In the battle to win over users with artificial intelligence, another giant has broken out between Microsoft and Google. Meta wants to take a place on the podium of this fast-paced race and already shows the surprising results of CM3leon (pronounced chameleon, chameleon in English). The image generator of Mark Zuckerberg’s company also interprets them and can edit them in the sense that it is asked. It works from the images or creates them, as requested by the user. Meta’s week in the AI ??field is rounded off with the announcement that its LLaMA 2 model is becoming open source, so anyone can download it to their computer and give it a specific application.
Regarding CM3leon, the date on which it will be made available to the public is not yet known. Meta has shown its impressive results and revealed some of its technical characteristics. “CM3leon is the first multimodal model trained with a recipe adapted from text-only linguistic models”, its researchers explained to Meta. The authors state that it “achieves leading performance in text-to-image generation, despite being trained with five times less computation than previous methods based on transformers – a system devised by Google for AI language models”.
CM3leon also uses one of the common methods in AI image generation, which is to “add a separately trained super-resolution stage to produce higher resolution images from the original model results”. The results shown by Meta are of quality, but there are no clues as to when or under what conditions it will be available to the public.
In any case, Meta is leaning towards opening up access to its artificial intelligence models a lot. On Tuesday, it announced together with Microsoft – the largest investor in OpenAI, which will bring ChatGPT’s AI to all its commercial products – an agreement to make the artificial intelligence of Mark Zuckerberg’s company, LLaMa 2, available as open source for anyone who wants to use it to develop applications.
The announcement of the free availability of LLaMA 2 was made during Microsoft’s Inspire event, which the Redmond company dedicates to companies with which it partners around the world, because its AI model will be distributed through the Azure platform, although also from other important ones such as Amazon Web Services.
Meta’s move aims to make a high-volume language model available to developers for free, unlike closed ones like OpenAI’s GPT-4, which is only available to its partners.
Some of the features that Meta has revealed of its large AI language model far exceed its predecessor, LLaMA 1. The new model has been trained with 40% more data than the first and has twice as much context.
LLaMA 2 is now available to anyone in three different sizes that are scaled by the number of parameters the language model uses: 7,000, 13,000, and 70,000 million.
This latest version is comparable to GPT-3.5, the model with which OpenAI launched ChatGPT in November, although it now also has a much improved model, of which hardly any features are known, GPT-4, which has the ability to be deeper in its responses and is multimodal, capable of using text and image.
Meta assures that its AI models “have undergone internal and external security tests”. The company also provides developers with guidance on responsible practices and security assessments. The EU is preparing a law that will regulate this technology.
