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

GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

threads.net

Post reply on HN
Print view View original

111–120 of 134 posts

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#111

Post by gradascent »

krasin wrote: post #78

Finetuning LLMs is currently the most promising way for next-gen robotics. One of such works (PaLM-e) among other things measured the impact of finetuning on general purpose tasks: https://twitter.com/DannyDriess/status/1632904698108821505 In short, an 8B model could degrade almost 10x after being finetuned on robotics tasks, while 500B model experiences a very minor degradation (~4%) and there's a hope that with a l…

Interesting! I would like to learn more about how AI is being applied to robotics. Do you have any suggestions for how to keep up with developments/ideas in this field?

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#112

Post by krasin »

gradascent wrote: post #111
krasin wrote: post #78

Finetuning LLMs is currently the most promising way for next-gen robotics. One of such works (PaLM-e) among other things measured the impact of finetuning on general purpose tasks: https://twitter.com/DannyDriess/status/1632904698108821505 In short, an 8B model could degrade almost 10x after being finetuned on robotics tasks, while 500B model experiences a very minor degradation (~4%) and there's a hope that with a l…

Interesting! I would like to learn more about how AI is being applied to robotics. Do you have any suggestions for how to keep up with developments/ideas in this field?

These two links could be a good start:

ALOHA-2: https://aloha-2.github.io/

RT-X: https://robotics-transformer-x.github.io/

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#113

Post by simonw »

danielmarkbruce wrote: post #108
simonw wrote: post #105

Earlier quoted context omitted.

For "citation needed", read "please link me to a blog post showing how, don't just tell me to Google for one". The internet is full of blog posts about this. That doesn't mean they're actually good - I'd love to be pointed at one that has proven itself useful for someone (and definitely isn't just LLM blog-spam). I don't care if it's trivial to fine-tune and get crap results - I care about fine-tuning where the resul…

It's an internet forum, not an academic journal. Water tight arguments are not needed. If one wants to call bs, they can just do it, no need to dance around the topic by asking for a citation.

OK, I call BS. Fine-tuning an LLM is not "trivial" - especially if you want to get useful results, as opposed to just being able to say "look, I fine-tuned an LLM".

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#114

Post by simonw »

redox99 wrote: post #87
LASR wrote: post #42

I lead AI teams at my company. I've advised leadership against any kind of training / fine-tuning anything. We're not in the business of training models. We will never be as good as OpenAI / Anthropic etc. Where the real value in applications is smarter prompting techniques and RAG. There is a lot of room at the bottom in doing "dumb" things and simply feeding models with the right context to deliver customer value.

That's a pretty odd stance. I've finetuned llama/mistral models that greatly outperform GPT4 with just a prompt. You have to know when to RAG, finetune, or RAG+finetune.

"I've finetuned llama/mistral models that greatly outperform GPT4 with just a prompt"

If you write about your experiments with that in detail I guarantee you'll get a lot of interest. The community is crying out for good, well documented, replicable examples of this kind of thing.

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#115

Post by smallnamespace »

qeternity wrote: post #69
smallnamespace wrote: post #50

Earlier quoted context omitted.

You definitely can't in the general case (for example, your 7B model is never going to be able to help much with coding, fine tuning or no). It can make sense if you have a particularly simple use case.

By definition you wouldn’t fine tune a 7B model to be generally as good at GPT4. You would just be trying to overfit some small amount of functionality in a narrow domain.

Yes but from the context of this discussion, we’re trying to figure out the “sweet spot” model size where it’s worth attempting fine tuning. My guess is it’s only worthwhile for matching simple tasks with small models, and any sufficiently complicated task it’s better to do few/zero shot instead.

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#116

Post by potatoman22 »

bernawil wrote: post #106
potatoman22 wrote: post #32

Earlier quoted context omitted.

One major advantage of FinGPT or Bloomberg's LLM is that the embeddings produced by the model can be used for downstream prediction tasks. GPT-4 does not expose its embeddings so it cannot be used for this.

sorry, noob here trying to make sense of this: you mean you can extract embeddings from the model file or that the embeddings are available in the repo and you can just use those files?

Kind of. You feed the LLM the input text for your prediction, you extract the activations of the final layer of the LLM (so the weights * the input of the previous layers), then use that activation vector, or embedding, as the input for a separate model. This separate model that uses the embedding can be any classifier or regression. A common use case for this is document classification.

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#117

Post by port443 »

pbronez wrote: post #104
zapperdulchen wrote: post #40

Earlier quoted context omitted.

Seems like the bitter lesson is still right: http://www.incompleteideas.net/IncIdeas/BitterLesson.html

I can’t access the article there… SSL error and then timeout. Here’s a link to the most recent WayBackMachine snapshot: https://web.archive.org/web/20240321091803/https://www.incom...

There's no SSL at all on that site, since it's http not https. Your browser is breaking the link.

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#118

Post by Implicated »

simonw wrote: post #75
hallqv wrote: post #67

This discussion is so dumb - finetuning a base model costs ~$1 with LORA/QLORA and can yield same performance as gpt-4, but at 1/100 of the cost per token. What Bloomberg did for $10M was not finetuning..

"finetuning a base model costs ~$1 with LORA/QLORA and can yield same performance as gpt-4, but at 1/100 of the cost per token" That's a big claim - can you back that up with any examples?

I had opened a new tab back when this comment was just a few minutes old in hopes that when I came back there was some really great blog post linked with the details on the sorcery.

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#119

Post by redox99 »

simonw wrote: post #114
redox99 wrote: post #87

Earlier quoted context omitted.

That's a pretty odd stance. I've finetuned llama/mistral models that greatly outperform GPT4 with just a prompt. You have to know when to RAG, finetune, or RAG+finetune.

"I've finetuned llama/mistral models that greatly outperform GPT4 with just a prompt" If you write about your experiments with that in detail I guarantee you'll get a lot of interest. The community is crying out for good, well documented, replicable examples of this kind of thing.

I'm so behind in this area. I had finetuned a model that was SOTA and worth publishing about in October, but procrastinated. I'm scared to check if somebody else already published on this topic.

Re: GPT-4, without specialized training, beat a GPT-3.5 class model that cost $10M

#120

Post by hlfshell »

gradascent wrote: post #111
krasin wrote: post #78

Finetuning LLMs is currently the most promising way for next-gen robotics. One of such works (PaLM-e) among other things measured the impact of finetuning on general purpose tasks: https://twitter.com/DannyDriess/status/1632904698108821505 In short, an 8B model could degrade almost 10x after being finetuned on robotics tasks, while 500B model experiences a very minor degradation (~4%) and there's a hope that with a l…

Interesting! I would like to learn more about how AI is being applied to robotics. Do you have any suggestions for how to keep up with developments/ideas in this field?

In October I wrote a blogpost on this subject: https://hlfshell.ai/posts/llms-and-robotics-papers-2023/

..and plan to do an updated version soon for much of what's been released since. I've also done work related to LLM and robotics integration, also on that site.

Happy to chat about it.

Post reply on HN