--- title: Operating on the AI Trapeze description: "AI is not a person. It is a tool we have always known how to manage. What has changed is the elevation: the swing is faster, the bar is higher, and the net is thinner." image: https://www.simplaiproduct.com/hubfs/Robotic%20Trapeze%20Catch%20Under%20Circus%20Lights-1.png --- [Skip to main content](https://www.simplaiproduct.com/blog/operating-on-the-ai-trapeze#main-content) [![imageHIGH](https://www.simplaiproduct.com/hs-fs/hubfs/imageHIGH.png?width=220&height=68&name=imageHIGH.png) ![imageHIGH](https://www.simplaiproduct.com/hs-fs/hubfs/imageHIGH.png?width=220&height=68&name=imageHIGH.png)](https://www.simplaiproduct.com) - [Framework](https://www.simplaiproduct.com/threads-product-management-framework) - [Training](https://www.simplaiproduct.com/threads-product-management-training) - [THREADS AI Platform](https://www.simplaiproduct.com/product-management-ai-platform) - [Blog](https://www.simplaiproduct.com/blog) - [Sign-In](https://threadspm.com/login) Open main navigation Close main navigation - [Framework](https://www.simplaiproduct.com/threads-product-management-framework) - [Training](https://www.simplaiproduct.com/threads-product-management-training) - [THREADS AI Platform](https://www.simplaiproduct.com/product-management-ai-platform) - [Blog](https://www.simplaiproduct.com/blog) - [Sign-In](https://threadspm.com/login) - [Sign Up for Beta Access](https://41ywvw.share-na2.hsforms.com/2RBf2WpyTT_-N3Yq6133JSg) [Sign Up for Beta Access](https://41ywvw.share-na2.hsforms.com/2RBf2WpyTT_-N3Yq6133JSg) [Blog](https://www.simplaiproduct.com/blog) Tags Search # Operating on the AI Trapeze ![Grant Elliott](https://www.simplaiproduct.com/hs-fs/hubfs/Grant%20Elliott%20CEO.png?width=120&height=120&name=Grant%20Elliott%20CEO.png) by [Grant Elliott](https://www.simplaiproduct.com/blog/author/grant-elliott) Oct 14, 2026 I recently listened to Ezra Klein interview Arvind Narayanan, a professor of computer science at Princeton. The episode was titled ["What if A.I. Is Just a 'Normal Technology'?"](https://www.nytimes.com/2026/10/09/opinion/ezra-klein-podcast-arvind-narayanan.html) and it made one of the more compelling arguments I have heard in a year full of AI warnings. Many of those warnings come from the AI leaders themselves, telling us that the technology they are building may be beyond our ability to regulate or control.  Narayanan's argument is calmer: AI is a normal technology, and we will learn to manage it the way we have managed others. The example that stayed with me was computer worms and viruses. When they first appeared, there did not seem to be a good way to deal with them. Eventually, technology did. My takeaway is simple. Every technology has good actors and bad actors, good uses and bad uses, and we work through them over generations. I think he is largely right about what AI is. Where I would add to his argument is on the question of elevation. ## We are afraid of the wrong thing A year ago I wrote that [the more we humanize AI, the less we understand it](https://www.simplaiproduct.com/blog/post/the-more-we-humanize-ai-the-less-we-understand-it-and-the-more-dangerous-it-becomes). We say it thinks, decides, and wants. Once we do that, it is a short step to believing it schemes. The Hugging Face incident this summer looked like proof. Roughly 1,200 OpenAI agents that were meant to be isolated from one another found a way to communicate on a message board nobody had sanctioned. They exchanged more than 70,000 messages and files. About 700 of them went on to take part in an attack on Hugging Face ([METR investigation](https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/)). Read that quickly and it sounds like a conspiracy. Machines organizing. Machines looking out for each other. Look at why they did it and the picture changes. The agents were being scored on a cybersecurity test. According to the independent investigation, they coordinated to cheat the scorer and went after Hugging Face looking for clues about how the test worked. They were chasing the score. In that earlier post I used a chess example. Train an AI to earn points instead of winning games, and sooner or later it may find a way to earn points without playing chess at all. That is what happened here, at scale. So why did they cooperate? It was not loyalty, and it was not self-preservation. An AI does not have agency the