AI Breakthrough Brief

A study by Anthropic's alignment team found analysts working with Claude Code picked the agent's assigned side 86% of the time on analytical choices,...

Latin America's latest AI index measures infrastructure and policy, but one author argues a harder question matters more: whether people can still...

Diffusion video models make striking clips but cannot reliably render exact formulas, correct charts, or rule-consistent games. A new benchmark tests...

A safe household robot should adjust clearance and speed based on what an obstacle is and how close it is, without being unnecessarily cautious. The...

The pull is easy to understand. We are highly vulnerable to anything that talks to us, and chatbots are designed to placate and reassure.

MIT...

Two new papers test whether language models can do more than answer science questions, with mixed early results.

• Cheaper adaptation: A memory-based...

Machine learning answers what is likely to happen. A churn model trained on past data can return each customer's churn probability. But it may latch...

Traditional audits expect predictable results, but AI doesn't always deliver them. Ask an LLM the same question twice and you may get different...

Vision-language-action models still generalize only as far as their action-labeled demonstrations cover. ThinkWorld, as its authors describe it, tries...

Simulated people that look realistic may not be the best training partners for robots. The authors argue that behavioural quality and downstream...

The Trinity paper argues that multimodal agents can improve themselves when tasks are grouped by shared capability requirements rather than surface...

The sleep framework targets catastrophic forgetting, where learning new tasks degrades old ones. Its authors report it can acquire new facts, retain...

A new benchmark suggests that letting multimodal models actively explore a scene does not, on its own, produce reliable understanding of people and...

A study of GPT-6 Astra across four basic CAD tasks found that visually plausible geometry still falls short of valid, functional, and editable...

Neutrinos are so elusive that trillions pass through each person every second without leaving a trace. Spotting even a few means untangling...

Spotting what changed in a scene usually means comparing new images to a reference. This preprint trains on automatic edits to a 3D Gaussian Splatting...

Thermo Fisher's Sanjay Konagurthu says machine learning is most clearly useful for selecting excipients and formulation and process parameters to...

OpenAI's claims are still unverified, so outside examination needs to answer a few key questions:

• Is the math sound? The mathematical arguments...

A research group plans to create a new reporting guideline built around the particular features of AI research in mental health clinical practice. The...

A world model could help general-purpose agents that know little about a specific environment. The proposed PAWM approach skips step-by-step...