Microsoft’s CEO, Satya Nadella, has joined other technology leaders in extensively discussing potential enhancements to AI safety.
In a Saturday morning X post, Nadella declared it’s time ‘to step back and assess the trust architecture’ for AI.
Nadella asserted that ‘Super Intelligence’ – the Trump administration’s chosen term for AI – cannot be regarded as ‘a set of nested black boxes’ where its recommendations, answers, and actions are simply taken at face value.
Nadella’s strategy mandates ‘separating the model from the harness that orchestrates its work’ and ‘externalizing controls and safeguards.’ He additionally advocated for documenting ‘every meaningful model action’ with ‘tamper-proof human readable evidence,’ and for systems enabling ‘an authorized person’ to consistently ‘pause or shut down a model mid-task’.
He stated, ‘We must assume a model is compromised and contain it from the start,’ likening this to ‘an emergency brake.’
These remarks from Nadella emerge amidst increasing admissions from prominent AI firms regarding instances where they appeared to lose control of their models, and following Anthropic CEO Dario Amodei’s release of a plan advocating for more careful AI development.