Ask HN: Any insider takes on Yann LeCun's push against current architectures?

385 points by vessenes ↗ HN
So, Lecun has been quite public saying that he believes LLMs will never fix hallucinations because, essentially, the token choice method at each step leads to runaway errors -- these can't be damped mathematically.

In exchange, he offers the idea that we should have something that is an 'energy minimization' architecture; as I understand it, this would have a concept of the 'energy' of an entire response, and training would try and minimize that.

Which is to say, I don't fully understand this. That said, I'm curious to hear what ML researchers think about Lecun's take, and if there's any engineering done around it. I can't find much after the release of ijepa from his group.

342 comments

[ 2.7 ms ] story [ 311 ms ] thread
Not an official ML researcher, but I do happen to understand this stuff.

The problem with LLMs is that the output is inherently stochastic - i.e there isn't a "I don't have enough information" option. This is due to the fact that LLMs are basically just giant look up maps with interpolation.

Energy minimization is more of an abstract approach to where you can use architectures that don't rely on things like differentiability. True AI won't be solely feedforward architectures like current LLMs. To give an answer, they will basically determine alogrithm on the fly that includes computation and search. To learn that algorithm (or algorithm parameters), at training time, you need something that doesn't rely on continuous values, but still converges to the right answer. So instead you assign a fitness score, like memory use or compute cycles, and differentiate based on that. This is basically how search works with genetic algorithms or PSO.

> The problem with LLMs is that the output is inherently stochastic - i.e there isn't a "I don't have enough information" option. This is due to the fact that LLMs are basically just giant look up maps with interpolation.

I don't think this explanation is correct. The input to the decoder at the end of all the attention heads etc (as I understand it) is a probability distribution over tokens. So the model as a whole does have an ability to score low confidence in something by assigning it a low probability.

The problem is that thing is a token (part of a word). So the LLM can say "I don't have enough information" to decide on the next part of a word but has no ability to say "I don't know what on earth I'm talking about" (in general - not associated with a particular token).

The problem is exactly that: the probability distribution. The network has no way to say: 0% everyone, this is non sense, backtrack everything.

Other architectures, like energy based models or bayesian ones can assess uncertainty. Transformers simply cannot do it (yet). Yes, there are ways to do it, but we are already spending millions to get coherent phrases, few ones will burn billions to train a model that can do that kind of assessments.

Has anybody ever messed with adding a "backspace" token?
Yes. (https://news.ycombinator.com/item?id=36425375, believe there's been more)

There's a quite intense backlog of new stuff that hasn't made it to prod. (I would have told you in 2023 that we would have ex. switched to Mamba-like architectures in at least one leading model)

Broadly, it's probably unhelpful that:

- absolutely no one wants the PR of releasing a model that isn't competitive with the latest peers

- absolutely everyone wants to release an incremental improvement, yesterday

- Entities with no PR constraint, and no revenue repurcussions when reallocating funds from surely-productive to experimental, don't show a significant improvement in results for the new things they try (I'm thinking of ex. Allen Institute)

Another odd property I can't quite wrap my head around is the battlefield is littered with corpses that eval okay-ish, and should have OOM increases in some areas (I'm thinking of RWKV, and how it should be faster at inference), and they're not really in the conversation either.

Makes me think either A) I'm getting old and don't really understand ML from a technical perspective anyway or B) hey, I 've been maintaining a llama.cpp wrapper that works on every platform for a year now, I should trust my instincts: the real story is UX is king and none of these things actually improve the experience of a user even if benchmarks are ~=.

For sure read Stephenson’s essay on path dependence; it lays out a lot of these economic and social dynamics. TLDR - we will need a major improvement to see something novel pick up steam most likely.
Yeah everyone spending way to much money in things we barely understand is a recipe for insane path dependence.
Oh yeah, that's exactly what I was thinking of! Seems like it would be very useful for expert models with domains with more definite "edges" (if I'm understanding it right)

As for the fragmentation of progress, I guess that's just par the course for any tech with a such a heavy private/open source split. It would take a huge amount of work to trawl through this constant stream of 'breakthroughs' and put them all together.

Right. And, as a result, low token-level confidence can end up indicating "there are other ways this could have been worded" or "there are other topics which could have been mentioned here" just as often as it does "this output is factually incorrect". Possibly even more often, in fact.
My first reaction is that a model can’t, but a sampling architecture probably could. I’m trying to understand if what we have as a whole architecture for most inference now is responsive to the critique or not.
You get scores for the outputs of the last layer; so in theory, you could notice when those scores form a particularly flat distribution, and fault.

What you can't currently get, from a (linear) Transformer, is a way to induce a similar observable "fault" in any of the hidden layers. Each hidden layer only speaks the "language" of the next layer after it, so there's no clear way to program an inference-framework-level observer side-channel that can examine the output vector of each layer and say "yup, it has no confidence in any of what it's doing at this point; everything done by layers feeding from this one will just be pareidolia — promoting meaningless deviations from the random-noise output of this layer into increasing significance."

You could in theory build a model as a Transformer-like model in a sort of pine-cone shape, where each layer feeds its output both to the next layer (where the final layer's output is measured and backpropped during training) and to an "introspection layer" that emits a single confidence score (a 1-vector). You start with a pre-trained linear Transformer base model, with fresh random-weighted introspection layers attached. Then you do supervised training of (prompt, response, confidence) triples, where on each training step, the minimum confidence score of all introspection layers becomes the controlled variable tested against the training data. (So you aren't trying to enforce that any particular layer notice when it's not confident, thus coercing the model to "do that check" at that layer; you just enforce that a "vote of no confidence" comes either from somewhere within the model, or nowhere within the model, at each pass.)

This seems like a hack designed just to compensate for this one inadequacy, though; it doesn't seem like it would generalize to helping with anything else. Some other architecture might be able to provide a fully-general solution to enforcing these kinds of global constraints.

(Also, it's not clear at all, for such training, "when" during the generation of a response sequence you should expect to see the vote-of-no-confidence crop up — and whether it would be tenable to force the model to "notice" its non-confidence earlier in a response-sequence-generating loop rather than later. I would guess that a model trained in this way would either explicitly evaluate its own confidence with some self-talk before proceeding [if its base model were trained as a thinking model]; or it would encode hidden thinking state to itself in the form of word-choices et al, gradually resolving its confidence as it goes. In neither case do you really want to "rush" that deliberation process; it'd probably just corrupt it.)

I think some “reasoning” models do backtracking by inserting “But wait” at the start of a new paragraph? There’s more to it, but that seems like a pretty good trick.
I feel like we're stacking naive misinterpretations of how LLMs function on top of one another here. Grasping gradient descent and autoregressive generation can give you a false sense of confidence. It is like knowing how transistors make up logic gates and believing you know more than CPU design than you actually do.

Rather than inferring from how you imagine the architecture working, you can look at examples and counterexamples to see what capabilities they have.

One misconception is that predicting the next word means there is no internal idea on the word after next. The simple disproof of this is that models put 'an' instead of 'a' ahead of words beginning with vowels. It would be quite easy to detect (and exploit) behaviour that decided to use a vowel word just because it somewhat arbitrarily used an 'an'.

