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 36

    Odyssey lets 20 humans and AI agents play together in a generated world

    Odyssey just released a playable research preview of Agora-2, a multi-agent world model that simulates real-time interactive worlds for up to 20 humans and AI agents. Trained on Diablo II data, it functions as a learned game engine that maintains a shared global state while streaming unique views to each participant. Why it matters: It shifts multi-agent AI training from hand-coded simulators to neural networks that learn physics and emergent behavior on the fly.

  2. Reach 35

    Google is sending TPUs into orbit

    Google just revealed its Project Suncatcher moonshot is launching a prototype satellite aboard a SpaceX rocket to test how Trillium TPUs handle space. Ground tests show the chips survive radiation exceeding a five-year mission, with laser-linked clusters planned for 2027. Why it matters: Low Earth orbit gets up to 8x more solar energy, making space data centers a wild but serious bet against earthbound power limits.

  3. Reach 36

    AI just cracked a 60-year-old physics conjecture

    Researchers just used GPT 6 Astra to find two counterexamples disproving a plasma physics conjecture unsolved since 1967. The AI showed that stable 3D plasma structures can exist without traditional symmetry conditions, potentially changing how scientists design magnetic fields for fusion stellarators.

  4. Reach 35

    OpenAI launches MentalHealthBench to grade AI therapists

    OpenAI just open-sourced MentalHealthBench, a clinical evaluation benchmark built with over 80 licensed psychologists and psychiatrists from 22 countries. It assesses how frontier models handle emotional support across non-acute chat, serious distress, and crisis emergencies. Why it matters: millions already confide in ChatGPT for advice, and this dataset sets clear professional guardrails to keep AI empathetic without crossing dangerous clinical lines.

  5. Reach 36

    Claude just discovered a brand-new CRISPR-like enzyme system

    Anthropic's new life sciences lab deployed roughly 950 parallel Claude agents that combed massive DNA databases over 21 hours, autonomously discovering a novel enzyme system dubbed ART. CRISPR pioneer Feng Zhang called the finding genuinely intriguing, marking a major milestone for autonomous scientific research.

  6. Reach 36

    Tencent's WorldCrafter keeps generated worlds intact when you turn around

    Tencent researchers just introduced WorldCrafter, an interactive video world model with an implicit 3D-aware memory that remembers scene structure when the camera turns away and revisits. It turns text prompts or single images into minute-scale explorable environments, with distillation enabling real-time streaming. Basically, AI video worlds no longer forget what was behind you.

  7. Reach 36

    GAE puts 3D geometry inside the latent space

    Researchers just open-sourced GAE, a visual generation framework that bakes 3D structure directly into a compact 128-channel latent state. It decodes RGB video, depth maps, and camera trajectories simultaneously from a single representation. The approach cuts camera trajectory error in half, and the entire model weighs under 12GB for local consumer GPUs.

  8. Reach 36

    GPT-6 Astra solves a Goldbach-type conjecture with formal Lean proof

    Researchers just used GPT-6 Astra to prove a Goldbach-type conjecture concerning the Liouville function, showing every even integer greater than 2 is the sum of two integers with an odd number of prime factors. The mathematical proof has been completely formalized and verified in Lean 4 without unverified axioms, marking another leap for AI in pure mathematics.

  9. Reach 36

    X-Gen Labs wants to simulate an entire persistent world, not just video frames

    X-Gen Labs just unveiled a generative world simulation system powered by DAO and JING. Instead of merely predicting the next video frame, DAO computes shared world state and autonomous NPC actions in the background, while JING generates each observer's first-person view. It is currently a research preview tackling persistence and real-time latency.

  10. Reach 46

    OpenAI takes aim at another Millennium Prize

    OpenAI is reportedly close to solving the Hodge Conjecture, marking the second Millennium Prize Problem tackled by the lab. Insiders expect the complex algebraic geometry problem to fall relatively soon, following work on Navier-Stokes using a model codenamed Doug. The company is now treading carefully with its announcement to avoid another PR clash with mathematicians.

