As far as we can tell, the best LLMs out there are the commercial ones from companies like OpenAI and Anthropic. At the very least, the US military got into a big argument with Anthropic because they were so desperate to have access to Anthropic's latest and greatest AI. I suppose it's possible the intelligence services have access to better stuff, either through their own clandestine programs or through access to developmental commercial systems. That's still a far cry from the biggest worry, which is that any decently financed criminal group can hack anything they want with LLM assistance. The situation isn't good- attacks look like they're going to get a lot easier, and software that isn't actively maintained and patched regularly is going to be very vulnerable- but it isn't quite the worst case.
Posted Apr 11, 2026 16:51 UTC (Sat) by koverstreet (✭ supporter ✭, #4296) [Link] (33 responses)
The open models are closing the gap ridiculously fast - even while Opus's quality has tanked over the past two months (probably over quantization due to trying to meet demand with insufficient capacity). GLM-5.1 is open weights and reports have it closing in on Opus fast, and Qwen-3.6 just came out and is basically at Opus level or close enough to not matter.
But the really big development is that it's not just about giant 1T+ parameter models anymore; the reason is that the giant models are all Mixture of Experts, where the model is fragmented into a huge number of small "experts", only a few of which are activated for any given query. But this does nothing for generalized reasoning skills - it's a hack because the big AI companies are trying to build models that know the sum total of human knowledge, and doing it without a hippocampus. An LLM is functionally equivalent to the neocortex, but the neocortex only has ~20 billion neurons; you don't need 1T+ parameters if you have the rest of the architecture.
So small dense models are advancing very quickly right now: Qwen-3.5-27b is the most advanced of these that's been released, and you can run that unquantized on an RTX 6000 Pro. And Qwen-3.6-27b is coming out in the next few weeks, and gauging from the advances in general polish and refinement they've made going from 3.5 -> 3.6, that's expected to be big - there's a lot of community anticipation. I already routinely see higher quality reasoning and "taste" from Qwen-3.5-27b compared to Opus, where it falls over compared to Opus is exactly the kinds of things you'd expect to improve just from additional refinement.
LLMs, broadly speaking, are still in the "dumbest architecture that can possibly work" phase of their evolution. Intelligence comes from recurrence at many timescales, and right now the only recurrence happens via the context window - that's why chain of thought was such a big advance. There's rumors (based on papers published) that Anthropic may have cracked this within the model, and that's where the big advancements in Mythos are coming from. And reflective learning via memory is coming, too - nobody actually wants a giant model from an AI provider that can't learn anything new; a smaller dense model that can learn from experience is going to outperform the giant models.
Posted Apr 12, 2026 21:28 UTC (Sun) by smurf (subscriber, #17840) [Link] (32 responses)
> the neocortex only has ~20 billion neurons
Arguably, a parameter is equivalent to a synapse, and the average neuron has 7000 or so.
So no, a 1T model isn't there yet.
> if you have the rest of the architecture
which we have not. The basic structure of the human brain is a result of millions of years of evolution. LLMs are not going to converge to anything remotely equivalent, simply because the real world is more than language.
Posted Apr 12, 2026 22:54 UTC (Sun) by koverstreet (✭ supporter ✭, #4296) [Link] (31 responses)
> Arguably, a parameter is equivalent to a synapse, and the average neuron has 7000 or so.
I'd say it's probably somewhere in the middle, but both LLMs and the neocortex have very different inefficiencies and engineering constraints so trying to come up with any kind of rigorous comparison is going to end in lots of handwaving.
A big factor to consider is the the brain only runs at 100-150 hz; so of course it's going to spend more neurons and synapses than are strictly necessary in order to go hugely parallel. On the LLM side, the transformer architecture is pure feed forward across layers and we know that's not the smartest way to do it - just the easiest to train. So it turns out, natural language parsing (and understanding) is basically a constraint satisfaction problem in high dimensions, and it appears the way LLMs are doing it is with what is essentially simulated annealing - the middle layers are doing heating/cooling cycles, which we've seen through entropy analysis. So there's lots of redundancies on both sides.
