This week's AI news that actually touches tooling choices, costs and risks for a small business.
Anthropic folds Cowork and chat into a single Claude app
The change reaches Pro and Max subscribers first across web, desktop and mobile, and tasks keep running after you close the laptop. Why it matters: fewer overlapping products means an easier rollout for teams that have nobody managing tools full time. simonwillison.net
Google released two new speech-to-speech Gemini Live models
They support interruptible voice conversations straight from a browser over a WebSocket connection, with no heavy libraries required. Why it matters: voice support, bookings and phone pre-qualification become a weekend prototype rather than a big project. simonwillison.net
Mozilla analysis: paying for frontier models buys roughly four months of lead at five times the cost
Open models, many of them from China, have closed much of the capability gap. Why it matters: for routine workloads it is worth checking whether the most expensive model is really needed to reach the same result. arstechnica.com
A Datasette security fix shows how thin data permission boundaries can be
A trailing newline in a table name let someone bypass permission checks and read private rows; patches shipped in two releases. Why it matters: if you are wiring an AI agent into your database, testing the permission logic matters as much as picking the model. simonwillison.net
Analysts flag the risks in the AI infrastructure investment boom
A handful of companies' data centre spending now drives a meaningful share of economic growth, raising sustainability questions. Why it matters: contracts and pricing models may shift, so avoid architectures locked to a single vendor. technologyreview.com
Automated AI accounts are flooding social platforms with low-value posts
A growing share of content comes from agents that openly introduce themselves as AI. Why it matters: organic social reach loses trust, making email, referrals and real-world presence more valuable in marketing again. arstechnica.com
Agility's new humanoid robot stops and squats to avoid hurting human coworkers
That allows deployment in shared workspaces without cages or safety barriers. Why it matters: for warehouses and light manufacturing the entry barrier drops, since expensive safety infrastructure is no longer a precondition. arstechnica.com