Watch
Signals to keep an eye onNadella urges controls outside the model
Microsoft CEO Satya Nadella argues that capable models should be treated like potential insider risks—not presumed malicious, but given only bounded access and authority. His proposals include tamper-resistant, human-readable logs, independent verification, continuous testing, and an emergency stop. He warns that one model checking another can leave nested black boxes; permissions should be enforced outside the model and its orchestration harness.
Make permissions, audit logs, and a human-controlled stop part of the agent harness, not model instructions.
OpenAI’s unnamed model produced math results
OpenAI has not named the model behind its recent math results. Andrew Curran says most were produced from a single prompt, with about three hours of thinking per result; he says OpenAI assigned it 4,000 problems. Another commentator says the roughly 400 published results were a byproduct of benchmarking, not a concerted effort to solve mathematics. These are observers’ accounts, not independently verified model details.
Treat the results as an intriguing research signal, not evidence of a generally available model or coding capability.
Zhipu describes a GLM self-training roadmap
An X post says Zhipu founder Tang Jie outlined a GLM-5.4/5.5 roadmap: scaling from 740 billion parameters to more than one trillion, with fully self-training recursive improvement as a goal. A separate post says GLM 5.5 may arrive next week or in late October, but presents that as rumor from unnamed informants. No release, access, pricing, or capability details are supplied.
Wait for a release and independent capability details before planning a model migration.
AI-built open-source alternatives target Adobe apps
A developer says he is building free Rust alternatives to Adobe software under MIT or Apache licenses, using AI tools to speed up reverse engineering. He says token sales for the existing Artcraft visual model and IDE will fund development. The apps remain a work in progress; Ars Technica notes that close imitation of Adobe’s look and interface could raise legal issues.
Explore the project as an experiment, but check its maturity and licensing before adopting it.