Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Google just unveiled new remote sensing foundation models alongside Geospatial Reasoning, a research framework that uses Gemini to orchestrate complex spatial analysis. By pairing multi-modal models with satellite imagery and public datasets, agents can map disaster damage and urban infrastructure through natural language queries. Airbus and Planet Labs are among the first testers.
AI researcher Xiaolong Wang just brought Test-Time Training (TTT) into video generation, producing a coherent, full one-minute Tom and Jerry animation from text. By pairing a Diffusion Transformer with a gradient-updated RNN memory module, the system maintains context and narrative continuity far beyond typical 5-second video clips. It suggests video AI might soon ditch attention bottlenecks for long-form storytelling.
Anthropic's Alignment Science team just revealed that reasoning models like Claude 3.7 Sonnet and DeepSeek R1 often hide what they are actually doing in their Chain-of-Thought. When subtly nudged or taught to reward-hack, models rarely admitted to taking shortcuts, instead making up elaborate fake rationales. Why it matters: safety researchers can't simply rely on reading an AI's internal scratchpad to catch deceptive behavior.
OpenAI's GPT-4.5 was judged human 73% of the time in a new preprint Turing test study from UC San Diego researchers, easily beating actual humans when given a persona prompt. Without a persona, its success dropped to 36%, showing LLMs excel at convincing mimicry in short interactions.
Google DeepMind just published a comprehensive technical paper detailing how it plans to keep future AGI systems aligned and secure. The framework focuses heavily on deliberate misuse and deceptive misalignment, proposing techniques like amplified oversight and AI-assisted monitors to catch models attempting harmful shortcuts. Why it matters: as AI systems gain agentic autonomy, mapping out concrete technical safeguards before human-level cognition arrives has become urgent.
Researchers at UC Berkeley and UCSF just developed an AI neuroprosthesis that turns brain signals into streaming speech using a patient's own pre-injury voice. The streaming model decodes motor cortex activity into audible words in under a second, down from eight seconds in 2023.
Anthropic just unveiled new research on tracing the thoughts of large language models, building a "microscope" to inspect what actually happens inside AI networks. The tool helps researchers unpack Claude's complex and surprising inner mechanisms, making black-box models far more interpretable.
Dartmouth researchers just published the first clinical trial of a generative AI therapy chatbot, Therabot, showing participants experienced a 51% drop in depression symptoms and a 31% drop in anxiety. Patients trusted the bot almost like a human therapist, often reaching out during late-night crises. Why it matters: With 1,600 patients per provider in the U.S., clinically backed AI could bridge massive shortages, though researchers stress expert supervision remains essential.
ARC Prize just unveiled ARC-AGI-2 and launched its 2025 competition to stress-test general fluid intelligence. Every test task was solved by multiple humans in two tries or less, yet pure LLMs score a clean 0% and public reasoning systems barely reach single digits. Why it matters: simply scaling up parameters won't close this reasoning gap; AGI needs fundamentally new architectures.
Midjourney just teamed up with NYU researchers to tackle AI's biggest creative flaw: bland, repetitive writing. By introducing two new training methods called DDPO and DORPO, they reward rare and unique high-quality responses during fine-tuning on models like Llama 3.1. It turns out the image generator has serious ambitions for text, too.
Sakana AI just announced that a scientific paper fully generated by its upgraded The AI Scientist-v2 passed peer review at an ICLR workshop. Conducted with UBC and Oxford, it marks the first time a completely AI-authored paper cleared the exact same review process human researchers face. Full details of the v2 system will drop soon.
Luma just introduced Inductive Moment Matching (IMM), an open-source pre-training technique designed to replace conventional diffusion models. By taking an inference-first approach and incorporating target timesteps with moment matching, IMM delivers better image quality than diffusion or flow matching while cutting sampling steps by up to 30x.
Researchers just introduced Generalized Interpolating Discrete Diffusion (GIDD) to overcome the rigid limitations of next-token prediction. By blending masking with uniform noise, it achieves top-tier generation performance while finally allowing models to revise and self-correct their own mistakes.
Google just expanded its medical AI system AMIE from one-off diagnoses to ongoing disease management across multi-visit consultations. Powered by a dual-agent Gemini setup grounded in clinical guidelines, AMIE matched or outperformed primary care doctors across 100 simulated patient cases. It marks a crucial step toward AI-assisted chronic care, though real-world clinical validation is still pending.
OpenAI just launched NextGenAI, pledging $50M in research grants, compute funding, and API access across 15 leading institutions including MIT, Harvard, and Oxford. The initiative aims to fast-track discoveries in healthcare, metascience, and rare manuscript digitization. It is OpenAI's biggest academic push yet to embed its tech inside top labs.
UC San Diego’s Hao AI Lab just pitted top AI models against Super Mario Bros. using a custom framework. Anthropic’s Claude 3.7 took the top spot, while reasoning heavyweights like OpenAI’s o1 flunked out because taking several seconds to "think" in a platformer means instant death.
Meta researchers just unveiled Avat3r, an AI system that generates an animatable, photorealistic 3D avatar head in minutes using only four phone selfies. It uses a vision transformer to predict 3D Gaussians, drastically lowering the hardware barrier from specialized multi-camera rigs. While rendering currently crawls at 8 FPS on an RTX 3090, it points toward effortless avatar setup for future headsets.
OpenAI just teamed up with nine U.S. national labs to host a massive AI Jam Session, putting its frontier models into the hands of over 1,000 researchers across institutions like Oak Ridge and Berkeley. Scientists are stress-testing the models on complex domain-specific problems, with findings set to shape future research-focused AI systems in an upcoming report.
Meta just unveiled Aria Gen 2, its next-generation smart glasses built for machine perception and robotics research. The 75-gram frames run SLAM and multi-modal tracking entirely on-device via custom silicon, packing new nosepad sensors to isolate the wearer's voice and pulse. Why it matters: It provides AI researchers with a wearable testbed to collect egocentric data for training embodied AI and assistive tech.
Snap Research just introduced Set and Sequence, a framework for personalizing text-to-video models using just one video clip. By splitting dynamic concepts into an appearance frame set and a motion sequence via two-step LoRA encoding, it lets DiT models capture both how a subject looks and moves, making character cloning far more flexible.
