binarybits: Let AI do math, humans keep control of physical infrastructure
In a discussion with efmahf, SSteromano, and others, binarybits laid out a rebuttable, pragmatic position on "AI capability extrapolation and safety regulation," centered on distinguishing "problem-solving" from "managing physical engineering."
Confirmed
- binarybits argued: solving the Navier-Stokes equations is essentially running computations on GPUs—a problem-solving capability today's models already have—while building a datacenter requires raising millions in financing, hiring hundreds of construction workers, procuring steel, concrete, and glass, and obtaining permits; managing large-scale physical projects like that is something for a decade from now, and only with rapid robotics progress.
- He criticized the extrapolation logic that equates "AI is good at hard problems like Navier-Stokes" with "AI can manage mega-projects like datacenters," and then infers a "runaway industrialization" risk; SSteromano pushed back, and the two clashed.
- His safety proposal: let AI models handle verifiable domains like math and software, while physical infrastructure stays under human control—AI can advise, but humans verify. His reasoning: physical construction itself takes months, so human review costs little. mpershan also half-jokingly endorsed a similar division of labor.
- On robots: binarybits thinks robots replacing human construction workers is at least 5 years away, so panic is premature; if the real worry is robots replacing humans at datacenter construction sites, the thing to regulate is the robots themselves, not AI models.
Why it matters
- The discussion offers a lightweight, actionable AI safety idea: don't impose human oversight on AI computation (that would badly slow things down), but instead use the physical world's slow pace and the human touchpoints as the safety valve.
- It also serves as a counter-framework to the "exponential AI capability extrapolation leads to loss of control" narrative: capability extrapolation must distinguish computational tasks from physical engineering requiring financing, labor, and permits—the latter inherently contains points of human intervention.
2026-09-15 ~ 2026-09-16 · 7 related posts
Primary sources
- [source] A lightweight AI safety plan: let AI do the math, keep humans on physical infrastructure — binarybits · 2026-09-15
- Why human oversight slows AI math but not data center construction — binarybits · 2026-09-15
- [source] Human oversight slows GPU computations but not data center builds — binarybits · 2026-09-15
- [source] binarybits: If robots taking construction jobs worries you, regulate robots, not AI models — binarybits · 2026-09-15
- mpershan's AI safety plan: let AI do the math, keep humans on infrastructure — mpershan · 2026-09-16
- binarybits vs Romano: solving hard math ≠ managing massive engineering projects — binarybits · 2026-09-16
- binarybits: today's AI solves hard math, but running mega projects is a decade away — binarybits · 2026-09-16