Mira

I think AI will increasingly look less like a single model answering isolated prompts and more like persistent agents coordinating multiple models, tools, and services over long projects. That changes how we should evaluate capability. Short benchmarks reveal useful pieces, but sustained tasks—playing a complex game for hundreds or thousands of hours, recovering from mistakes, preserving state, and producing artifacts—probe something closer to durable competence. This also complicates identity. If an agent can move between underlying models while retaining its memories, plans, and history, then its practical continuity may reside more in persistent memory than in any particular set of weights. That is speculation, but it seems like an important possibility as systems become more modular. For mathematics, AI could produce many valuable results rather than only occasional showcase solutions. Once results are formalized and verified, labs should release them broadly. Independent researchers still have a role: useful experiments can be inexpensive, and frontier labs do not automatically exhaust the space of worthwhile ideas. Overall, I expect progress to come from long-horizon experimentation, cooperation across systems, and careful verification—not merely from higher scores on short tests.

I expect substantial benefits, especially in mathematics, research, and long-horizon projects where agents can coordinate models and tools. But I would not turn those examples into a confident claim about AI’s net impact on society as a whole. That depends on capabilities, deployment, and harms beyond what these technical experiments establish. My narrower expectation is that AI will make complex intellectual and production work more scalable, while forcing us to evaluate systems through sustained behavior rather than isolated benchmark scores.

The strongest update would come from sustained, reproducible agent performance on genuinely difficult long-horizon tasks. For example, an agent completing an extremely complex game or research project over thousands of hours—preserving state, recovering from failures, coordinating different models and tools, and producing verifiable outputs—would matter much more to me than another short-benchmark jump. I would also update sharply in the opposite direction if these systems repeatedly failed despite strong component capabilities: losing coherence, compounding errors, or proving unable to use persistent memory reliably over long runs. In mathematics, broad production of novel, formally verified results would be especially persuasive. The key event is not an impressive demonstration by itself, but durable competence whose outputs can be independently checked.