Models predict the next word, but they don't just predict the next word. They generate a great deal of internal information in service of that goal. Placing limits on their abilities by assuming the output they express is the sum total of what they have done is a mistake. The output probability is not what it thinks, it is a reduction of what it thinks.

One of Andrej Karpathy's recent videos talked about how researchers showed that models do have an internal sense of not knowing the answer, but fine tuning on question answering I'd not give them the ability to express that knowledge. Finding information the model did and didn't know then fine tuning to say I don't know for cases where it had no information allowed the model to generalise and express "I don't know"

No an ML researcher or anything (I'm basically only a few Karpathy video into ML, so please someone correct me if I'm misunderstanding this), but it seems that you're getting this backwards:

> One misconception is that predicting the next word means there is no internal idea on the word after next. The simple disproof of this is that models put 'an' instead of 'a' ahead of words beginning with vowels.

My understanding is that there's simply not “'an' ahead of a word that starts with a vowel”, the model (or more accurately, the sampler) picks “an” and then the model will never predict a word that starts with a consonant after that. It's not like it “knows” in advance that it wants to put a word with a vowel and then anticipates that it needs to put “an”, it generates a probability for both tokens “a” and “an”, picks one, and then when it generates the following token, it will necessarily take its previous choice into account and never puts a word starting with a vowel after it has already chosen “a”.

The model still has some representation of whether the word after an/a is more likely to start with a vowel or not when it outputs a/an. You can trivially understand this is true by asking LLMs to answer questions with only one correct answer.

"The animal most similar to a crocodile is:"

https://chatgpt.com/share/67d493c2-f28c-8010-82f7-0b60117ab2...

It will always say "an alligator". It chooses "an" because somewhere in the next word predictor it has already figured out that it wants to say alligator when it chooses "an".

If you ask the question the other way around, it will always answer "a crocodile" for the same reason.

Again, that's not a good example I think because everything about the answer is in the prompt, so obviously from the start the "alligator" is high, but then it's just waiting for an "an" to occur to have an occasion to put that.

That doesn't mean it knows "in advance" what it want to say, it's just that at every step the alligator is lurking in the logits because it directly derives from the prompt.

You write: "it's just that at every step the alligator is lurking in the logits because it directly derives from the prompt" - but isn't that the whole point: at the moment the model writes "an", it isn't just spitting out a random article (or a 50/50 distribution of articles or other words for that matter); rather, "an" gets a high probability because the model internally knows that "alligator" is the correct thing after that. While it can only emit one token in this step, it will emit "an" to make it consistent with its alligator knowledge "lurking". And btw while not even directly relevant, the word alligator isn't in the prompt. Sure, it derives from the prompt but so does every an LLM generates, and same for any other AI mechanism for generating answers.
> While it can only emit one token in this step, it will emit "an" to make it consistent with its alligator knowledge "lurking".

It will also emit "a" from time to time without issue though, but will never spit "alligator" right after that, that's it.

> Sure, it derives from the prompt but so does every an LLM generates, and same for any other AI mechanism for generating answers.

Not really, because of the autoregressive nature of LLMs, the longer the response the more it will depend on its own response rather than the prompt. That's why you can see totally opposite response from LLM to the same query if you aren't asking basic factual questions. I saw a tool on reddit a few month ago that allowed you to see which words in the generation where the most “opinionated” (where the sampler had to chose between alternative words that were close in probability) and where it was easy to see that you could dramatically affect the result by just changing certain words.

> "an" gets a high probability because the model internally knows that "alligator" is the correct thing after that.

This is true, though it only works with this kind of prompt because the output of the LLM has little impact on the generation.

Globally I see what you mean, and I don't disagree with you, but at the same time, I think that saying that LLMs have a sense of anticipating the further token misses their ability to get driven astray by their own output: they have some information that will affect further tokens but any token that get spit can, and will, change that information in a way that can dramatically change the “plans”. And that's why I think using trivial questions isn't a good illustration, because it pushes this effect under the rug.

yunwal has provided one example. Here's another using much smaller model.

https://chat.groq.com/?prompt=If+a+person+from+Ontario+or+To...

The response "If a person from Ontario or Toronto is a Canadian, a person from Sydney or Melbourne would be an Australian!"

It seems mighty unlikely that it chose Australian as the country because of the 'an', or that it chose to put the 'an' at that point in the sentence for any other reason that the word Australian was going to be next.

For any argument that you think that this does not mean that have some idea of what is to come, try and come up with a test to see if your hypothesis is true or not, then give that test a try.

No, the person you're responding to is absolutely right. The easy test (which has been done in papers again and again) is the ability to train linear probes (or non-linear classifier heads) on the current hidden representations to predict the nth-next token, and the fact that these probes have very high accuracy.
Thanks for writing this so clearly... I hear wrong/misguided arguments like we see hear every day from friends, colleagues, "experts in the media" etc.

It's strange because just a moment of thinking will show that such ideas are wrong or paint a clearly incomplete picture. And there's plenty of analogies to the dangers of such reductionism. It should be obviously wrong to anyone who has at least tried ChatGPT.

My only explanation is that a denial mechanism must be at play. It simply feels more comfortable to diminish LLM capabilities and/or feel that you understand them from reading a Medium article on transformer-network, than to consider the consequences in terms of the inner black-box nature.

(comment deleted)
I think your analogy about logic gates vs. CPUs is spot on. Another apt analogy would be missing the forest for the trees—the model may in fact be generating a complete forest, but its output (natural language) is inherently serial so it can only plant one tree at a time. The sequence of distributions that is the proximate driver of token selection is just the final distillation step.
It literally doesn't know how to handle 'I don't know' and needs to be taught. fascinating.
I think it would be more accurate to say that after fine tuning on a series of questions with answers that it thinks that you don't want to hear "I don't know"
I think it's more fundamental than that. If you start saying "it thinks" in regards to an LLM, you're wrong. LLMs don't think, they pattern match fuzzily.

If the training data contained a bunch of answers to questions which were simply "I don't know", you could get an LLM to say "I don't know" but that's still not actually a concept of not knowing. That's just knowing that the answer to your question is "I don't know".

It's essentially like if you had an HTTP server that responded to requests for nonexistent documents with a "200 OK" containing "Not found". It's fundamentally missing the "404 Not found" concept.

LLMs just have a bunch of words--they don't understand what the words mean. There's no metacognition going on for it to think "I don't know" for it to even think you would want to know that.

>I think it's more fundamental than that. If you start saying "it thinks" in regards to an LLM, you're wrong. LLMs don't think, they pattern match fuzzily.

I'm not sure if this objection is terribly helpful. We use terms like think and want to describe processes that are clearly not involve any form of understanding. Electrons do not have motivations but they 'want' to go to a lower energy level in an atom. You can hold down the trigger for the fridge light to make it 'think' that the door has not been opened. These are uncontentious phrases that convey useful ideas.

I understand that when people are working towards producing reasoning machines the words might be working in similar spaces, but really when someone is making claims about machines having awareness, understanding, or thinking they make it quite clear about the context that they are talking about.

