Image: © Lemonsoup14/Stock.adobe.com
Meta claims Llama 3 is a ‘major leap’ from its predecessor and ‘on par’ with the best options on the market.
Meta is attempting to gain an edge in the AI rat race with the release of Llama 3, the company’s latest batch of large language models (LLMs).
The social media giant has revealed two models of this new generation of AI models, which are trained on 8bn and 70bn parameters to support a “broad range of use cases”. Meta claims these models show “state-of-the-art” performance on various industry benchmarks and come with new capabilities such as “improved reasoning”.
The new models come less than a year after Meta released Llama 2 for both research and commercial use. This was the successor to the company’s research-focused model released earlier that year.
Meta gave a teaser about the power of Llama 3 earlier this year, when the company said it was using two “data centre scale” clusters that both contain more than 24,000 Nvidia H100 GPUs to develop Llama 3.
Meta claims the latest models are a “major leap” from Llama 2 and that the company’s goal was to create open models that are “on par with the best proprietary models available today”.
“Improvements in our post-training procedures substantially reduced false refusal rates, improved alignment and increased diversity in model responses,” Meta said in a blogpost. “We also saw greatly improved capabilities like reasoning, code generation and instruction following making Llama 3 more steerable.”
The two models being released by Meta are just the start of Llama 3’s journey according to Meta, as the company plans to make Llama 3 multilingual and multimodal in the near future.
The company is planning to make Llama 3 widely available through various partnerships, as it will soon be accessible on AWS, Databricks, Google Cloud’s Vertex AI, Hugging Face, Kaggle, IBM WatsonX, Microsoft Azure, Nvidia and Snowflake.
Llama 3 is also being integrated into existing Meta products such as Meta AI to give users a more effective AI assistant. This AI service is being rolled out to multiple countries – though EU countries are not included for now.
To back the claim that Llama 3 is “the most capable openly available LLM to date”, Meta developed an evaluation set to compare its own models to that of rival products of similar sizes. This evaluation set contains 1,800 prompts that cover various key use cases.
But the latest Stanford AI Index recently claimed robust evaluations for large language models are “seriously lacking” and that there is a lack standardisation in responsible AI reporting.
“Leading developers, including OpenAI, Google and Anthropic, primarily test their models against different responsible AI benchmarks,” the report said. “This practice complicates efforts to systematically compare the risks and limitations of top AI models.”
Find out how emerging tech trends are transforming tomorrow with our new podcast, Future Human: The Series. Listen now on Spotify, on Apple or wherever you get your podcasts.
