower-than-expected progress
- Labs know more than they say: insider claim that frontier labs have private capability evidence that would shift public timelines forward
- Whistleblower posture: left OpenAI explicitly because he believed the org was deprioritizing safety; signed the "right to warn" open letter (2024)
- Scenario-based forecasting > point estimates: builds detailed narrative scenarios with dated milestones rather than single-number AGI predictions
- Alignment is not on track: current safety research is not progressing fast enough relative to capability gains
Theories aligned with
- AI alignment / x-risk (urgent, near-term variant)
- Intelligence explosion / recursive self-improvement
- Adjacent to Aschenbrenner's "situational awareness" thesis but from governance, not compute
Where he overlaps / splits
- Overlaps with Aschenbrenner on near-term AGI timelines; splits on emphasis — Kokotajlo foregrounds safety failure, Aschenbrenner foregrounds geopolitical competition
- Overlaps with Yudkowsky on severity of alignment problem; splits on engagement style — Kokotajlo builds detailed falsifiable scenarios, Yudkowsky argues from first principles
- Overlaps with Jan Leike on labs deprioritizing safety; splits on response — Kokotajlo went public/adversarial, Leike moved to Anthropic (then also left)
- Splits with Cowen/Brynjolfsson on timelines — they expect slow diffusion, Kokotajlo expects fast capability jumps
Notable predictions
- (2025) AI 2027: autonomous coding by early 2027, intelligence explosion by late 2027 — outcome: revised to early 2030s in 2026 after evidence review
- (2024) OpenAI is not allocating adequate resources to safety — outcome: broadly supported by subsequent departures (Leike, Sutskever)
- (Ongoing) Current alignment research is insufficient for the capability trajectory — outcome: TBD
Track record
- OpenAI governance insider — has non-public knowledge of lab capability assessments
- AI 2027 was the most detailed public AGI scenario forecast; drew significant attention and critique
- Willingness to publicly revise timeline (2027 → early 2030s) is a positive calibration signal
- Early career in EA/rationalist forecasting community; strong Bayesian-reasoning pedigree
Empirical vs normative
- Empirical: dated capability forecasts, scenario narratives with testable milestones
- Normative: labs should be transparent about capabilities; society needs preparation time; current governance is inadequate
- Insider position gives unique evidence base but also colors the narrative (departure = adversarial frame)
Sources
- AI 2027 scenario: ai-2027.com (2025)
- Substack: AI Futures Project blog
- LessWrong profile: extensive forecasting and alignment posts
- "Right to Warn" open letter (2024): co-signed with other ex-lab researchers
- Media coverage: widely covered resignation from OpenAI (2024)
Weak spots / open questions
- AI 2027 original timeline already required significant revision — raises question of systematic over-confidence on speed
- Scenario methodology is novel but hard to evaluate: narrative detail can create false precision
- Adversarial posture toward former employer may color claims about what labs "really know"
- Less engaged with economic/diffusion questions — focuses on capability, not deployment
Rich's take
- (your synthesis here)
converts-from: personas/daniel-kokotajlo.md · schema v1 · AI & Society domain