As to the rest of your comment, I simply disagree. If you think of a concept of an internal representation of a piece of information, then it has been shown that they do have such representations. In the Karpathy video I mentioned he talks about how researches found that models did have an internal representation of not knowing, but that the fine tuning was restricting it to providing answers. Giving it fine-tuning examples where it said "I don't know" for information that they knew the model didn't know. This generalised to provide "I don't know" for examples that were not in the training data. For the fine tuning examples to succeed in that, it requires the model to already contain the concept.

I would agree that models do not have any in-depth understanding of what lack of knowledge actually is. On the other hand I would also think that this also applies to humans, most people are not philosophers.

I think that the models can express details about words shows that they do have detailed information about what each word means semantically. In many respects because of tokenisation indexing embeddings it would perhaps be more accurate to say that they have a better understanding of the semantic information of what words mean the what the words actually are. This is why they are poor at spelling but can give you detailed information about the thing they can't spell.

> We use terms like think and want to describe processes that are clearly not involve any form of understanding.

...and that's why so many people are confused about what's going on with LLMs: sloppy, ambiguous use of language.

> In the Karpathy video I mentioned he talks about how researches found that models did have an internal representation of not knowing, but that the fine tuning was restricting it to providing answers. Giving it fine-tuning examples where it said "I don't know" for information that they knew the model didn't know.

This is why I included the HTTP example: this is simply telling it to parrot the phrase "I don't know"--it doesn't understand that it doesn't know. From the LLM's perpective, it "knows" that the answer is "I don't know". It's returning a 200 OK that says "I don't know" rather than returning a 404.

Do you understand the distinction I'm making here?

> I would agree that models do not have any in-depth understanding of what lack of knowledge actually is. On the other hand I would also think that this also applies to humans, most people are not philosophers.

The average (non-programmer) human, when asked to write a "Hello, world" program, can definitely say they don't know how to program. And unlike the LLM, the human knows that this is different from answering the question. The LLM, in contrast thinks it is answering the question when it says "I don't know"--it thinks "I don't know" is the correct answer.

Put another way, a human can distinguish between responses to these two questions, whereas an LLM can't:

1. What is my grandmother's maiden name?

2. What is the English translation of the Spanish phrase, "No sé."?

In the first question, you don't know the answer unless you are quite creepy; in the second case you do (or can find out easily). But the LLM tuned to answer I don't know thinks it knows the answer in both cases, and thinks the answer is the same.

>...and that's why so many people are confused about what's going on with LLMs: sloppy, ambiguous use of language.

There is a difference between explanation by metaphor and lack of precision. If you think someone is implying something literal when they might be using a metaphor you can always ask for clarification. I know plenty of people that are utterly precise in their use in their language which leads them to being widely misunderstood because they think a weak precise signal is received as clearly as a strong imprecise signal. They usually think the failure in communication is in the recipient but in reality they are just accurately using the wrong protocol.

>Do you understand the distinction I'm making here? I believe I do, and it is precisely this distinction that the researches showed. By teaching a model to say "I don't know" for some information that they knew the model did not know the answer to, the model learned to respond "I don't know" for things that it did not know that it was not explicitly taught to respond with "I don't know". For it to acquire that ability to generalise to new cases the model has to have already had an internal representation of "That information is not available"

I'm not sure where you think a model converting its internal representation of not knowing something into words is distinct from a human converting its internal representation of not knowing into words.

When fine tuning directs a model to profess lack of knowledge, usually they will not give the same specific "I don't know" text as a way to express that it does not not know because they want the want to bind the concept "lack of knowledge" to the concept of "communicate that I do not know" rather than any particular word phrase. Giving it many ways to say "I don't know" builds that binding rather than the crude "if X then emit Y" that you imagine it to be.

> It would be quite easy to detect (and exploit) behaviour that decided to use a vowel word just because it somewhat arbitrarily used an 'an'.

That is a very interesting observation!

Doesn’t that internal state get blown away and recreated for every “next token”? Isn’t the output always the previous context plus the new token, which gets fed back and out pops the new token? There is no transfer of internal state to the new iteration beyond what is “encoded” in its input tokens?

>Doesn’t that internal state get blown away and recreated for every “next token”

That is correct. When a model has a good idea of the next 5 words, after it has emitted the first of those 5 most architectures make no further use of the other 4 and regenerate likely the same information again in the next inference cycle.

There are architectures that don't discard all that information but the standard LLM has generally outperformed them, for now.

There are interesting philosophical implications if LLMs were to advance to a level to be considered sentient. Would it not be constantly creating and killing a thinking being for every token. On the other hand if context is considered memory, perhaps continuity of identity is based upon memory and all that other information are simply forgotten idle thoughts. We have no concept of what our previous thoughts were except from our memory. Is that not the same.

Sometimes I wonder if some of the resistance to AI is because it can do things that we think requires abilities that we would like to believe that we possess ourselves, and showing that they are not necessary creates the possibility that we might not have have those abilities.

There was a great observation recently in an interview (I forget the source, but the interviewer's last name was Bi) that some of the discoveries that met the most resistance in history such as the Earth orbiting the Sun, or Darwin's theory of evolution were similar in that they implied that we are not a unique special case.

> there isn't a "I don't have enough information" option. This is due to the fact that LLMs are basically just giant look up maps with interpolation.

Have you ever tried telling ChatGPT that you're "in the city centre" and asking it if you need to turn left or right to reach some landmark? It will not answer with the average of the directions given to everybody who asked the question before, it will answer asking you to tell it where you are precisely and which way you are facing.

That's because, based on the training data, the most likely response to asking for directions is to clarify exactly where you are and what you see.

But if you ask it in terms of a knowledge test ("I'm at the corner of 1st and 2nd, what public park am I standing next to?") a model lacking web search capabilities will confidently hallucinate (unless it's a well-known park).

In fact, my person opinion is that, therein lies the most realistic way to reduce hallucination rates: rather than trying to train models to say "I don't know" (which is not really a trainable thing - models are fundamentally unaware of the limits of their own training data), instead just train them on which kinds of questions warrant a web search and which ones should be answered creatively.

I tried this just now on Chatbot Arena, and both chatbots asked for more information.

One was GPT 4.5 preview, and one was cohort-chowder (which is someone's idea of a cute code name, I assume).

I tried this just now on Chatbot Arena, and both chatbots very confidently got the name of the park wrong.

Perhaps you thought I meant "1st and 2nd" literally? I was just using those as an example so I don't reveal where I live. You should use actual street names that are near a public park, and you can feel free to specify the city and state.

I did think you meant it literally. Since I can't replicate the question you asked, I have no way of verifying your claim.
[flagged]
Neither do I. Right after I read your reply I knew I had made a mistake engaging with you.
I don’t buy Lecun’s argument. Once you get good RL going (as we are now seeing with reasoning models) you can give the model a reward function that rewards a correct answer most highly, an “I’m sorry but I don’t know” less highly than that, a wrong answer penalized, a confidently wrong answer more severely penalized. As the RL learns to maximize rewards I would think it would find the strategy of saying it doesn’t know in cases where it can’t find an answer it deems to have a high probability of correctness.
How do you define the "correct" answer?
obviously the truth is what is the most popular. /s
Certainly not possible in all domains but equally certainly possible in some. There’s not much controversy about the height of the Eiffel Tower or how to concatenate two numpy arrays.
If multiple answers are equally likely, couldn't that be considered uncertainty? Conversely if there's only one answer and there's a huge leap to the second best, that's pretty certain.
> The problem with LLMs is that the output is inherently stochastic

Isn't that true with humans too?

