Pomerado (YC F26) enables agents to interact with any website by creating deterministic, self-healing APIs in minutes. This solves the problem of slow and unreliable browser automation, which is often restricted by security and compliance policies.
The FDTest method identified that future brain activity depends on past information, confirmed by deviations from the generalized Fluctuation–dissipation theorem. The analysis covered whole-brain models during wakefulness and deep sleep.
A team of 48 AI agents analyzed NASA lidar data to locate a previously unknown Maya site in Mexico. The process from initial idea to official government confirmation took less than 72 hours.
Maximum Inner Product Search (MIPS) is the core mechanism for recommendation and retrieval-augmented generation systems. New approximate methods are proposed to improve speed when handling millions of vector keys.
Researchers demonstrated that virus-like particles function in cells, ex vivo, and chicken embryos. The method enables high-precision gene editing without using Cas9.
The mdlARC model with 100 million parameters achieved 37.08% on the ARC-AGI-2 leaderboard, ranking 6th and scoring three times higher than the next best submission. This result outperforms last year's winning LLM fine-tuned on generated puzzles.
The PhysInfer framework addresses parameter unidentifiability in single-cell RNA-seq by integrating a physics-constrained neural network. It reparameterizes optimization from microscopic rates to interpretable burst size and frequency metrics.
Unsloth released Qwen3.8-27B with optimized streaming performance for long conversations and enhanced UI features. The update includes improved image generation, HTML/PDF chat exports, and new coding agent capabilities via MLX models.
AlphaGenome 2 predicts how genetic variants affect gene expression across hundreds of human cell types, bridging single-cell biology and human genetics. It serves as a tool to uncover cellular drivers of disease.
A new study reveals that the blood-brain barrier actively transports 92 specific proteins from the blood to the brain, acting as an intelligent supply line. This discovery could aid in developing treatments for Alzheimer's disease.
The company will showcase results from KaiNETIC-1, KaiNETIC-2, and KaiNETIC-3 trials, alongside an analysis of unmet clinical needs for adults with class II and III obesity.
An analytic model and Monte Carlo simulations predict state-detection fidelities exceeding 99% within 100 microseconds.
The startup uses simulations based on its own digital systems to prepare models for difficult edge cases, aiming to boost task accuracy beyond current benchmarks.
The author describes developing InfiniteDiffusion, a method for infinite terrain generation using diffusion models, which was accepted at SIGGRAPH 2026. The project was launched as a Minecraft mod and received over a million views.
The PRISM transformer-based decoder matches integer-programming accuracy while offering significantly higher throughput for fault-tolerant protocols with erasure rates up to 50% of the threshold.
Researchers at PSI used the Fable 5.2 model to autoformalize and prove the Aaronson-Ambainis conjecture, confirming that quantum computers can outperform classical ones.
A new Nature study suggests replacing the standard BMI metric with an alternative framework could reclassify obesity status for millions, potentially improving patient care.
Air ambulance firm PHI Health filed a lawsuit against a contractor for nearly $27,000 in medical helicopter bills linked to a retired employee's heart attack. Health care providers are increasingly using litigation to collect on out-of-network charges.
Researchers from the University of New South Wales detected differences in arm movement in people who later received a Parkinson's diagnosis using wrist sensors from the UK Biobank study.
The authors prove that two complete rulebooks cannot exist for certifying safe programs. Any finite set of rules omits at least one compliant program.