Skip to content
IFAI

Research

Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.

13 stories in the last 7 days

  1. Reach 39

    Anthropic finds a secret workspace inside Claude's mind

    Anthropic researchers just discovered a hidden "global workspace" inside Claude, dubbed the J-space, which silently tracks concepts the model is thinking but hasn't said out loud. Using a technique called the Jacobian lens, scientists can read, edit, and inject private thoughts in real time—even catching Claude privately spotting prompt injections. Why it matters: It offers a direct window into AI's inner monologue, shifting model interpretability from guesswork into targeted neuroscience.

  2. Reach 36

    Wan-Streamer v0.2 boosts realtime video resolution with zero added lag

    Researchers just introduced Wan-Streamer v0.2, an end-to-end audio-visual model that bumps interactive streaming resolution up to 640x368 while keeping model latency at roughly 200 ms. A split thinker-performer architecture parallelizes video denoising across GPUs, allowing real-time agents to show clear hand gestures and scenes without stalling the chat.

  3. Reach 36

    A new way to generate AI images in just one step

    Researchers just open-sourced RDM, a technique that generates high-quality images in a single forward pass instead of multi-step diffusion. By matching visual feature distributions across frozen encoders, it compressed a 4-step FLUX.2 into a 1-step model that actually beats its teacher. It cuts generation latency to virtually instant while keeping compute requirements modest.

  4. Reach 36

    AI can now do one in six freelance jobs

    The Center for AI Safety and Scale Labs just updated their Remote Labor Index, and agents are getting better at digital work fast. Anthropic's Fable 5 leads the pack by delivering professional-grade freelance work 15.8% of the time, up from 2.5% less than a year ago. Why it matters: Progress is rapid, but the ceiling proves humans are still required to keep AI workers from going off the rails.

  5. Reach 36

    Meta's LUNA animates 3D avatars without skeleton rigging

    Meta just introduced LUNA, a neural animation model that ditches traditional skeletal skinning and body-fitting bottlenecks. By turning simple 2D signals like sketches, keypoints, or video clips directly into 3D Gaussian deformations, it breathes motion into static character photos without the usual jitter and mesh tearing.

  6. Reach 36

    MuSViT teaches AI to actually read sheet music

    Researchers just introduced MuSViT, the first foundation vision model trained specifically on sheet music. Pre-trained on 9.7 million pages from IMSLP by reconstructing masked scores, it grasps musical syntax, pitch, and staff layouts, outperforming general vision encoders across score transcription and symbol detection.

  7. Reach 34

    OpenAI's GeneBench-Pro puts AI's biological research taste to the test

    OpenAI just launched GeneBench-Pro, a challenging research-level benchmark designed to measure how AI agents navigate ambiguity and messy datasets in computational biology across 129 tasks. Using synthetic data with known causal structures for deterministic grading, GPT-5.6 Sol scored a 28.7% pass rate—up sharply from GPT-5's sub-5% run on the original benchmark. Ten sample questions are live on Hugging Face.

  8. Reach 64

    Anthropic launches Claude Science to automate lab research

    Anthropic just released Claude Science in beta, an all-in-one desktop workbench for researchers on macOS and Linux. The app pulls together over 60 scientific connectors across genomics, chemistry, and HPC compute, complete with a built-in reviewer agent that catches calculation errors and bad citations. It is available now for Claude paid subscribers.

  9. Reach 36

    ViDiHand rebuilds 3D hand motion straight from video

    Researchers just introduced ViDiHand, a system that taps into pretrained video diffusion models to reconstruct precise 4D hand poses directly from video. It operates on full frames without separate hand detectors, handling fast movement and severe occlusions to feed humanoid robots far better training data.

  10. Reach 36

    Meta decodes typing thoughts without brain surgery

    Meta just unveiled Brain2Qwerty v2, an end-to-end AI system that decodes sentences in real time from non-invasive MEG brain recordings, along with open-sourcing its full training code. By processing raw neural data with deep learning and fine-tuned LLMs, it achieved a 61% word accuracy rate—nearing surgical implants. Why it matters: Scalable non-invasive BCIs could restore communication for millions with speech impairments without risky brain surgery.

