Short takes on new AI tools, what they do, who they’re for, and the catch.

Aleph Alpha released Kolibri, an open weight English and German model with 78B parameters and about 3B active. Full weights are free to run on your own machines when the data can't leave the building.
52
Apple will make macOS Full Disk Access require a very explicit yes, saying AI agents substantially raise the risk. Once an agent can see files and messages, that quiet permission really matters.
1
1
90
Ai2's Olmo-core 3 trains mixture of experts models past a trillion parameters about 2.7x faster than Megatron-core. Open labs finally get a realistic path to frontier-scale training without enterprise money.
79
Black Forest Labs released Flux 3 Image, which can edit one part of a picture and leave the rest alone. That's the fix for AI images where one small change ruins everything else.
101
Runway's Praxis-1 learns robot control mostly from third-person video instead of scarce robot demos. Partners are already testing it on their own arms and humanoids, and open weights are planned in the coming months.
93
Cloudflare shipped Clef on Workers AI: decision models that return probabilities, not prose. You pass a state plus yes/no, pick-one, or score questions, and get a probability per option. Clef-flash's median latency is about 39ms, fast enough to gate a tool call.
1
80
Anthropic is putting $100M into training 10,000 engineers to deploy Claude inside companies like McKinsey and Deloitte. It runs like a medical residency. The part I like: you only graduate after leading a real project. anthropic.com/news/claude-fr…
77
OpenAI fired three safety researchers for allegedly sharing confidential company information with a third-party AI safety organization. OpenAI says they mishandled sensitive information outside company procedures. foxbusiness.com/technology/o…
57
AI agents made over 200,000 requests to a U.S. Education Dept site while hunting school stats, including a failed SQL injection. No non-public data was accessed. If you send agents to research public sites, check logs for injection probes. securityweek.com/ai-agents-a…
59
Google's Project Suncatcher prototype launched into orbit today on SpaceX Transporter-18. The team already has contact. They're testing how TPUs handle radiation and heat in space, a first step toward solar-powered AI compute off Earth. blog.google/innovation-and-a…
2
88
Cloudflare AI Search is generally available. Point it at your docs and you get hybrid search, OCR, and an MCP endpoint without building retrieval yourself. Billing starts November. The free month is doing a lot of work. developers.cloudflare.com/ai…
1
1
77
OpenAI logged 16,000 attempts in two days to extract its models' hidden reasoning. The operators didn't breach a database. They just moved the encrypted thoughts into another chat and asked the model to decrypt them.
1
64
Cloudflare's Auto Router cut internal AI spend up to 30% versus always using Sol or Opus. Most prompts never needed the frontier model. We just kept picking the expensive one.
49
Gemini 4 Argon raised its output limit to 1 million tokens, up from 64K. It's rolling out first to cyber defenders with the guardrails off. The rest of us can wait.
77
Shareable profiles bring your Sites and plugins together in ChatGPT, so others can find and reuse what you’ve made. Built it? Show it off.
85
52
1,150
89,932
The biggest AI labs signed a White House deal to self-police. Internal controls, outside auditors, board committees. Enforcement is vibes.
52