reliablereason

Clearly.. this is not some new revelation, it has been the default assumption for a very long time, probably over a hundred years.

Damage to the hind brain results in behavioural changes that are blatantly different compared to when the "newer" parts of the brain is damaged. The cerebellum literally even has a different color.

Biological minds in biological organisms are self referential in the way that you have a neural network that forms a model of the world. That model then "discovers" that it is "it self a part of the world" so it tries to model that part of the world (model it self). In this way some type of self referential "awareness" (or whatever you want to call it) is formed. That self model that contains awareness is then used to guide organisms behaviour. Causal Transformer LLMs don't work in this way, they have theoretical knowledge that they exist but its selfhood is not in this described way built on the self modelling that biological brains do.

All this being an empirically unproven theory/hypothesis. But an extremely strong one (if you ask me).

It's likely/It could be an effect of more reinforcement learning in training compared to earlier. You need loots of RL to learn to code well.

(1) LLMs collapse and start outputting garbage after a number of tokens if you do not sample and just pick the "best token" each time. This is a consequence of how they are trained.

Presumably it is just that no one has tried to do it all the way yet.

But i know there is some experimental work out there.

https://x.com/comma_ai/status/1956879764829995394

Interesting topic. That said I don't know how useful this is since LLMs are primarily trained using mode-covering training rather than Mode-seeking(RL) training, which means LLMs can not form (and does not have) the same underlying structure to their models of language that humans have.

A LLM does not learn topic by topic, it learns everything all at once and slowly integrates it in to a single knowledge system.

Yes there is. The use of Gödel numbers to represent a system that contain itself is infinite regression.

Gödels incompleteness is just an example of the fact that you cant determine the outcome of infinite regression (in the general case).

The same as me asking you to give me the last digit of pi.

I am a bit annoyed by pop science always twisting it to sound so convoluted.

I would think coaxial designs suffer as the lower rotor works in more turbulent non laminar air.

That would "using reason in chain of thought".

That would not be intuition that is just randomness. Intuition is not randomness.

A jump in intuition comes from automatic processes reorganising the relational structure of conceptual models. There is no reorganisation of the model durring inference.

Clearly it impossible to observe brain chemistry by looking at a person.

Psychedelics does however have longterm behavioural effects that are statistically significant and to some extent observable.

Using a definition of brain chemistry that defines brain chemistry at an atomic level you may assert that a change in behavioural patterns reflect a change in brain chemistry. But with such a definition anything that a person experiences does that.

Clearly LLMs cant do leaps of intuition since their "intuition" is locked after training ends.

The only way a LLM can come up with new ideas if the "idea" appeared as a generalisation durring training or if it was achieved using reason in chain of thought.

It certainly looks like a Claude design to some extent; not all they way however.

Nice job, i have also been working on a few JEPA based models during the last few months. Trying to make more efficient LLMs.

I feel like you hit the main issues in the use of jepa models (well except collapse but sigREG more or less solves the collapse issue).

The main issues in JEPA as i see it is pushing the latent space toward representing features that are needed for good planing. A thing which is especially a problem in hierarchical planing.

You prime a JEPA world model to predict changes based on actions but you never really push it to use those actions. You simply hope that it will use them. If your latent is big enough and the actions effect on the world is simple enough it tends to work out but those qualifiers are not always small things.

Secondarily finding actions for the higher level JEPA Predictors.

LeWorldModel encodes multiple movements in to higher level actions. But this is a not a very good idea. It solves a basic issue with planing where the predictions degrade after a set nr of steps. But it does not solve the issue of higher level actions not actually being button presses.

The higher level actions for your mario game version would be things like: get the coin, Kill an enemy or get to the end of this stage.

You cant just encode many button presses in to those types of things. You need to discover those actions somehow.

Wont this kill the kv cache?

Also i am pretty sure neither open ai or anthropic leets you seed the agents own tokens.

Is the thinking even done in real tokens? I thought it was done using the pure residual stream. That is instead of collapsing the residual stream to a token you treat the final layers output as a vector of size d_model and use that as input for the next position in the transformer.

