Meta just announced major updates to its Meta Advantage automated ad suite. Advertisers can now switch manual campaigns to Advantage+ shopping campaigns with a single click in Ads Manager, while catalog ads are getting support for video creatives dynamically served by AI across Feed and Reels. A new Advantage+ audience feature also turns rigid targeting constraints into flexible suggestions.
Meta just open-sourced ImageBind, a research model linking text, audio, visual, thermal, depth, and IMU motion data into a single embedding space. It aims to let AI cross-reference physical movement and sounds just like current models do with text prompts. While still an early research demo with no immediate consumer app, it keeps Meta firmly planted in open source.
A leaked internal memo from a Google researcher claims that neither Google nor OpenAI has a moat against open source. The document argues that open-source contributors are out-innovating tech giants using efficient methods like LoRA and cheap compute. Why it matters: The era of relying solely on massive closed models may be ending as fast, nimble, community-driven alternatives close the gap.
Meta CEO Mark Zuckerberg just announced that generative AI will touch every single one of the company's products. Meta is testing chat agents for WhatsApp and Messenger alongside visual creation tools for Instagram, with plans to roll out features in the coming months. The goal is to bring helpful AI assistants to regular users and businesses at scale.
Meta just open-sourced DINOv2, a self-supervised computer vision model that matches or beats standard approaches without relying on text captions or labeled data. Its features can be plugged straight into basic linear classifiers for tasks like depth estimation and segmentation, serving as an off-the-shelf multipurpose visual backbone.
Meta just outlined how virtual and augmented reality are reshaping education, highlighting schools in Japan and the US already using Meta Quest 2 headsets. Case studies showed VR chemistry labs boosting test scores and easing social anxiety. Why it matters: True immersion could bridge geographic education divides, provided schools get policy and training support.
The PyTorch team just rolled out PyTorch 2.0, introducing a major compiler revamp that boosts performance without breaking backward compatibility. The marquee feature, torch.compile, leans on OpenAI Triton to deliver an average 36% speedup under AMP precision across 165 models. Native scaled dot product attention is also baked in, eliminating the need for external FlashAttention packages.
Local LLMs just became real on everyday hardware. Thanks to Georgi Gerganov’s new llama.cpp and 4-bit quantization, Meta’s LLaMA can now run directly on personal laptops like a MacBook. It brings GPT-3-class capabilities straight to individual developers, breaking AI out of the corporate cloud garden.
Meta CEO Mark Zuckerberg just announced a new top-level product group dedicated to generative AI. The team plans to build creative tools first, followed by long-term "AI personas" for WhatsApp, Messenger, and Instagram to assist users. Zuckerberg noted there is still plenty of foundational work to do before shipping fully fledged features.
Meta just launched LLaMA, a foundational large language model ranging from 7B to 65B parameters designed to run on far less computing power. The 13B version was trained on 1 trillion tokens, aiming to give academic researchers a lightweight foundation to study model biases and toxicity. For now, access is strictly limited to noncommercial research use.
Shutterstock just rolled out a text-to-image generator powered by OpenAI's DALL-E 2 inside its Creative Flow platform. Crucially, the AI-generated images are immediately ready for commercial licensing, and the company plans to pay contributors whose works helped train the model.
Meta’s chief AI scientist Yann LeCun just pushed back on ChatGPT hype, arguing the viral bot is not particularly innovative in terms of underlying science. LeCun credited OpenAI with doing a stellar engineering job combining pre-existing technologies like Transformers and self-supervised learning, noting that established giants hold back because they have much more to lose when models invent facts.
Meta just expanded its partnership with Shutterstock to train and evaluate AI models using millions of licensed images, videos, and music tracks. Shutterstock says it will continue compensating contributing artists through its contributor fund. The move gives Meta a clean, copyright-safe training pipeline as generative AI faces mounting legal battles.
Meta just launched its Variance Reduction System (VRS) in the US to stamp out algorithmic bias in ad targeting. Built as part of a settlement with the Department of Justice, the machine-learning tech measures and balances demographic reach against target audiences, starting with housing ads before expanding to credit and employment.
Microsoft just unveiled VALL-E, a text-to-speech model that can replicate a voice from a mere three-second sample while preserving emotion and room acoustics. Trained on 60,000 hours of speech, some outputs are remarkably realistic, though Microsoft is withholding the code over deepfake concerns.
OpenAI president Greg Brockman just warned that 2023 will make 2022 look like a sleepy year for AI advancement. Between GPT-4 waiting in the wings, Microsoft hooking ChatGPT into Bing, and coding tools taking off, automation is accelerating across creative and technical work. Why it matters: If you thought artists were the only ones facing disruption, practically every desk job is next in line.
Deep learning is quietly becoming the engine behind the metaverse, driving everything from eye tracking and 3D reconstruction to Meta's real-time speech translation. Instead of rigid scripted interactions, platforms are moving toward generative AI brand representatives that hold fluid conversations. Why it matters: Virtual worlds won't feel truly immersive until human-machine communication feels completely natural.
Meta just introduced Habitat 3.0, a simulator built to train embodied AI alongside human avatars in everyday indoor tasks. Robots can now learn social behaviors like yielding in narrow hallways and sharing chore duties in realistic 3D homes. Why it matters: it lets researchers evaluate social human-robot collaboration safely and cheaply before deploying to physical hardware.
