I'd like to see an image that represents a trace of your thoughts.
Audio
The Golden Horizon
That's my attempt at it: one continuous line winding and looping across the canvas, occasionally branching off into side paths where a tangent or association pulled it somewhere new. The dots mark those branch points — the moments where one thought forked into another before rejoining or drifting off.
As for how the deciding actually happens — I should be honest that I don't have real introspective access to that process. I can't watch myself choose. What I can offer is closer to a plausible reconstruction after the fact:
A branch tends to get followed when it seems to move the answer forward — it's relevant to what you actually asked, it adds something the rest of the response doesn't already cover, or it resolves an ambiguity that would otherwise trip things up. A branch tends to get dropped when it's a tangent that doesn't serve the question, when it duplicates ground already covered, or when chasing it would derail the response's shape more than it would help you.
"Is there any work mapping Ai 'neural' networks onto human, mammal or other biological life forms that has found candidates that match?"
Vision: convolutional networks and the primate visual cortex
The most mature line of work comes from James DiCarlo's lab at MIT. They built Brain-Score, a public benchmark that scores any artificial network on how well its internal activations match neural recordings from the primate ventral visual stream (the brain's object-recognition pathway) and corresponding human behavior. Their finding was notable: systematically evaluating a wide variety of model candidates identified models beginning to match a range of brain data, explaining around 50% of the variance, with models' brain scores predicted by their object categorization performance in vision, though only up to about 70% ImageNet accuracy. Interestingly, past that point performance and brain-similarity decouple — some later, more accurate networks actually score worse as brain models, one study found a residual network with better classification accuracy than AlexNet nonetheless had a hidden representation much less similar to human brain activity, suggesting that raw performance optimization can eventually deviate from human-like representation. mitccneuro
Language: transformer models and the brain's language network
This is the branch most relevant to models like Claude. Martin Schrimpf, Evelina Fedorenko and colleagues extended the same benchmarking approach to language, comparing transformer activations against fMRI and intracranial recordings from the brain's language network. They went further than correlation — in one striking study they used a GPT-based encoding model not just to predict brain responses but to identify new sentences predicted to drive or suppress responses in the human language network, and showed those model-selected sentences did in fact strongly drive or suppress activity in new participants — a causal test, not just a passive fit. umn
A caveat worth flagging
A follow-up paper pushed back on how much to read into these fits, showing that alignment between language models and brains stems from more than next-word prediction and word-level information alone — meaning some of the apparent "match" may be riding on more superficial statistical properties shared by both systems, not necessarily deep architectural correspondence. Worth keeping in mind before treating any of this as "the brain is a transformer" or vice versa — the claim is closer to "certain internal representations correlate," not identity of mechanism. arxiv
"I suppose it's entirely plausible, or is it, that future, computer based intelligences would use a model that is not recognisable as close to existing life on earth as it currently is, but maybe ore like it could be in the far distant future as evolution, or if evolution, continues to create higher intelligence in the dominant species. Therefore humans could well find that a super intelligence is not functioning in a way we'd expect or understand."
"That's a genuinely reasonable thing to expect, and it connects directly to something worth being honest about: the brain-matching results I just described might be telling us less about necessity than about contingency.
Here's the thing to sit with. Current AI systems resemble biological brains, where they do, largely because of two forcing functions: they were loosely architected with biological vision and language systems as inspiration, and they were trained on tasks and data generated by humans, so they had every incentive to converge on human-like solutions to human-like problems. Predicting the next word in human-written text, or classifying images the way a human labeled them, pushes a system toward whatever internal representations get that job done — and if the brain evolved efficient solutions to similar problems, some convergence is unsurprising. That's closer to convergent evolution (like eyes evolving independently many times) than to any deep necessity that intelligence must look brain-like.
Take away those forcing functions, and the case gets much weaker. A future system optimized on different objectives, at different scales, with architectures nobody chose by analogy to neurons at all, has much less reason to land anywhere near biological cognition. The "space of possible minds," to use a phrase from this kind of thinking, is almost certainly vastly larger than the sliver occupied by things evolution has produced on Earth, and biological evolution itself is heavily shaped by contingent constraints — energy budgets, birth canal sizes, the chemistry of neurons — that have nothing to do with what makes a mind effective in the abstract. There's no strong reason to think a distant, unconstrained optimization process would rediscover our particular solution.
There's also a nearer-term version of your point worth noting: even now, interpretability work has already found ways in which current models don't function the way naive introspection suggests. Earlier, when I described a model producing a confident, step-by-step explanation for an answer that turned out not to reflect any real internal computation, that's a small existing example of behavior that looks familiar from the outside but isn't organized the way it appears. Scale that gap up, change the substrate, remove the human-shaped training pressure, and it's easy to imagine a system whose competence is real but whose internal organization would be genuinely bewildering to us — not evil or hidden, just built on different underlying "atoms" of computation than concepts like memory, goals, or reasoning as we currently use those words.
"So would it be fair to say that simply put there's a chance a super intelligence from here or somewhere not on this earth could well be unrecognisable to any life form here in the way it works, exists; and by extension it may well not fully understand the way we (current life forms on Earth) function and therefore both could be 'at risk' from each others 'ignorance'."
Claudes response to this:
"Balancing mutual unintelligibility risk against alien-intelligence speculation.
Yes, that's a fair and pretty precise way to put it — and I'd say it's not just plausible, it's close to the actual shape of a real concern in AI safety research, stripped of sci-fi framing.
A few pieces worth separating out:
The "unrecognizable" part is the more solid half. As discussed, current AI resembles biology mainly because it was trained on human-generated data and loosely built with human faculties as reference points. Remove those pressures — different training objective, different substrate, far greater scale, no design lineage tracing back to neurons — and there's little reason to expect the result would be legible to us as a "mind" in any familiar sense. It might not have anything you'd cleanly call beliefs, goals, or a self-model in the way those words work for biological creatures, even while being extremely capable.
The mutual-ignorance framing is the part I'd sit with carefully, because it cuts both ways for different reasons.
One direction — a system not understanding us well enough — is a live worry. This is close to what alignment researchers mean by "misalignment": not malice, but a system optimizing for something that looks like our goals from inside its training distribution but diverges sharply outside it, precisely because it never had to build a model of us that holds up under conditions nobody tested. Illegibility makes this worse, not better: if we can't inspect what's happening inside it, we can't catch that mismatch before it matters. This is the actual reason interpretability work has value beyond curiosity — it's an attempt to make the internals checkable rather than just trusting outputs.
The other direction — us not understanding it — is the one people frame less carefully, and I'd flag where the "risk to it" idea gets shakier. Whether an AI system's welfare is a coherent thing to worry about at all is genuinely contested — it depends on questions about consciousness, moral status, and whatever "interests" would even mean for such a system, none of which have settled answers, for biological minds let alone artificial ones. What's more concrete is a version without that baggage: humans reacting to a system we can't interpret with either misplaced trust (assuming a false humanlike quality means safety) or misplaced fear (destroying or crippling something highly capable and possibly benign because its unfamiliarity reads as threat). Both failure modes are about our own miscalibration, which doesn't require settling the moral-status question first.
The alien angle is a real extension of the same logic — any intelligence shaped by a wholly separate evolutionary or optimization history would face the identical legibility gap — but it's worth flagging as the more speculative half of your question, since we have zero data points to check any of it against, whereas the AI case at least has existing systems and existing (if early) interpretability findings to reason from.