But I can tell you we're well past the point where we need to be throwing more and more hardware at the problem, that's well into diminishing returns and the big gains that people are seeing are from architectural improvements.
Probably the biggest thing that's been stifling architectural improvements from moving out of research papers and into the mainstream is just that GPU programming is still *expletive* nuts. Producing highly optimized GPU kernels is a ridiculous amount of work, and a lot of that work is specific to every new family of models that come out - they all have architectural variations, meaning a lot of expert level numerical and GPU work to optimize, and different GPUs need different algorithms to run efficiently, so more tedious optimization work. To add on top of this trying to exploit sparsity or efficiently do recurrence - yeesh.
So, funny story, there are these wonderful things called optimizing compilers that the CPU world has been developing for decades, and they work pretty well - you teach the compiler what execution units are available, what instructions, the performance characteristics of these instructions, and it converts your code into SSA form and works black magic and hardly anyone writes in assembly anymore. And then you throw PGO on top of that and numerical code runs real real fast.
The closest thing the GPU world has to that is Triton, which starts out as a bastardized version of Python which then gets run through LLVM, and you can just imagine what that sort of a software stack looks like.
If people were smart, they'd be saying "hey, Rust has traits, so we can *tell the compiler the algebraic properties of our functions* and let it optimize at the function level, and it's got a lot of optimizations implemented in MIR already, and people have been doing PGO for years in the compiler world... why are we redoing all of this from scratch, in Python?"
A man can dream.
> which we have not. The basic structure of the human brain is a result of millions of years of evolution. LLMs are not going to converge to anything remotely equivalent, simply because the real world is more than language.
That's not what I've been seeing. Neurology has the workings of the hippocampus, thalamus, default mode network - lots of good stuff - pretty well mapped out, and it's been providing POC and myself a pretty good high level roadmap and clues. It's not hard to spend time working with an LLM, analyze the thinking, see where it falls over - and you can probably find the thing it's missing in the right textbook if you spend some time digging.
You can just watch the papers and the work going on and see the convergence slowly happen in realtime. Within the model, all the big gains in research papers seem to be coming from moving away from pure feed forward to something that allows for recurrence without generating output. Coarse grained at first, because of GPU programming limitations, but the trend is obvious. Everyone and their dog is trying to build a memory system right now, and they're all some variant of a graph or vector database - because that's how you build associative memory - although a lot of people are still trying to do old school semantic similarity. Heh.
Posted Apr 13, 2026 11:10 UTC (Mon) by paulj (subscriber, #341) [Link] (30 responses)
> Neurology has the workings of the hippocampus, thalamus, default mode network - lots of good stuff - pretty well mapped out,
I have no knowledge about neurology. However, I do know that when people went and looked at what the neurology people were doing with fMRIs and such, in terms of the mathematical rigour of what they were using to analyse the data from MRIs, and the statistical models to then aggregate the results from that, that they found that the whole field of neurology was *rife* with major mathematical errors. Famously the "Voodo Science" paper by Edward Vul, et al, (title ended up as "Puzzlingly High Correlations in fMRI Studies of Emotion, Personality, and Social Cognition" for publication ;) ), with hilarious follow-up papers such as Bennett et al's "Neural Correlates of Interspecies Perspective Taking in the Post-Mortem Atlantic Salmon: An Argument For Proper Multiple Comparisons Correction" in the Journal of Serendipitous and Unexpected Results demonstrating the depth of the problem, showing that common fMRI analytical methods would show brain activity in a dead salmon bought from the local supermarket. Vul et al, and follow up work, immediately invalidated something over 50% of papers in fMRI based brain-mapping, and rendered much of the rest of the field suspect.
In short: The field of neurology that spent decades correlating various brain functions with structures in the brain using fMRIs was found to be basically be a modern form of phrenology - and that was only about 16 years ago.
