AI
AI models, agents and the tools people use at work.
Latest news
View all
Akamai Launches Agent Sandbox and Managed Inference in Beta
Akamai's Agent Sandbox isolates code that AI agents write; Akamai Inference adds managed model APIs. Both are in beta.
Google Meet Speech Translation Expands to 63 Languages
Meet now detects the spoken language and translates between any of 63 languages, set per participant.
Googlebook Linux Terminal Preview Puts AI Agents in a Sandboxed VM
Google's Googlebook Linux terminal preview runs commands, and AI agents given sudo, inside a Debian VM.
Microsoft Introduces Copilot Home, Code and Autopilot
The new experiences have different rollout schedules, with Autopilot still in private preview.
OPAQUE Introduces Agent Manifest and Confidential MCP
OPAQUE's June announcement focuses on verifying agent identity, permissions and tool execution.
Google Releases Gemma 4 Under Apache 2.0
Gemma 4 arrives in four model sizes, from smaller device-focused versions to larger workstation models.
Explainers & guides
View all
AI Temperature Explained: Why Answers Vary
How temperature changes token sampling, how top-p differs, and why consistent wording does not prove a fact.
Embeddings Explained: Meaning as Coordinates
How text becomes a vector, how semantic search ranks neighbors, and why similarity still needs evaluation.
Content Credentials Explained: What a C2PA Label Proves
What a signed C2PA record on an image or video contains, what valid and trusted mean, and what it cannot prove.
Prompt Injection Explained: When an AI Tool Obeys a Stranger
How text in an email, page or file can redirect an AI tool, and which controls limit the damage.
Retrieval-Augmented Generation Explained: How AI Answers From Your Documents
How RAG tools search your documents before answering, and how to check what they say.
Context Window Explained: What an AI Tool Can Keep in View
What fills an AI tool's context window, how tokens are counted and why long chats lose track.
Why AI Hallucinates: The Mechanics Behind Confident Wrong Answers
How AI tools learn and how they are scored push them to guess. Where the errors come from and how to catch them.
Before You Give an AI Agent Access to Your Inbox
Reading mail and sending it are different permissions. Start with the job you actually want done.
Why an AI Benchmark Win May Not Help With Your Work
A leaderboard measures a test. Your workflow adds documents, tools, delays and costly mistakes.
About this desk
A model announcement is the beginning of a story. We follow what becomes available, which tasks a product can perform and what access it needs to perform them. Releases and previews are kept distinct, and company claims remain attributed.
Start with our guide to inbox agents if you are deciding what an assistant should be allowed to read or change. The benchmark explainer shows why a leaderboard result may not predict performance on your own documents.
This desk also covers local models, practical workflows and the infrastructure behind AI services. We look for useful details about memory, cost, permissions and verification. Coverage based on documentation is identified as such; it is not presented as a hands-on test. The latest reports and longer guides below give you both the announcement and the context needed to assess it.