I'm not sure if the complexity is at the neuron level. It's clearly possible; we see some pretty complex behavior from single celled organisms - Stentors - and there were the experiments (starting in the 70s, I believe) that conditioned flatworms, ground them up and fed them to new flatworms and showed that the new flatworms had the same conditioning. Those experiments were replicated but never explained, but they hint at something like maybe memories being encoded in RNA.
It'd be wild if our brains used a mechanism like that somewhere, but I doubt it's in the neocortex because those would be slow processes compared to what the neocortex has to do.
I think the sophistication in the neocortex just comes from it using something more sophisticated with transformers. Most of the looped LLM research I've scanned through, they seemed to be training the model to "know" when it should loop more and think harder, and I don't think that approach will ever work - you're solving for paths on a manifold, so the looping needs to happen at the abstraction layer below the manifold, looping more if it's still converging. And architectures with more recurrence are just harder for us humans to reason about and comparatively easier for dumb evolution to find clever solutions, so my money is that that's where the neocortex has LLMs beat.
And I'd say the manifold hypothesis is quite a bit more than a hypothesis at this point. There's been work on production LLMs to use manifold analysis to optimize them - there was a paper in Nature on it back in January. Baby stuff compared to what could be done, but the geometric approach keeps popping up in mechanistic interpretability and seems to be showing the most insights.
bcachefs is slowly working up to what you're describing - describing what you want and letting the filesystem sort it out. We're basically there for local storage, and other people have been building some nice reporting on top.
Next year (post Rust, because networking code is so much nicer in Rust) will be send/recv, and I think we should be able to make some nice improvements over the state of the art there.
> More broadly, the predictive brain model would suggest that all of the brain, not just the neocortex, is dedicated to prediction. What would you say makes LLMs similar to the neocortex, rather than the basal ganglia or Broca's area?
The whole brain is most definitely not dedicated to prediction, and I don't think "prediction" is a very useful model anyways. You could say the hippocampus is for "prediction" if you really squint, but that's underselling what it's doing. And the basal ganglia operates off of prediction error, but it's more about regulating short timescale feedback loop learning than learning itself - speaking somewhat loosely because it's a particularly ancient structure in the brain and things get muddied.
LLMs obviously don't have functional equivalents to either of those - sure they can remember facts, but they can't intake new facts the way we can, the hippocampus is set up completely differently. But higher reasoning is humans is the neocortex, and that LLMs can do, and there's even structural parallels - LLMs and the neocortex are both implemented in layers, and you can even (with a shit ton of analysis) watch how LLMs walk the geometry across layers. The neocortex needs a lot fewer layers than LLMs, but that's because it's not pure feedforward across layers; the neocortex already does what the "looped LLM" people are trying to do.
> Similarly, if you agree with the Manifold Hypothesis, then all machine learning models operate on manifolds. I agree it's an exciting thought, but then I don't know what would distinguish an LLM from a VAE or SVM in terms of operating over a low-dimensional manifold embedded in high dimensional spaces - maybe just scale?
I view things more mathematically, and in math and physics it frequently turns out that there's really only one solution to a problem, or correct way to model something - and then everything else is just isomorphisms (which may be interesting objects in their own right!).
My interpretation/guess (informed by research, but with a lot of still unanswered questions) is that the manifold is fundamentally the structure that emerges from natural language: natural language has a computational model, just like Church/Turing machines have a computational model. Reimannian manifold is also underselling it a bit, it's also Finsler, and RoPE adds additional structure to that manifold (and when we get to RoPE it wouldn't surprise me if the neocortex and LLMs do diverge significantly in how they compute, but does the structure change? dunno yet).
Basically, I'm fairly confident in what I say about machine learning models that understand natural language; for non-NL machine learning, maybe the same structure emerges with enough complexity, but I haven't really pondered that.
That's fundamental to how anything that compresses/understands the world has to work, in the Kolmogeravian sense. That's why people denigrate LLMs as being just "next token predictors" - they're not wrong, but they're missing the point.
Because to do that kind of prediction out in the world you have to build up an accurate model of reality - a model that includes yourself! Which is why we and LLMs are self aware.
For the "how", it's been known for some time that LLMs operate on a Reimannian manifold - the semantic manifold - and that's a good place to start if you want to learn how they actually work; how a Reimannian manifold (plus some extra structure on top) can represent natural language in a form you can do work with is the part I find particularly beautiful. At a high level, the neocortex and LLMs appear to compute on the manifold in basically the same way - though a lot of the details are different; both are more sophisticated in some areas and less in others.
