Meta aims for AI dominance with the release of Llama 3.1

Meta has announced Llama 3.1, the most advanced open source Large Language Model AI to date, but EU users will have to wait.

Llama 3.1 405B is the largest and most advanced open source AI model released to date, and Meta claims it eclipses current giants such as ChatGPT 4o and Claude 3.5 Sonnet in performance benchmarks.

The release hasn't come as a surprise, since Meta has been teasing its existence for several months now. What's significant here though, aside from its performance, is the business model surrounding it. While ChapGPT and Claude 3.5 Sonnet are privately developed models, Meta hopes its open source approach will lead to faster development, lower cost, and lower resource requirements. In fact, Meta is claiming that its approach costs roughly half that of rivals to run.

Indeed, Mark Zuckerberg is comparing the development of AI to that of Linux, which as we know, pretty much everybody has installed on their computer <cough>. Joking aside, Zuckerberg points out in a blog post related to the release that Linux is the bedrock for modern cloud computer and mobile device operating systems, and its open source nature is why it became such a secure, trusted and reliable system to use.

He went on to make the point that Llama 2 was only comparable to older AI models last year, and that open source development has lead to Llama 3.1 now being comparable to, and in some case better than, some of the leading proprietary LLM AI systems in record time.

Benchmarks comparing Llama 3.1 405B to rival AI systems. Images: Meta.

Having an open source AI system would appear to make much more sense for developers who don't want to be locked into a proprietary system that could change at any given time, or can't be trained on custom data. The reduced running costs, too are a notable advantage in favour of the Llama system.

How safe is Llama 3.1?

Zuckerberg goes on to detail a lot about the safety of AI systems in terms of potential intentional harm, data integrity and security. Most would be forgiven for scoffing, given the proliferation of disinformation that has propagated throughout the likes of Facebook over the years, along with accusations of biased algorithms. But, regardless of who is making these points, it would be difficult to argue against the idea of an open source system vs a private proprietary one in this regard. The fact is that open source software is much more open to scrutiny than a closed proprietary system.

Meta has been red teaming Llama 3.1 both with internal and external experts to stress test the system to find out any unexpected use case scenarios that could be used for nefarious purposes. It is also sharing model weights, recipes, and safety tools, as well as including components such as Llama Guard, which is a multilingual safety model and Prompt Guard, a prompt injection filter.

Of course, despite all the talk of transparency, nothing specific has been said about precisely what data Llama 3.1 has been trained on. Although in terms of stats it apparently now has 405-billion parameters and was trained on 16,000 Nvidia H100 GPUs on over 15 trillion tokens. This works out at around the $400,000,000 mark for the processors alone, putting aside any Walmart volume discount Meta might have negotiated. At least we know now who has been hogging all the H100s recently.

Meta is working with over 25 partners, including Amazon, Microsoft, Google, Nvidia, and Databricks to bring Llama 3.1 out into the world, and users will be able to try it in chatbot form initially in WhatsApp and the Meta AI website in the US. The important key words there being "in the US".

Unfortunately, just as with Apple Intelligence, Llama 3.1 will not be rolling out in EU countries any time soon. Whilst Apple Intelligence has fallen foul of the Digital Markets Act, Llama rollout in the EU is being restricted by the GDPR act, which is designed to protect the data of internet users in the EU, and to give them control over how that data is used and stored. This isn't to say EU users will never gain access to Llama 3.1 or Apple Intelligence, but that they may have to wait a while as Apple and Meta negotiate and develop solutions to the problem.

In terms of overall usage, Meta sees Llama 3.1 being used to train smaller, less resource intensive AI systems that can be rolled out more cost effectively. Given the overall clout of Meta, along with the open source nature of Llama, it's difficult to see how it could become anything other than the industry standard AI system. And, of course, that brings questions of its own, despite the purported open and transparent nature of it all.

tl;dr cheat sheet

• Meta has released Llama 3.1 405B. The open source approach aims to accelerate development, reduce costs, and lower resource requirements. 
• Mark Zuckerberg has likened the development of Llama 3.1 to that of Linux, emphasizing the benefits of an open source system. This has raised eyebrows.
• The model will be initially available in chatbot form in WhatsApp and the Meta AI website in the US. However, the rollout of Llama 3.1 in the EU is restricted by the GDPR, designed to protect user data.
• Meta aims for Llama 3.1 to be used to train smaller, less resource-intensive AI systems and hopes that it will become the industry standard AI system.