Amnesty International Coach from Meta from the current artificial intelligence lacks “smart behavior” – a report
The head of artificial intelligence in Meta said that the current artificial intelligence models lack the main features of human intelligence, and he was invited that the company’s last model will solve this problem.
Business Insider said on May 26 that at the AI Action Summit in Paris earlier this year, the chief of Amnesty International, Yan Lacon, President of Amnesty International, said, “There are four basic characteristics of smart behavior that every animal, or relatively smart animal, and certainly humans can do,” said Business Insider on May 26.
“Understand the material world, the ongoing memory, the ability to think and the ability to plan complicated actions, especially hierarchical planning,” said Liu.
He said that the current large language models (LLMS) that Power Popular Ai Chatbots did not reach this threshold, and “merging these capabilities will require a shift in how to train them.”
Lacon said that some of the largest Amnesty International and Technology Giants are “the possibilities of the horse” on the current models in their race to control the game of artificial intelligence.
Meta is already experimenting with a system called Retrieval General (RAG), a way to enhance LLM outputs using external knowledge sources.
In February, V-Jepa, a non-denial model learned by predicting missing or shaded parts of the video.
Related to: Meta gets the gesture of the European Union organizer to train artificial intelligence with social media content
Lecun believes that “global models” will be a better approach because they will be trained in real life scenarios and have a higher awareness of the current style intelligence.
The concept includes models that can imagine any action and prediction of the resulting global state. Since the world has unexpected capabilities, Lecun believes that training should happen through abstraction, which reflects how humans understand the material world.
Date of artificial intelligence brain migration
Meanwhile, Meta suffers from a great loss of talents from the artificial intelligence research team, especially among researchers who created the original Llama model in 2023, from the inside. I mentioned On May 26.
Only three of the original Llama authors are still in Meta, and many joined Mistral, a Paris -based emerging company in which former researchers participated in Meta and the main Llama engineers.
The latest version of Meta, Llama 4, got a lukewarm reception from developers, many of whom are now looking forward to their faster movement in the movement who have allocated thinking models such as GPT-4O from Openai, the report from Google Gemini 2.5 Pro, and Claude 4 Sonnet recently from Anthropic.
May 15, The Wall Street Journal I mentioned This Meta was delaying its pioneering offer AI LLM, Llama 4 “Beheemoth”.
magazine: Amnesty International heals blind