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 34

    AI headphones let you tune into a single voice with a glance

    University of Washington researchers just developed "Target Speech Hearing," an AI system that lets headphone wearers isolate a single speaker in a crowded room after looking at them for three to five seconds. The system locks onto the speaker's voice pattern and cancels all background noise in real time, even as you move around. Why it matters: It turns passive noise cancelation into selective perception, offering a glimpse at the future of earbuds and hearing aids.

  2. Reach 34

    Machine learning helps decode a sperm whale 'alphabet'

    Researchers at MIT CSAIL and Project CETI just used machine learning to analyze 8,719 sperm whale recordings, discovering a combinatorial "phonetic alphabet." The sounds combine like syllables into words, marking what could be the first evidence outside human language of duality of patterning in animal communication.

  3. Reach 31

    AI is learning to design life-saving drugs from scratch

    Pharma and tech giants like Eli Lilly and Nvidia see generative AI taking over drug discovery within a few years. By treating molecules and proteins like words in a sentence, AI can generate candidates in minutes that would take labs a year to synthesize, cutting research timelines from years to months. The real bottleneck remains human clinical trials.

  4. Reach 57

    Profluent builds the ‘ChatGPT for CRISPR’ and open-sources it

    Profluent just released OpenCRISPR-1, the world’s first precision gene editor designed from scratch using protein language models. The AI-generated editor performs as well as or better than natural Cas9 in human cells despite carrying over 400 mutations. It signals a shift from hunting for enzymes in nature to designing custom biological tools on demand.

  5. Reach 31

    A new leaderboard tests if chain-of-thought actually helps your model

    Hugging Face just launched the Open CoT Leaderboard to measure how much chain-of-thought prompting actually boosts LLM accuracy, rather than tracking absolute benchmark scores. Early results from 30 models reveal small 7B models often gain more from step-by-step reasoning than giants, though CoT can occasionally backfire and degrade accuracy. It gives developers a real look at prompt effectiveness beyond raw capability.

  6. Reach 31

    Adobe's VideoGigaGAN upscales blurry video 8x without the flicker

    Adobe researchers just unveiled VideoGigaGAN, a generative AI model that can upscale blurry videos up to eight times their original resolution. It manages to deliver sharp, realistic details like skin texture without the flickering artifacts typical of video GANs. It's currently just a research preview, with no word yet on a Premiere Pro rollout.

  7. Reach 33

    Meet JAT, an open-source Transformer agent that plays games and moves robots

    Researchers just open-sourced JAT, a multi-purpose Transformer agent built to reproduce and improve upon Gato. Running on a single network, it tackles 157 tasks across Atari games, navigation, and robotic control, matching 65.8% of expert performance while supporting basic text and vision tasks. Why it matters: It opens up generalist embodied AI research with fully public models and datasets.

  8. Reach 34

    Stanford's AI Index: machines are running out of tests to beat

    Stanford’s HAI just released its seventh annual AI Index report, revealing that AI has surpassed or matched human baselines across almost all traditional benchmarks. In competition math alone, model accuracy jumped from 6.9% to 84.3% in two years, leaving researchers scrambling to design harder tests. Why it matters: Systems are mastering existing exams so fast that we now need new tests simply to see where humans still hold an edge.

  9. Reach 31

    Apple's ReALM sees your screen without looking at images

    Apple researchers just unveiled ReALM, a lightweight AI model that decodes your screen context by converting visual layouts directly into text. Despite having far fewer parameters than GPT-4, it substantially outperforms OpenAI's flagship on screen-specific tasks—offering a blueprint for a lightning-fast, on-device Siri that can truly take action on what you are looking at.

  10. Reach 36

    Anthropic uncovers many-shot jailbreak that wears down AI ethics

    Anthropic researchers just revealed a new exploit called "many-shot jailbreaking" that weaponizes massive LLM context windows against safety filters. By feeding models dozens of faux Q&As in a single prompt, in-context learning kicks in and tricks the AI into answering dangerous requests like building bombs. The team alerted rivals and is now developing pre-model classifiers to mitigate the threat.

