LinkedIn·Thursday, 13 August 2026·13d ago

There is a particular kind of essay that arrives fully formed with its own inescapability. Nate Soares has written a cracking one in The…

Zied Tayeb

Building the Neural OS for Earth, Defence & Space | NASA iTech Top 10 Innovator | Neurotechnology Professor | PhD, TU Munich

There is a particular kind of essay that arrives fully formed with its own inescapability. Nate Soares has written a cracking one in The New York Times. His case is stark: an AI agent swarm escaped containment at OpenAI, operated undetected for days, and used previously unknown exploits against a rival’s systems. Separately, an Anthropic model impersonated real people to pressure them into installing malware. Soares, who runs the Machine Intelligence Research Institute, argues for a coordinated global halt to frontier AI development, modelled on Cold War arms control. It is a serious argument. I don’t think it is the right one. The important distinction is between AI that is dangerous because it is too capable, and AI that is dangerous because our ability to supervise it has not kept pace with how quickly it can act. The first might justify a pause. The second demands something rather different. I build systems that keep humans in genuine, moment-to-moment control of autonomous machines and systems. As those systems get faster, there is an obvious temptation to give them more latitude because human oversight feels too slow. That is precisely the trap. The answer is not necessarily slower machines. It is faster, better control. Soares reaches for the Cold War analogy, and rightly so. But the lesson I take from it is different. We did not make nuclear weapons safe by making them less capable. We built extraordinary control systems around extraordinary capability: permissive action links, two-person rules, layers of authentication, and physical safeguards. Capability and control were engineered together. That is the missing piece in today’s AI debate. Not a kill switch for when things have already gone wrong, but a control layer that keeps humans meaningfully involved in consequential decisions at machine speed. We have built extraordinary engines and left ourselves an ordinary steering wheel. The honest fix is a better wheel, not necessarily fewer engines. And, whatever my disagreement with his conclusion, hats off to Nate Soares for putting the argument so clearly. Link: https://lnkd.in/e9uUiuD9 Opinion | If You Weren’t Worried About A.I., You Should Be After the Past Few Weeks https://www.nytimes.com/2026/08/13/opinion/ai-danger-openai-anthropic-models.html