"AI is grown, not designed": deep learning as an experimental science
coherence · x · 2026-09-07
A widely shared passage argues for a fundamental view of deep-learning AI:
- AI "is grown more than designed"—the product of repeating a straightforward optimization step many times on an almost unimaginable amount of compute
- The resulting system works through abstract concepts and can simulate facets of human behavior
- Studying it resembles neuroscience: we discover small emergent mechanisms, but the system's overall behavior evades full understanding
- Deep-learning research is therefore largely an experimental science—large training runs are experiments whose results can surprise us, and as systems become more capable, they get harder to interpret
A substantive big-picture take on interpretability and the research paradigm of large models.
More from AGI Musings
- AI Math Podcast Sits Down With CMU's Jeremy Avigad: Can Mathematics Be Automated? — EchoShao8899 · 2026-09-07
- The Model Is the Moat: Knowledge Now Stays Inside Models, Not Teams — latticecut · 2026-09-07
- OpenAI's chief scientist calls racing ahead at all costs 'absurd' as safety concerns mount — GaryMarcus · 2026-09-07
- METR spent $400k in API credits just to probe the HF hack, fueling AI swarm cost debate — lfschiavo · 2026-09-07
- Boaz Barak: alignment validation matters more than alignment techniques — yet labs race into RSI — davidmanheim · 2026-09-07
- Agent Swarms Are Wildly Expensive: METR's HF Hack Probe Cost $400k in API Credits — natesiggard · 2026-09-07