Meta strengthened its open source stance on Tuesday by releasing the Meta Llama 3.1 family of large language models.
Llama 3.1 includes Meta’s largest generative AI model to date, 405B, and updates to versions 70B and 8B.
Llama 3.1 Expands Models context window up to 128K, which increases the amount of information that can be transmitted by the AI system. It also supports eight languages.
Meta also changed its license, allowing developers to use the outputs of Llama models to improve other models.
Against a current
The introduction of the large model represents a reversal of the recent trend in the AI market towards small language models.
“Interestingly, this goes against all the trends we’ve seen with language models,” said Mark Beccue, an analyst in TechTarget’s enterprise strategy group.
One possible reason why Meta chose to do this is because Llama 3.1 405B is the first customizable Master in Open Source LawBeccue said.
“There’s now an option for a lot of these companies that are moving to open source – and there are a lot of them – to be able to take a very large model and do whatever they want with it,” he said.
However, creating a large model like the 405B comes at a high cost.
Meta said that to train Llama 3.1 405B, it used more than 16,000 Nvidia H100 GPUs. These AI chips cost between $25,000 and $40,000 each depending on the configuration, meaning Meta spent up to $640 million to train the new model.
As a result, Llama 3.1 405B may be too expensive for some companies to deploy and maintain, said Paul Nashawaty, an analyst with Futurum Group.
“The 405 billion parameter model requires immense computing resources, including high-performance GPUs and significant storage,” he said in a statement to the media. “This translates into significant upfront costs for hardware as well as ongoing expenses for power and cooling.”
So for smaller businesses, generative AI tools already available on major cloud platforms could be cheaper, he continued.
Useful for business
However, a large open source model This unlocks value for businesses, said William McKeon-White, an analyst at Forrester Research.
“Companies are starting to understand that GenAI is still quite difficult to implement,” he said. “Having an open source model or a tool that you don’t pay for per transaction can be very useful for companies that are looking to use these models in a very customized way.”
Additionally, models like these enable more complex reasoning. Many applications for which organizations are considering using these models, such as fraud detection and medical diagnosis, will require complex reasoning, he added.
Having an open source model or something that you don’t pay per transaction for can be very helpful for organizations looking to use these models in a very customized way.
William McKeon-WhiteAnalyst, Forrester Research
However, controversy surrounds what Meta considers open source despite Meta CEO Mark ZuckerbergMeta’s blog, also published Tuesday, details his company’s commitment to open source. Some argue that the Llama models aren’t fully open source because Meta hasn’t released its training data.
“It’s open source software in that you can custom train it, you can customize the model. But you still don’t know what the data sources are,” Beccue said. “To me, that leads to this inherent problem with these models, which is that you can’t trace the accuracy of them.”
Additionally, it is unclear whether Llama 3.1 is more accurate than Lama 3added Beccue.
Regardless, Meta differs from Google and OpenAI in that it is willing to make its larger model customizable so others can build on it, said Arun Chandrasekaran, an analyst at Gartner.
“Models are still invaluable to businesses. And you still have access to model weightwhich I think is very important for enterprise customers,” Chandrasekaran said.
Scaling Security
In addition to the significant size of Llama 3.1 405B’s parameters, Meta revealed that it is also expanding its AI security capabilities. The social media giant introduced two new security tools.
Llama Guard 3 is an input and output moderation model that helps developers detect violating content.
Prompt Guard is another tool that helps developers respond to rapid injection and prison escape contributions.
Flash injections use data from untrusted sources to make a model work in unexpected ways. Jailbreaks are instructions that override a model’s security features.
Meta also revealed that it has conducted both human and AI-assisted investigations. red team test to understand how its models work against various opposing actors and in different situations.
“They’re starting to take security a little more seriously and they’re starting to make efforts in that area,” Chandrasekaran said.
Monetization and other news
One of the biggest challenges for Meta and other vendors as they continue to innovate is how to make money from their products, Chandrasekaran said.
“We continue to see innovation in this space. But at the same time, I think more and more companies are realizing that they’re creating very powerful products, but they’re also struggling to monetize them,” he said.
Llama 3.1 is now available on AWS.
Scale AI also revealed that it has partnered with Meta to help businesses customize, evaluate, and release Llama 3.1.
AI hardware and software vendor Nvidia also revealed that businesses can use Nvidia AI Foundry to customize Nvidia’s open models and third-party open models, including Llama 3.1.
Meta also revealed that its AI assistant, Meta AI, is now available in seven new languages and in more countries around the world.
Meta AI users now have the option to use Llama 3.1 405B on WhatsApp and meta.ai.
The Assistant is now also more creative, according to Meta, with new “picture me” prompts that let users create images.
Esther Ajao is a TechTarget editorial writer and podcast host covering artificial intelligence software and systems.
We use cookies to ensure that we give you the best experience on our website. If you continue to use this site we will assume that you are happy with it.Ok
