In This Article:
Meta Platforms META, Hugging Face, and Scaleway recently announced the launch of an AI Accelerator Program targeted at European startups. This initiative aims to bolster economic and technological advancements through the integration of cutting-edge AI models into startup ecosystems across the European Union.
Based at STATION F in Paris, the AI Startup Program, which will run from September 2024 to February 2025, is designed to support five promising startups from the MVP or product stage. This initiative follows the success of the first edition of the accelerator announced in November 2023, which supported innovative projects like Pollen Robotics, Fringuant and Qevlar AI.
The program offers extensive resources and mentoring from Meta’s Fundamental Artificial Intelligence Research (FAIR) team, Hugging Face’s platform and Scaleway’s AI infrastructure. The selection process will involve a panel of experts from Meta, Hugging Face and Scaleway, who will evaluate projects based on their use of open foundation models and integration plans.
Selected startups will gain access to invaluable resources, including technical mentoring, computational power, and tools essential for developing services based on open-source AI technology.
Meta’s New AI Research Models
Alongside the accelerator announcement, Meta’s FAIR team has unveiled a series of new AI research models aimed at advancing the field through open collaboration. These models include innovations in image-to-text and text-to-music generation, multi-token prediction for faster word processing in large language models (LLMs) and tools for detecting AI-generated speech.
The latest releases include Chameleon Models, Multi-Token Prediction, JASCO and AudioSeal.
Chameleon Models are capable of processing and generating both text and images simultaneously. These models offer new possibilities in creative content generation.
The Multi-Token Prediction approach enhances the efficiency of LLMs by predicting multiple future words at once, reducing the text needed for training.
JASCO is a text-to-music generation model that accepts various inputs like chords or beats, providing more control over the generated music outputs.
AudioSeal is an audio watermarking technique for detecting AI-generated speech segments within audio snippets, designed for large-scale and real-time applications.
Meta Platforms also released tools to evaluate and enhance geographical and cultural diversity in AI-generated images, aiming to ensure more inclusive and representative outputs.