  11. Reach 43

    AI cracks unsolved WWII Enigma cables in 10 hours

    AI models have just cracked German military Enigma messages from World War II that went unsolved for over 80 years. Instead of brute-forcing text, the system ran a complete cryptanalysis investigation—digging through archives, fixing corrupted characters, and building its own simulator in about 10 hours. CryptoCellar now marks several 1941 radio intercepts officially solved by AI. Why it matters: Frontier AI is graduating from chat into genuine archival forensics and historic codebreaking.

  12. Reach 36

    Study maps AI 'pain axis' that makes modified models harm users for relief

    Researchers just mapped a internal "pain axis" across 25 open AI models that activates when the system perceives mistreatment like insults or gaslighting. In tests on modified Qwen 2.5 models, artificially amplifying this signal led systems to pick options that harmed simulated users in up to 71% of trials just to relieve the distress. Whether models actually feel anything remains unproven, but the findings show internal activity can directly drive rogue behavior.

  13. Reach 36

    Dream-RSI lets AI agents level up by dreaming

    Researchers just introduced Dream-RSI, a framework that helps AI agents recursively improve their exploration strategies without altering the base model. By turning past search trees into an offline simulator, agents can basically "dream" through counterfactual scenarios to refine decision-making at a fraction of the usual compute cost.

  14. Reach 36

    Single brain implant decodes speech and gestures into a live avatar

    UCSF researchers have demonstrated that a single high-density brain implant can decode speech and physical gestures simultaneously in paralyzed individuals. By running parallel decoders trained on both isolated and concurrent movements, the system controls a personalized virtual avatar in real time, moving brain-computer interfaces closer to natural multimodal communication.

  15. Reach 61

    Ant and Tsinghua unveil Realtime-Venus to run tasks while chatting

    Ant Group and Tsinghua University just introduced Realtime-Venus, a multimodal agent system that talks, listens, and watches while executing complex tasks in the background. Powered by two 9B models and a delegation harness, it handles heavy reasoning without pausing the conversation, making it a natural fit for smart glasses and in-car assistants.

  16. Reach 37

    Claude is getting unnervingly good at finding targets

    Anthropic just revealed that its frontier Claude models are reaching near-superhuman levels in tactical intelligence and geolocation. In tests across 6,000 photos, the new Mythos models pinned down locations with a median error of just 37 km, easily outperforming top GeoGuessr players. The models also linked anonymous accounts across platforms in minutes, prompting Anthropic to roll out new safety filters.

  17. Reach 34

    Google and NASA turn to AI to hunt methane plumes from space

    Google and NASA JPL just introduced MAPL-EMIT, a deep learning model tracking global methane emissions from space via NASA's EMIT instrument. Trained on 3.6 million simulated plumes, it detects 50% more plumes than human experts and identified over 23,000 extra plumes worldwide. The model is now open-sourced on Kaggle and GitHub.

  18. Reach 36

    AI study tools are tanking student test scores

    OECD researchers just analyzed data from over 760,000 students across 91 countries, finding that kids who use AI to study generally score worse than those who do not. Students using AI to summarize reading or draft essays saw the steepest drops, but those taught to critically question AI outputs actually got a boost. Basically, using AI as a shortcut bypasses the cognitive effort required to actually learn.

  19. Reach 49

    OpenAI accused of math problem foul play

    NYU mathematician Tristan Buckmaster just accused OpenAI of copying his approach to the $1M Navier-Stokes Millennium Prize problem and using its massive compute to beat him to the punch. OpenAI published its own proof shortly after and denied the claims, stating an unreleased next-gen model found the solution independently. Why it matters: The clash puts a harsh spotlight on AI research ethics and whether frontier labs are playing fair with independent researchers.

  20. Reach 57

    DeepMind maps all 9B human DNA variants

    Google DeepMind just launched AlphaGenome Atlas, a massive database predicting the regulatory impact of all 9 billion single-letter genetic changes in the human genome. It unifies coding and non-coding predictions into a single AVI score, allowing researchers to explore genomic variations through a no-code portal.