AI at Meta — searchable YouTube transcripts & AI Q&A

Introducing Meta Omnilingual Automatic Speech Recognition | Transcription for 1,600+ languages

Meta's Omnilingual ASR model supports automatic speech recognition for over 1,600 languages, including many underserved ones, by leveraging community-collected data. The model can perform zero-shot transcription on unseen languages given just five speech-transcription pairs, and the team open-sources it to empower fine-tuning for individual languages. The project aims for omnilinguality, where even languages with a single speaker can be transcribed.

Knowledge Distillation with Llama 3.1 405B | Llama for Developers

Fine-tuning remains essential for adapting pre-trained models to specific domains, despite advances in prompting techniques, as it fundamentally reshapes model behavior and can reduce long-term costs. Synthetic data generation from large models like Llama 3.1 405B is a cost-effective way to create training data, but it must be mixed with real data to avoid model collapse and ensure groundedness. Knowledge distillation from large teacher models to smaller student models enables efficient deployment, especially for domain-specific tasks, though it is still experimental and requires careful quality control.

Meta PARTNR: Unlocking Human-Robot Collaboration

Meta FAIR argues that robots will soon collaborate with humans in everyday settings, not just work in isolation. They introduce a large planning model trained on over 100,000 natural language instructions and simulated in Habitat 3.0, which enables robots to dynamically replan when humans interfere, such as when a toy is removed during cleanup. This approach allows robots to adapt to novel environments without retraining from scratch.

Run Llama 3 on Linux | Build with Meta Llama

Meta's Llama 3 models can be run locally on Linux with an NVIDIA GPU of at least 16 GB VRAM, as demonstrated by downloading the weights and using the provided scripts. The tutorial shows how to set up a Python environment, clone the Llama repo, download the 8B and 8B-instruct models, and run text completion and chat examples via torchrun. This approach keeps data local and independent of external APIs, though it requires accepting the license and using a URL valid for 24 hours.

10 Years of Advancing the State-of-the-Art through Open Science at FAIR | AI at Meta

Open science and bottom-up research culture at FAIR have driven major AI advances, argues Wesley and colleagues. PyTorch, built out of necessity, became a dominant framework after open-sourcing, and self-supervised learning underpins modern LLMs like Llama. The next frontier involves world models and physical understanding, with responsible AI as a key focus.