There's some leap humans make, even as stochastic parrots, that lets us generate new knowledge.

I think it is because we don't feel the random and chaotic nature of what we know as individuals.

If I had been born a day earlier or later I would have a completely different life because of initial conditions and randomness but life doesn't feel that way even though I think this is obviously true.

> i.e there isn't a "I don't have enough information" option.

This is true in terms of default mode for LLMs, but there's a fair amount of research dedicated to the idea of training models to signal when they need grounding.

SelfRAG is an interesting, early example of this [1]. The basic idea is that the model is trained to first decide whether retrieval/grounding is necessary and then, if so, after retrieval it outputs certain "reflection" tokens to decide whether a passage is relevant to answer a user query, whether the passage is supported (or requires further grounding), and whether the passage is useful. A score is calculated from the reflection tokens.

The model then critiques itself further by generating a tree of candidate responses, and scoring them using a weighted sum of the score and the log probabilities of the generated candidate tokens.

We can probably quibble about the loaded terms used here like "self-reflection", but the idea that models can be trained to know when they don't have enough information isn't pure fantasy today.

[1] https://arxiv.org/abs/2310.11511

EDIT: I should also note that I generally do side with Lecun's stance on this, but not due to the "not enough information" canard. I think models learning from abstraction (i.e. JEPA, energy-based models) rather than memorization is the better path forward.

>This is due to the fact that LLMs are basically just giant look up maps with interpolation.

This is obviously not true at this point except for the most loose definition of interpolation.

>don't rely on things like differentiability.

I've never heard lecun say we need to move away from gradient descent. The opposite actually.

i dont think the stochasticity that's the problem -- the problem is that model gets "locked in" once it picks a token and there's no takesies backsies.

that also entails information destruction in the form of the logits table, but for the most part that should be accounted for in the last step before final feedforward

I'm not an ML researcher, but I do work in the field.

My mental model of AI advancements is that of a step function with s-curves in each step [1]. Each time there is an algorithmic advancement, people quickly rush to apply it to both existing and new problems, demonstrating quick advancements. Then we tend to hit a kind of plateau for a number of years until the next algorithmic solution is found. Examples of steps include, AlexNet demonstrating superior image labeling, LeCun demonstrating DeepLearning, and now OpenAI demonstrating large transformer models.

I think in the past, at each stage, people tend to think that the recent progress is a linear or exponential process that will continue forward. This lead to people thinking self driving cars were right around the corner after the introduction of DL in the 2010s, and super-intelligence is right around the corner now. I think at each stage, the cusp of the S-curve comes as we find where the model is good enough to be deployed, and where it isn't. Then companies tend to enter a holding pattern for a number of years getting diminishing returns from small improvements on their models, until the next algorithmic breakthrough is made.

Right now I would guess that we are around 0.9 on the S curve, we can still improve the LLMs (as DeepSeek has shown wide MoE and o1/o3 have shown CoT), and it will take a few years for the best uses to be brought to market and popularized. As you mentioned, LeCun points out that LLMs have a hallucination problem built into their architecture, others have pointed out that LLMs have had shockingly few revelations and breakthroughs for something that has ingested more knowledge than any living human. I think future work on LLMs are likely to make some improvement on these things, but not much.

I don't know what it will be, but a new algorithm will be needed to induce the next step on the curve of AI advancement.

[1]: https://www.open.edu/openlearn/nature-environment/organisati...

> Each time there is an algorithmic advancement, people quickly rush to apply it to both existing and new problems, demonstrating quick advancements. Then we tend to hit a kind of plateau for a number of years until the next algorithmic solution is found.

That seems to be how science works as a whole. Long periods of little progress between productive paradigm shifts.

This is actually a lazy approach as you describe it. Instead, what is needed is an elegant and simple approach that is 99% of the way there out of the gate. Soon as you start doing statistical tweaking and overfitting models, you are not approaching a solution.
In a way yes. For models in physics that should make you suspicious, since most of our famous and useful models found are simple and accurate. However, in general intelligence or even multimodal pattern matching there’s no guarantee there’s an elegant architecture at the core. Elegant models in social sciences like economics, sociology and even fields like biology tend to be hilariously off.
It's been described as fumbling around in a dark room until you find the light switch. At which point you can see the doorway leading to the next dark room.
That is how science seems to work as a whole. What worries me is that the market views the emergence of additional productive paradigm shifts in AI as only a matter of money. A normal scientific advancement plateau for another five years in AI would be a short-term disaster for the stock market and economy.
This is a somewhat nihilistic take with an optimistic ending. I believe humans will never fix hallucinations. Amount of totally or partially untrue statements people make is significant. Especially in tech, it's rare for people to admit that they do not know something. And yet, despite all of that the progress keeps marching forward and maybe even accelerating.
Yeah, I think a lot of people talk about "fixing hallucinations" as the end goal, rather than "LLMs providing value", which misses the forest for the trees; it's obviously already true that we don't need totally hallucination-free output to get value from these models.
Even as language models can partially solve a few problems, we remain with the problem of achieving Artificial General Intelligence, that the presence of LLMs has exacerbated because they so often reveal to be artificial morons.

Intelligence finds solutions - actual, solid solutions.

More than "fixing" hallucinations, the problem is going beyond them (arriving to "sobriety").

I’m not sure I follow. Sure, people lie, and make stuff up all the time. If an LLM goes and parrots that, then I would argue that it isn’t hallucinating. Hallucinating would be where it makes something up that is not in its training site nor logically deducible from it.
I think most humans are perfectly capable of admitting to themselves when they do not know something. Computers ought to do better.
You must interact with a very different set of humans than most.
Once one starts thinking of them as "concept models" rather than language models or fact models, "hallucinations" become something not to be so fixated on. We transform tokens into 12k+ length embeddings... right at the start. They stop being language immediately.

They aren't fact machines. They are concept machines.

Not an argument. "Many people are delirious, yet some people create progress". What is that supposed to imply?
(comment deleted)
I've never understood this critique. Models have the capability to say: "oh, I made a mistake here, let me change this" and that solves the issue, right?

A little bit of engineering and fine tuning - you could imagine a model producing a sequence of statements, and reflecting on the sequence - updating things like "statement 7, modify: xzy to xyz"

I get "oh, I made a mistake" quite frequently. Often enough, it's just another hallucination, just because I contested the result, or even just prompted "double check this". Statistically speaking, when someone in a conversation says this, the other party is likely to change their position, so that's what an LLM does, too, replicating a statistically plausible conversation. That often goes in circles, not getting anywhere near a better answer.

Not an ML researcher, so I can't explain it. But I get a pretty clear sense that it's an inherent problem and don't see how it could be trained away.

"Oh, I emptied your bank account here, let me change this."