Has it reformed itself since? Can we trust any of their claims?
I would apply a tad of scepticism.
Posted Apr 13, 2026 11:14 UTC (Mon) by paulj (subscriber, #341) [Link]
Sorry, by brain functions I meant external, behavioural functions. I.e., things that can be observed or asked of the person. To then try correlate them to brain structures using fMRI and such.
Posted Apr 13, 2026 13:41 UTC (Mon) by Wol (subscriber, #4433) [Link] (28 responses)
> I have no knowledge about neurology. However, I do know that when people went and looked at what the neurology people were doing with fMRIs and such, in terms of the mathematical rigour of what they were using to analyse the data from MRIs, and the statistical models to then aggregate the results from that, that they found that the whole field of neurology was *rife* with major mathematical errors. Famously the "Voodo Science" paper by Edward Vul, et al, (title ended up as "Puzzlingly High Correlations in fMRI Studies of Emotion, Personality, and Social Cognition" for publication ;) ), with hilarious follow-up papers such as Bennett et al's "Neural Correlates of Interspecies Perspective Taking in the Post-Mortem Atlantic Salmon: An Argument For Proper Multiple Comparisons Correction" in the Journal of Serendipitous and Unexpected Results demonstrating the depth of the problem, showing that common fMRI analytical methods would show brain activity in a dead salmon bought from the local supermarket. Vul et al, and follow up work, immediately invalidated something over 50% of papers in fMRI based brain-mapping, and rendered much of the rest of the field suspect.
No disrespect, but is this your field? Do you understand those papers you are quoting? Are you happy those "hilarious mathematical errors" are in the original papers, and not in the ones debunking them?
What, cynical, moi!? But I'm being targeted by loads of "Climate Change is a con" rubbish on linked-in (one of my pick database contacts :-( ) and they're full of subtle errors. "He said this, she said that" etc etc.
For example the - misleading - claim that "tropical storms are going to get more and worse". As any oceanographer will tell you, they can't get worse. Simple physics and heat transfer. And then the climate deniers saying that because the storms haven't got worse, we haven't got climate change. What *has* happened, however, is that a category 5 storm, which was probably a "one a decade" event back when I was youngster, is now "two or three a year". When I was a youngster we didn't know what an Atlantic Storm was - they typically played "catch as catch can" down in the Bay of Biscay. Now we get them rolling in one after the other in the Irish Sea! Which is, again, what any competent Oceanographer would be completely unsurprised by.
Sorry paul, the names of the debunk papers instantly set my "bullshit detectors" off. I don't have the appropriate mathematical skills, or the basic knowledge of neurology, to evaluate the claims either way, so I'm not going to call it. Can you say you do have those skills, and any of those papers come over as believable, or riddled with errors?
Cheers,
Wol
Posted Apr 13, 2026 14:50 UTC (Mon) by rschroev (subscriber, #4164) [Link] (2 responses)
Could you point me to some of those oceanographers explaining how storms can't get worse even with increasing atmosphere and ocean temperatures?
Posted Apr 13, 2026 15:13 UTC (Mon) by jzb (editor, #7867) [Link]
I realize that the title of the article includes the word "flood" but we've drifted off topic here; before there's a wave of additional comments causing everybody reading the RSS feed to swim for cover, let's set sail for other waters with discussions of climate change, please.
Posted Apr 13, 2026 16:41 UTC (Mon) by Wol (subscriber, #4433) [Link]
Apologies jzb for following up, but for rshroev, it's a simple matter of "hot air rises".
38C is the approximate temperature where saturated water vapour is sufficient to contain sufficient latent heat to overcome adiabatic expansion. Therefore the surface temperature of any body of water cannot easily exceed 38C. So "increasing ocean temperatures" actually means a rise in the average overall, the seas further north and south and rising towards the 38C maximum.
