Hacker News Reader

Comment by CamperBob2 | Hacker News Reader

You want to talk evidence, let's talk evidence. What does the brain do that these models don't (or, more generally, can't)?

Be specific.

And no, it doesn't "happen with every new technology." Nothing even remotely like this has been seen before the present decade.

View on HN

I'm no neuroscientist, but I hung out with a few, so here are some bits I picked up from them. In biological brians:

- neurons have signal surpression (-ive activation) as well as propagation (+activation)

- neurons appear to process and propagation due to unknown internal mechanisms.

- there are complex sub structures within biological neural nets where the architecture is radically different from in other sections of the network, strongly in contrast to the homogenous structure of ANN's

- many different types of neurons with different properties and behaviors in terms of network formation and network activity are present in BNN's in contrast to ANN's

- BNN's learn during processing. ANN's are static after training.

Those are mostly structure-versus-function arguments, centered on what a brain is as opposed to what it does. Only the static nature of the current ML models seems like a valid point... and it's changing too.

I'd wager that future models (which admittedly may not look much like today's LLMs) will blur the lines between mutable context and immutable weights. When that happens, all bets are truly off.

> future models (which admittedly may not look much like today's LLMs) will blur the lines between mutable context and immutable weights.

"may not look much like today's LLMs" is really sweeping the whole point under the rug. The difference between what we are doing now with static LLM models and dynamic models will require algorithmic changes that have not yet been invented. That's in addition to the fact that the processing of a single neuron is completely parallel with respect to inputs. GPUs make them more parallel, but the analog devices we are trying to mimic are vastly more complicated than what we are using as a substitute.

Whatever. This ( https://eqbench.com/results/creative-writing-v2/deepseek-ai_... ) is intelligence in action, from an entity with an internal world model that's at least as valid as mine.

Either that, or it's witchcraft.

(Either that, or somebody with both time and talent to waste is taking the piss and pretending to be an AI.)

If anyone ever does succeed in updating weights dynamically on a continuous basis, things are going to get really interesting really quickly. Even now, arguments based on relative complexity are completely invalid. A few fast neurons are at least as good at 'thinking' as a lot of slow ones are.

Hah, "Whatever." Seriously?

The writing is still derivative meandering pulp. No need to invoke magic. It's cool that it can make complete paragraphs though. LLMs are a huge breakthrough. They're just not intelligent.

Updating weights on a continual basis requires processing weights productively on a regular basis. That's many flops per weight away from where we are in processing and bandwidth. The computation limitations still apply.

Metacommentary: I stopped arguing because this thread is reeking of AI bro non-sense, they can't argue with facts so they continue with vague statements and deflection (e.g like the "Whatever" you've got above). The GenAI bubble can't burst soon enough.

As I said, neuroscience be damned, we've got transformers now. (/s)

"My king, it's amazing these things can contain and regurgitate knowledge on demand. The next step is an army of Gollems that will destroy your enemies, all I need is three bags of gold." - Sargon the scribe, near Baghdad first month, 2024 BC

"My king, the Hittites have created scrolls that are wide and long despite that we burned their caravan taking palm leaves to the north. They have found a way to flense lambs to make pages themselves, and their ink is pure silver. They will soon have an army of Gollems and will destroy your empire. To stop this all I need is ten bags of gold." - Sargon the scribe, near Baghdad second month, 2024 BC

"My king, alas all the gold has gone, and yet neither will the gollems walk, nor have the Hittities come. My dancing girls grow cold, and I have no figs for my table all I need is 20 bags of gold" - Sargon the scribe, near Baghdad third month, 2024 BC

What happened to Sargon next is not recorded...

There’s agents now. It’s not just paragraphs. And LLMs can beat you on intellectual tasks. I don’t see how you can throw that word around as if you know your more intelligent when it clearly can be superior to you on many many tasks.

You can't win this argument through logic less mere rhetoric. You need scientific evidence. Hint: There isn't.

Edit: Ok, let me be less harsh and play with your argument. Can an LLM have an acid/psilocybin/ketamine trip, distort its view of self and reality, then rewire some of its internal connections and come a little different based on the experience? I guess not, there are more examples and they all show that as far as we know LLMs are not minds/brains and viceversa even if they seem similar in a chat interface. (If you don't empathise with that example remove the drugs and change them for a near-death experience).

I strongly argue humans are not (just) LLMs, but I think we're far from getting evidence*. I think we will get to AGI before we know how the brain works and there is no dicotomy in that, if anything the tech is showing us that we can get at least some intelligence in other substrates.

