Hacker News Reader

Comment by abraxas | Hacker News Reader

I have a hard time imagining anything that comes even close to ChatGPT being able to run on consumer hardware in the next couple of years

View on HN

That's what I thought 2 weeks ago. I figured it'd be ~5 years before I could do anything at home. But already people have the leaked facebook llama weights running on CPU w/under 32 GB of system ram doing a token a second or so.

that llama is likely much smaller than chatgtp

The whole point of the LLaMa paper is that large models are undertrained and oversized.

not sure where did you get it, but they trained 65b llama too, which outperformed llama 7b on their benchmark.

Here is the paper and synopsis:

https://arxiv.org/abs/2302.13971

> We introduce LLaMA, a collection of foundation language models ranging from 7B to 65B parameters. We train our models on trillions of tokens, and show that it is possible to train state-of-the-art models using publicly available datasets exclusively, without resorting to proprietary and inaccessible datasets. In particular, LLaMA-13B outperforms GPT-3 (175B) on most benchmarks, and LLaMA-65B is competitive with the best models, Chinchilla-70B and PaLM-540B. We release all our models to the research community.

what's your point? If you read actual paper you will see that 65B outperforms that 7B which was discussed above, and size matters.

I don’t think anyone said size doesn’t matter. The point is that by training for longer, models with fewer parameters will perform similarly to GPT-3 models with more parameters.

probably not if gap is so large: 7B vs 175B.

Also, all those benchmark are trash because they can't track data leaks in training data. For example they trained llama on github, where GSM8k eval data is located, of course model will perform well on GSM8K, because it memorized answers.

You don’t have to guess, the information is in the provided synopsis.

I’m sure there are issues similar to your description. Nevertheless, you seem to be a staunch defender of GPT-3, which to me indicates some kind of bias? Like, who cares if LLaMA is better - in fact, isn’t that a good indicator of progress?

> You don’t have to guess, the information is in the provided synopsis.

yes, I checked benchmarks in paper, and there are many where gpt won over 7b llama. Also, it is not clean experiment, because models were trained on different datasets.

> I’m sure there are issues similar to your description. Nevertheless, you seem to be a staunch defender of GPT-3, which to me indicates some kind of bias? Like, who cares if LLaMA is better - in fact, isn’t that a good indicator of progress?

personality rants have been ignored.

:shrugging:

Okay, then.

You can run llama-30b right now on high-end consumer hardware (RTX 3090+) using int4 quantization. With two GPUs, llama-65b is within reach. And even 30b is surprisingly good, although it's clearly not as well trained as ChatGPT specifically for dialog-like task setting.

I can certainly imagine it after seeing https://github.com/ggerganov/llama.cpp

Still a couple years out but moving way faster than I would have expected.

ChatGPT has 175B weights if I'm not mistaken. Llama 7B would not be in any way comparable.

For people who think the number of parameters determines LLM coherence, well, that's a good rule of thumb. But there's an optimal training set data size to parameters count and gpt3 was trained on too little data. The LLM coming out now trained on more data with fewer parameters and achieve something close.

Sure, the 7 billion parameter can't do long outputs. But the 13 billion one is not too bad. They're not a full replacement by any means but for many use cases a local service that is stupider is far preferable to a paid cloud service.

One finding in the LLaMA paper [1] is that our current large models are undertrained. LLaMA with 13B params outperforms GPT-3 175B (not ChatGPT), but an "instruct" version of LLaMA was finetuned over the 65B model and did quite well.

[1] https://arxiv.org/pdf/2302.13971.pdf

It's crazy, but it seems to be happening already. Granted, that's probably still a far-cry from Chat-GPT, but it seems inevitable a few years down the line.

https://news.ycombinator.com/item?id=35100086

Moore's law isn't quite beaten yet, so the (hypothetical future) RTX 5090 and 6090 is gonna be insane. Combined with software optimization and refinement of the techniques, along with training != inference, means I think we'll see something better, runnable locally, in a couple of years. The leaps and bounds Stable Diffusion has gone is insane. Facebook's LLaMA is also seeing a similar growth from just having the model available.

I hear that disco music is also experiencing tremendous growth, and is poised to make a comeback any day now.

I could perceive something like in 1 or less.

M3 Macbook with eGPU functionality restored in conjunction with more efficient programming would mean having enough memory available to all the processors. This would definitely count as consumer hardware.

Custom built GPU-like devices with tons of RAM could become vogue. Kind of like the Nvidia A100 but even more purpose built for running LLMs or whatever models come next.

Reply on news.ycombinator.com