Daniel Kokotajlo

Use only Daniel’s labeled answers in the publisher transcript. He puts fully automated AI research around end-2028, rejects racing through an intelligence explosion, and argues that even a pause at human-level AI would radically transform the economy. Distinguish his forecasts from the host’s incident claims.

Coauthored latest forecast update: slightly shorter timelines, better evidence and modeling; combines coding uplift, revenue and time horizons. Estimates remain conditional on moving as fast as technically feasible. Distinguish Daniel’s parameters from Eli’s and Brendan’s.

Coauthored policy scenario advocating a verified international slowdown, transparent AI research and distributed power. It is a recommendation, not a prediction of AI arriving in 2040. He expects development sooner absent intervention; the concrete scenario uses another author’s timeline.

Team clarification: a slower transparent frontier can reduce power concentration and allow safety progress. China verification and government competence remain challenges. Sympathetic to full shutdown but concerned it may buy less alignment progress before agreements fail.

Thomas Larsen’s supplement to the coauthored plan, not a personal Daniel forecast. Explains public visibility into research and training activity while protecting model weights, reciprocal verification, outside scrutiny and checks against power abuses.

Thomas Larsen’s explicit assumptions, not Daniel’s personal numerical estimates. Separates confident high-level predictions and recommendations from uncertain dates, takeoff speed, alignment difficulty and ability to detect covert projects.

Historical update: Daniel moved Automated Coder median from late-2029 to mid-2028 after agentic-coding evidence and revised time-horizon estimates. Shows genuine updating; current answers should prioritize the subsequent August model and interview.

Coauthored self-evaluation grades concrete predictions rather than treating the scenario as established fact. Initial quantitative progress was slower than predicted; July amendment raises the estimated pace. Coding uplift and valuation lagged while revenue was stronger.

Coauthored correction of reporting that confused scenario years, modes, medians, raw model trajectories and different authors’ forecasts. They never claimed certainty about 2027. Superseded numerically by later quarterly updates.

Coauthored scenario linking coding automation to automated research, rapidly accelerating capabilities, misalignment and concentrated power. The scenario is a forecast exercise with branches, not an account of actual events. Later forecast updates supersede its dates.

Publisher’s speaker-labeled interview with Daniel and Scott Alexander. Use Daniel’s answers only: coding automation can remove research bottlenecks, government oversight and transparency counter secrecy and power concentration, and physical deployment still has bottlenecks. Timeline references are historical.

Third-party speaker-labeled transcript; use only Daniel’s answers, not Steven Bartlett’s framing. Asked whether we are heading somewhere bad if things don’t change, he says yes but he is not confident: something like 70%, because the current default path heads somewhere very scary. He corrects the host’s “70% chance of human extinction”: the figure is for AIs taking over or a comparably very big catastrophe, and AIs might take over without killing everyone. He does not think we are definitely doomed and could see it working out well.