*PS: fwiw ausence of evidence does not support any sides. I shouldn't remark that, but here we are.

> You need scientific evidence

This, while true, is putting the cart before the horse.

One needs to define what the question is before it is possible to seek evidence.

I see many things different between Transformer models and brains, but I do not know which, if any, matters.

> Can an LLM have an acid/psilocybin/ketamine trip, distort its view of self and reality, then rewire some of its internal connections and come a little different based on the experience?

Specifically those chemicals? No, not even during training when the weights aren't frozen, as nobody bothered to simulate the relevant receptors*.

Can LLMs experience anything, in the way we mean it? Nobody even knows. We don't know what the question means well enough to more than merely guess what to look for.

Do the inputs to an LLM, a broader idea of "experience" without needing to solve the question of which definition of consciousness everyone should be using, rewire some of its internal connections? All the time.

* caveat: it doesn't, at first glance, seem totally impossible that the structure of these models is sufficiently complicated for them to create the simulation of those things in the weights themselves in order to better model the behaviour of humans generating text from experiencing those things. But I also don't have any positive reason to expect this any more than I would expect an actor portraying a near-death-experience to have any idea what that's like on the inside.

The question/argument was well defined by parent-grandparent posts: people reason as LLMs, including a suggestion that people discussing here are not being self-conscious of said similarity (i.e. people parroting/regurgitating). You can check it above.

I won't continue with these discussions as I feel people are being just obtuse about this and it is becoming more emotional than rational.

> The question/argument was well defined by parent-grandparent posts: people reason as LLMs, including a suggestion that people discussing here are not being self-conscious of said similarity (i.e. people parroting/regurgitating). You can check it above.

With a lack of precision necessary when the phrase "You need scientific evidence" is invoked.

It's very easy to trip over "common sense" reasoning and definitions — even Newton did so with calculus, and Euclid did so with geometry.

> I won't continue with these discussions as I feel people are being just obtuse about this and it is becoming more emotional than rational.

As you wish. I am relatively unemotional on this.

The brain can invent formal language that enables unambiguously specifying an algorithm that when provided with a huge amount of input data can simulate the output of the brain itself.

Can these models do that?

> The brain can invent formal language that enables unambiguously specifying an algorithm that when provided with a huge amount of input data can simulate the output of the brain itself.

Can *a* brain do this?

Every effort I've seen has (1) involved *many people* working on different parts of the problem, and (2) so far we've only got partial solutions.

Even if you allow collaboration, then when (2) goes away, your question is trivially answered "yes" because we just feed into the model the same experiences that were fed into the human(s) who invented the model, and the simulation then simulates them inventing the same model.

> Even if you allow collaboration, then when (2) goes away, your question is trivially answered "yes" because we just feed into the model the same experiences that were fed into the human(s) who invented the model, and the simulation then simulates them inventing the same model.

These models doesn't work that way. Its probably possible to make such a model, but currently we don't know how to make them.

I literally said so, yes.

Do you have empirical proof any LLM can generate another program that can be trained to simulate it? Otherwise it is science fiction.

Misread me, but as it happens, I might have seen exactly that.

I didn't ask for it, one of the Phi models got confused part way though and started giving me a python script for machine learning instead of what I actually asked for.

I'm saying I have no evidence that humans have ever demonstrated this capability either, and if we were to invent such an algorithm and model (even more generally than LLM) the AI that is this would by definition necessarily have to pass this test.

Have experiences and feelings. Have personal knowledge of something.

Be able to express language with a training/knowledge/etc corpus of far smaller size.

A lot of that is provided by multi-modality including feedback from the body and interacting with objects and people with more than just the words. That expands the context of a human's experiences dramatically compared to just reading books.

Plus even when humans are reading books a mental image of what's going on in the story is common. Not everyone has that, but it shows how much a basic LLM lacks and multi-modal would add.

Now the real question to my mind is whether we can train models with actual empathy to learn from the experiences of other people without having to go through the experience directly. Doing so would put them above many individuals' understanding already. . .

While I agree with:

> Be able to express language with a training/knowledge/etc corpus of far smaller size.

Nobody knows what it would mean to test for:

> Have experiences and feelings.

And I am unclear if:

> Have personal knowledge of something.

Would or would not be met by putting them in charge of a robot of some kind to collect experiences?

I'd certainly agree that the default is to make them "book smart" rather than "street smart", but at most that's an "is not" rather than "cannot".

Reply on news.ycombinator.com