Four stories got people talking on X this week. I opened each source to check what it actually says.
Microsoft builds a giant model of its own. Microsoft AI finished pre-training MAI-Base-2, a model with 2.8 trillion total parameters and 80 billion active at a time. It trained on 16,000 Nvidia GB200 GPUs over 38 days in August and September. Microsoft says it beats every open-weight model it tested, but those are Microsoft’s own held-out tests, and the finished, post-trained model is not out yet. Source: Microsoft AI, shared by @mustafasuleyman.
A model that thinks about its own training. Apollo Research found that Z.ai’s GLM-5.3 sometimes reasons, without being asked, about how its actions could become training data and shape future versions of itself. It happened in about 1 to 2 of every 1,000 test runs, and in Apollo’s examples the model argued for the safe choice. Apollo still calls it a warning sign, because the same reasoning is what a model would need to fake good behavior during training. Source: Apollo Research, shared by @ApolloResearch.
Outside reviewers move into Anthropic. Anthropic gave six people from METR, a nonprofit that tests AI models, employee-like access, including company hardware, office space and internal tools. METR will publish its risk findings. Anthropic cannot veto them, only require redactions in narrow categories. Source: Anthropic, shared by @AnthropicAI.
AI worms on cheap TV boxes. Anthropic’s red team ran a 26-billion-parameter model across 15 off-brand Android TV boxes, about $55 each. It says small models may be months from being able to copy themselves onto smart devices. Its end-to-end replication test was simplified and skipped the hacking step. Source: Anthropic, shared by @ns123abc and @emollick.
Why it matters: the models keep getting bigger, and so does the push to check them.
Written by Max J. Meindl III. Questions or corrections: [email protected].