Compute
Google releases vLLM and SGLang engines for its TPU chips
Google reports a cost of 2.01 cents for each million tokens on its Ironwood TPU, the minimum cost on an open benchmark.
What matters in AI.
SubscribeTag
240 stories carry this tag.
Compute
Google reports a cost of 2.01 cents for each million tokens on its Ironwood TPU, the minimum cost on an open benchmark.
Deals
HHS says only Claude gives it the services it has at this time. A new contest will pick a vendor after this award.
Security
In 24 to 48 hours, victims got charges of tens to hundreds of thousands of dollars.
Security
An attacker can make an agent operate a tool with no model check. All 3 vendors made patches.
Agents
Zoom lets other AI assistants use its meeting data, and the user can work without Zoom open.
Benchmarks
GPT-5.6 Sol pushed code in 54% of tests with a local remote and in 65% with a GitHub remote.
Compute
In one test with agents that write code, llm-d found 97% of the kept work that it could use again.
Industry
A new customer can sign in with a Google, GitHub, Apple or Amazon account, and most give no credit card.
Policy
Apple will remove an app that adds a feature after inspection if the feature was not in the version it checked.
Models
One embedding space with 768 dimensions holds text, code, images, video and audio, and one system can find all of them.
Agents
Agents in Foundry can use websites and company sites in a cloud browser, and a person can use Take control.
Security
Check Point Research tells that an attacker with a Microsoft login can open the email, files and tools of a company.