Hacker News Reader

Comment by mattnewton | Hacker News Reader

It’s an apples to oranges comparison- humans are getting massive amounts of data from their nervous system far richer than the token stream of Wikipedia. The model has to learn everything about the world solely from statistical patterns of text whereas humans are seeing, hearing and touching the world in full resolution. Humans also take an active role in seaking novel stimuli that they can learn from.

But that does I think shed light on why multi-modality is so promising; beyond additional use cases, it could make the models better at existing text tasks too when they train on all that video data.

And I think it also reinforces why everyone is so hyped for reinforcement-learning techniques like the self play that alpha zero used to come LLMs.

View on HN

While that could be an explanation I am not fully sold on that idea. For example we don't get continuous streams of information on something like abstract math but we still dona decent job with it without going through millions of math problems.

If multi modality gets us through this phase then you are right in your analysis. Let's see what come out in the next 5 years

I don’t mean to say that is _all_ that is separating us from the transformer architecture. I do think we’re more efficient than gradient descent, and sgd+attention is a function of how modern parallel processing works more than it is a function of any deep insight into the brain’s computation or learning methods.

I just think we’re really making a silly comparison when we compare a stream of text to the multi sensory experiences humans have directly, reducing what we are doing to just reading words off a screen or page.

> humans are getting massive amounts of data from their nervous system far richer than the token stream of Wikipedia.

And their tokenizers are better - that's the primary reason why ChatGPT is so bad at math, its network literally doesn't (and can't!) comprehend numbers longer than the token size by definition.

Additionally, humans can cross-reference information with information they already know at learning time. For example, an AI "reads" some climate change denier politician's speech - how is it supposed to know that the politician is a climate change denier and downrank their opinion on climate change as a result, at least without a human having pre-classified the data?

Humans tokenize numbers too. Which is easier to read? 100000000 or 100,000,000? You only need tokens for 000 to 999

Yes we do, but we have the contextual knowledge available as well, so we know that the "100","000","000" are all belonging to one "supertoken".

That’s exactly what transformers do. They parse things in context. If it can parse C++ well enough to write C++ code, then it can parse thousand-separated numbers.

sure, you can get the model to understand with enough examples, but OP of this thread was saying LLMs need many many more steps than humans do. Mschuster was saying this could partially be an artifact of the tokenization - humans visually seeing 1,000,000 and 1000001 for the first time can see that they are similar based on how they look. A transformer must learn two token sequences, which may not have any tokens in common, are different by 1.* And it must do this based entirely on context and how likely one is to appear in the same place as another. That takes a massive amount of data in practice and could be one reason why humans appear to learn from much sparser data and faster.

* in gpt-2’s case, you get 1,000,000 -> 16, 11, 830, 11, 830, 198, 49 And 1000001 -> 388, 486 which means the model will see a sequence of noise embeddings entirely different when training. That means seeing 1,000,000 any number of times will not really help the model prepare for when it sees 1000001 for the first time. You can play around with https://tiktokenizer.vercel.app/?model=gpt2 to see how tokenization of numbers has been improved by different models to try and mitigate this.

> And it must do this based entirely on context and how likely one is to appear in the same place as another. That takes a massive amount of data in practice and could be one reason why humans appear to learn from much sparser data and faster.

That, and another thing: A LLM fundamentally boils down to a statistical repetition engine. That means, if it has never seen the input - say, some random large numbers -, it will hallucinate garbage if you ask it a question ("multiply 245358943543548 by 438574589675486789654"), whereas a human would recognize where the boundaries of the numbers are and plug it into a calculator (or, some particularly big brained people would actually do it in their mind) to figure out the result.

Some AI or pseudo-AI (dunno what Wolfram Alpha uses under the hood) has to have a "front" AI to recognize such questions and redirect them to a calculator, but in my eyes that's just cheating for a very specific case and not enabling general intelligence.

I don't buy this explanation, yes we are getting "constant data", but most of that is completely useless like, traffic noise, looking at the wall etc, and most of it isn't novel at all.

I.e. of the GB of data I get every day, most of it is visual data of my apartment which I've already seen 10000 times and audio data of the same shit my girlfriend says every day. I'm not ingesting the whole of Wikipedia or every book written, which is far richer and more varied.

> the same shit my girlfriend says every day.

Have you tried increasing the temperature?

On the richness of the data- When you were a child every time you looked at any wall, how many wikipedia articles would it have taken to describe that entire experience? The model needs to discover what is “noise” by rereading something as sparse as Wikipedia over and over. Think of it as understanding films purely from the script. Or learning what sand is and how it behaves physically from Wikipedia rather than picking it up and having it run through your fingers.

The experience of breathing and staring at a wall has an incredible amount of data that has to be learned before it is noise and is boring, and I’d argue it would take many Wikipedia articles to describe it fully.

And on the novelty function- SGD does update things based on a loss and how unexpected things were to the model. But that’s very crude compared to what humans are doing, where everything gets the same learning rate multiple, and we feed nearly random data to the model. There is a lot of evidence that if you pick your training set carefully you can train a lot faster, and I think humans are picking their own training set.

Humans no longer have the big picture. Our knowledge is being atomised because there's so much of it. You can see this in medicine: a cardiologist spends all of his time and effort on a single organ. In the 19th century it was possible for a doctor to keep up with the entire field.

That's depth-versus-breadth anxiety. Is the body of human knowledge, everything there is to know, losing deep meaning due to becoming too broad, making it impossible to grasp everything? No. Consolidation happens. Knowledge as a whole should appear deep, not broad, you're just looking at it from the wrong end.

Humans also remember and update their model in real-time (with various caveats obviously). LLM’s forget when you close the window and need 1000 examples before they learn anything.

Reply on news.ycombinator.com