Why Llama 3.1 is Important | Hacker News Reader

Why Llama 3.1 is Important

timkellogg.me

56 pointstkellogg23 comments

Stating independence as the advantage of Llama 3.1 is a bit funny. Without the huge amount of computational resources from Meta, Llama 3.1 won't be possible. We are still dependent on certain big companys' "good" willness to be able to enjoy the benefits of open source.

May as well enjoy it while it lasts, but yeah it echoes the precarious state that VR is currently in, where Meta is more or less propping up the entire consumer market with their willingness to burn billions of dollars on Reality Lab every quarter, and anyone investing in VR development just has to hope that rug doesn't get pulled from under them any time soon. Llama 3.1 isn't going away of course, but how much are you willing on gamble on Llama 3.1 possibly being the last SOTA model to be released this freely?

> [...] the precarious state that VR is currently in, where Meta is more or less propping up the entire consumer market with their willingness to burn billions of dollars on Reality Lab every quarter,

They don't quite seem so happy to do that for much longer.

https://www.businessinsider.com/mark-zuckerberg-cuts-spendin...

They could buy a lot of H100s for the roughly $4 billion they dump into VR/Metaverse each quarter...

They are complimentary technologies. Think holodeck.

For the release, yes. But from now onwards everyone will benefit from Llama 3.1 regardless of Meta's involvement. That's the advantage; I can build something with Llama 3.1 and it will keep working indefinitely without any dependence on another party.

Yeah, that's a good point. On the other hand, the model is out there, irrevocably so, and you can use knowledge distillation techniques to train other weaker models. So yeah, we're not fully independent, but the situation is significantly improved.

eh synthetic data sets based on llama 3.1 will train some fun models

I’ve always wondered if it’s possible to pool together small amounts of resources from users e.g. similar to torrenting with networking, but with compute and use it to host a large model like this one. If the community wants open source to be the way, I believe this would be the only way to get these large models accessible to users while the community stay in semi-control of the platform.

Edit: would also work with training I'm assuming?

I just got Llama 3.1 GGUFs working on my Mac laptop with a new plugin for my LLM CLI tool: https://llm.datasette.io/

Here's information on the new plugin: https://simonwillison.net/2024/Jul/23/llm-gguf/

Once you've installed LLM ("brew install llm" or "pipx install llm" or "pip install llm") you can try the new plugin like this:

    llm install llm-gguf
    llm gguf download-model \
      https://huggingface.co/lmstudio-community/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/main/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf \
      --alias llama-3.1-8b-instruct --alias l31i

    llm -m l31i "five great names for a pet lemur"

This is using the GGUF version of Llama 3.1 8B Instruct from here: https://huggingface.co/lmstudio-community/Meta-Llama-3.1-8B-...

Simon, when do you sleep?

Does the community license let companies fine-tune it or retrain it for their use cases?

There are significant restrictions on it so it's not fully open-source, but maybe it's only a real problem for Google and OpenAI and Microsoft.

Open source has turned into a game of, what's the most commercial value I can retain, while still calling it open-source and benefiting from the trust and marketing value of the 'open source' branding.

The last section is the most important. There’s a massive difference between what you can do with the text output of an LLM versus being able to know and play with the individual weights, depending on your use case.

Written by llama3 or chatgpt?

I'll choose to take that as a compliment

nice writeup! and witty reparté!

Excited about this - though probably more the 70B than 405B because its also really good & will be accessible for cheap & in bulk.

btw pretty sure nobody is creating adapters for a 405B with a laptop and a weekend ;)

haha yeah i was chuckling at that too

I used to think it was cheaper. But according to https://llama.meta.com/ GPT-4o Mini is actually cheaper most of the time.

The Llama models tend to spark a race-to-the-bottom in terms of pricing - I fully expect Llama 3.1 to end up cheaper than GPT-4o mini after a bunch of providers spin up endpoints that compete on price.

they dropped the price because of the llama release

This is competing with 4o not 4o mini, right?

The 405B model does. It's a bit more competitive on price but not that much better.

Reply on news.ycombinator.com