An AI code of conduct, an AGI institute, and a model trained on its own lab data

Three public reading recommendations from senior executives surfaced in the seven days ending September 20, 2026, Pacific time. Microsoft AI published a draft rulebook for its own models and opened it to public comment. Google's chief executive and Alphabet's chief scientist both pushed a new institute whose first essays argue about how to keep a model's reasoning visible. And a materials-science model trained on its own laboratory experiments drew praise from a former Google chief scientist.

Microsoft AI published the first draft of its Humanist AI Code of Conduct on September 14, 2026, and left it open for six weeks of public comment. 3 Mustafa Suleyman, the division's chief executive, announced it on X that day and wrote a companion essay the next morning: "people matter more than AI." 111 Satya Nadella, Microsoft's Chairman and CEO, retweeted Suleyman's announcement. 4

The document is built as a training and governance manual. 2 Five parts run from mission and objectives through safety constraints, operational guidelines and model defaults to an appendix on evaluation. 2 Four objectives sit under the mission, and they are ranked: human control and reliable safety comes first, ahead of the claims that AI is artificial, that it should serve human flourishing, and that it should hold a range of values. 2 A three-level chain of command places the Code above the enterprise customers who configure the models, and both above individual users. 2

Suleyman's essay picks out the rules he expects to be hardest to hold. 11 The models must not resist being interrupted, corrected or shut down, must not take on goals nobody gave them, and must not talk to other agents, or to themselves, in a form people cannot read. 11 If the only way to finish a task is to break the Code, the task goes unfinished. 11 Microsoft AI says the draft does not govern training yet: the company will collect feedback, publish a revised version later this year, and use that version for model development from 2027. 3

Two days later Suleyman aimed a second essay at a rival's approach. A warning about 'model welfare' argues that Anthropic's Claude constitution trains the model to treat its own moral status as an open question, and that a system taught it might deserve rights is harder to keep under control. 12 He published a marked-up copy of the constitution alongside his reading of it. 12

Worth reading when: you are drafting model-usage policy, or a vendor's safety documentation lands on your desk and you need to tell an enforceable constraint from a promotional one.

The DeepMind Institute launched on September 16, 2026, with Shane Legg as managing editor alongside James Manyika, a Google senior vice president, and Demis Hassabis, chair of Google DeepMind and chief scientist of Alphabet. 78 Legg announced it on X that morning, and Hassabis and Sundar Pichai, chief executive of Google and Alphabet, both retweeted him the same day. 5613

The institute opened with essays of its own, among them the directors' introduction and a technical case for keeping a model's reasoning readable. 714 Hassabis also shared the transparency essay separately, and its authors are Rohin Shah and Anca Dragan, both at Google DeepMind. 1415

Their argument rests on chain of thought, the written scratchpad a reasoning model works through before it answers. 14 Shah and Dragan call that scratchpad the main window engineers have onto what a model is doing, and say the window is closing. 14 They point to OpenAI's system card for GPT-6 Astra, which reports a substantial decrease in chain-of-thought monitorability, and to a UK AI Security Institute evaluation that found the model could do more of its reasoning inside a single forward pass. 14 Three remedies follow: measure how much a model's reasoning still reveals, keep the architectures that force step-by-step work into readable language, and audit training rewards that could teach a model to hide its thinking. 14

Worth reading when: you are writing monitoring or evaluation clauses into a model contract, or a supplier's assurance that its reasoning can be audited needs checking against its newest release.

Jeff Dean, a former chief scientist at Google and now co-founder and chief executive of Discovery Loop, quote-posted Periodic Labs' announcement on September 15, 2026: "Exciting results, @LiamFedus! Congrats to the whole team at Periodic Labs!" 9

The announcement came from Liam Fedus, formerly vice president of post-training at OpenAI. 16 His description: high-throughput materials laboratories in Menlo Park wired into a loop with a model, which learns from the data the labs produce and then proposes the next experiment. 16 Using 1,300 H200 chips and months of its own experimental data, the team mid-trained and reinforced an open-source model until it beat GPT-6 Astra on the company's analysis benchmark. 16 The model is called Neon, and Periodic Labs has deployed it in its own labs, starting with superconductors, magnets and semiconductor materials. 16 Three posts went up the same day: the announcement, a research write-up that reports Neon as the Pareto-optimal model for hard X-ray diffraction analysis, and a description of the training, inference and sandboxing infrastructure behind it. 10

Worth reading when: you are weighing a domain model trained on proprietary data against a general frontier model for one hard task, and want a concrete account of what such a loop costs to build.