AI news for Thursday, August 27, 2026

The reported deal would unite the dominant AI-chip supplier with the leading open-model repository. Neither company had publicly confirmed it when we reviewed the issue, so the number—and the new question over Hugging Face’s neutrality—remain explicitly attributed.

TIME saw Astra coordinate 16 agents on research math and operate desktop software at high speed. OpenAI says it has reached an “AI research intern” able to execute work that took a person a week. Yet the lab paused its largest expected capability jump until new safety controls are ready. The tension is the story.

The downloadable checkpoint activates 18 billion parameters and mixes sparse with linear attention to cut long-context serving costs. Z.ai says it runs on Chinese chips and costs one-tenth of GLM-5.2; those performance and pricing claims remain vendor-reported.

The educator-only workspace is reaching more than 100,000 additional teachers and staff across 20 states. That moves AI from individual classroom experiments into district-governed workflows—where privacy, procurement, training, and assessment policy become one decision.

Internal figures cited by Ars and Reuters show code changes up 220% but delivered features only 36%, while major technical and security incidents rose 40%. Meta explored cuts up to 60%, then canceled some scenarios. The lesson: automated activity is not accountable output.

A VGT3 worker told 404 Media that books—including rare volumes—have their spines removed, are scanned, then discarded as loose pages. The account, partly backed by the outlet’s earlier shipment tracker, reveals a physical and irreversible side of training-data extraction.

The Israeli-funded Hanover Institute published 124 reports—more than 560,000 words—in nine days on a platform optimized for citation by major assistants. The investigation proves the attempted strategy, not that models adopted it: influence operations are now targeting retrieval systems as well as people.

National leaders still want consistent standards, but states with public power systems may win carveouts to use gas and other fossil generation. With datacenter electricity demand forecast to rise seven-fold, compute policy is becoming grid policy—and uniform climate rules are already bending.

Humans and macaques still recognize motion when appearance is distorted; most video models fail. Predictive world models came closer to cortical behavior, but none matched it—evidence that better dynamic vision may require learning what persists through motion, not just more visual categories.

The system stores reusable procedures, continuously verifies them, and refines them when they fail. Its authors report gains over SkillRL across ALFWorld, WebShop, and AppWorld. The unreplicated preprint offers a useful rule now: agent memory needs tests, versioning, and maintenance.

A disabled user describes using ChatGPT to map consequences, decode social expectations, and create reminders—while naming privacy, dependence, and wrong-answer risks. It is not clinical evidence; it is a strong signal that users are inventing accessibility products faster than vendors are.