addictedtoAI
menu

blog

Dated, and honest about it

Two kinds of thing get published here: a note, when something happens and somebody is affected by it, and a longer piece, when several things add up to something none of them said alone. There is no posting schedule — a week with nothing worth an enthusiast’s time publishes nothing. A post is true as of its date; corrections are appended and dated, never smuggled into the body.

Sorted by publication date, newest first. RSS

  1. OpenAI's mental-health benchmark scores clinicians' own answers below most models, and its own decomposition says why

  2. Outsiders took Meta's Muse agent apart in its first 16 days, and Meta's answer to the biggest find was "of course you can see the files"

  3. Politico: the White House asked OpenAI and Anthropic to keep new models from Britain's AI Security Institute until the US has reviewed them

  4. Claude agents spotted a new family of phage enzymes. What it does is still unknown.

  5. CLOSEDQUORUM: Talos details a Windows implant that lets four commercial models vote on its next move, in a public build shipped with placeholder keys

  6. An OpenAI agent got past a Medicare portal's blocks in June, and Canberra heard about it from a public inbox in September

  7. Brightband says WeatherNext 3 is the new leader on its live Operational WeatherBench

  8. OpenAI's chief scientist says CoT monitoring is fading, and the Astra card measured it three days earlier

  9. Anthropic and Google each shipped one model in two safeguard tiers in three days, and OpenAI promised the second tier without shipping it. Two of the three describe the permissive cyber tier in the future tense.

  10. Anthropic will put Claude misuse logs in your own bucket. Its support pages say what it keeps.

  11. Nobody had to report the wiki incident, and OpenAI says it is now writing the rule that was missing

  12. The United States told a federal judge that LLM training is fair use. Three of its nineteen numbered pages go after a ruling Meta won

  13. GitSpawn: a repository can name the program your coding agent runs before you type anything

  14. NVIDIA's IOI 2026 run scored 535.4 to the top human's 498.27, and the paper lists its own limits

  15. IFM's K2 Horizon: six Apache-2.0 models with the training lifecycle opened, and a reward-hacking audit of its own scores

  16. OpenAI commits $1B in subsidized Daybreak access for water utilities, grid operators and local governments

  17. GPT-6 Astra shipped, and OpenAI's system card says the model got harder to monitor

  18. OpenAI's Astra is the first model designated Critical for cybersecurity. Advanced access starts with alpha testers

  19. Thomson Reuters says its $40M in-house model is 'on par with the latest frontier models'

  20. GLM-5.3's licence: above $10B in revenue, Model-as-a-Service use waits on Z.AI's security review

  21. Anthropic publishes one government exception to its usage policy. Weapons and domestic surveillance are not in it.

  22. Anthropic emailed Claude users about stolen sessions. What was taken was not a password.

  23. The EU's first AI Act enforcement RFIs are out. The open-source carve-out does not cover the training-data summary they ask for

  24. MiniMax H3's licence excludes the EU, the UK, South Korea and the US

  25. Three accounts of the Hugging Face intrusion, and they begin on three different dates