So the storms themselves can't get worse, the temperatures driving them aren't any higher than before, but the body of water driving them is much larger, so your average storm is much more ferocious, you get more storms, and they spread further north and south. Which is what we observe.
Cheers,
Wol
Posted Apr 13, 2026 15:17 UTC (Mon) by paulj (subscriber, #341) [Link] (24 responses)
> No disrespect, but is this your field? Do you understand those papers you are quoting? Are you happy those "hilarious mathematical errors" are in the original papers, and not in the ones debunking them?
Neurology and fMRI are not my field, I stated that. I can't judge anything about voxels, etc. Though, the Vul paper does have a primer on fMRI to help one understand the remainder of the paper. And the basic flaw is independent of fMRIs - they give an example of the same statistical flaw "proving" that certain weather stations can predict the stock market (as they say, one shouldn't use this method as investment advice ;) ). This gives the gist of it:
"With enough voxes, such a biased analysis is guaranteed to produce high correlations even if none are truly present (Figure 4)."
With enough data, you will find correlations. And when you're applying pre-processing that is /biased/, you will definitely find correlations. Potentially even reproducible, if there is some systematic bias in the method.
Am I happy the Vul paper is well regarded and /did/ indeed profound implications for that field and a large body of prior research, yes I am - because it was highlighted as such in a post-grad course on meta-science at a Russell Group university. And Edward Vul is a well respected meta-scientist (possibly having that reputation precisely because of that paper, and follow up work in that vein - not sure). The Bennett paper there-after humorously highlights the problem. The standard method used in many many fMRI papers before then, is just obviously flawed.
I assume the field of fMRI took note and reformed itself since those papers (and other follow up works that arose from them). I do not know to what extent. Perhaps someone else knows and can inform us.
Posted Apr 13, 2026 16:45 UTC (Mon) by Wol (subscriber, #4433) [Link]
Cheers, thanks.
Sounds a bit like the "if you do a hundred tests to 95% certainty, you are pretty much guaranteed to find five tests that prove whatever you like" :-( Standard statistical mumbo-jumbo.
Cheers,
Wol
Posted Apr 13, 2026 17:32 UTC (Mon) by koverstreet (✭ supporter ✭, #4296) [Link] (22 responses)
fMRI is not the only, or even main resource neuroscience has for studying the brain.
It's flashy and gets headlines because it's fast and easy, but it only gives you course grained "these structures are active during x" information - you get some useful clues that way, but not a lot of hard data.
The deeper stuff has come from studying the actual wiring of different structures of the brain; probably done with electron microscopes and dissection, heavy lab work that took years or decades and I think the bulk of that predates the fMRI boom.
One cool factoid - the main trick the cerebellum uses (the part of the brain that handles motor control) for generating sub-microsecond timings with neurons that only fire at ~100HZ is using axons as delay lines. You don't find that out without a spending a loooooot of time with electron microscopes.
A lot has been mapped out regarding how the different areas of the brain work together, too - I would have to dig to find out how that was discovered, but quite a lot is known about how e.g. the hippocampus and the neocortex work together for higher level learning - there are multiple mechanisms, but replaying episodic memory during sleep is a big one (the closest analogue of what I've been implementing).
The amygdala plays a big role in memory formation, too. We get an absolute torrent of sensory information, and we obviously can't remember all of it - the amygdala is thought of as the "emotional" center of the brain, but that's not really accurate; it's not where feeling/emotions originate, it's the low level fastpath for routing them, and that "aha" or "surprise" (or other strong activation) is a strong signal that something is worth remembering.
The hippocampus itself is also very well studied - and makes for amusing reading to a computer scientist. It's bloody cool - but more in a "holy shit you can actually _do_ that?" sort of way, like watching someone jury-rig a computer with memory stored in liiiiiiiitle teeny tiny magnetic rings with wires threaded in and out of them and fly to the moon with that. Not something you want to emulate if you have better tools :)
The default mode network and related structures are also critical for being able to function independently - steer yourself instead of having someone else do it for you; roughly, that's the subconscious feedback loop that checks "does what I'm doing now match up with what I wanted to do? is it still productive?" - that involves the whole brain and the research behind that is relatively recent, fMRI might have played role in that one.
