Keep pulling the thread on Daniel Kokotajlo.
The AI Futures Model predicts timelines to full coding automation that are 3-5 years longer than the previous AI 2027 model.
Representatives from Anthropic and OpenAI have suggested AGI could arrive as early as 2027 or 2028.
Frontier AI company revenue is approximately $20 billion and growing at a rate of about 4.1x per year, according to an Epoch dataset.
The AI Futures Model predicts a median arrival date for a superhuman coder (SC) in 2032, compared to predictions of 2027-2028 from the previous model.
Researchers at Anthropic and OpenAI are reporting that AI-assisted labor is beginning to accelerate their work.
The AI Futures Model, with Eli Lifland's parameters, calculates a 26% probability of the transition from Automated Coder (AC) to Artificial Superintelligence (ASI) occurring in less than one year.
METR may cease measuring its time horizon benchmark by 2027 due to the exponentially increasing cost of building tasks and collecting human baselines.
The rate of progress on the METR benchmark trend has been notably higher since 2024 compared to previous years.
Daniel Kokotajlo's personal forecast assigns a 9% probability to AGI arriving by the end of 2026.
The new AI Futures Model, with median parameters, predicts the arrival of a Superhuman Coder (SC) in December 2031.
In the median case of the AI Futures Model, a "taste-only singularity" does not occur; each capability doubling from research taste improvements alone would take 20% longer than the previous one.
The AI Futures Model's longer timelines for full coding automation are primarily due to being less optimistic about the speed of AI R&D acceleration before full automation is achieved.