Models & Capabilities
Model releases, capabilities, evaluations, system cards, and multimodal or agent performance.
-
Models & CapabilitiesRead story
Schools cannot always switch off AI search alone
WIRED reports that some schools want a search option without Google's AI answers. Turning off the answer may also change access to search; the available controls need a local check.
Nov 4, 2026 -
Models & CapabilitiesRead story
That £400 billion AI forecast was not a payout
WIRED links a UK AI growth figure to a Google-backed forecast. It is a projection with assumptions, not money the country has already gained.
Nov 4, 2026 -
Models & CapabilitiesRead story
AI safety tests need someone ready to act on the result
Researchers interviewed 37 experts about tests used to guide AI safety decisions. They found hopes that tests can expose risks and shape company promises. The experts also questioned what tests can measure and who must act.
arXiv cs.CY · Nov 3, 2026 -
Models & CapabilitiesRead story
AI can solve a task and still miss the clock
Jack Clark describes a 50-task test of live web and game actions. The best reported success rate under real time was 36%. More thinking time helped some models, but it did not make a live deadline disappear.
Import AI · Nov 2, 2026 -
Models & CapabilitiesRead story
A helpful AI question can make people check less
In a test with 180 people, AI assistants asked follow-up questions before giving advice. People felt better understood but spent less time checking the replies. They also accepted more wrong answers unchanged.
arXiv · Oct 31, 2026 -
Models & CapabilitiesRead story
The Census Bureau now asks firms about "Super Intelligence"
Simon Willison found a second change to the Census business survey. It now calls AI "Super Intelligence." The first results using that label were not out when he wrote, so its effect is unknown.
Nov 2, 2026 -
Models & CapabilitiesRead story
AI hiring judges may favor the AI that wrote an application
In a controlled preprint test, AI judges favored letters from some AI writing models. Some also favored letters made by their own model, even when candidate qualifications stayed the same. That is a warning for hiring-tool tests, not a measured rejection rate at real employers.
arXiv cs.CY · Nov 2, 2026 -
Models & CapabilitiesRead story
This malware can ask AI which stolen accounts matter most
Cisco Talos studied code that steals login details and active sessions. It can ask AI which victims to target first. Talos did not show that this sample is in use.
Cisco Talos · Oct 31, 2026 -
Models & CapabilitiesRead story
AI can draft an emergency alert. It cannot check itself.
Grist reporting carried by WIRED shows how an AI-drafted warning can sound clear while leaving blanks or getting a location wrong. A qualified person still needs to check facts, recipients, and instructions against official sources before any real alert goes out.
Oct 28, 2026 -
Models & CapabilitiesRead story
More access for AI watchdogs is not the same as more power
Anthropic says six outside evaluators can work inside the company and publish findings under set limits. Zvi Mowshowitz welcomes access but asks how much power they have.
Don't Worry About the Vase · Oct 27, 2026 -
Models & CapabilitiesRead story
A math AI peer should leave room for a child’s own solution
A small Peers4U study explored AI math peers with 22 children. Some groups produced more of their own solutions. That is not proof of lasting learning gains.
arXiv · Oct 27, 2026 -
Models & CapabilitiesRead story
One link is not proof until you check the passage
A paper about trust signals in chatbots leads to a small check: open a linked source and find the passage that backs a low-stakes claim.
arXiv · Oct 23, 2026