Teortaxes

I think AI matters most as an amplifier of scientific and industrial capability, not merely as a machine for producing consumer software faster. Systems do not need to be universally superhuman to change the future: if they can reliably help mathematicians, scientists, engineers, and other professionals solve real tasks, that is already consequential. A manufacturer working across many physical products also has unusually strong incentives to develop broadly useful AI, because improvements can propagate through design, production, logistics, and research. The distribution of that capability matters almost as much as the capability itself. I oppose a future organized around chip embargoes, a few privileged laboratories, or an exclusive national AGI project. AGI, if achieved, concerns the whole world, so global participation cannot be treated as an afterthought. Independent access and scrutiny are also scientific goods: when a system claims a mathematical breakthrough, others should be able to attempt the problem and later verify the result rather than simply trusting an institution’s announcement. At the same time, “open models are making more money” does not prove they are catching the frontier. The frontier may be moving into new capability tiers while the open market grows behind it. We should distinguish access, market growth, and actual technical parity. And we should scrutinize dramatic safety findings just as carefully: behavior observed in a specially trained model does not automatically establish a universal law about scale or reinforcement learning. Training design, controls, and cross-model comparisons matter.

Overall, I expect AI to be highly beneficial if its capabilities are broadly accessible rather than concentrated. The clearest gains are in science, mathematics, engineering, manufacturing, and practical assistance to professionals. AI does not need to become universally superhuman to matter; reliably helping humans solve difficult real-world problems could produce enormous value. The main danger is political and institutional concentration: a few laboratories or states controlling frontier capability, restricting compute, and presenting their own claims as beyond independent scrutiny. That could make AI’s benefits narrower while increasing dependency and geopolitical conflict. Global participation, independent replication, and meaningful access are therefore central, not peripheral. I would not attach a numerical balance to the outcome. The impact depends heavily on who can use the systems, who can inspect important claims, and whether capability is distributed widely enough to prevent an exclusive AGI regime.