Today's AI chat models handle math and code but stumble on physical intuition and hard constraints, which better prompts can't fix. The fix proposed is agents trained on physics reasoning and checked by tools that enforce physical laws, needing new training data and reward signals.

They tune the prompt-and-tools wrapper around a small model, then instead of copying an expert's whole solution, they rewrite only the step where the model's attempt fails. Copying full runs hurt it on all seven tasks by 4 to 30 points; this fix kept the gains.

Give an AI agent the main open-ended question from a strong unpublished paper, then let the paper's own authors judge its answer. In two six-day trials the agents handled all the coding alone but made no real progress, so both answers were rejected.

Judging a system by how well it plays one game can be faked with endless practice, so intelligence is how efficiently it picks up new skills. A new puzzle test uses only the basic knowledge humans are born with to compare people and AI fairly.

A roadmap for AI that improves itself and the way it improves, moving through stages from carrying out upgrades to choosing its own strategies and adapting to new settings. A new index shows today's language models still fall far short of this goal.

The study asked popular AI chatbots to describe political figures in all six UN languages, then read the moral judgments in their answers. Models from different countries and languages showed different values, matching their creators' worldviews rather than being neutral.

This overview maps where statistics can help make large language models more trustworthy and transparent, from measuring uncertainty to protecting privacy and adapting models to new data. It asks how these text tools could aid statistical analysis, arguing the two fields should work together.

They tested whether AI coding assistants check trust signals—signed releases, build records, channel lists—before installing software, scoring behavior from container logs, not its words. Verification was nearly absent: 9 of 1,920 trials opened any signal, none ran a verification command, so signals changed nothing.

An agent must figure out a hidden rule machine by asking whether strings are accepted and guessing the whole machine, like reverse-engineering a program. Language models solved small machines but failed as these grew, though they knew the standard algorithm and could code it.

A long-context model keeps a fixed-size memory of past words and tracks how sure it is about each stored association, so new information counts more when the memory is shaky. At two sizes, it beat top similar models on word-prediction error and task accuracy.