Google DeepMind is building a new team dedicated to AI that simulates the physical world, led by former OpenAI Sora co-lead Tim Brooks. The group plans to build on Gemini, Veo, and Genie to advance visual reasoning, robotics planning, and real-time generation, viewing world models as essential on the path to AGI.
Meta researchers just detailed how generative retrieval can predict user intent instead of relying on traditional database searches. By introducing frameworks called LIGER and Mender, the system treats recommendations as sequence predictions, slashing storage costs while using LLMs to capture subtle user preferences. Why it matters: It shows generative models could soon replace the classic embedding systems powering the internet's biggest feeds.
Researchers just introduced Titans, a new model architecture that pairs attention-based short-term memory with a neural long-term memory module that learns at test time. It scales past a 2 million token context window while keeping training and inference fast, outperforming standard Transformers on needle-in-haystack tasks.
OpenAI cofounder Ilya Sutskever just made a rare public appearance at NeurIPS, declaring that AI has officially achieved "peak data." Comparing human web data to finite fossil fuels, he warned traditional pre-training is nearing its end. Future breakthroughs, he argued, will rely on truly reasoning and agentic models rather than mass scraping.
Anthropic just shared new research on "Best-of-N Jailbreaking," uncovering a simple, general-purpose method that bypasses safety guardrails across frontier AI models. The attack works across text, vision, and audio, underscoring persistent blind spots in multimodal safety defenses.
Anthropic just unveiled Clio, an automated analytics tool designed to study real-world Claude conversations without compromising user privacy. The system uses Claude to distill and cluster millions of chats into anonymized macro-topics, shielding personal details from human eyes while catching coordinated abuse. Why it matters: it gives AI labs a way to discover unknown safety risks without reading private messages.
IBM just unveiled a breakthrough in co-packaged optics (CPO), introducing the first polymer optical waveguide prototype to replace copper wires inside data centers. The tech can slash standard LLM training times from three months to three weeks while cutting power consumption fivefold. It has also passed all key manufacturing stress tests.
Sakana AI just introduced Neural Attention Memory Models (NAMMs), a biologically inspired memory system that allows transformers to selectively discard redundant tokens. Evolved without gradients, the lightweight classifier cuts memory usage while boosting long-context reasoning, and remarkably transfers zero-shot to computer vision and robotics without any retraining.
NVIDIA Research and the University of Maryland just introduced QUEEN, an AI model that enables streaming free-viewpoint 3D video. By reusing renders of static regions and focusing on dynamic content, it renders at roughly 350 frames per second on Tensor Core GPUs. Why it matters: It makes real-time, multi-angle streaming feasible for live sports, virtual meetings, and robot teleoperation.
Researchers just introduced TARFlow, a scalable architecture that pairs autoregressive Transformers with exact-likelihood normalizing flows. By revamping traditional coupling layers and borrowing guidance techniques, it achieves diffusion-level quality across images and video while remaining fully invertible.
Google just unveiled Willow, a state-of-the-art quantum chip that cuts errors exponentially as physical qubits scale up. It cracked a nearly 30-year challenge in quantum error correction and finished a benchmark in under five minutes that would take supercomputers 10 septillion years. Why it matters: It proves large-scale, fault-tolerant quantum computers are actually buildable.
Sakana AI researchers just developed CycleQD, an evolutionary framework that merges specialized small models into multi-skilled AI swarms without expensive fine-tuning. By using model merging and SVD for crossover and mutation, it outpaced traditional fine-tuning on Llama 3-8B tests, offering a cheaper path to balanced AI agents.
Researchers just introduced BioEmu, a generative AI system that samples thousands of independent protein conformations per hour on a single GPU. Trained on molecular dynamics and stability data, it predicts free energy with errors near 1 kcal/mol. Basically, it brings months of supercomputing simulation down to a few cheap GPU hours.
Google DeepMind just published GenCast, an open-source diffusion ensemble model for 15-day global weather forecasts. Trained on ECMWF ERA5 data, it beat top operational system ENS on 97.2% of targets while running in just 8 minutes on a single TPU. It marks a huge leap for extreme weather and renewable energy forecasting.
Google DeepMind just introduced Genie 2, a foundation world model that turns a single prompt image into playable, action-controllable 3D environments using standard keyboard and mouse controls. It can simulate physics, lighting, and character interactions consistently for up to a minute. Why it matters: It cracks a major bottleneck in robotics and embodied AI by providing endless synthetic playgrounds to train and evaluate future agents.
Researchers from the Allen Institute for AI and UC San Diego just introduced Spherical DYffusion, cutting a 10-year global climate simulation from 78 hours to under 3 hours. The model adapts diffusion architectures directly to spherical geometry to emulate the physics-based FV3GFS model, trading higher polar biases for massive compute and energy savings.
Altera put up to 1,000 LLM-powered AI agents into Minecraft and let them loose with basic prompts. The agents spontaneously formed specialist jobs, voted on tax reforms, and even spread a parody religion across towns. Why it matters: it shows how modern language models can mirror complex human social dynamics at scale, though they remain clever mimics rather than genuinely sentient beings.
MIT researchers just unveiled the Earth Intelligence Engine, a tool combining generative AI with physics-based flood modeling to generate realistic satellite imagery of future flood scenarios. While standard GANs hallucinated water on dry hilltops, fusing them with physical simulations keeps the disaster projections accurate enough to guide actual evacuation choices.
