The Open Source Initiative (OSI) just called out Meta for misusing the term "open source" with LLaMA 2. Because Meta restricts certain commercial uses and application areas, the license contradicts the Open Source Definition. OSI argues real open source gives developers full sovereignty, and plans to release a dedicated "Open Source AI" standard draft on October 17.
Meta just signed an MoU with the Indian government's 'India AI' initiative to bring its open-source AI models to local researchers and startups. The partnership focuses on building datasets for India's diverse low-resource languages using models like Llama, MMS, and NLLB, alongside plans for a joint Center of Excellence.
Thousands of writers just signed an open letter calling out OpenAI, Meta, and other AI developers for training language models on millions of copyrighted works without permission or compensation. The letter demands companies obtain licensing and fairly pay authors for both past and ongoing use, warning that unlicensed scraping threatens creators' livelihoods.
Qualcomm just announced a partnership with Meta to run its new open-source Llama 2 model directly on phone and PC chips starting in 2024. Instead of relying entirely on massive cloud data centers, running models locally on device could dramatically slash operating costs and unlock snappier on-device virtual assistants.
Microsoft just expanded its AI partnership with Meta, becoming the preferred cloud partner as Meta makes Llama 2 commercially available. Azure developers can now deploy and fine-tune 7B, 13B, and 70B models with built-in safety tools, while Windows devs get native optimization via ONNX Runtime and DirectML to run models locally.
Meta just unveiled CM3leon, a single foundation model capable of both text-to-image and image-to-text generation. By adapting recipes from text-only language models, it achieved a state-of-the-art 4.88 FID score on zero-shot MS-COCO while requiring five times less compute than previous transformer methods. It marks a big step toward unified autoregressive multimodality.
Comedian Sarah Silverman and two authors just slapped OpenAI and Meta with copyright infringement lawsuits in US District Court. The complaint alleges ChatGPT and LLaMA were trained on pirated books grabbed from "shadow libraries" like Bibliotik without consent, pointing to ChatGPT's book summaries as proof. The battle over AI training datasets is officially heating up.
UK finance expert Martin Lewis just sounded the alarm over an AI deepfake video circulating on Facebook that clones his likeness and voice to promote an investment scam. Lewis called the development terrifying and demanded regulatory crackdowns on big tech ad platforms, while Meta stated it proactively removed the offending videos.
Meta just published 22 "system cards" explaining how AI algorithms rank content across Facebook and Instagram. The cards break down predictive models and ranking signals used in Feed, Stories, and Reels, alongside expanded controls for users to tune recommendations. Why it matters: It offers a rare peek inside the recommendation engines driving global social media.
Users are taking Meta's open-source LLaMA model to build explicit, guardrail-free sexbots, with some generating graphic and violent role-play scenarios. The trend reignites a fierce debate over frontier AI: does open-sourcing democratize innovation, or hand uncontrollable tools to bad actors?
Meta just introduced Voicebox, a generative AI speech model that matches a voice from an audio clip as short as two seconds. It generates natural text-to-speech across six languages, enables cross-lingual style transfer, and lets creators cleanly erase background noise like dog barks and re-synthesize speech.
Meta just open-sourced I-JEPA, a computer vision model that predicts missing parts of images using abstract concepts rather than calculating pixel by pixel. Trained in under 72 hours on just 16 A100 GPUs, it uses up to 10x less compute than traditional methods. Yann LeCun sees it as a major proof of concept for building common-sense world models.
French startup Mistral AI just raised a $113 million seed round at a $260 million valuation led by Lightspeed Venture Partners, only four weeks after launching. Founded by former DeepMind and Meta researchers, the team plans to build open-source generative models using publicly available data starting in 2024. Why it matters: Mistral wants to prove open-source infrastructure can challenge proprietary giants in the enterprise market.
Meta just unveiled MusicGen, a generative AI model that turns text prompts and melody inputs into original songs. Trained on 20,000 hours of music and instrument samples, it can theoretically expand a simple hum into a full track. It opens music creation to anyone, though copyright fights with the record industry are likely looming.
Meta just gave employees an internal look at a suite of upcoming generative AI features, including customizable chatbots for WhatsApp and Messenger, prompt-based photo editing on Instagram, and an internal assistant dubbed Metamate. Mark Zuckerberg said the tech will eventually touch every single product.
Instagram is reportedly testing an in-app AI chatbot that lets users pick from 30 distinct personalities to answer questions, give advice, and draft messages. The leaked feature aligns with Meta's push into customizable AI personas, though balancing playful social engagement with user safety remains a tricky hurdle.
Meta just announced its next-gen Quest 3 VR headset, launching this fall starting at $499.99 for 128GB. It features a new Snapdragon chip with twice the graphics power, a 40% slimmer optic profile, and improved color passthrough ahead of Apple's expected mixed-reality debut.
Meta just open-sourced its Massively Multilingual Speech (MMS) models, expanding text-to-speech and speech-to-text capabilities to over 1,100 languages—a tenfold leap over current tools. It can also identify more than 4,000 spoken languages, trained on diverse recordings of translated religious texts. It is a critical push toward saving endangered dialects and bringing speech AI to underserved communities.
Meta just unveiled CodeCompose, an internal AI coding tool designed to rival GitHub Copilot. The largest 6.7B model is tuned on Meta’s own repositories and integrated directly into engineers' IDEs, with thousands of employees adopting suggestions at a 20%+ acceptance rate. Why it matters: Meta is prioritizing proprietary internal developer tooling over external cloud assistants, keeping its proprietary codebase strictly in-house.
Meta just introduced its first custom silicon for artificial intelligence, the Meta Training and Inference Accelerator (MTIA). Designed specifically for deep learning recommendation models and integrated with PyTorch, the chip is up to three times more power-efficient than GPUs per watt. While it trails Nvidia GPUs on more complex neural networks, it signals Meta's serious push to design its own compute infrastructure.
