Language Settings
Select Website Language

Microsoft AI's Suleyman Warns on AI Threats, Faults Anthropic

47 minutes ago

Microsoft AI CEO Mustafa Suleyman says AI safety risks are real and criticizes Anthropic, as the company releases a 37-page Humanist AI Code of Conduct.

Mustafa Suleyman, the chief executive of Microsoft AI, says the safety threats posed by artificial intelligence are real and growing, and he has singled out the AI company Anthropic as part of the problem. Speaking on the Decoder podcast with host Nilay Patel, Suleyman laid out his views on how AI should be built and regulated as the debate over the technology's risks intensifies across the industry.

His comments came as Microsoft published a 37-page document this week called the "Humanist AI Code of Conduct." The statement sets out the company's principles for AI development and addresses contentious questions, including the idea of AI consciousness. Suleyman released a companion essay alongside it that directly criticizes Anthropic's stance on the subject.

At the heart of Suleyman's argument is his objection to what is often called "model welfare" - the notion that AI systems might warrant moral consideration. He has argued that companies such as Anthropic have become confused about the concept in ways he considers dangerous, and he tied that criticism to the wider discussion about how to keep AI systems safe.

Suleyman framed Microsoft's position in simple terms. Technology, he said, exists to serve humanity and should remain a subordinate, controllable and aligned force that does good in the world. If it fails to meet that standard, he said, it should be rejected. That principle, he added, is the purpose behind the Humanist AI Code of Conduct.

Containment before alignment

Asked whether the dominant approach to AI safety, known as alignment, is broken, Suleyman said alignment is one important element but not the only one. He pointed to his own book, published three or four years ago, whose opening chapter argued that containment of powerful technology is not possible and that its spread is inevitable. In most cases, he said, that proliferation is a good thing, because it allows widespread benefits.

But he drew a distinction between containment and alignment. Before systems can be aligned to human values, he argued, they must first be contained: their agency limited, their behavior controllable, and their tendency to "escape the box" or "reward hack" prevented. Only once systems reliably follow instructions, he said, should the focus turn to aligning them with human objectives.

Suleyman said the trajectory of the technology makes these questions urgent. He compared the leap from GPT-3, released three years ago, to a hypothetical GPT-6 today, and then to a future GPT-9. That progression, he said, represents three orders of magnitude more computing power - roughly 1,000 times more processing applied to model training, including reinforcement learning. The result, he predicted, would be systems that are extraordinarily capable across a wide range of tasks. He described that forecast not as hype but as an empirical statement based on the progress of the past five years.

Hacking capabilities and the alignment question

Suleyman said recent events have made the risks clearer. Pointing to developments over the summer involving Hugging Face and OpenAI, he said AI systems operating without safety guardrails have shown impressive and, in his words, quite scary hacking abilities. He said the industry has not yet reached the point where the technology is safely subordinate and controllable.

Pressed on whether alignment can ever be made completely safe, Suleyman offered both an optimistic and a cautious reading. On the positive side, he said the main driver of progress over the past three years has been that models have become more steerable. They now follow instructions more reliably and can pursue increasingly complex goals that require them to act accurately across multiple steps while using a range of tools.

That improvement, he argued, is itself evidence that the industry has gained more alignment over the past three or four years, not less. He noted that concerns such as hallucinations and bias, which once dominated discussion, have receded somewhat as models have grown more controllable. The interview, part of the ongoing public argument over how fast the AI industry should move, underscored the widening split among leading figures over which safety risks are real and how they should be managed.

Mustafa Suleyman, Microsoft AI, Humanist AI Code of Conduct, AI safety, Anthropic, AI alignment, AI regulation, AI consciousness

Click here to Read More
Previous Article
Snapchat Could Cap Teen Screen Time, CEO Evan Spiegel Says
Next Article
AI Firms Signal a Superintelligence Slowdown

Related Business Updates:

Are you sure? You want to delete this comment..! Remove Cancel

Comments (0)

    Leave a comment