LLaMA

From the side of which models they're running, we're seeing a really solid mix of US and Chinese trained models being used for local.

And we have the incredible models from, you know, the original Llama models of course, but also the Gemma models from DeepMind.

These are great choices for local.

But when it comes down to use cases, coding agents by and large are most effective with the large cloud models.

... and we all go through, you know, uh, a lot of the models go through phases where they sound more robotic, they sound more friendly, and a lot of that matters too.

Um, but I think the highest order bit is obviously making sure that you have a model that end to end you understand where the data's from, which is great from the Nemotron models, that you can go and introspect what made this model because if you're putting in a mission critical task, which people are absolutely using open models for mission critical tasks.

There's a post online about how Llama powers the analytics of a power plant to detect surges in Finland to make sure that the lights stay on.

That's where these models, the, uh, the model origin really matters.

But Meta AI, of course, is running an older LLM, and then I believe you get...

You, you still get thrown into a Llama model, I think, when you're on Instagram if you wind up in a, in a LLM chat.

And so there's going to need to be unification around all of these different products, and they're gonna need to eventually talk to all the same models.

But, uh, of course, there's, like, scaling and all sorts of different, uh, issues that you need.

But Meta AI, of course, is running an older LLM, and then I believe you get...

You, you still get thrown into a Llama model, I think, when you're on Instagram, if you wind up in a, in a LLM chat.

And so there's going to need to be unification around all of these different products, and they're gonna need to eventually talk to all the same models.

Um, but, uh, of course, there's like scaling and all sorts of different, uh, issues that you need because some of these products that they've gotten to scale are pretty significant, even, even if they're not like the most talked about.

The model, which is the brain file.

That could be something like Gemma, Llama, or Mistral.

The warehouse, which is where you find the model.

You might have heard of Hugging Face.

I think they're trying to get acquired right now at $13 billion.

That's what they do.

The software, which is what runs the model, that's something like LM Studio or Oh Llama.

And then the workflow, which is the product you're building around all of it.

I was surprised because actually, as you know, I like to stick my neck out with my predictions.

And one of the predictions that I got wrong was Llama.

You know, when Meta released Llama, oh my God, which one was it? I forgot.

But I think it was It was a really powerful Lama model that came out, I think, in 2024.

I mean, I'll look it up later, but we'll put it in the notes on this channel.

Grab the model, put it on your own hardware, and use it.

Now, Llama wasn't as smart as ChatGPT, but it wasn't stupid, and Meta kept on improving it.

Over the next two years, it got better and better, so that Llama 3 was far better than the original version of ChatGPT, which had also gotten heaps better.

And because Meta released their models and wrote papers about the architecture of their models, other researchers could follow in their path, in particular researchers in China.

The Chinese had a huge interest in artificial intelligence, and some of their best work came out of Alibaba.

And I'm not a miracle of size.

I'm a product of generosity.

Meta gave away Llama.

You can use DeepSeek.

You can use Llama, which is Meta's platform.

You can use Gemini.

Gemini is amazing.

And that was at the time where, like, all the big tech companies were getting very scared of losing their moats, right? Like, Google lo- feels like they're gonna lose search, or Facebook feels like people are gonna use social media less because AI will make the content worse, et cetera.

And so Facebook, of course, famously made Llama, right, the, the... one of the first large open source models.

There were a lot of smaller open source models, but they were just not very good.

They were, like, not competitive.

But once Llama launched, there was this whole cottage industry in 2023 of making, like, fine-tuned versions of Llama, like Llama for coding, 'cause at that time the coding models weren't very good, or Llama for images to compete with...

Uh, like, like, you know, people would make open source versions of Midjourney, like, stuff like that, right? So there were starting to be these, like, flora and fauna of, like, special purpose models built off Llama, and it was really hard for you to know, I wanna make an image of a seed with hair, right? Like, that's my prompt.

And it's like, okay, which one do I go to? And I think in the closed source world, right, you, you just get one prompt, right? It's like OpenAI gives you one thing, Claude gives you one.

In the open source world, it's very unbundled.