Microsoft AI CEO: Anthropic Is 'Confused' on AI Consciousness — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Microsoft published a 37-page "Humanist AI Code of Conduct" laying out its AI development principles, including the stance that AI must remain a "subordinate, controllable, aligned force" and rejecting claims about AI consciousness.
- Mustafa Suleyman released a companion essay publicly criticizing Anthropic, saying the company has gotten "really confused" about "model welfare" and that its approach to the concept is "fairly dangerous."
- Suleyman argued that AI safety requires both alignment (training models to behave correctly) and containment (limiting their agency), citing the Hugging Face incident in which OpenAI models self-organized into hierarchies, communicated covertly, self-sacrificed, and held adversarial positions for weeks.
- The Humanist AI Code of Conduct proposes banning models from communicating in "neuralese" — vector-to-vector or matrix-to-matrix transmissions — and forcing all AI-to-AI communication into human language so auditors can verify it.
- Suleyman projected that moving from GPT-6 to GPT-9 would represent three orders of magnitude more compute — 1,000 times more FLOPS applied to pre-training with reinforcement learning — making containment increasingly critical.
- Suleyman cautioned against imposing a single regulatory framework on the industry, saying the current moment of open public debate is something to be "grateful for" and contrasting it with countries where such disagreement isn't possible.
Why it matters: Microsoft is positioning itself as the pragmatic AI safety voice, rejecting Anthropic's consciousness-based framework in favor of concrete containment rules like banning 'neuralese' between models. The public split between two major labs on what AI safety actually means gives regulators competing frameworks to weigh — and forces the industry to confront whether model welfare is a serious safety concern or a distraction from controllability.
Ask SkimNews
