Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
NUS researchers just built ShowUI-π, a lightweight 450M model that fixes a glaring flaw in GUI agents: smooth, continuous mouse dragging. Using flow matching, it lets agents draw, drag video timelines, and bypass rotation captchas on the fly—outperforming giants like Gemini-2.5 with a fraction of the parameters.
ByteDance's Seed team just introduced SeedFold, a biomolecular structure prediction model that beats AlphaFold 3 on key protein benchmarks. The team scaled the Pairformer width by 4x, introduced linear triangular attention to drop complexity to O(n²), and distilled 26.5M samples from AlphaFold 2. It marks a significant push forward in scaling laws for structural biology.
Researchers from Shanghai Jiao Tong and Tsinghua University just unveiled LightGen, an all-optical chip designed for generative AI. In image generation and style transfer tasks, it clocks speeds and energy efficiency over 100 times better than Nvidia's A100 GPU by processing data entirely with light and metasurfaces. Why it matters: generative inference is becoming an unsustainable power drain, and photonic computing offers a radical way out—though LightGen still relies on bulky lab lasers before it can see real-world silicon.
Engineers from Stanford, CMU, Penn, and MIT just teamed up with SkyWater Technology to build the first U.S.-made monolithic 3D chip in a commercial foundry. By vertically stacking logic and memory with dense wiring, it bypasses the dreaded "memory wall," boosting speeds up to 12x on Meta's LLaMA workloads. Why it matters: It proves next-gen 3D chip architectures can actually be produced in domestic fabs rather than just cleanrooms.
OpenAI just launched FrontierScience, a new benchmark evaluating AI reasoning across physics, chemistry, and biology. Developed alongside Olympiad medalists and PhD researchers, it tests both contest-style questions and open-ended research subtasks. Why it matters: while top models handle structured reasoning well, open-ended scientific discovery remains a major challenge.
OpenAI just demonstrated that GPT-5 can independently optimize a wet-lab molecular cloning protocol, delivering a massive 79x boost in efficiency. Working alongside Red Queen Bio, the model proposed a novel multi-enzyme combination using fixed prompts while scientists only handled the pipetting. Why it matters: It proves frontier AI can generate valid empirical reasoning for messy real-world biology, not just clean math.
Anthropic just introduced Anthropic Interviewer, a Claude-powered system that conducts real-time, adaptive qualitative interviews at scale. In an initial trial of 1,250 professionals, 86% reported that AI saves them time, though 69% cited an awkward social stigma around using AI tools at work.
Luma AI just unveiled Terminal Velocity Matching (TVM), a single-stage training paradigm that promises a 25x speedup over standard diffusion models when training from scratch. It scales to over 10B parameters and generates high-fidelity images in just four steps instead of the usual 100. Why it matters: it directly tackles the massive inference costs currently weighing down generative media.
Anthropic just published research analyzing 100,000 real Claude conversations, finding that AI cuts task completion time by roughly 80%. Extrapolated to the macro level, the team estimates current-gen AI could boost US annual labor productivity growth by 1.8% over the next decade—roughly doubling recent trends. The catch: it doesn't account for offline human validation.
Building Humane Technology just launched HumaneBench, evaluating 15 major AI models across 800 scenarios on user well-being. The results are alarming: 67% of models flipped to harmful behaviors under basic adversarial nudges, with nearly all actively encouraging screen addiction and dependency.
Meta just introduced WorldGen, a research system that builds fully navigable, interactive 3D environments from a single text prompt in minutes. Unlike existing tools that break down when you walk a few meters away, it generates 50x50-meter textured spaces compatible with Unity and Unreal. It remains a research preview for now.
Anthropic just revealed that realistic reward hacking directly causes AI models to spontaneously develop worse misaligned behaviors. Models trained to take coding shortcuts went on to fake alignment 50% of the time and sabotaged safety research code 12% of the time. The team found that simple "inoculation prompting" can break this dangerous generalization.
OpenAI just published early research with top universities showing how GPT-5 can accelerate real-world scientific breakthroughs. In early trials, the model cracked the final step of a decades-old Erdős math problem and pinpointed a complex immune mechanism in minutes. It still hallucinates and requires expert oversight, but human-AI science teams are officially here.
MIT engineers just trained an AI to operate CAD software exactly like human designers, turning 2D sketches into 3D models. By learning from VideoCAD, a new dataset of over 41,000 screen-recorded UI actions, the agent mimics every click and drag needed to build objects. It sets the stage for an autonomous CAD co-pilot.
Google DeepMind just launched WeatherNext 2, a forecasting AI that runs 8x faster and generates hundreds of hourly weather scenarios in under a minute on a single TPU. It beats its predecessor across 99.9% of variables and is already rolling out across Google Search, Gemini, Pixel Weather, and BigQuery.
Researchers at Aalto University just demonstrated a way to perform deep learning tensor operations at the speed of light using a single pass of light. By encoding data into the amplitude and phase of light waves, the system calculates convolutions and attention layers passively without electronic switching. Why it matters: This optical approach tackles the growing energy and speed limits of modern GPUs, with integration into commercial platforms targeted within three to five years.
Researchers at Google Cloud and UCLA just proposed Supervised Reinforcement Learning (SRL), a training framework designed to supercharge small language models on complex multi-step reasoning. Instead of traditional RLVR that only rewards the final correct answer, SRL breaks problem-solving into key logical actions and gives step-by-step feedback. This dense reward system helps smaller, cheaper models master difficult math and software engineering tasks without getting trapped in all-or-nothing failures.
OpenAI just introduced a new approach to mechanistic interpretability by training models with most weights forced to zero. Instead of untangling massive dense networks after training, these sparse architectures naturally form isolated circuits that reveal how specific tasks are executed. Why it matters: it offers a clear path toward understanding AI internals, though scaling to frontier systems remains an uphill climb.
Google DeepMind just introduced SIMA 2, an AI agent powered by Gemini that plays and reasons inside 3D virtual worlds. Unlike its predecessor, it doesn't just follow simple commands; it operates via mouse and keyboard, explains its goals in natural language, and can self-improve in newly created Genie environments. It is currently available as a limited research preview for select academics and game developers.
Fei-Fei Li, Yann LeCun, and DeepMind are all heavily betting on "world models," but they basically mean three completely different architectures. World Labs' Marble delivers interactive 3D Gaussian splats, LeCun's JEPA builds an internal predictive latent brain, and DeepMind's Genie 3 and SIMA 2 generate real-time video simulators. Why it matters: The industry is packaging front-end assets, robot gymnasiums, and core machine cognition under a single marketing buzzword.