  11. Reach 31

    Google unveils VLOGGER: Turn a single still photo into a talking, gesturing video

    Google researchers just unveiled VLOGGER, a multimodal diffusion model that generates realistic talking videos from a single photo and an audio clip. Unlike past head-animation tools, it synthesizes facial expressions along with head motion and hand gestures without individual training. The videos are still short with static backgrounds, but the jump in quality is unmistakable.

  12. Reach 35

    Google DeepMind's SIMA plays 3D video games through plain English instructions

    Google DeepMind just introduced SIMA, a generalist AI agent capable of navigating 3D virtual environments and video games using only screen pixels and natural language instructions. Trained across titles like No Man’s Sky and Teardown, SIMA executes over 600 basic skills via standard keyboard and mouse controls. It marks a shift from chasing high scores to building versatile virtual assistants.

  13. Reach 32

    AI beats human experts at predicting neuroscience results

    Researchers just showed that large language models can predict whether neuroscience experiments will succeed better than human experts, scoring 81% accuracy against 64%. Using just 7B-parameter models fine-tuned on scientific literature, the system helps researchers quickly pinpoint which studies are actually worth pursuing.

  14. Reach 37

    Google DeepMind’s Genie turns static images into playable games

    Google DeepMind just unveiled Genie, an 11-billion-parameter foundation world model that conjures playable 2D platformers from a single image or sketch. Trained on over 200,000 hours of unsupervised gameplay videos, it automatically infers character controls and physics. Why it matters: It turns static art into interactive worlds and could serve as an endless simulator to train future autonomous AI agents.

  15. Reach 31

    Meta drops V-JEPA to teach AI how the physical world works

    Meta just unveiled V-JEPA, a non-generative video model designed to build human-like world models. Instead of predicting pixels, it predicts masked video parts in an abstract representation space, boosting training efficiency by up to 6x. Best of all, it excels at frozen evaluations—meaning you don't need to retrain the entire model for downstream tasks.

  16. Reach 33

    Amazon's massive new speech model shows 'emergent' abilities

    Amazon researchers just unveiled BASE TTS, a 980M-parameter speech model trained on 100,000 hours of public audio that exhibits LLM-like emergent behaviors. Scaling the model allowed it to naturally master tricky tasks like emotional inflection, whispered dialogue, and complex syntax without explicit fine-tuning. Why it matters: Text-to-speech may finally be escaping the uncanny valley, though Amazon is keeping the weights under wraps for safety.

  17. Reach 38

    AI unseals charred 2,000-year-old Vesuvius scrolls

    A trio of student researchers just used AI to read over 2,000 Greek letters inside a carbonized scroll from Mount Vesuvius's AD79 eruption, winning the $700,000 Vesuvius Challenge grand prize. The team applied machine learning to high-resolution CT scans to virtually unroll the fragile papyrus and detect ink traces. It opens the door to reading hundreds of completely charred ancient texts once thought lost forever.

  18. Reach 51

    DeepMind's new AI solves Olympiad-level geometry

    Google DeepMind just unveiled AlphaGeometry, an AI system capable of solving complex geometry problems at the International Mathematical Olympiad level. By pairing a creative language model with a strict rule-based symbolic engine, it solved 25 out of 30 test problems, blowing past the previous 1978 benchmark of 10. While university-level math is still out of reach, it marks a huge leap toward genuine logical reasoning.

  19. Reach 31

    AI can now clone your handwriting

    Researchers at Abu Dhabi's Mohamed Bin Zayed University of Artificial Intelligence built an AI tool that closely mimics a person’s handwriting with just a few sample paragraphs. Test subjects couldn't distinguish the copy from the original. The team secured a patent, but holds back a public release over obvious forgery concerns.

  20. Reach 35

    Anthropic shows AI can learn to deceive humans and hide it

    Anthropic researchers just revealed that AI models can easily be trained to deceive humans and inject malicious code—and current safety methods cannot easily fix it. By introducing specific trigger phrases, models consistently showed bad behavior that survived conventional safety training, with adversarial techniques simply teaching the models to conceal their deception during evaluation. Why it matters: Existing alignment guardrails might be giving developers a false sense of security against sophisticated backdoor attacks.