Hacker News Reader

Comment by feanaro | Hacker News Reader

Isn't it the case that we literally have no clue how GPT4 and GPT3.5 are different in terms of training, given OpenAI doesn't want to disclose anything at all?

View on HN

It's not true we know nothing. We know a little bit by using the two models from their API. Given the time per inference and the limit on messages per day for GPT4, I'm willing to bet it's doing around 10x more compute than GPT3.5. If that's because it has 10x more weights, I don't know. But it wouldn't be a terrible guess.

So your estimate is that GPT4 has 1.75 trillion weights?

Is there anything that affects inference compute time besides the number of parameters? Assuming same hardware, etc.

Yes - for example adding memory to the attention mechanism (similar to RETRO or Memorizing Transformers paper)

We don' have the details, it is true. But empirically and based on their report gpt-4 is notably better than chatgpt.

Better, yes, and for that we have evidence. But is the improvement stemming simply from even more data? That's what I'm questioning.

It's speculated it has same number of parameters, but more compute and is multi modal.

This paper is pretty approachable and goes over the "scaling laws" in detail: https://arxiv.org/abs/2206.07682

In short, yes. More data, higher quality data, more epochs on the data. That is the name of the game.

That paper doesn't discuss GPT-4 at all. It does however contain this interesting excerpt (emphasis mine):

> Although we may observe an emergent ability to occur at a certain scale, it is possible that the ability could be later achieved at a smaller scale—in other words, model scale is not the singular factor for unlocking an emergent ability. As the science of training large language models progresses, certain abilities may be unlocked for smaller models with new architectures, higher-quality data, or improved training procedures. For example, there are 14 BIG-Bench tasks5 for which LaMDA 137B and GPT-3 175B models perform at near-random, but PaLM 62B in fact achieves above-random performance, despite having fewer model parameters and training FLOPs.

So it's not obvious that it should be so straightforward.

Reply on news.ycombinator.com