%201.png)
Meta casts Llama 3 and reinforces his AI assistant, Meta AI. The objective is clear: beat ChatGPT.
A year ago, the company created by Mark Zuckerberg did not seem particularly interested in the field of artificial intelligence. The metaverse seemed to focus all his attention, but in reality his AI division was hard at work. In recent months those efforts have borne fruit, and today the company has made an announcement that may put it at the forefront in the popular segment of large language models in which OpenAI, Microsoft or Google also compete.
With the flame
The first version of LlaMA, released in February 2023, went somewhat unnoticed, but just the opposite happened in July 2023, when Meta engineers introduced LlaMA 2 (later spelled Llama 2). The company differentiated itself from the competition by using an Open Source model—although it is not entirely open source—and became a pillar for a large number of parallel developments based on that model. In fact, it was also especially interesting because anyone could download it to have their "own ChatGPT" running from their PC.
Mark Zuckerberg announces Meta's new big goal: creating artificial general intelligence
Llama 3 arrives. Meta's efforts have been increasing and now they present Llama 3, the latest major version of their foundational "Open Source" model. According to the company, Llama 3 outperforms other similar models in various benchmarks and stands out especially as a programming assistant. The "Open Source" quotes are important: the development uses a "community license" which, as in previous cases, is not exactly the same—although it is quite similar—to Open Source.
Two initial versions. Llama 3 is currently available in two versions, 8B and 70B with 8,000 and 70,000 million parameters respectively. Both can be downloaded from Meta with prior registration, and both versions are in turn available in two versions: a pre-trained one (the raw model that predicts the next token to be generated) and a polished one to follow the users' instructions. Both have a context limit of 8,192 tokens, modest when compared to proposals such as Claude 3 Opus, which supports context windows of 200,000 tokens. The model can generate images in addition to text.
Voracious training
Meta trained both models on a gigantic cluster with 24,000 GPUs. In the case of the 70B model, Zuckerberg commented in an interview with Dwarkesh Patel that this training could have continued and that variant could have been better, but they preferred to move on to training new developments and offer that 70B version that despite everything was not "saturated " for the training after "ingesting" 15 billion tokens. "We probably could have fed it more tokens and it would have become a little better," Zuckerberg explained.
A gigantic multimodal 400B model in sight. In addition to the 8B and 70B Llama 3 models, Meta is preparing a spectacular 400B version with 400,000 million parameters that according to some experts will be on par with GPT-4 Turbo, Claude 3 Opus or Gemini Ultra and various benchmarks. That model will support multimodal input and output, that is, both text and images.
Meta brags about performance
Although it is difficult to know if one AI model is really better than another, Meta has provided the results of the behavior of its 8B and 70B Llama 3 models in various synthetic performance tests, and in almost all of them these models outperform comparable versions. such as Mistral 7B, Gemini Pro 1.5 or Claude 3 Sonnet.
Meta AI is reinforced with Llama 3 and launches a web version. The first big beneficiary of this launch is the company's chatbot, called Meta AI. This development was launched in September 2023, and debuted as an integrated assistant in WhatsApp, Facebook, or Instagram. Now the web version arrives at www.meta.ai, and both in it and in those integrated assistants the chatbot is based on Llama 3.
The new Meta AI assistant already has a web version, and its operation is modeled on other rivals such as ChatGPT, Copilot, Gemini or Claude.
At the moment Meta AI is not in Spain (or in Spanish). It is possible to use the chatbot on the web without registration, but for example for some options such as generating images it will be necessary to log in with a Facebook account. Meta.ai is available in English in the US and a dozen other countries, but as Meta says, "we're just getting started" and the option is expected to soon expand to other regions. At the moment it is not available in Spain, and in our brief tests - it can be accessed with VPN, as was the case with other similar launches in recent months - we were able to verify that its behavior was comparable to that of ChatGPT.
The web version answered our brief questions about the World Cup champions quite accurately, although it made a few slips. France, for example, was not runner-up in 1986 (it was Germany).
Objective: beat ChatGPT (and the rest). The AI chatbot segment is more animated now than ever and we have before us a new and interesting proposal that is more precisely so because of its "almost" Open Source philosophy and because it can be used as a basis for parallel developments. The possibility of having a web version to use both from the PC and from the mobile - and not only as an assistant integrated into the Meta apps - is the other highlight of a launch that increases the pressure against ChatGPT, an absolute benchmark in this market but little by little it is being cornered by alternatives that are improving by leaps and bounds.
By Javier Pastor 3.