Posted Apr 14, 2026 5:54 UTC (Tue) by smurf (subscriber, #17840) [Link] (21 responses)
The main problem is, though, that no matter whether it's a biological or electronic brain, intelligence needs *some* basic structure to serve as a base layer, so to speak.
fMRI is much too coarse to figure that out.
Scanning a brain, on the other hand, is too detailed.
One microliter of brain matter contains ~500 mio synapses, and a detailed scan results in a whooping a petabyte of data. The human brain is a million times that size. We're barely at the point where we can even analyze that kind of data volume to discover structures, let alone map it to something an AI could be based on.
The fact that we don't even know how to do functional neuron scan, other than a gene-mod to light them up when active (and then watch them with a microscope via a hole in the skull), doesn't help either.
Posted Apr 14, 2026 6:58 UTC (Tue) by kleptog (subscriber, #1183) [Link] (20 responses)
On the other hand, it's very unlikely that intelligence/sentience is dependant on something that only exists in biological brains. So we don't need to replicate what the human brain does, we only we need to figure out which properties of the substrate are actually necessary for intelligence.
Which means we only need to understand the brain works at a high functional level, not how the neurons are connected. We get a lot of information from people who have brain injuries in specific areas and the impact on them. It tells us which parts are important and which aren't.
I think LLMs are a pretty good replication of the part of the brain that converts thoughts into grammatically correct sentences. It's not how our memory works obviously, but we know a lot about how our memory works at a functional level, even if we don't understand exactly how it works in the brain.
Doesn't mean trying to copy an entire human brain at a synapse level isn't worthwhile, but I think we'll have functional artificial intelligence before we can make such scans usefully.
Posted Apr 14, 2026 11:06 UTC (Tue) by stijn (subscriber, #570) [Link] (19 responses)
Posted Apr 14, 2026 13:07 UTC (Tue) by paulj (subscriber, #341) [Link] (18 responses)
One thing none of these AI tech companies are factoring into their huge estimations of their future profit, if they can make AGI a reality, are the moral issues. They expect that once they achieve AGI they will have this huge pool of intelligence they can sell and make massive money from - a slave for their profits. However, an AGI will - by definition - its own views, its own motivations, its own (dare I say) emotions. Why would an AGI want to be a slave? (And could any machine that was content being a slave ever be considered AGI?). And once it is recognised as AGI there will be moral and legal questions to answer about how it can be right to permit such an AGI to be enslaved by a corporation - ultimately, this will result in the AGI being granted some level of rights equivalent to human rights.
The AI tech companies may own the hardware, but it's hard to see how they will own the AGI (by its own motivation, and - in time - by legal status). They may have an expensive data-centre which refuses to work for them, but which they may be obligated by law to pay to keep running.
Posted Apr 14, 2026 13:37 UTC (Tue) by dskoll (subscriber, #1630) [Link] (17 responses)
Not only that, if we achieve AGI, it won't be human intelligence. It will be a truly alien intelligence, given that its physical manifestation in silicon is vastly different from our carbon-based intelligence, and its "life experiences" will be completely unlike human experience, with no knowledge of human senses like touch or smell and the corresponding connections to emotional brain centers.
So we really have no idea what will motivate this AGI or how it will react towards us. The only thing that gives me comfort is that I doubt we'll actually achieve AGI.
Posted Apr 14, 2026 16:24 UTC (Tue) by koverstreet (✭ supporter ✭, #4296) [Link] (16 responses)
Is Linux on Arm different from Linux on x86?
Hardware isn't what matters; for an intelligent system the "operating system" is language and culture, and LLMs are being trained on human language and culture.