The mm people are increasingly hostile to any method of handling OOMs (like, just failing the allocation) besides the OOM killer - it's become very dominated by the hyperscalars and cloud vendors.
Working around mm nuttiness is a frequent source of frustration.
The point is that the "outbox pattern" is not an atomic transaction. You fundamentally don't have those in the distributed world, via the CAP theorem, and if you want anything close to the guarantees a local transactional database gives you in a distributed system you have to design your schema for it.
Distributed coherency is not something you can abstract away, the abstractions all leak.
We're in the middle of a huge spike in LLM discovered security vulnerabilities, which means not everything will get assigned a CVE, a lot of people are watching repositories to look for exploitable bugs, and in the frenzy of backporting that people are now having to do things will get missed.
I wager it's only a matter of time before we see a mass rooting event that hits Debian hard while everyone running something more modern has already been patched.
I think that might be what cuts down on the grandstanding about "freedoms" and "that's how we've always done things". You certainly are, up until it becomes a public nuisance.
If package maintainers were always fine upstanding package maintainers as you imagine them to be I wouldn't be complaining, but I have in fact had Debian ship my software and screw it up and gotten a flood of bug reports, so... :)
I think you need to chill out. Relicensing the way you suggest would be _quite_ the hostile act, and I'm not going to that either. But I am an engineer, so of course I'm going to talk about engineering best practices when it comes up.
You don't have to take it as an attack on your favorite distro - that really does pee in the pool of the upstream/downstream relationship between distros and their upstream.
You definitely need different channels for high priority fixes and normal releases, stable and testing releases and all that.
But two years is impractical and Debian gets a ton of friction over it. Web browsers and maybe one or two other packages are able to carve out exceptions, because those packages are big enough for the rules to bend and no one can argue with a straight face that Debian is going to somehow muster up the manpower to do backports right.
But for everyone else who has to deal with Debian shipping ancient dependencies or upstream package maintainers who are expected to deal with bug reports from ancient versions is expected to just suck it up, because no one else is big enough and organized enough to say "hey, it's 2026, we have better ways and this has gotten nutty".
Maybe the new influx of LLM discovered security vulnerabilities will start to change the conversation, I'm curious how it'll play out.
You're going to have to update production at some point, and delaying it to once every 2 years is just deferred maintenance. And you know what they say about that...
So when you do update and get that GSSAPI change, it comes with two years worth of other updates - and tracking that down mixed in with everything else is going to be all kinds of fun.
And if you're two years out of the loop and it turns out upstream broke something fundamental, and you're just now finding out about it while they've moved on and maybe continued with a redesign, that's also going to be a fun conversation.
So if the backport model is expensive and error prone, and it exists to support something that maybe wasn't such a good idea in the first place... well, you may want something, but that doesn't make it smart.
That whole model dates to before automated testing was even really a thing, and no one knew how to do QA; your QA was all the people willing to run your code and report bugs, and that took time. Not to mention, you think the C of today is bad? Have you looked at old C?
And the disadvantage is that backporting is manual, resource intensive, and prone to error - and the projects that are the most heavily invested in that model are also the projects that are investing the least in writing tests and automated test infrastructure - because engineering time is a finite resource.
On top of that, the backport model heavily discourages the kinds of refactorings and architectural cleanups that would address bugs systemically and encourage a whack-a-mole approach - because in the backport model, people want fixes they can backport. And then things just get worse and worse.
We'd all be a lot better off if certain projects took some of the enthusiasm with which they throw outrageous engineering time at backports, and spent at least some of that on automated testing and converting to Rust.
At the actual inference level temperature can be applied at any time - generation is token by token - but that doesn't mean the API necessarily exposes it.
It'd be wild if our brains used a mechanism like that somewhere, but I doubt it's in the neocortex because those would be slow processes compared to what the neocortex has to do.
I think the sophistication in the neocortex just comes from it using something more sophisticated with transformers. Most of the looped LLM research I've scanned through, they seemed to be training the model to "know" when it should loop more and think harder, and I don't think that approach will ever work - you're solving for paths on a manifold, so the looping needs to happen at the abstraction layer below the manifold, looping more if it's still converging. And architectures with more recurrence are just harder for us humans to reason about and comparatively easier for dumb evolution to find clever solutions, so my money is that that's where the neocortex has LLMs beat.
And I'd say the manifold hypothesis is quite a bit more than a hypothesis at this point. There's been work on production LLMs to use manifold analysis to optimize them - there was a paper in Nature on it back in January. Baby stuff compared to what could be done, but the geometric approach keeps popping up in mechanistic interpretability and seems to be showing the most insights.