Meta’s Llama 3 AI model includes an image generation feature that updates as the user types prompts. Photo / FT montage, Meta
Meta’s Llama 3 AI model includes an image generation feature that updates as the user types prompts. Photo / FT montage, Meta
Meta has released a new and less “sanctimonious” version of its artificial intelligence model, as the social media group races against Microsoft-backed OpenAI and Google to develop the cutting-edge technology.
Facebook’s parent company said on Thursday its latest model, Llama 3, has “vastly improved capabilities”, including the abilityto reason, in a step towards human-level intelligence.
The news comes as Silicon Valley giants have been unveiling their next generation of powerful AI models, having poured billions of dollars into the technology off the back of the generative AI hype caused by ChatGPT.
However, large language models, the underlying technology that powers AI products such as OpenAI’s ChatGPT and Google’s Gemini, have come under increasing scrutiny as consumers become aware of its flaws.
Nick Clegg, Meta’s president of global affairs, told the Financial Times that it has sought to respond to criticisms that Llama 2 refused to answer “innocent and innocuous” questions.
“People, I think, found [Llama 2] a little bit sanctimonious, a little bit reluctant to engage,” he said. “So we’ve really, really tried to work hard to reduce what we call ‘false refusals’ to make the underlying models more helpful and more and more responsive.”
Nick Clegg at the Meta AI Day in London on April 9. Meta, Google, OpenAI and leading startups are churning out new AI language models and trying to persuade customers they've got the smartest or fastest or cheapest chatbot technology. Photo / Kirsty Wigglesworth, AP
Clegg added that Llama 3 is designed to be apolitical. The issue of bias of AI models built by Silicon Valley giants has created a backlash from users. Google paused its Gemini image generation tool over historical inaccuracies and depictions of different ethnicities and genders.
Llama 3 will initially be offered in two sizes, one with 8 billion parameters and one with 70 billion parameters — the number of variables used to train a system and ultimately shape its output.
Larger models with more parameters often perform better than smaller models. Google and OpenAI also offer varying model sizes to cater to different demands.
Unlike Google and OpenAI’s models, Llama 3 will be open source, meaning that developers can readily build products using the model for free. However, big cloud giants such as Apple and Google will be charged for training their models on the system.
“We are delighted to open source so that developers, entrepreneurs and researchers can feed off Llama 3 for new innovation [but not for] very large platforms who might just want to piggyback on our investments to sort of short circuit their own,” said Clegg. “We’re not pouring billions of dollars into this stuff in order to save Apple money.”
Llama 3 will be integrated into Meta AI, its chatbot that works across its suite of products such as Facebook Messenger, WhatsApp and its line of Ray-Ban smart glasses. Meta AI is also expanding its access to more than a dozen English-speaking countries, including Australia, Pakistan, Zambia and Canada.
However, consumers in the EU and UK have yet to access Meta AI.
The EU has some of the strictest rules in the world for policing AI, and the UK is drafting legislation aimed at regulating the most powerful models.
“We are still proceeding [across Europe] but [are doing so] slightly at the pace of the discussions we have with regulators and others,” said Clegg.
Llama 3 is seven times larger than its predecessor. Meta is building a 405 billion parameter version that is still being trained and will be released subject to “safety evaluations”, which are being conducted over the next few months.
Meta claims its performance is comparable to the latest models from rivals, including French-start up Mistral, US-based Anthropic and Google.
One example of how Meta AI will improve its offering is through its messaging service WhatsApp, introducing an image generation feature that edits a picture as a user writes.
For example, if you write “horse galloping”, the system generates that image, but then updates as you type other prompts, such as asking for a beach in the background.