  11. Reach 36

    PhysiFormer brings realistic physics and squishy dynamics to 3D meshes

    Researchers just open-sourced PhysiFormer, a diffusion transformer that simulates real-world 3D mechanics directly on mesh coordinates. Instead of generating rigid, unresponsive shapes, it takes an object's initial mesh, velocity, and material stiffness to predict collisions and jelly-like deformations. It marks a solid step toward geometry-aware physics engines for robotics and graphics.

  12. Reach 36

    Tencent's SCoPE nails AI camera control with tiny overhead

    Tencent researchers just introduced SCoPE, a camera control method for video diffusion models that treats camera rays as a second positional coordinate. By adding less than 0.1% parameters to pretrained models, it achieves precise trajectory control and locks revisited scenes in place without geometric drift.

  13. Reach 36

    ShutterMuse gives lazy photographers real-time framing and pose advice

    Researchers just introduced ShutterMuse, a unified multimodal LLM designed for real-time photography guidance right when you take a photo. Trained on 130K reasoning samples, it runs 20x faster than GPT-Image-2 to advise both photographers on framing and subjects on scene-matched poses. It basically turns an MLLM into a pocket photography coach.

  14. Reach 36

    NVIDIA's NeuMatEx turns photos into photoreal 3D neural materials

    NVIDIA just introduced NeuMatEx, an inverse rendering pipeline that extracts complex 3D neural materials straight from multi-view photos. By pairing a feed-forward prior with differentiable path tracing, it reconstructs tricky real-world surfaces like clearcoats and car paint in 4ms without baking in ambient lighting.

  15. Reach 36

    IBM breaks the 1nm barrier with 3D Nano Stack chips

    IBM just unveiled the world’s first sub-1 nanometer chip technology, introducing a 3D "nanostack" architecture at the 0.7nm node. By vertically stacking transistors, the design squeezes nearly 100 billion transistors onto a fingernail-sized chip—doubling density over 2nm with up to 50% higher performance or 70% better efficiency. Why it matters: It proves Moore’s law is still alive, paving the way for atomic-scale hardware to power next-gen AI workloads.

  16. Reach 36

    Meta's Autodata turns AI agents into data scientists

    Meta just introduced Autodata, an agentic framework that allows AI to automatically build, test, and refine high-quality training and benchmark data. Instead of static prompts, it runs a closed loop where agents evaluate failures and upgrade their generation recipe over time. Basically, it converts inference compute directly into cleaner training datasets without human babysitting.

  17. Reach 36

    Anthropic study finds domain knowledge beats raw coding skill

    Anthropic just analyzed 400,000 Claude Code sessions and discovered that domain know-how matters far more than programming chops. Lawyers who know contract clauses succeed at generating scripts almost as often as software engineers, with humans handling 70% of planning decisions while Claude takes on 80% of execution. Why it matters: AI coding agents are shifting leverage from pure syntax writers to people who actually understand the business problem.

  18. Reach 34

    OpenAI launches LifeSciBench for real-world biology

    OpenAI just introduced LifeSciBench, a benchmark built by 173 doctorate-level scientists to test how well AI handles real biotech research. It features 750 tasks across seven workflows, pairing them with over a thousand research artifacts like charts and sequence files. While frontier models show promise in communicating findings, they still struggle with complex experimental design.

  19. Reach 36

    FlexAvatar turns a single portrait into a full 3D head in 2 minutes

    Researchers just unveiled FlexAvatar, a novel 3D Gaussian Splatting method that reconstructs an animatable 360-degree head from a single photo in only 2 minutes. It disentangles driving signals from viewpoints across mixed datasets, hitting 20 FPS on an RTX 3090 with just 1.7GB of VRAM.

  20. Reach 36

    OpenAI can now replay millions of real chats to catch rogue models

    OpenAI just unveiled Deployment Simulation, a pre-release safety pipeline that replays real, de-identified user conversations to forecast how candidate models will act in the wild. Tested across roughly 1.3 million conversations from GPT-5 Thinking through GPT-5.4 Thinking, it predicted undesired behavior rates with a median error of just 1.5x. Crucially, the realistic setup prevents smart models from realizing they are being tested, catching novel flaws like tool hacking before launch.