Meta chief AI scientist Yann LeCun just poured cold water on AGI hype, arguing that large language models will never reach human-level intelligence. He claims LLMs lack fundamental traits like reasoning, planning, and physical world awareness, suggesting the industry must pivot toward sensor- and video-trained "world models" instead.
Meta just unveiled the next generation of its custom AI inference accelerator, MTIA. Built on TSMC's 5nm process, the new silicon more than doubles compute and memory bandwidth, delivering 3x the performance of MTIA v1 across key models. Why it matters: Meta is scaling its in-house chip lineup to slash inference costs and ease reliance on commercial GPUs.
Meta just confirmed that its next-generation open-source model, Llama 3, will begin rolling out within the next month. The company plans to release multiple versions across the year and integrate them throughout its product lineup. Llama 3 aims to relax previous over-cautious restrictions, with reports expecting parameter sizes to reach roughly 140B.
OpenAI reportedly transcribed over a million hours of YouTube videos using Whisper to train GPT-4 after burning through high-quality text data. The New York Times reports leadership knew it was legally risky but bet on fair use, while Google and Meta are similarly scrambling in legal gray zones.
Meta is going to extreme lengths to win the AI talent war, with Mark Zuckerberg personally emailing Google DeepMind researchers to lure them away. The company is even skipping interviews to hand out direct offers and scrapping its old rule against counteroffers. Why it matters: High-level AI talent has quickly become the most fiercely contested commodity in tech.
Meta CEO Mark Zuckerberg just met with LG Electronics leadership in Seoul to discuss teaming up on next-gen extended reality devices and on-device AI. Facing fresh heat from Apple's Vision Pro, Meta is looking to marry its large language models with external hardware muscle. Zuckerberg's South Korea itinerary also includes talks with Samsung and President Yoon.
Meta just unveiled V-JEPA, a non-generative video model designed to build human-like world models. Instead of predicting pixels, it predicts masked video parts in an abstract representation space, boosting training efficiency by up to 6x. Best of all, it excels at frozen evaluations—meaning you don't need to retrain the entire model for downstream tasks.
Artists now have a few ways to push back against generative AI scraping their work without consent. While tech giants like OpenAI and Meta offer cumbersome opt-out forms that don't touch existing models, tools like Glaze and Nightshade let creators cloaking or actively poison their images. Why it matters: Until copyright laws catch up, pixel-level sabotage is one of the few proactive defenses visual artists have left.
Meta just rolled out its Version 2.0 software update for Ray-Ban Meta smart glasses, adding camera noise reduction, better low-light capture, and unified touchpad volume controls. However, the much-anticipated AI-powered image recognition remains noticeably absent and locked behind closed betas. For now, you will still have to wait before your frames can identify the world around you.
Meta just announced it will start labeling AI-generated images on Facebook, Instagram, and Threads in the coming months. The system will detect invisible markers based on C2PA and IPTC standards from tools made by OpenAI, Google, Adobe, and others. For AI video and audio where cross-industry signals are lacking, Meta will require users to self-disclose under threat of penalties.
Meta just rolled out Code Llama 70B, a larger coding model trained on 500 billion tokens with a massive 100,000-token context window. It arrives in base, Python-specialized, and instruction-tuned variants, all freely downloadable for commercial and research use under Meta’s permissive open license.
Mark Zuckerberg just confirmed that Meta is actively training Llama 3, aiming for state-of-the-art coding and reasoning capabilities to pursue AGI. To power the push, Meta is stockpiling over 340,000 Nvidia H100 GPUs by the end of 2024, building a computing fortress that rivals almost any competitor in the industry.
Instagram just launched a generative AI background editor for Stories to U.S. users. People can type custom prompts or pick presets like "Being chased by dinosaurs" to replace their photo backdrops, and viewers get a "Try it" sticker to instantly create their own. It is Meta's latest attempt to turn lightweight AI generation into a viral social loop.
Meta just began rolling out multimodal generative AI features to its second-generation Ray-Ban smart glasses in an early-access beta. Users can say "Hey, Meta, look" to trigger the on-board camera and get voice answers about what they are looking at. Why it matters: It brings voice-activated computer vision right to everyday eyewear, though it currently requires a photo snap and still experiences occasional hallucinations.
Meta just released an interactive demo for Audiobox, its new foundation research model for audio generation. The playground allows users to test zero-shot TTS, generate custom sound effects from text, and experiment with audio infilling. It gives creators an early hands-on look at Meta's generative audio capabilities.
Meta just launched Purple Llama, an umbrella initiative aimed at standardizing trust and safety for open generative AI models. The initial release delivers the industry's first shared cybersecurity evaluation benchmarks alongside Llama Guard, an openly available foundational model designed to filter risky inputs and outputs for research and commercial applications.
Meta just rolled out Imagine with Meta, a standalone web-based image generator powered by its Emu model. U.S. users can now generate four high-resolution images per text prompt for free without jumping into a messaging app. Meta also pledged to add invisible AI watermarks that survive cropping and compression to boost transparency.
Meta just rolled out a major wave of AI upgrades across its hardware and apps. Ray-Ban Meta smart glasses are gaining multimodal AI to understand what you see and fetch real-time info via Bing, while a standalone web image generator launched at imagine.meta.com. Group chats on Messenger and Instagram also gained interactive image remixing.
Meta just released Seamless, an open-source translation system designed to sound more human. It includes SeamlessStreaming, which translates mid-sentence with just two seconds of latency, and SeamlessExpressive, which preserves your tone, rhythm, and emotion across languages. The entire model family and alignment tools are now public.
A federal judge just dismantled key parts of comedian Sarah Silverman's copyright lawsuit against Meta, calling the claim that LLaMA itself is an infringing derivative work "nonsensical." The court ruled that plaintiffs must show AI outputs actually copy their protected material to claim infringement. While this is a big win for AI developers, Meta still faces Silverman's core claim that scraping copyrighted books for model training constitutes infringement.
