When will computer hardware match the human brain? (1998)
Predictions from a 1998 essay on when computer hardware would match human brain power are reexamined in light of today’s GPUs, supercomputers, and AI models. Commenters debate whether metrics like MIPS and FLOPS meaningfully capture brain-like computation, pointing to architectural differences, unknowns in neuroscience, and the vast energy efficiency gap between silicon and biology. Many conclude that while raw compute may now be in the right ballpark for some brain-scale simulations, the limiting factors are data, theory, and algorithms rather than sheer hardware capability.
Hardware Progress and Moravec’s Predictions
- Many note Moravec’s late‑90s plots still roughly fit modern trends if you count GPUs/FLOPs, not just CPU MIPS.
- MIPS scaling largely stalled, but GPU FLOPs exploded; modern GPUs and supercomputers approach or exceed his “brain-equivalent” ballpark on some metrics.
- Some argue you must normalize by cost ($/compute) as Moravec did, which makes supercomputer comparisons tricky.
- Others point out his scale implies we could simulate a nematode in 1998; since that’s still hard, his simple ops/sec mapping is suspect.
What Sustains Moore’s Law–Like Growth?
- Suggested drivers: huge private investment, positive feedback from using current chips to design the next, and market pressure for yearly performance gains.
- Skeptics say this “the market did it” answer just shifts the question of how we keep finding advances.
Brain vs Computer: Architecture, Energy, and Limits
- Repeated emphasis that brain ≠ von Neumann CPU: massively parallel, largely analog, 3D, with complex cell types (neurons, astrocytes), synapse dynamics, and brain waves.
- The brain runs on ~20 W; GPUs need hundreds of watts, though some argue electricity is cheaper and human “training” takes decades.
- Disagreement over whether we can treat neurons/synapses as simple weights vs needing detailed biophysics. Analog vs digital nature of the brain is called unresolved.
Defining and Detecting Intelligence
- One camp aligns with a performance-based view: intelligence is what systems do, not what they’re made of.
- Others argue internals matter: a lookup table or shallow chatbot can pass a Turing test without “real” intelligence.
- Chollet’s definition (skill acquisition efficiency over many tasks with minimal priors/data) is cited; critics say it misses human-like logical reasoning and “executive” control.
- LLMs are seen by some as powerful yet sample-inefficient, lacking robust reasoning and true memory; others note similarities to predictive brains plus missing working-memory mechanisms.
Do We Already Have Brain-Scale Compute?
- Some rough estimates: billions–trillions of GPU FLOPs may already rival or beat brain-level ops for some abstractions.
- Counterarguments stress missing pieces: unknown brain algorithms, data, memory bandwidth, and realistic neuron/synapse models.
- Back-of-envelope transistor-per-synapse and neuron-count calculations are heavily disputed as oversimplified.
Simulation Efforts and Timelines
- One commenter involved in brain modeling claims a full mouse brain multiscale simulation fits on a modest supercomputer, with room for simplification.
- They hypothesize top supercomputers could simulate a human brain if we had adequate data and theory.
- Others speculate real-time full human-brain analogs might appear late this century, while some insist brains aren’t computers and such equivalence may never be meaningful.