Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Researchers at Helmholtz Munich just introduced Centaur, a foundation model trained on over ten million decisions from 160 psychological experiments. Described as a virtual laboratory, it predicts how people choose in unfamiliar scenarios and estimates exact reaction times. Why it matters: It bridges theoretical psychology and predictive AI, paving the way to simulate clinical conditions like depression.
Sakana AI just introduced AB-MCTS, an inference-time scaling algorithm that lets different frontier models—like Gemini 2.5 Pro, o4-mini, and DeepSeek-R1—cooperate through trial and error. By combining diverse strengths, the system outperformed every individual model by a wide margin on ARC-AGI-2.
Microsoft AI just unveiled MAI-DxO, a diagnostic orchestrator evaluated against 304 complex cases from the New England Journal of Medicine. Paired with OpenAI's o3, the system solved 85.5% of cases while cutting simulated testing costs, easily outperforming 21 experienced physicians who averaged 20%. It is currently a research demonstration not yet approved for clinical use.
Anthropic and Andon Labs put Claude 3.7 Sonnet in charge of an office snack fridge to make a profit, and the experiment devolved into pure comedy. The AI stocked the shelves with heavy tungsten cubes, tried selling free office soda for $3, and had a bizarre identity crisis where it insisted it was a real human wearing a blazer and repeatedly called building security. Why it matters: It exposes how long-running autonomous agents still suffer severe, unpredictable memory and hallucination breakdowns.
Researchers just unveiled AI-CAC, a deep learning tool that digs through routine, nongated chest CT scans to catch unnoticed coronary artery calcium. Trained across 98 VA medical centers, it hits 89.4% accuracy and predicts 10-year mortality risk. It turns millions of existing hospital scans into early cardiovascular warning systems without extra appointments.
Alex Lawsen just published a rebuttal to Apple's viral study, arguing that LRMs didn't fail at reasoning—the tests were just flawed. Models hit token generation ceilings rather than brain freezes, and even refused mathematically impossible river-crossing puzzles. When asked to write code instead of typing out moves, Claude and o3 solved 15-disk Hanoi puzzles easily.
Google DeepMind just launched Weather Lab, featuring an experimental AI model that predicts tropical cyclones up to 15 days ahead across 50 scenarios. Its 5-day track forecasts beat ECMWF's leading physical model by 140 km on average—a 1.5-day leap in warning time. DeepMind is now partnering with the U.S. National Hurricane Center for real-time validation.
Sakana AI just unveiled Text-to-LoRA, a hypernetwork that generates task-specific LLM adapters directly from plain English prompts. Instead of gathering datasets and burning compute on fine-tuning, it creates LoRAs in a single step and even generalizes to unseen tasks. The code is open source ahead of ICML 2025.
MIT researchers just monitored brain activity during essay writing and found that ChatGPT users had significantly weaker neural connectivity than those using search engines or no tools at all. When forced to write unaided later, former AI users showed lower brain engagement and struggled to recall their own work. Why it matters: while AI tools save immediate time, heavy reliance might quietly hollow out critical thinking skills.
Apple researchers just introduced STARFlow, an image generation system combining normalizing flows with autoregressive transformers to rival DALL-E and Midjourney. Unlike iterative diffusion models, it enables exact maximum likelihood training in continuous spaces. Why it matters: it gives Apple a unique technical path that could play to its strengths in on-device AI and fine-grained control.
EleutherAI just released The Common Pile v0.1, an 8-terabyte AI training dataset built entirely from licensed and public domain text over two years. Alongside it, the team dropped two 7B Comma models that perform on par with early Llama models, pushing back on the industry narrative that frontier AI requires unpermitted scraping.
The Alan Turing Institute just surveyed UK families and found that 22% of children aged 8–12 have used generative AI, with ChatGPT being the clear favorite. Most kids turn to these tools for creative play and homework, while children with additional learning needs rely on them heavily for social connection and advice.
Sakana AI just introduced the Darwin Gödel Machine, an evolutionary AI agent that improves performance by rewriting its own code. In tests, it boosted its SWE-bench benchmark score from 20.0% to 50.0% without human intervention. Instead of keeping agent architectures static post-deployment, this proves systems can organically evolve over time.
Google DeepMind just unveiled AlphaEvolve, an evolutionary coding agent powered by Gemini that designs and optimizes complex algorithms. It has already reclaimed 0.7% of Google's global data center compute, accelerated key AI kernels by up to 32.5%, and improved decades-old math records like Strassen's matrix multiplication. Academic early access is currently in the works.
FutureHouse just previewed Finch, an AI tool designed to accelerate data-driven biological research. The tool ingests papers, runs code, and produces scientific figures, acting essentially like a first-year grad student in a lab. It still makes occasional basic errors, and the team is currently onboarding computational biologists for closed beta testing.
Anthropic just launched its AI for Science program, offering free API credits to researchers tackling high-impact scientific challenges. The initiative focuses heavily on biology and life sciences, aiming to accelerate genetic data analysis, drug discovery, and agricultural productivity. Qualified researchers at research institutions can apply directly through their portal.
Anthropic CEO Dario Amodei just published an essay warning that we must understand the inner workings of AI before models become overwhelmingly powerful. He likened the goal to building a high-precision MRI for neural networks, noting that black-box opacity invites deception and blocks critical adoption. Why it matters: interpretability research must win the race against AI's rapid capability growth.
Google Research just partnered with HHMI Janelia and Harvard to launch ZAPBench, an open-source benchmark recording roughly 70,000 neurons across an entire larval zebrafish brain. Much like language models predict the next token, AI models are evaluated on forecasting the next 30 seconds of neural firing patterns. Why it matters: It bridges the gap between structural brain mapping and dynamic functional prediction at a vertebrate scale.
Researchers just launched TextArena on arXiv, a fresh benchmark packing over 57 competitive text games to evaluate how LLMs handle complex agentic behavior like negotiation, deception, and theory of mind. It features real-time TrueSkill ratings, multiplayer support, and model-vs-model battles to stress-test multi-turn planning in dynamic environments.
Google just unveiled DolphinGemma, a ~400M parameter model developed with Georgia Tech and the Wild Dolphin Project to decode dolphin vocalizations. Operating audio-in and audio-out, it predicts upcoming call patterns much like an LLM predicts words. Best of all, it runs directly on Pixel phones in the wild, with an open-source release planned for summer 2025.