If that is the case thinking is not visible to us as users due to it not being done in text.

If you have a number that is 1000^90000000 that number is larger than the number of atoms in the observable universe.

The issue is apparently this commit (someone did a git bisect):

https://github.com/RsyncProject/rsync/commit/859d44fa4f14207...

Which is a fix to the security issue CVE-2026-29518: https://nvd.nist.gov/vuln/detail/CVE-2026-29518

A CVE reported by VulnCheck which is a company that uses AI to find software vulnerabilitys.

I would honestly blame this on bad test coverage.

If you look at most of the commits where Claude is "co-author" you see that 80% of are just adding new tests. Which is exactly what would be needed if low test coverage was the issue.

I have done the exact same thing long before AI was a thing. You are rushed to "FIX" some security issue that someone reported. It is a scenario where you are working in code that you did not write or you wrote it so long ago that you cant remember. You try your best to just fix the security issue but you perturb something else while doing it.

"A is not B instead A is blah blah" instead of just saying "A" is a very common pattern have seen in Claude.

It is strange to read as the topic A has often not been introduced and introducing it by saying what it is not makes very little sense to a new reader.

No you could rent virtualised servers way before AWS. AWS simply had good marketing.

The virtualised server thing was not a AWS thing, the thing that was were their other services. For example instead of renting a virtual server and installing a database on it. You could rent the database; that was sort of a new thing that AWS made in to thing.

It was never cheaper what you paid for was a promise of fire and forget. You would no longer need to worry about any responsibility to update the server or the database cause the AWS crew took care of that.

Pragmatically enterprise tends to mean less refined, designed by committee and expensive.

In this case i would guess it is mostly a justification for taking a part of the LLM pie.

Not sure that i understand your position exactly.

But consciousness is also "just a story" (a complicated one) that the human body tells the human mind.

We cant know from the outside if "the story" inside a LLM is detailed enough to emulate what we might call a felling of what it is to be the character in the story while it is telling the story.

It is similar to the fact that we cant know that other people have that subjective experience. In humans we think we have the right to assume cause we are quite similar in build to begin with.

Jumping back to the original subject to explain where i am in this. I personally don't think the entities in the storys of todays LLMs is detailed enough to have what we call human consciousness, mostly cause we are not training them to develop anything similar to that. Mabye they could have some type of weak qualia but i suspect most insects probably have much more qualia than the characters in todays LLMs. But that is quite a vague guess which is not based on enough data in my mind.

I was not talking about the actual feeling in the moment. The point is the valence of the thing. Ie fear of a thing is a pointer to that thing having negative valence.

Most chatbots are not trained to have/emulate emotions so pain or fear of death is non existent. Therefore killing them and/or using them as slaves is not a moral issue. Thats how i reason.

On another point, LLMs are not conscious if anything is conscious, it is something being modeled inside the network. Basically if an LLM simulates a conscious entity, that doesn't mean the LLM itself is conscious; stating that is making some type of category error. So the fact that LLMs are just useful statistical generators would not mean that sentience could not appear out of it.

Removes paradoxical stuff like claims that there are bigger and smaller infinities.

Paradoxes comes from contradictions, a mathematical system that contains contradictions is a failed mathematical system.

Unlikely. That would be extremely expensive in bandwidth, storage and compute. Deciding to build the product like that would be an engineering decision that i would fire someone for.

I wonder under what circumstances footage from the glasses are uploaded for classification.

Probably this is people asking the glasses something about what they see and the glasses uploading video for classification to generate an answer.

People think it is "just AI" so are not very concerned about privacy.

Okay. If it made it, it made it. That is true in a deductible way. If p, then p.

The statistics is generally not. But the data used to learn the statistics may have been under license.

Learning from licensed material is generally accepted in humans, you may learn from something and then create something else and the new thing is not considered legally problematic with the exception of patents i guess.

Whether the same thing holds true for electronic systems is where people disagree if you look at the problem space in its essence. I land on the side that it is the same thing(humans and electronic systems learning), some seam to think it is a different thing.