If you’re anything like me, you’ve heard of Meta. You’re also at least vaguely aware of Meta’s Frontier LLM called “” “Llama” which Meta markets as “open source” but for our purposes we’ll call this “open weight.” So, for our purposes here, what does “open weight” mean and how is it different from “open source”?
Open weight LLMs differ from open source in that they keep their training data and training code secret (if you want to dive deeper on this, the folks at opensource.org did a more detailed writeup). Open source, on the other hand, actually makes that training data and training code public (one example of an open source LLM is OLMo). As a side note, in case you’re curious, open source LLMs, in this definition, aren’t considered Frontier models.
Open weight LLMs (like Llama) do allow you to download and run them locally and allow for significant customization and fine tuning. There’s a lot to unpack in the words “significant customization and find tuning” and as such it’s out of scope for these short posts, but think of things like domain-specific training or removing or adding bias or restrictions. Also, we shouldn’t underestimate the ability to download and run locally to privacy- or localization-focused organizations.
And that brings us to the LLMs that have become household names in just the last few years, namely OpenAI’s GPT and Anthropic’s Claude. These are proprietary models, meaning you don’t download them to run locally, you can’t customize the models themselves (more on what you can do to enhance GPT or Claude in a future post), and you (or, more commonly, your application) interact with them via APIs.
I know this post is long but we have to tackle one more thing. What happens when a hyperscaler (some examples are Amazon AWS, Microsoft Azure) provides OpenAI or Anthropic’s models through offerings like Bedrock or AI Foundry? In these cases, the hyperscaler actually runs the model on their infrastructure, isolating your data from the LLM provider (e.g. Anthropic or OpenAI) and providing certain assurances of privacy and security. This is a common way for ISVs to access proprietary models for their applications.
I hope you’ve enjoyed this short post written by a human. Next time we’ll tackle Training vs Inference and Tokens!
You may also like
