Nathan Benaich has released the ninth annual State of AI report, framing this edition around AI helping build better AI, world models, physical AI, geopolitics, cyber defense and predictions for the next 12 months.
One headline number comes from Anthropic rather than an independent audit. Benaich cited the company’s internal index, which measured Claude leading 26% of model research and development work in August while humans supervised. He also said agent use in knowledge work is climbing.
For Benaich, the next question is not simply whether agents can complete research tasks. It is whether they can develop “scientific taste”: choosing worthwhile experiments, recognizing elegant approaches and abandoning an unproductive line of work.
Robotics gets its own scaling story
Benaich argued that robotics is approaching its “GPT-2 moment”, while acknowledging that some observers would already call it GPT-3. His case rests on broader pretraining helping robots generalize to unfamiliar tasks and world models letting them simulate possible outcomes before acting in the physical world.
Wayve co-founder and CEO Alex Kendall highlighted one production example. He said the report includes updates on GAIA-4, Wayve’s world model, which the company has paired with a full simulation harness.
The attached Sources establish the report’s release and these selected themes, but they do not include the full report or its complete prediction list. Claims about specific forecasts therefore remain outside this summary.
