Benjamin Franklin wrote once nothing was certain, except death and taxes. Let me modify this sentence to reflect the current gold rush in gold: nothing is certain, except death, taxes and new models of AI, the last of these three arriving at an always accelerating rate.
Earlier this week, Google has published improved Gemini models, and, earlier in the month, Openai has unveiled its O1 model. But Wednesday, it was Meta’s turn to trot his last to the annual company of the company Meta Connect 2024 Developer conference at Menlo Park.
Multimodality of Llama
Meta’s multilingual family has reached version 3.2, With the bump of 3.1 meaning that several models of lama are now multimodal. LLAMA 3.2 11B – A compact – and 90B model, which is a larger and more capable model, can interpret graphics and graphics, legendary images and identity objects in images given a simple description.
Given a card of a park, for example, Llama 3.2 11b and 90b could answer questions such as “when will the land become more steep?” ” And “What is the distance from this path?” Or, provided a graph showing the income of a company during a year, the models could quickly highlight the most efficient months of the group.
For developers who wish to strictly use models for text applications, Meta says that Llama 3.2 11b and 90b have been designed to be replacements “without appointment” for 3.1. 11b and 90b can be deployed with or without new safety tool, Llama Guard Vision, designed to detect potentially harmful (that is to say biased Or toxic) Text and images fed or generated by models.
In most of the world, Multimodal Llama models can be downloaded and used on a large number of cloud platforms, including Face, Microsoft Azure, Google Cloud and AWS. Meta also welcomes them to the official website of Llama, llama.com, and uses them to feed her AI, Meta Ai, on Whatsapp, Instagram and Facebook.
But Llama 3.2 11b and 90b cannot be accessible in Europe. Consequently, several Meta AI features available elsewhere, such as image analysis, are disabled for European users. Meta once again Blame the “unpredictable” nature of the regulatory environment of the block.
Meta expressed her concerns about – and rejected A voluntary security commitment linked to – The AI ActEU law which establishes a legal and regulatory framework for AI. Among other requirements, the AI Act requires that companies developing an AI in the EU undertake to trace if their models are likely to be deployed in “high -risk” situations, such as the police. Meta fears that the “Open” nature of its modelsThis does not give an overview of how models are used, could make it difficult to adhere to the rules of the AI Act.
The Meta are also involved in the provisions of the GDPR, the Large Law on Privacy of the EU, concerning the formation of AI. Meta Trains of Public Data Models of Instagram and Facebook users who have not withdrawn – data that in Europe is subject to GDPR guarantees. EU regulators earlier this year have requested meta-stop training on European user data while they were asking the company’s RGPD compliance.
Meta gave in, while at the same time approving a open letter Call “a modern interpretation” of the GDPR which does not “reject progress”.
Earlier this month, Meta said it would resume training on British user data After “(incorporate) regulatory feedback” in a revised recovery process. But the company has not yet shared an update on training throughout the block.
More compact models
Other new LAMA models – models that have not been trained on European user data – are Launch in Europe (and around the world) on Wednesday.
LLAMA 3.2 1B and 3B, two light and textual models designed to operate on smartphones and other on-board devices, can be applied to tasks such as the summary and rewriting of paragraphs (for example in an email). Optimized for ARM equipment from Qualcomm and Mediatek, 1B and 3B can also press tools such as calendar applications with a little configuration, says Meta, allowing them to take actions independently.
There is no follow -up, multimodal or not, at Llama 3.1 405B Model released in August. Given the massive size of 405b – it took months to train – these are probably contained calculation resources. We asked Meta if there were other factors at stake and update this story if we hear.
Meta’s new Llama battery, a Lama -focused development tool costume, can be used to refine all Llama 3.2: 1b, 3b, 11b and 90b models. Whichever way they are personalized, the models can treat up to around 100,000 words at a time, says Meta.
A room for Mindshare
The meta-PDG Mark Zuckerberg often talks about ensuring that everyone has access to the “advantages and opportunities” of the AI. Implicit in this rhetoric, however, is a desire that these tools and models be meta creation.
Expenses on models which can then merchant oblige competition (for example OPENAI, Anthropic) to reduce prices, distributes the META version of AI largely and allows META to incorporate improvements from the open source community. Meta says its Llama models have been downloaded more than 350 million times and are used by large companies, including Zoom, AT&T and Goldman Sachs.
For many of these developers and businesses, it is intangible that the Lama models are not “open” in the strictest sense. Meta license constraints How some developers can use them; Platforms with more than 700 million monthly users must request a special license in Meta that the company will grant its discretion.
Admittedly, there are not many platforms of this size without their own internal models. But Meta is not particularly transparent on the process. When I asked the company this month if he had further approved a discretionary Lama license for a platform, a spokesperson told me that Meta “had nothing to share on the subject”.
Don’t be mistaken, the meta play to keep themselves. Is spending millions put pressure on It’s plowing billion in servers, data centers and network infrastructure To form future models.
None of the Llama 3.2 models solves the primordial problems with today’s AI, such as its tendency to invent things And regurgitate problematic training data (For example E-books protected by copyright This could have been used without authorization, the subject of a collective appeal against Meta). But, as I wrote before, they advance one of Meta’s key objectives: to become synonymous with AI, and in particular the generative AI.