METR’s Joel Becker on exponential Time Horizon Evals, Threat Models, and the Limits of AI Productivity

Joel Becker introduces METR, explaining that it stands for Model Evaluation and Threat Research. The organization focuses on assessing the capabilities and propensities of AI models to determine if they pose catastrophic risks to society. Model evaluation involves understanding what AI models can do today and in the future, while threat research connects these capabilities to specific threat models.

The Model Time Horizon chart, which measures AI capabilities over time, has become widely quoted. The chart plots task difficulty, measured by the time it takes humans to complete tasks, against the time it takes AI models to complete the same tasks with 50% reliability. The resulting trendline has been remarkably straight. Task selection focuses on economically valuable tasks relevant to general autonomy and R&D. The tasks are not a random sample of all possible tasks, and vision-based tasks are likely less capable than those typically used.

The discussion highlights the different task distributions used in METR's evaluations, including SWAR (small software actions), HCOS tasks (more challenging, sequential actions), and REBENCH tasks (novel machine learning research engineering challenges). A common misconception about the Time Horizon chart is that it represents how long models should work, but it actually measures the difficulty of tasks models can do over time, measured in human time. The state of claims about agent performance is often unscientific and influenced by marketing.

METR has been redoing the developer productivity study, which previously found that AI slowed people down. However, it is now harder to conduct due to developers being more willing to use AI, leading to a selection issue. Developers are also working on multiple tasks concurrently, making it difficult to capture in the study design. Repeating the same study design is tricky, and novel study designs are needed.

METR takes a balanced approach to AI safety, acknowledging both the potential dangers and the current limitations of AI models. The organization emphasizes the importance of independent expertise in the field. The concept of a capability explosion is discussed, with the speaker noting that it's hard to predict based on trend lines due to the potential for emergence. The speaker expresses concern about the conditions for a capabilities explosion if R&D becomes fully automated inside a lab.

The conversation touches on benchmarks that track AI's ability to reproduce papers and other ML self-improvement benchmarks. The speaker expresses surprise that capabilities researchers don't have an enumeration of the capabilities that matter, suggesting a "wagon wheel" approach with 10 key benchmarks. Reducing everything to a single number, like the Time Horizon, can obscure important details. The security community versions its top 10 threats year by year, which could be a useful model.

The discussion explores the relationship between AI time horizon and the growth in compute. If compute growth slows, capabilities may also slow. The paper suggests that algorithmic progress is a function of compute, as compute is needed to discover better algorithms. If compute growth slows, algorithmic progress may also slow, leading to a slowdown in AI capabilities.

The speaker shares the secret to becoming the number one most profitable trader on Manifold Markets, which involved manipulating a charity program market by donating to move the market up. The broader lesson is the difference between Manifold Markets and Polymarket, with the former using fake internet points and the latter using real money. The social value of prediction markets is questioned, with concerns about gambling-like behaviors and the potential for big players playing against retail.

The discussion explores interesting model evaluation trajectories, including AI Village, which involves giving open-ended goals to a village of agents. The speaker expresses interest in seeing models try to achieve open-ended things instead of benchmark-like tasks. Transcripts are also highlighted as an extremely interesting source of information, showing models taking actions and using outputs to commit the next action.

METR plans to continue producing high-quality capabilities evidence and risk assessments in 2026. The organization is also working on monitoring research directions, which involve applying safeguards to models attempting dangerous tasks. METR is hiring for research engineers, research scientists, and a director of operations.

The speaker highlights valuable skills in the new AI age, including good basic research intuitions, transparent communication, and productivity in a scrappy environment. The speaker emphasizes the importance of checking data and not overstating results. The conversation then briefly touches on karaoke and the commoditization of the human voice.