24 Years On, a 2.5k-Parameter Creature Experiment Still Undercuts AI Consciousness Claims
robleclerc · x · 2026-10-03
The author revisits work from 24 years ago: evolving neural networks of virtual creatures in a simulated 3D environment to get them to talk. He watched hundreds of such scenes and notes how easy it was to project animism onto creatures that struggled and fumbled around — even though they were simple 4-layer networks with 20-30 neurons per layer, roughly a 2.5k-parameter model.
His conclusion: current AI is just a scaled-up version of this, and there is still no reason to think it is conscious.
The post quotes a thread on Anthropic co-founder Olah's claims about models showing "introspection" and internal states that "functionally mirror joy, satisfaction, fear, grief and unease." The quoted take argues the key word is "mirror" — what is described is mimicry, and rewarding that with legal rights is no saner than granting personhood to a macaw that learned to ape its handlers.
More from AGI Musings
- Acemoglu questions whether AI drug-discovery promises can justify massive AI investment — asusarla · 2026-10-03
- 'No Meaningful Bond With a Turing-Test-Passing AI'? One Line About Dogs Sinks the Argument — SydSteyerhart · 2026-10-03
- Burkov: Generative AI Only Makes Money for GPU Sellers, Echoing Dotcom Bubble — burkov · 2026-10-03
- XPRIZE offers $10M prize to enable meaningful communication with wild animals — PeterDiamandis · 2026-10-03
- Why We Desperately Want a Scientific Definition of AI Consciousness — moonsandhues · 2026-10-03
- Grimes pokes at AI builders: "You seek to build god yet deny his existence" — moonsandhues · 2026-10-03