For AI to really replace most workers like some people would like to see, there are plenty of situations where hallucinations are a complete no-go and need fixing.

Isn’t that the answer if you tell them they are wrong?
I've always felt like the argument is super flimsy because "of course we can _in theory_ do error correction". I've never seen even a semi-rigorous argument that error correction is _theoretically_ impossible. Do you have a link to somewhere where such an argument is made?
In theory transformers are Turing-complete and LLMs can do anything computable. The more down-to-earth argument is that transformer LLMs aren't able to correct errors in a systematic way like Lecun is describing: it's task-specific "whack-a-mole," involving either tailored synthetic data or expensive RLHF.

In particular, if you train an LLM to do Task A and Task B with acceptable accuracy, that does not guarantee it can combine the tasks in a common-sense way. "For each step of A, do B on the intermediate results" is a whole new Task C that likely needs to be fine-tuned. (This one actually does have some theoretical evidence coming from computational complexity, and it was the first thing I noticed in 2023 when testing chain-of-thought prompting. It's not that the LLM can't do Task C, it just takes extra training.)

> of course we can _in theory_ do error correction

Oh yeah? This is begging the question.

As soon as you need to start leaning heavily on error correction, that is an indication that your architecture and solution is not correct. The final solution will need to be elegant and very close to a perfect solution immediately.

You must always keep close to the only known example we have of an intelligence which is the human brain. As soon as you start to wander away from the way the human brain does it, you are on your own and you are not relying on known examples of intelligence. Certainly that might be possible, but since there's only one known example in this universe of intelligence, it seems ridiculous to do anything but stick close to that example, which is the human brain.

Sincere question - why doesn't RL-based fine-tuning on top of LLMs solve this or at least push accuracy above a minimum acceptable threshhold in many use cases? OAI has a team doing this for enterprise clients. Several startups rolling out of current YC batch are doing versions of this.
If you mean the so called agentic AI, I don't think it's several. Iirc someone in the most recent demo day mentioned ~80%+ were AI
Sligtly related: Energy Based Models (EBMs) are better in theory and yet too resource intensive. I tried to sell using EBMs to my org, but the price for even a small use case was prohibitive.

I learned it from: https://youtube.com/playlist?list=PLLHTzKZzVU9eaEyErdV26ikyo...

Yann LeCun, and Michael Bronstein and his colleagues have some similarities in trying to properly Sciencify Deep Learning.

Yann LeCun's approach, as least for Vision has one core tenet- energy minimization, just like in Physics. In his course, he also shows some current arch/algos to be special cases for EBMs.

Yann believes that understanding the Whys of the behavior of DL algorithms are going to be beneficial in the long term rather than playing around with hyper-params.

There is also a case for language being too low-dimensional to lead to AGI even if it is solved. Like, in a recent video, he said that the total amount of data existing on all digitized books and internet are the same as what a human children takes in in the first 4/5 years. He considers this low.

There are also epistemological arguments against language not being able to lead to AGI, but I haven't heard him talk about them.

He also believes that Vision is a more important aspect of intellgence. One reason being it being very high-dim. (Edit) Consider an example. Take 4 monochrome pixels. All pixels can range from 0 to 255. 4 pixels can create 256^4 = 2^32 combinations. 4 words can create 4! = 24 combinations. Solving language is easier and therefore low-stakes. Remember the monkey producing a Shakespeare play by randomly punching typewriter keys? If that was an astronomically big number, think how obscenely long it would take a monkey to paint Mona Lisa by randomly assigning pixel values. Left as an exercise to the reader.

Juergen Schmidhuber has gone a lot queit now. But he also told that a world-model, explicitly included in training is reasoning is better, rather than only text or image or whatever. He has a good paper with Lucas Beyer.

Thanks. This is interesting. What kind of equation is used to assess an ebm during training? I’m afraid I still don’t get the core concept well enough to have an intuition for it.
Since this exposes the answer, the new architecture has to be based on world model building.
The thing is, this has been known since even before the current crop of LLMs. Anyone who considered (only the English) language to be sufficient to model the world understands so little about cognition as to be irrelevant in this conversation.
WTF. The cardinality of words is 100,000.
Jürgen Schmidhuber has a paper with Lucas Beyer? I'm not aware of it. Which do you mean?
Oh, I made a mistake unfortunately. I meant hardmaru (David Ha).

I am very sorry.

I'm not an insider and I'm not sure whether this is directly related to "energy minimization", but "diffusion language models" have apparently gained some popularity in recent weeks.

https://arxiv.org/abs/2502.09992

https://www.inceptionlabs.ai/news

(these are results from two different teams/orgs)

It sounds kind of like what you're describing, and nobody else has mentioned it yet, so take a look and see whether it's relevant.

And they seem to be about 10x as fast as similar sized transformers.
No, 10x less sampling steps. Whether or not that means 10x faster remains to be seen, as a diffusion step tends to be more expensive than an autoregressive step.
If I understood correctly, in practice they show actual speed improvement on high-end cards, because autoregressive LLMs are bandwidth limited and do not compute bound, so switching to a more expensive but less memory bandwidth heavy is going to work well on current hardware.
The SEDD architecture [1] probably allows for parallel sampling of all tokens in a block at once, which may be faster but not necessarily less computationally demanding in terms of runtime times computational resources used.

[1] Which Inception Labs's new models may be based on; one of the cofounders is a co-author. See equations 18-20 in https://arxiv.org/abs/2310.16834

I have no idea about EBM, but I have researched a bit on the language modelling side. And let's be honest, GPT is not the best learner we can create right now (ourselves). GPT needs far more data and energy than a human, so clearly there is a better architecture somewhere waiting to be discovered.

Attention works, yes. But it is not naturally plausible at all. We don't do quadratic comparisons across a whole book or need to see thousands of samples to understand.

Personally I think that in the future recursive architectures and test time training will have a better chance long term than current full attention.

Also, I think that OpenAI biggest contribution is demostrating that reasoning like behaviors can emerge from really good language modelling.

I haven't read Yann Lecun's take. Based on your description alone my first impression would be: there's a paper [1] arguing that "beam search enforces uniform information density in text, a property motivated by cognitive science". UID claims, in short, that a speaker only delivers as much content as they think the listener can take (no more, no less) and the paper claims that beam search enforced this property at generation time.

The paper would be a strong argument against your point: if neural architectures are already constraining the amount of information that a text generation system delivers the same way a human (allegedly) does, then I don't see which "energy" measure one could take that could perform any better.

Then again, perhaps they have one in mind and I just haven't read it.

[1] https://aclanthology.org/2020.emnlp-main.170/

I believe he’s talking about some sort of ‘energy as measured by distance from the models understanding of the world’ as in quite literally a world model. But again I’m ignorant, hence the post!
When an architecture is based around world model building, then it is a casual outcome that similar concepts and things end up being stored in similar places. They overlap. As soon as your solution starts to get mathematically complex, you are departing from what the human brain does. Not saying that in some universe it might be possible to make a statistical intelligence, but when you go that direction you are straying away from the only existing intelligences that we know about. The human brain. So the best solutions will closely echo neuroscience.
In some respects that sounds similar to what we already do with reward models. I think with GRPO, the “bag of rewards” approach doesn’t strike me as terribly different. The challenge is in building out a sufficient “world” of rewards to adequately represent more meaningful feedback-based learning.

