Two years after having opened the code of its Llama models, Meta brought together its community to Lamacon assert its position: the future of artificial intelligence will not be that of an omniscient assistant but it will be that of thousands of specialized agents, trained for specific uses, optimized for speed and cost, and adapted to the field.
From the first minutes of the conferencethe tone is set. Chris Cox, Chief Product Officer at Meta, does not speak of generalization or market domination. He talks about modularity, personalization, “regain control” on AI. He evokes not an intelligence that can do everything, but AI that make One thing good, quickly, and without additional cost.
“We went from the dream of the super assistant to the reality of the useful agent,” he sums up.
On stage, the examples are linked. An Llama model now works aboard the international space stationwithout network connection, to help astronauts navigate in technical textbooks.
At & tfor its part, uses a specialized AI to analyze thousands of hours of customer calls every day, and synthesize the ten most frequent bugs to correct. In Sub-Saharan Africathe Pharmachat application provides agricultural recommendations in local languages, without going through an overly expensive or unsuitable generalist AI.
Each time, the model has been Fine-tuna, distilled, reducedto meet a specific business, technical, or linguistic need.
Behind this strategy, a strong technological conviction: large models are not useless, but their raw shape is rarely useable directly. Meta therefore relies on a triptych:
-
- Behemotha giant model that serves as a teacher;
- Mavericka more compact model used in production;
- And Scoutthe ultra-light version, designed to turn on a single GPU.
These models can be distilled To maintain useful intelligence, and eliminate the superfluous. Clearly: do more with less.
Meta claims to have built a full infrastructure : a simplified API, accessible fine-tuning tools, an integrated evaluation system, an interoperability with the Openai ecosystem … Llama is not limited to a model and wants to be a real platform.
Mark Zuckerberg, during an exchange with Ali Ghodsi (Databricks), once again sums up the ambition of the group which is to allow each company to deploy His own agentadapted to its data, customers, language.
“An AI assistant who sells the competitor’s products?” No business wants that. »»
The complexity of this transition requires training, fine-tuner, because deploying a specialized agent remains a technical approach. And not all companies have the means to do so. Hence the importance of an active community, open documentation, and tools that are easy to handle.
“It is not a minimalist vision of AI,” said Angela Phan, Llama engineer. “It is a more distributed, more appropriate, more useful vision. »»
What to remember
-
- Meta considers that the unique and general model has reached its operational limits.
- The company relies on a Constellation of specialized AI agentsadapted to specific uses.
- Llama 4 was designed to facilitate the distillationthere personalization and the local deployment.
- Concrete use cases (spatial, health, agriculture, customer service) illustrate the rise of these vertical AIs.
- Meta now offers a full infrastructurefrom API to exporting fine-tunated models.
- Meta’s bet is to return the AI Simpler, faster, cheaper.
