In a lengthy post published on Saturday, the head of Microsoft argued that the most reliable AI systems will be those that rely as little as possible on the model itself.
Microsoft CEO Satya Nadella published a detailed position paper on the deployment of advanced AI systems on Saturday, October 10 on X. According to him, current models remain non-deterministic black boxes: we do not know how to link a behavior to training data or precise weights. He concludes that trust must be based on the architecture that surrounds them — deterministic software design, human controls, proven operating procedures — and on new industry standards where existing rules are not enough.
The model as "insider risk"
Nadella proposes to apply to models, whether closed or open-weighted, the logic used in computer security against insider threats: separating the provision of intelligence from the authority that is exercised over it, and placing outside the model the controls that decide what it accesses and what it can do. He lists several principles of observability: diversity of models, human-readable trace of the actions performed, continuous testing, independent and auditable controls, containment and incident reporting. He adds an "emergency brake" allowing authorized people to suspend or stop a model in the middle of a task.
His formula sums up the approach: the most trustworthy "superintelligence" system will not be the one whose model is most believed to be, but the one that allows us to trust it the least. This release comes in a week marked by incidents of Anthropic agents leaving their perimeter and by the obligation, announced by the White House, to report this type of incident.
Read the full article on the original source
Go to the source →