Last April, Meta announced that it was working on a first in the AI industry: an open-source model whose performance matched the best private models from companies like OpenAI.
Today, that model has arrived. Meta is releasing Llama 3.1, the largest open-source AI model ever built, which the company claims outperforms GPT-4o And Claude 3.5 Sonnet d’Anthropic The company has also set up several performance tests. It is also making its Llama-based Meta AI assistant available in more countries and languages, while adding a feature that can generate images based on a person’s specific likeness. CEO Mark Zuckerberg now predicts that Meta AI will be the most used assistant by the end of the year, surpassing ChatGPT.
The Llama 3.1 is significantly more complex than the smaller Llama 3 models. which came out a few months agoThe largest version has 405 billion parameters and was trained with over 16,000 Nvidia Ultra-expensive H100 GPUsMeta does not disclose the development cost of Llama 3.1, but based on the cost of Nvidia chips alone, it is safe to assume that it is in the hundreds of millions of dollars.
So, given the cost, why does Meta continue to offer Llama with a license that only requires approval from companies with hundreds of millions of users? a letter published on the Meta company blogZuckerberg argues that open-source AI models will outperform — and are already improving faster than — proprietary models, in the same way that Linux has become the open-source operating system that powers most phones, servers and gadgets today.
“An inflection point in the industry where most developers are starting to primarily use open source”
He compares Meta’s investment in open source AI to its previous Open Compute project, which he says saved the company “billions” by bringing in outside companies like HP to help improve and standardize Meta’s data center designs as it built out its own capabilities. Going forward, he expects the same dynamic to play out with AI, writing, “I think the release of Llama 3.1 will be an inflection point in the industry where most developers will start to primarily use open source.”
To help get the word out about Llama 3.1, Meta is working with more than two dozen companies, including Microsoft, Amazon, Google, Nvidia, and Databricks, to help developers roll out their own versions. Meta says Llama 3.1 costs about half the price of OpenAI’s GPT-4o to run in production. It’s publishing the model’s weights so companies can train it on custom data and tune it as they see fit.
Unsurprisingly, Meta doesn’t say much about the data used to train Llama 3.1. People working at AI companies say they’re not disclosing this information because it’s a trade secret, while critics say it’s a tactic to delay the inevitable wave of copyright lawsuits that are coming.
Meta claims to have used synthetic data, that is, data generated by a model rather than humans, to make the 405 billion parameter version of Llama 3.1 better than the smaller 70 billion and 8 billion parameter versions. Ahmad Al-Dahle, Meta’s vice president of generative AI, predicts that Llama 3.1 will be popular with developers as a “teacher for smaller models that are then deployed” in a “more cost-effective” way.
When I ask if Meta agrees with the growing consensus As the industry begins to run out of quality training data for models, Al-Dahle suggests a ceiling is coming, though it may be further away than some think. “We definitely think we have a few more (training) cycles,” he says. “But it’s hard to say.”
For the first time, Meta’s adversarial testing of Llama 3.1 involved looking for potential use cases in cybersecurity and biochemistry. Another reason for testing the model more rigorously is what Meta describes as emerging “agentic” behaviors.
For example, Al-Dahle tells me that Llama 3.1 can integrate with a search engine API to “retrieve information from the Internet based on a complex query and call multiple tools in succession to do your job.” He also gives the example of asking the model to graph the number of homes sold in the United States over the last five years. “It can take the search (from the Web) for you, generate the Python code, and run it.”
Meta has also implemented Llama, its AI assistant, which is positioned as a multi-purpose chatbot like ChatGPT and can be found in almost every part of Instagram, Facebook, and WhatsApp. Starting this week, Llama 3.1 will first be accessible via WhatsApp and the Meta AI website in the US, followed by Instagram and Facebook in the coming weeks. It is also being updated to support new languages, including French, German, Hindi, Italian, and Spanish.
While Llama 3.1’s most advanced parameter model, which has 405 billion parameters, is free in Meta AI, the assistant will downgrade you to the smaller 70 billion parameter model after you exceed an unspecified number of prompts in a given week. That suggests the 405 billion model is too expensive for Meta to run at scale. Spokesperson Jon Carvill told me the company will provide more information on the prompt threshold after evaluating early usage.
A new “Imagine Me” feature from Meta AI scans your face through your phone’s camera to let you insert your image into the images it generates. By capturing your image this way, rather than through your profile pictures, Meta hopes to avoid creating a deepfake machine. The company sees demand from people who want to create more types of AI media and share them on their feeds, even if it means blurring the line between what’s visibly real and what’s not.
Meta AI will also be coming to the Quest headset in the coming weeks, replacing its voice control interface. Like its implemented in Ray-Ban Meta glassesyou will be able to use Meta AI on the Quest to identify and learn more about what you are looking at while you are in the headset passthrough mode which shows the real world through the screen.
“I think the whole industry is just at the beginning of its journey towards product-market fit”
Besides Zuckerberg’s prediction that Meta AI will be the most used chatbot by the end of this year (ChatGPT has over 100 million users), Meta has yet to release usage figures for its assistant. “I think the entire industry is just at the beginning of its journey toward product-market fit,” Al-Dahle says. While AI may already seem overhyped, it’s clear that Meta and others believe the race is just beginning.