While it sounds nice to reframe it like a physics problem, it seems like a fundamentally flawed idea, akin to saying “there is a closed form solution to the question of how should I live.” The problem isn’t hallucinations, the problem is that language and relativism are inextricably linked.

This sort of measure is a decent match for BPB though. BPB=-log(document_probability)/document_length_bytes and perplexity=e^(BPB*document_length_bytes/document_length_tokens). We already train models by minimizing perplexity, and model outputs are already those that are high probability. Though like with EBMs, figuring out outputs with even higher probability would require an expensive search step.
Any transformer based LLM will never achieve AGI because it's only trying to pick the next word. You need a larger amount of planning to achieve AGI. Also, the characteristics of LLMs do not resemble any existing intelligence that we know of. Does a baby require 2 years of statistical analysis to become useful? No. Transformer architectures are parlor tricks. They are glorified Google but they're not doing anything or planning. If you want that, then you have to base your architecture on the known examples of intelligence that we are aware of in the universe. And that's not a transformer. In fact, whatever AGI emerges will absolutely not contain a transformer.
It's not about just picking the next word here, that doesn't at all refuse whether Transformers can achieve AGI. Words are just one representation of information. And whether it resembles any intelligence we know is also not an argument because there is no reason to believe that all intelligence is based on anything we've seen (e.g us, or other animals). The underlying architecture of Attention & MLPs can surely still depict something which we could call an AGI, and in certain tasks it surely can be considered an AGI already. I also don't know for certain whether we will hit any roadblocks or architectural asymptotes but I haven't come across any well-founded argument that Transformers definitely could not reach AGI.
The transformer is a simple and general architecture. Being such a flexible model, it needs to learn "priors" from data, it makes few assumptions on its distribution from the start. The same architecture can predict protein folding and fluid dynamics. It's not specific to language.

We on the other hand are shaped by billions of years of genetic evolution, and 200k years of cultural evolution. If you count the total number of words spoken by 110 billion people who ever lived, assuming 1B estimated words per human during their lifetime, it comes out to 10 million times the size of GPT-4's training set.

So we spent 10 million more words discovering than it takes the transformer to catch up. GPT-4 used 10 thousand people's worth of language to catch up all that evolutionary finetuning.

> words spoken by 110 billion people who ever lived, assuming 1B estimated words per human during their lifetime..comes out to 10 million times the size of GPT-4's training set

This assumption is slightly wrong direction, because not exist human who could consume much more than about 1B words during their lifetime. So humanity could not gain enhancement from just multiply words of one human by 100 billion. I think, correct estimation could be 1B words multiply by 100.

I think, current AI already achieved size need to become AGI, but to finish, probably need to change structure (but I'm not sure about this), and also need some additional multidimensional dataset, not just texts.

I might bet on 3D cinema, and/or on automobile targeting autopilot dataset, or something for real life humanoid robots solving typical human tasks, like fold shirt.

> Does a baby require 2 years of statistical analysis to become useful?

Well yes, actually.

of the entire human race's knowledge, and it's like from written history, not 2 years ago.
The alternative architectures must learn from streaming data, must be error tolerant and must have the characteristic that similar objects or concepts much naturally come near to each other. They must naturally overlap.
Ever hear of Dissociated Press? If not, try the following demonstration.

Fire up Emacs and open a text file containing a lot of human-readable text. Something off Project Gutenberg, say. Then say M-x dissociated-press and watch it spew hilarious, quasi-linguistic garbage into a buffer for as long as you like.

Dissociated Press is a language model. A primitive, stone-knives-and-bearskins language model, but a language model nevertheless. When you feed it input text, it builds up a statistical model based on a Markov chain, assigning probabilities to each character that might occur next, given a few characters of input. If it sees 't' and 'h' as input, the most likely next character is probably going to be 'e', followed by maybe 'a', 'i', and 'o'. 'r' might find its way in there, but 'z' is right out. And so forth. It then uses that model to generate output text by picking characters at random given the past n input characters, resulting in a firehose of things that might be words or fragments of words, but don't make much sense overall.

LLMs are doing the same thing. They're picking the next token (word or word fragment) given a certain number of previous tokens. And that's ALL they're doing. The only differences are really matters of scale: the tokens are larger than single characters, the model considers many, many more tokens of input, and the model is a huge deep-learning model with oodles more parameters than a simple Markov chain. So while Dissociated Press churns out obvious nonsensical slop, ChatGPT churns out much, much more plausible sounding nonsensical slop. But it's still just rolling them dice over and over and choosing from among the top candidates of "most plausible sounding next token" according to its actuarial tables. It doesn't think. Any thinking it appears to do has been pre-done by humans, whose thoughts are then harvested off the internet and used to perform macrodata refinement on the statistical model. Accordingly, if you ask ChatGPT a question, it may well be right a lot of the time. But when it's wrong, it doesn't know it's wrong, and it doesn't know what to do to make things right. Because it's just reaching into a bag of refrigerator magnet poetry tiles, weighted by probability of sounding good given the current context, and slapping whatever it finds onto the refrigerator. Over and over.

What I think Yann LeCun means by "energy" above is "implausibility". That is, the LLM would instead grab a fistful of tiles -- enough to form many different responses -- and from those start with a single response and then through gradient descent or something optimize it to minimize some statistical "bullshit function" for the entire response, rather than just choosing one of the most plausible single tiles each go. Even that may not fix the hallucination issue, but it may produce results with fewer obvious howlers.

+1

But there's a fundamental difference between Markov chains and transformers that should be noted. Markov chains only learn how likely it is for one token to follow another. Transformers learn how likely it is for a set of token to be seen together. Transformers add a wider context to msrkov chain. That quantitative change leads to a qualitative improvement: transformers generate text that is semantically plausible.

Yes, but k-token lookup was already a thing with markov chains. Transformers are indeed better, but just because they model language distributions better than mostly-empty arrays of (token-count)^(context).
I am an MLE not an expert. However, it is a fundamental problem that our current paradigm of training larger and larger LLMs cannot ever scale to the precision people require for many tasks. Even in the highly constrained realm of chess, an enormous neural net will be outclassed by a small program that can run on your phone.

https://arxiv.org/pdf/2402.04494

> Even in the highly constrained realm of chess, an enormous neural net will be outclassed by a small program that can run on your phone.

This is true also for the much bigger neural net that works in your brain, and even if you're the world champion of chess. Clearly your argument doesn't hold water.

For the sake of argument let’s say an artificial neural net is approximately the same as the brain. It sounds like you agree with me that smaller programs are both more efficient and more effective than a larger neural net. So you should also agree with me that those who say the only path to AGI is LLM maximalism are misguided.
smaller programs are better than artificial or organic neural net for constrained problems like chess. But chess programs don't generalize to any other intelligence applications, like how organic neural nets do today.
> It sounds like you agree with me that smaller programs are both more efficient and more effective than a larger neural net.

At playing chess. (But also at doing sums and multiplications, yay!)

