Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

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

news.ycombinator.com

Post reply on HN
Print view View original

291–300 of 343 posts

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

#291

Post by mft_ »

chriskanan wrote: post #259
Nimitz14 wrote: post #254

Earlier quoted context omitted.

Any continual learning papers you're a fan of?

Depends on what angle you are interested in. If you are interested in continual learning for something like mitigating model drift such that a model can stay up-to-date where the goal is attain speed ups during training see these works: Compared to other methods for continual learning on ImageNet-1K, SIESTA requires 7x-60x less compute than other methods and achieves the same performance as a model trained in an offl…

> working on continual learning to explain memory consolidation works in mammals and why we have REM phases during sleep

That's a nice model: human short-term memory is akin to the context window, and REM sleep consolidating longer-term memories is akin to updating the model itself.

How difficult would it be to perform limited focused re-training based on what's been learnt (e.g. new information, new connections, corrections of errors, etc.) within a context window?

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

#292

Post by CamperBob2 »

nextts wrote: post #164
throw310822 wrote: post #159

Earlier quoted context omitted.

I don't get it. 1) Yes it's true, learning from text is very hard. But LLMs are multimodal now. 2) That "size of a lion" paper is from 2019, which is a geological era from now. The SOTA was GPT2 which was barely able to spit out coherent text. 3) Have you tried asking a mouse to play chess or reason its way through some physics problem or to write some code? I'm really curious in which benchmark are mice surpassing c…

Mice can survive, forage, reproduce. Reproduce a mammal. There is a whole load of capability not available in an LLM. An LLM is essentially a search over a compressed dataset with a tiny bit of reasoning as emergent behaviour. Because it is a parrot that is why you get "hallucinations". The search failed (like when you get a bad result in Google) or the lossy compression failed or it's reasoning failed. Obviously the…

Mice can survive, forage, reproduce. Reproduce a mammal. There is a whole load of capability not available in an LLM.

And if it stood for "Large Literal Mouse", that might be a meaningful point. The subject is artificial intelligence, and a brief glance at your newspaper, TV, or nearest window will remind you that it doesn't take intelligence to survive, forage, or reproduce.

The mouse comparison is absurd. You might as well criticize an LLM for being bad at putting out a fire, fixing a flat, or holding a door open.

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

#293

Post by Lerc »

kerkeslager wrote: post #285
Lerc wrote: post #240

Earlier quoted context omitted.

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…

>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.

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

#294

Post by calmbell »

Matthyze wrote: post #11
jawiggins wrote: post #4

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…

> 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.

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.

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

#295

Post by eximius »

randomNumber7 wrote: post #229
eximius wrote: post #95

I believe that so long as weights are fixed at inference time, we'll be at a dead end. Will Titans be sufficiently "neuroplastic" to escape that? Maybe, I'm not sure. Ultimately, I think an architecture around "looping" where the model outputs are both some form of "self update" and "optional actionality" such that interacting with the model is more "sampling from a thought space" will be required.

Why, even animals sleep? And if you for example learn an instrument you will notice that a lot of the learning of the muscel memory happens during sleep.

I guess you're saying that non-inference time training can be that "sleep period"?

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

#296

Post by chongli »

mietek wrote: post #288
chongli wrote: post #283

Earlier quoted context omitted.

Even if you’re a pure Dennettian functionalist you still commit to a functional difference between signals in transit (or at rest) and signals being processed and interpreted. Holding a cassette tape with a recording of a symphony is not the same as hearing the symphony. Applying this case to AI gives rise to the Chinese Room argument. LLMs’ propensity for hallucinations invite this comparison.

Are LLMs having subjective experiences? Surely not. But if you claim that human subjective experiences are not the result of electrical signals in the brain, then what exactly is your position? Dualism? Personally, I think the Chinese room argument is invalid. In order for the person in the room to respond to any possible query by looking up the query in a book, the book would need to be infinite and therefore imposs…

The Chinese Room is a perfect analogy for what's going on with LLMs. The book is not infinite, it's flawed. And that's the point: we keep bumping into the rough edges of LLMs with their hallucinations and faulty reasoning because the book can never be complete. Thus we keep getting responses that make us realize the LLM is not intelligent and has no idea what it's saying.

The only part where the book analogy falls down has to do with the technical implementation of LLMs, with their tokenization and their vast sets of weights. But that is merely an encoding for the training data. Books can be encoded similarly by using traditional compression algorithms (like LZMA).

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

#297

Post by kryogen1c »

sturza wrote: post #289
kryogen1c wrote: post #279

Earlier quoted context omitted.

> You can encode anything into the "language Im just a layman here, but i don't think this is true. Language is an abstraction, an interpreative mechanism of reality. A reproduction of reality, like a picture, by definition holds more information than it's abstraction does.

A picture is also an abstraction. If you take a picture of a tree, you have more details than the word "tree". What i think the parent is saying, is that all the information in a picture of a tree can be encoded in language, for example a description of a tree, using words. Both are abstractions but if you describe the tree well enough with text(and comprehend the description) it might have the same "value" as a pict…

> all the information in a picture of a tree can be encoded in language

What words would you write that would as uniquely identify this tree from any other tree in the world, like a picture would?

Now repeat for everything in the picture, like the time of day, weather, dirt on the ground, etc.

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

#299

Post by inimino »

Etheryte wrote: post #125
inimino wrote: post #75

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.

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'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.

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

#300

Post by inimino »

rsfern wrote: post #129
Etheryte wrote: post #125

Earlier quoted context omitted.

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

Right.
Post reply on HN