In what domains have AI systems matched skilled humans? Arguably: mathematics, software engineering, call customer support triage, pattern matching/identification, standardised exams (e.g. bar exam), translation, paralegal work (discovery),
In which domains and at what tasks do they still usually fail? Resisting prompt injection, real-time conversation (audio uncanny valley), physical world manipulation (though progress seems to be rapid here), out of distribution reasoning.
Where does it look like scaling might not hold?
When will physical limits to AI development: hit? be overcome?
Will such limits to AI development impact all frontier AI companies uniformly? If not, how will the impacts be spread? The four key factors driving AI progress are underpinned by increasing computational power to run and train AI systems, each of which is likely to continue until at least 2028/9.
How long can AI systems operate without human help?METR’s time horizons measure task length by reference the time a human expert software engineer would need to complete a given coding task. A 1-hr horizon at 50% reliability means an AI is likely to succeed half the time on software tasks that take human experts about an hour. The 80% reliability measure is often preferred. Horizons in domains outside software engineering will differ.
What economically valuable work (outputs) can AI systems automate? What do they already automate? Note that the AI system would not necessarily need to do the work as well as the average human. Work is liable to automation for several reasons. Relevant factors: capabilities of the AI model, reliability, level of human help needed, cost. Note that probabilities compound (a chain of five 90%-reliability-tasks is only 60% reliable - this impacts long horizon task planning/execution).
How much of the AI research and development pipeline can be done by AI systems, i.e. how close are we to recursive self improvement? E.g. data collection/processing, compute, algorithms, pre- and post-training methods, system deployment.
What are the impacts if a model with X capabilities is adopted at scale? What happens when billions of people are able to use current AI systems to their full potential?
What if it’s kept under the control of a single person/company/nation, or small number of actors?
How can we distinguish between genuine bottlenecks (ceilings) versus those that increase cost, or slow progress? What factors fall in which category?
