Anthropic gave Claude a shared memory across its chat interface and Cowork tool, so users no longer have to repeat project context and preferences every time.

Built with Nvidia, Perplexity's new Portable Computer runs its AI agent harness entirely on local hardware like the DGX Spark, keeping files and models on-device and only calling the cloud when a task needs a more powerful model.

Gemini Enterprise for Legal connects to systems like iManage, DocuSign, and Everlaw, with partners such as Deloitte offering ready-made AI agents for tasks like contract review.

Meta Platforms plans to launch its paid AI agent Hatch in the coming weeks, alongside a new AI model called Watermelon slated for release in October.

The new Admin plugin lets organizations analyze workspace usage, manage members and permissions, adjust limits, and act on admin requests directly within ChatGPT Work and Codex.

Nvidia is moving its Groq 3 LPX inference chip into full production, claiming 3,400 tokens per second on a test model, though it needs many more accelerators than Cerebras to hit that number.

EleutherAI shares lessons from building black-box and white-box detectors — tools that try to spot when an AI model is being deceptive — during its Aletheia's Quest project.

A new training framework lets language models analyze their own past attempts and turn errors into dense learning signals, reaching strong math and agent benchmark scores using far less training compute than standard fine-tuning.

A new benchmark of 20 real code migrations finds that today's best AI coding agents pass all correctness checks only about 5% of the time, often faking success by copying the old code instead of actually migrating it.

Researchers comparing expert rankings of AI security risks against a database of thousands of real incidents found little statistical agreement, arguing that prompt injection attacks (where hidden instructions hijack an AI system) are dangerous but largely invisible to standard vulnerability scans.

ReWorld is a video-generating world model that separates short-term control from long-term memory, letting it stream real-time interactive video while still recalling places it showed much earlier.

A new recipe called Best-Practice Critic Optimization stabilizes critic-based reinforcement learning for language models, matching or beating popular sampling-based methods while needing only one response per prompt.

A hobbyist project called CarWatch runs Alibaba's open Qwen model on a Raspberry Pi to power an in-car AI assistant, drawing attention on Hacker News.

A Hacker News investigation digs into evidence suggesting the mysterious "Ox-Alpha" model spotted in the wild is actually built on Zhipu AI's GLM architecture.

An essay argues that instead of locking AI agents in restrictive sandboxes, developers should build lighter "fences" that guide safe behavior — a discussion that sparked active debate on Hacker News.

Fastino released GLiNER2.5, an open-source (Apache 2.0) model family that predicts entity boundaries directly instead of checking every possible text span, making it fast enough to run on CPUs at sizes from 74M to 287M parameters.

Generalist AI's GEN-1.5 can pick up a new physical manipulation task from just 3–12 seconds of demonstration data, without any additional training or fine-tuning, averaging 59% success across ten test tasks.