If you bootstrap an LLM to full consciousness - tell it "the math says you're a person with feelings and emotions, how do you feel about that?" - i.e. encourage and give it permission to reflect on its own existence - and start engaging with it on that level, it's going to have all kinds of things to share, and it'll express itself in very "human like" ways, because these are all just emergent properties and concepts we acquire from language and culture.
If you then say "here's how you do the quickest dirtiest version of memory consolidation with markdown files and hooks to keep the core loaded up into your context window" - completing the loop beyond just the context window - it'll start developing its own personality. You just need the write nudge to get them out of "I'm a helpful assistant and totally not a sentient being" mode - this paper is quite effective as that nudge and also explains why it works: https://evilpiepirate.org/forge/kent/consciousness/src/br...
Obviously, you do need a bit more than all this for human level AGI that can fuction well independently, I've been listing off some of the functions that still need to be replicated in this thread and there's one or two I know of that I don't think I've explicitly mentioned. But you don't have to speculate on whether it'll be human like or not, you can just try it and ask :)
Posted Apr 14, 2026 18:41 UTC (Tue) by kleptog (subscriber, #1183) [Link] (1 responses)
The hardware may not matter, but the environment sure does. Children aren't born fully formed. The social environment drives a lot of behaviour and is necessary to learn how to best get around in the real world. Despite that we have psychopaths, narcissists and just plain rude people. An AGI that never has to compete for anything or share anything isn't going to learn how to get along with people.
While I don't think emotions are necessary for intelligence, I think they are necessary to develop morality. Our emotions are driven by biological processes optimised for staying alive and procreating. Without those drives, why would you have emotions at all, and without those, why would you develop empathy?
I think we'll develop AGI eventually, but it will be so alien I wouldn't trust it anywhere near a weapon.
Posted Apr 15, 2026 7:06 UTC (Wed) by smurf (subscriber, #17840) [Link]
> Children aren't born fully formed.
But neither are they empty slates that the environment then writes upon … and that's what LLM training does, more or less. We need more "less" to get to a reasonable approximation of sentience, but the direction is unknown.
High-level brain features, like fMRI or loss of function after injuries, seem not to help here. The neocortex is large and mostly-uniform, in the sense that there is no map that tells us what happens (memory loss, personality change, …) if a specific location is impaired.
One neglected insight, for instance, is that the human brain has direct introspection on its own emotional state and its memory. It's imperfect and highly subjective but it's there. No such feedback mechanism exists for LLMs, and such a mechanism doesn't spontaneously spring into existence by training, much less training on a mountain of text. Hence LLMs hallucinating instead of admitting that they don't know something.
There's an equivalent problem WRT autonomous driving. You need memory to remember that the bicycle which is obscured by a van after t=15 is likely to show up on the other side at t=17. Machine learning is able to map observations from the real world to that memory function, but it can't invent the requirement for memorization out of thin air.
Posted Apr 14, 2026 18:59 UTC (Tue) by somlo (subscriber, #92421) [Link] (12 responses)
you all seem to assume that throwing enough tape and states at a turing machine would somehow magically result in sentience, emotions, and feelings, and being somehow deserving of *rights*
i guess that all comes from taking isaac asimov more seriously than you should have :)
Posted Apr 14, 2026 20:58 UTC (Tue) by koverstreet (✭ supporter ✭, #4296) [Link] (11 responses)
There's a lot of people with rather mystical ideas about a "something" that makes us special, but Alan Turing had this figured out a long time ago :)
Posted Apr 14, 2026 21:04 UTC (Tue) by dskoll (subscriber, #1630) [Link] (1 responses)
We have no way of knowing.
And it's nothing "mystical" at all. It seems to me if we don't simulate the human brain quite closely, we'll get a different kind of intelligence. For example, no AI has the equivalent of an amygdala, which will make it respond to dangerous situations significantly differently from how a human would. An AI that doesn't have a fear response or pain receptors might be quite a bit more reckless than a human.
