Tools & Demos 2
The MIT-licensed bridge launches the Computer Use runtime bundled with the ChatGPT desktop app and exposes it through MCP to Claude Code, Codex CLI and Pi. It supports Apple Silicon macOS and X11 Linux, inherits OpenAI's execution and permission policies, and still requires powerful screen-recording and accessibility access; the repository reached 490 stars.
`run.steer()` queues a message for the next turn instead of terminating the current action; an active subagent moves into the background and continues. The same interaction is available in Cursor's UI, where follow-ups wait for the next tool boundary.
HN Threads 1
In 18 Tennessee middle schools, assignment to Khan Academy with its coaching-style AI tutor raised mathematics achievement by 1.3 national percentile ranks per term, about 0.06–0.08 standard deviations over a year—similar to practice without AI. Although 96% tried Khanmigo, the median student messaged it on only one-third of practice days and rarely held substantive maths dialogue.
Industry 1
Later this month, a US test will place clearly labelled ads next to—not inside—generated images. OpenAI is also adding conversion-data integrations and attribution partners, while claiming placement guardrails will keep ads away from emotionally vulnerable or sensitive conversations; the company says ChatGPT reaches 1.2 billion people weekly.
HF Trending Papers 5
The benchmark releases 27,218 messages from ten real human–AI relationships lasting up to 120 days. Only 3.4% of messages needed earlier context, often thousands of messages back; tested systems barely beat chance at detecting when history mattered and inferred more about people than they revealed (256 HF upvotes).
The model learns coloured point trajectories from human videos, then retargets predicted hand motion into robot actions. The authors report that human-video pretraining improved DexJoCo success by 56.9 percentage points and beat the previous state of the art across ten tasks by 11.7 points (19 HF upvotes).
Models must infer scene structure and dynamics from video, then implement deformation, fluids, fracture and other behaviour in code. The evaluation finds that strong static reconstruction does not yet translate into reliable dynamic simulation (19 HF upvotes).
A lightweight world-model branch learns intermediate images for 2D and 3D mental rotation, then uses them in reasoning. It improves by as much as 39.25 percentage points over its fine-tuned backbone; corrupting the generated states sharply reduces performance (9 HF upvotes).
Software-in-the-loop reconstruction asks an agent to reproduce hidden behaviour from public input-output examples, with semantic, structural and anti-shortcut verification. Training Qwen3.8-27B on verified trajectories raised Terminal-Bench 2 from 47.94% to 53.56% (7 HF upvotes).