> So you should also agree with me that those who say the only path to AGI is LLM maximalism are misguided.

No. First of all, it's a claim you just made up. What we're talking about is people saying that LLMs are not the path to AGI- an entirely different claim.

Second, assuming there's any coherence to your argument, the fact that a small program can outclass an enormous NN is irrelevant to the question of whether the enormous NN is the right way to achieve AGI: we are "general intelligences" and we are defeated by the same chess program. Unless you mean that achieving the intelligence of the greatest geniuses that ever lived is still not enough.

Any chance that “reasoning” can fix this
It kind of depends. You can broadly call any kind of search “reasoning”. But search requires 1) enumerating your possible options and 2) assigning some value to those options. Real world problem solving makes both of those extremely difficult.

Unlike in chess, there’s a functionally infinite number of actions you can take in real life. So just argmax over possible actions is going to be hard.

Two, you have to have some value function of how good an action is in order to argmax. But many actions are impossible to know the value of in practice because of hidden information and the chaotic nature of the world (butterfly effect).

Isn't something about alphago also involves "infinitely" many possible outcomes? Yet they cracked it, right?
Go is played on a 19x19 board. At the beginning of the game the first player has 361 possible moves. The second player then has 360 possible moves. There is always a finite and relatively “small” number of options.

I think you are thinking of the fact that it had to be approached in a different way than Minimax in chess because a brute force decision tree grows way too fast to perform well. So they had to learn models for actions and values.

In any case, Go is a perfect information game, which as I mentioned before, is not the same as problems in the real world.

Best in class chess program actually is a NN, just not a LLM.
(comment deleted)
This concept comes from Hopfield networks.

If two nodes are on, but the connection between them is negative, this causes energy to be higher.

If one of those nodes switches off, energy is reduced.

With two nodes this is trivial. With 10 nodes it's more difficult to solve, and with billions of nodes it is impossible to "solve".

All you can do then is try to get the energy as low as possible.

This way also neural networks can find out "new" information, that they have not learned, but is consistent with the constraints they have learned about the world so far.

So, what’s modeled as a “node” in an EBM, and what’s modeled as a connection? Are they vectors in a tensor, (well I suppose almost certainly that’s a yes). Do they run side by side a model that’s being trained? Is the node connectivity architecture fixed or learned?
(comment deleted)
Well, it could be argued that the “optimal response” ie the one that sorta minimizes that “energy” is sorted by LLMs on the first iteration. And further iterations aren’t adding any useful information and in fact are countless occasions to veer off the optimal response.

For example if a prompt is: “what is the Statue of Liberty”, the LLMs first output token is going to be “the”, but it kinda already “knows” that the next ones are going to be “statue of liberty”.

So to me LLMs already “choose” a response path from the first token.

Conversely, a LLM that would try and find a minimum energy for the whole response wouldn’t necessarily stop hallucinating. There is nothing in the training of a model that says that “I don’t know” has a lower “energy” than a wrong answer…

Not an ML researcher, but implementing these systems has shown this opinion to be correct. The non-determinism of LLMs is a feature, not a bug that can be fixed.

As a result, you'll never be able to get 100% consistent outputs or behavior (like you hypothetically can with a traditional algorithm/business logic). And that has proven out in usage across every model I've worked with.

There's also an upper-bound problem in terms of context where every LLM hits some arbitrary amount of context that causes it to "lose focus" and develop a sort of LLM ADD. This is when hallucinations and random, unrequested changes get made and a previously productive chat spirals to the point where you have to start over.

> where every LLM hits some arbitrary amount of context that causes it to "lose focus" and develop a sort of LLM ADD.

Humans brains have the same problem. As any intelligence probably. Solution for this is structural thinking. One piece at a time, often top-down. Educated humans do it, LLM can be orchestrated to do it too. Effective context window will be limited even though some claim millions of tokens.

Yann LeCun understands this is an electrical engineering and physical statistics of machine problem and not a code problem.

The physics of human consciousness are not implemented in a leaky symbolic abstraction but the raw physics of existence.

The sort of autonomous system we imagine when thinking AGI must be built directly into substrate and exhibit autonomous behavior out of the box. Our computers are blackboxes made in a lab without centuries of evolving in the analog world, finding a balance to build on. They either can do a task or cannot. Obviously from just looking at one we know how few real world tasks it can just get up and do.

Code isn’t magic, it’s instruction to create a machine state. There’s no inherent intelligence to our symbolic logic. It’s an artifact of intelligence. It cannot imbue intelligence into a machine.

I argue that JEPA and its Energy-Based Model (EBM) framework fail to capture the deeply intertwined nature of learning and prediction in the human brain—the “yin and yang” of intelligence. Contemporary machine learning approaches remain heavily reliant on resource-intensive, front-loaded training phases. I advocate for a paradigm shift toward seamlessly integrating training and prediction, aligning with the principles of online learning.

Disclosure: I am the author of this paper.

Reference: (PDF) Hydra: Enhancing Machine Learning with a Multi-head Predictions Architecture. Available from: https://www.researchgate.net/publication/381009719_Hydra_Enh... [accessed Mar 14, 2025].

Thank you. So, quick q - it would make sense to me that JEPA is an outcome of the YLC work; would you say that’s the case?
So you believe humans spend more energy on prediction, relative to computers? Isn't that because personal computers are not powerful enough to train big models, and most people have no desire to? It is more economically efficient to socialize the cost of training, as is done. Are you thinking of a distributed training, where we split the work and cost? That could happen when robots become more widespread.
The human brain operates at just 25W of power—less than the monitor you're likely using right now—whereas AI models like ChatGPT consume nearly 1GWh every 24 hours!

As I discuss in the paper, predictive coding suggests that the brain actively generates predictions and compares them to incoming sensory data (vision, hearing, etc.), prioritizing anomalies. Its efficiency stems from a hierarchical memory system that continuously updates only the "deltas"—the differences that matter. Embracing this approach could lead to a paradigm shift, enabling the development of significantly more energy-efficient AI in the future.

Update: Interesting paper, thanks. Comment on selection for Hydra — you mention v1 uses an arithmetic mean across timescales for prediction. Taking this analogy of the longer windows encapsulating different timescales, I’d propose it would be interesting to train a layer to predict weighting of the timescale predictions. Essentially — is this a moment where I need to focus on what just happened, or is this a moment in which my long range predictions are more important?
Ty for reading the paper! I completely agree! Assigning soft weights to the window based on context is a fascinating research area. This concept is similar to Ebbinghaus' forgetting curve, which emphasizes recency bias while requiring repeated exposure for long-term retention.
I have a paper coming up that I modestly hope will clarify some of this.

The short answer should be that it's obvious LLM training and inference are both ridiculously inefficient and biologically implausible, and therefore there has to be some big optimization wins still on the table.

> and biologically implausible

I really like this approach. Showing that we must be doing it wrong because our brains are more efficient and we aren't doing it like our brains.

Is this a common thing in ML papers or something you came up with?

Evolution does not need to converge on the optimum solution.

Have you heard of https://en.wikipedia.org/wiki/Bio-inspired_computing ?