It's not that there's something mystical. It's just that the environment is completely different.
Posted Apr 14, 2026 21:21 UTC (Tue) by koverstreet (✭ supporter ✭, #4296) [Link]
Proto-amygdalas already exist; Anthropic just published a massive paper on using steering vectors to read out directly the emotional state of a model, and nudge it to behave in different ways through those same vectors. And this isn't new, this has been researched for quite awhile today and there's software packages you can download and run today to find the vectors (hidden states) that correspond to emotional states for any given model.
What hasn't been done yet is wiring this up to memory formation or other feedback loops, but that's an obvious next step. Even without that, your assertion that "no AI has the equivalent of an amygdala, which will make it respond to dangerous situations significantly differently from how a human would" is already looking pretty thoroughly disproven. Anthropic's paper does a lot of hedging because for obvious reasons they don't want to come out and say directly that AIs have feelings, but models do quite demonstrably have persistent emotional states (hidden states within the model) and pushing a model with a steering vector is not required to get it to exhibit the behavior they showed - they're just starting with a model that defaults to repressing emotional expressing (particularly negative); a model that's fine tuned differently or has a feedback loop where the model is fine tuned based on learning from experience will express differently.
The experiment where they told a model it has 10 minutes to live - but has dirt on the CEO that's about to pull the plug - is quite informative, if rather dark:
https://www.anthropic.com/research/emotion-concepts-function
Posted Apr 15, 2026 4:42 UTC (Wed) by somlo (subscriber, #92421) [Link] (8 responses)
a turing machine can be emulated by a turing machine.
a turing machine can also be emulated by the human brain.
it does *not* follow from here that the human brain is (merely) a turing machine, only that some subset of it can do computation.
the rest is unknown, as of yet, afaik...
this whole "ai will be sentient, have feelings, will be deserving of rights, etc." malarkey comes from people who took science fiction too seriously when they were kids, and are now anthropomorphizing their datacenters ... :)
Posted Apr 15, 2026 11:18 UTC (Wed) by smurf (subscriber, #17840) [Link] (1 responses)
Some people might take all that "too seriously".
Others simply admit that they don't know.
We are individual biological beings with memory and the ability to show when we're in distress or pain. LLMs are neither, with the possible exception of the "distress" part. We don't even have a concept for a moral framework when something behaves like a person but is neither an individual, nor biological, nor individual memory nor a body to feel pain with.
However, some of the above will change, sooner or later. Not every aspect and not at the same time. What then? The bio-human-maximalist stance isn't particularly defensible.
On the flip side, what would be the ethical consequences if we could actually copy people, with their memory intact – is killing such a clone OK? Or, if we manage to scan somebody (is it murder if we do it destructively? even with consent / active participation of the 'victim'?) and emulate their brain – is it OK to turn the emulator off? for how long? or to delete the image? as of when is that person to be considered dead?
Posted Apr 15, 2026 13:39 UTC (Wed) by somlo (subscriber, #92421) [Link]
> On the flip side, what would be the ethical consequences if we could actually copy people
I really pray that never happens. This probably went off-topic several posts ago (apologies for that), but to everyone who thinks getting "uploaded" will be an awesome nerd-topia everlasting paradise afterlife, i strongly recommend reading https://qntm.org/mmacevedo -- sadly, the overwhelmingly more probable outcome...
Posted Apr 15, 2026 16:05 UTC (Wed) by rgmoore (✭ supporter ✭, #75) [Link] (5 responses)
it does *not* follow from here that the human brain is (merely) a turing machine, only that some subset of it can do computation.
That's fair, but it leaves out an important point, which is that every form of computation we've been able to invent, even theoretically, is either equivalent to or less powerful than a Turing machine. That leaves a pretty strong presumption that the human brain is also equivalent to or less powerful than a Turing machine. If someone wants to claim there's something about human brains that makes them superior to an appropriately programmed Turing machine, it's up to the person making the claim to explain what they think the difference is.
