Is LLM Consciousness Undecidable? A Pushback on 'Just Token Statistics'
burny_tech · x · 2026-10-01
Responding to the claim that an LLM is merely a data model using token statistics and thus cannot be conscious, burnytech argues next-token prediction is just an objective—you must consider the whole complex system. He notes researchers, including the mechanistic interpretability subfield, try to empirically test consciousness theories (e.g. global workspace theory) on emergent representations inside weight matrices, though he concedes such theories are immature and hard to falsify.
His core stance: absolute certainty in either direction—conscious or not—is neither scientifically nor philosophically justified, and the debate presupposes agreement on philosophy-of-mind assumptions like physicalism, which doesn't exist.
Related event: Debate Flares Over Whether LLMs Are Conscious(4 posts)→
More from AGI Musings
- AI x Bio startups still have to do the hard stuff: labs, trials, proprietary data — anshulkundaje · 2026-10-02
- Raw intelligence is not a stock of tokens: the habit of internalizing structure — yunta_tsai · 2026-10-02
- Why AI progress confuses everyone: daily coding-model users can't explain it to others — paulnovosad · 2026-10-02
- Ben Todd: EA was always a mix of weird and sensible ideas, not a slide into weirdness — ben_j_todd · 2026-10-02
- Observer: personal agents are automating away exactly the tasks people enjoy most — signulll · 2026-10-02
- MIT report flags AI access gap as the new digital divide in education — LearnWithBishal · 2026-10-02