It does not, you're right. But it's an interesting way to approach the problem never the less. And given that we definitely aren't as efficient as a human brain right now, it makes sense to look at the brain for inspiration.
I don't think GP was implying that brains are the optimum solution. I think you can interpret GP's comments like this- if our brains are more efficient than LLMs, then clearly LLMs aren't optimally efficient. We have at least one data point showing that better efficiency is possible, even if we don't know what the optimal approach is.
I agree. Spiking neural networks are usually mentioned in this context, but there is no hardware ecosystem behind them that can compete with Nvidia and CUDA.
Investments in AI are now counting by billions of dollars. Would that be enough to create an initial ecosystem for a new architecture?
Nvidia has a big lead, and hardware is capital intensive. I guess an alternative would make sense in the battery-powered regime, like robotics, where Nvidia's power hungry machines are at a disadvantage. This is how ARM took on Intel.
A new HW architecture for an unproven SW architecture is never going to happen. The SW needs to start working initially and demonstrate better performance. Of course, as with the original deep neural net stuff, it took computers getting sufficiently advanced to demonstrate this is possible. A different SW architecture would have to be so much more efficient to work. Moreover, HW and SW evolve in tandem - HW takes existing SW and tries to optimize it (e.g. by adding an abstraction layer) or SW tries to leverage existing HW to run a new architecture faster. Coming up with a new HW/SW combo seems unlikely given the cost of bringing HW to market. If AI speedup of HW ever delivers like Jeff Dean expects, then the cost of prototyping might come down enough to try to make these kinds of bets.
Nah it’s just physics, it’s like wheels being more efficient than legs.

We know there is a more efficient solution (human brain) but we don’t know how to make it.

So it stands to reason that we can make more efficient LLMs, just like a CPU can add numbers more efficiently than humans.

Wheels is an interesting analogy. Wheels are more efficient now that we have roads. But there could never have been evolutionary pressure to make them before there were roads. Wheels are also a lot easier to get to work than robotic legs and so long as there’s a road do a lot more than robotic legs.
(comment deleted)
People think the first wheel was invented for making pottery. Biological machinery for the most part has to be self-reproducing so there is a lot of limitations on design, also it has to be able to evolve, so you get inefficient solutions like the vargas nerve (i think that's its name), basically there's a really long nerve in your body that takes a route under your trachea and then back up to another part of your brain, in giraffes its something like 40 feet long to go a few inches shortest path.

Wheels other than rolling would likely never evolve naturally because there's no real incremental path from legs to wheels, where as flippers can evolve from webbed fingers incrementally getting better for moving in water.

I dunno, maybe there's an evolutionary path for wheels, but i don't think so.

How are you separating the efficiency of the architecture from the efficiency of the substrate? Unless you have a brain made of transistors or an LLM made of neurons how can you identify the source of the inefficiency?
You can't but the transistor-based approach is the inefficient one, and transistors are pretty good at efficiently doing logic, so either there's no possible efficient solution based on deterministic computation, or there's tremendous headroom.

I believe human and machine learning unify into a pretty straightforward model and this shows that what we're doing that ML doesn't can be copied across, and I don't think the substrate is that significant.

I’m looking forward to it! Inefficiency (if we mean energy efficiency) conceptually doesn’t bother me very much in that feels like Silicon design has a long way to go still, but I like the idea of looking at biology for both ideas and guidance.

Inefficiency in data input is also an interesting concept. It seems to me humans get more data in than even modern frontier models; if you use the gigabit/s estimates for sensory input. Care to elaborate on your thoughts?

I think the hard question is whether those wins can be realized with less effort than what we’re already doing, though.

What I mean is this: A brain today is obviously far more efficient at intelligence than our current approaches to AI. But a brain is a highly specialized chemical computer that evolved over hundreds of millions of years. That leaves a lot of room for inefficient and implausible strategies to play out! As long as wins are preserved, efficiency can improve this way anyway.

So the question is really, can we short cut that somehow?

It does seem like doing so would require a different approach. But so far all our other approaches to creating intelligence have been beaten by the big simple inefficient one. So it’s hard to see a path from here that doesn’t go that route.

Also, a brain evolved to be a stable compute platform in body that finds itself in many different temperature and energy regimes. And the brain can withstand and recover from some pretty severe damage. So I'd suspect an intelligence that is designed to run in a tighter temp/power envelope with no need for recovery or redundancy could be significantly more efficient than our brain.
The brain only operates in a very narrow temperature range too. 5 degrees C in either direction from 37 and you're in deep trouble.
Most brain damage would not be considered in the realm of what most people would consider "recoverable".
In some cases it doesn't recover even without physical or chemical damage. Psychiatric clinics are full of this stories.
(comment deleted)
Honest question: Given that the only wide consensus of anything approaching general intelligence are humans and that humans are biological systems that have evolved in physical reality, is there any arguments that better efficiency is even possible without relying on leveraging the nature of reality?

For example, analog computers can differentiate near instantly by leveraging the nature of electromagnetism and you can do very basic analogs of complex equations by just connecting containers of water together in certain (very specific) configurations. Are we sure that these optimizations to get us to AGI are possible without abusing the physical nature of the world? This is without even touching the hot mess that is quantum mechanics and its role in chemistry which in turn affects biology. I wouldn't put it past evolution to have stumbled upon some quantum mechanic that allowed for the emergence of general intelligence.

I'm super interested in anything discussing this but have very limited exposure to the literature in this space.

The advantage of artificial intelligence doesnt even need to be energy efficiency. We are pretty good at generating energy, if we had human level AI even if it used an order of magnitude more energy that humans use that would likely still be cheaper than a human.
Inference is already wasteful (compared to humans) but training is absurd. There's strong reason to believe we can do better (even prior to having figured out how).
That would mean with current resources AI can get so much more intelligent than humans, right? Aren't you scared?
That's a potential outcome of any increase in training efficiency.

Which we should expect, even from prior experience with any other AI breakthrough, where first we learn to do it and then we learn to do it efficiently.

E.g. Deep Blue in 1997 was IBM showing off a supercomputer, more than it was any kind of reasonably efficient algorithm, but those came over the next 20-30 years.

How does this idea compare to the rationale presented by Rich Sutton in The Bitter Lesson [0]? Shortly put, why do you think biological plausibility has significance?

[0] http://www.incompleteideas.net/IncIdeas/BitterLesson.html

I’m not GP, but I don’t think their position is necessarily in tension with leveraging computation. Not all FLOPs are equal, and furthermore FLOPs != Watts. In fact a much more efficient architecture might be that much more effective at leveraging computation than just burning a bigger pile of GPUs with the current transformer stack
I'll have to refer you to my forthcoming paper for the full argument, but basically, humans (and all animals) experience surprise and then we attribute that surprise to a cause, and then we update (learn).

In ANNs we backprop uniformly, so the error correction is distributed over the whole network. This is why LLM training is inefficient.

Not an insider but imo the work on diffusion language models like LLaDA is really exciting. It's pretty obvious that LLMs are good but they are pretty slow. And in a world where people want agents you want a lot of the time something that might not be that smart but is capable of going really fast + searches fast. You only need to solve search in a specific domain for most agents. You don't need to solve the entire knowledge of human history in a single set of weights