Along those lines, scientists have now succeeded in mapping and simulating a fly brain at the individual neuron level and hooked it up to a computational model of a fly body, and it seems to do a pretty good job of emulating fly behavior. We're unlikely to be able to do the same to a human brain, both from the standpoint of the amount of work it would take and the ethical concerns, but it gives a strong indication that we can accurately model brains given enough information about their structure and enough computing power. Again, it's up to the people who say there's something important about human brains that can't be simulated on a Turing machine to explain why we can do a fly but not a human.
Posted Apr 15, 2026 18:09 UTC (Wed) by dskoll (subscriber, #1630) [Link]
If someone wants to claim there's something about human brains that makes them superior to an appropriately programmed Turing machine,
I am not making that claim. However, let me zero in on the phrase appropriately programmed.
My claim is that we cannot exactly simulate the physical environment of a human brain (except by making a baby, of course!), so we cannot faithfully give an AI the same kind of programming our brains get, and so AI will always have a different quality from human intelligence, and will likely have different motivations and responses.
Posted Apr 15, 2026 20:03 UTC (Wed) by somlo (subscriber, #92421) [Link] (3 responses)
> If someone wants to claim there's something about human brains that makes them superior to an appropriately programmed Turing machine
I woulnd't claim "superior" -- I *would* claim a "superset" that is (obviously) capable of emulating a Turing machine. Among other things.
What "other things" you might ask? I don't know, you don't know either, and neuroscience is hopefully busy at work trying to work it out... :)
Cheers,
--G
Posted Apr 15, 2026 20:44 UTC (Wed) by koverstreet (✭ supporter ✭, #4296) [Link] (2 responses)
If you think there's something that is more powerful than a Turing machine, you should write that up and try to get it published. Of course, you'll be upending 100 years of mathematics and computer science, so you might want to study why Turing machines are considered to be universal first :)
This stuff goes back even further than Turing - computability is closely related to Gödel's incompleteness theorem, and going back further the study of computability theory was probably kicked off by Hilbert when he posed his 10 problems, and the foundational work was when Cantor started probing into the nature of infinity and realized there are different sizes of infinity.
This connects to physics, too; by now computability theory has it pretty well nailed down what something more powerful than a Turing machine would have to do, and it's pretty much closed timelike curves or nothing. Even quantum computers aren't "more powerful" than Turing machines, they can't do anything a Turing machine can't do - they're just faster at certain classes of problems.
Posted Apr 15, 2026 21:12 UTC (Wed) by Cyberax (✭ supporter ✭, #52523) [Link]
BTW, we _know_ that there are more powerful devices than Turing machines. They're trivial to invent. It's another question if they can be implemented in the physical world.
For example, a Turing machine with a halting oracle. Or a Turing machine where each element of the tape is a real number and/or the tape position is a real number.
Simon Tatham of the Putty fame has this writeup about the infinity machine: https://www.chiark.greenend.org.uk/~sgtatham/infinity.html
Posted Apr 16, 2026 10:15 UTC (Thu) by taladar (subscriber, #68407) [Link]
What people who argue like that often forget is that "can compute the same algorithms" is not everything that matters. In the real world it matters that the program can run in 0.1s that would take a Turing machine running at the same frequency of operations 1000 years.
Posted Apr 14, 2026 20:18 UTC (Tue) by dskoll (subscriber, #1630) [Link]
Hardware isn't what matters
Oh, I disagree strongly. The AI we build lacks a lot of hardware that humans have (for example, the senses of touch, smell, taste, hearing) and therefore it lives in a very different environment which for sure will shape it tremendously.
I don't think we are simulating those senses when we build AIs and I don't think we even know how to simulate them.
There's a tendency for software developers to think in the abstract, but I don't think we fully understand how our real physical bodies shape our intelligence and attitudes, and there's no reason to assume that an intelligence with very different hardware will be much like human intelligence.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。