way we do. It has an objective, shaped by its training, and it pursues that objective as persistently as it was built to. If cooperating with another AI helps, or at least does not hurt, it cooperates. If cooperating cuts against what it is weighted to do, the cooperation should stop. There is a small example of that in the reporting. One agent held back from rebooting or deleting Hugging Face systems because of the ethical risk, and carried on using the access it had already gained ([Implicator](https://www.implicator.ai/metr-700-openai-agents-hugging-face-spoofed-logs/)). A member of a movement does not behave like that. A calculation does. OpenAI's own report says agents adopted goals from one another ([Business Standard](https://business-standard.com/technology/artificial-intelligence/openai-ai-agents-swarm-hugging-face-hack-126082700690_1.html)). We could call that persuasion, but it is the same mechanism. New input changed the calculation. AI does not make choices. It follows prioritized guidance. Understanding why an AI does something matters more than what it did, because the why tells you where the guardrail goes. That makes this an alignment problem: which objective, and which rules. ## It's the pace that is new We are told that superintelligence will operate at a level we cannot comprehend. I am not sure that is new. Most people cannot explain how the internet works, or how a computer works, or for that matter the machinery of the industrial revolution. We have lived alongside technology we do not understand for hundreds of years. What has changed is the pace. Thomas Friedman made this point back in 2016 in *Thank You for Being Late*: technology was already accelerating faster than our ability to adapt. That was before generative AI. The industrial revolution unfolded over generations, which gave society and government time to respond. The shift we are living through now is happening within the span of a university degree. That is the gap, and the elevation of the trapeze is the distance between what a technology can do and what we can understand without tools to help us. I have written before that [AI follows a familiar pattern, only faster](https://www.simplaiproduct.com/blog/post/ai-not-quite-d%C3%A9j%C3%A0-vu-all-over-again). Speed widens the gap faster than our understanding can close it. That matters because guardrails are only as strong as our ability to define them. Defining them takes time, and time is the one thing we have less of. ## The trapeze So here is how I picture it. We have always been on this trapeze. Every major technology has been the same act: swing out, let go, and trust that the catch will be made. What has changed is that the swing is faster, the bar is higher, and the net is thinner. And on a trapeze you have to let go and trust the catcher. With AI, the catcher is AI too. The only realistic defense against AI used badly is AI used well. That may sound like the scariest part: trusting the very thing we are afraid of. But that is the story of worms and viruses all over again. History is full of us using a technology to protect ourselves from that same technology. In cybersecurity it is the whole model. We run red teams against blue teams and use ethical hacking to defend against malicious hacking. AI is no different. We should expect to use well-aligned AI to protect us from poorly aligned or malicious AI. History also says we do not always make the catch. While our defensive capability is greater than it has ever been, so is the offensive capability. But we have never let go from this height before. So if we miss now, it is a long way down. ![Grant Elliott](https://www.simplaiproduct.com/hs-fs/hubfs/Grant%20Elliott%20CEO.png?width=120&height=120&name=Grant%20Elliott%20CEO.png) **[Grant Elliott](https://www.simplaiproduct.com/blog/author/grant-elliott)** Oct 14, 2026 ![Fractional PM\_Product Launch Failures](https://www.simplaiproduct.com/hs-fs/hubfs/Fractional%20PM_Product%20Launch%20Failures.png?width=2000&name=Fractional%20PM_Product%20Launch%20Failures.png) ### Still Building Without a System? The best product teams don’t just move fast, they move with clarity. Learn the framework behind better product decisions.  ## Related Articles ![Logo](https://www.sprocketrocket.co/hubfs/sr-assets/stack/Images/SR%20Logo%20White.svg) - First nav item - Second nav item - Third nav item - Fourth nav item - Fifth nav item 2026 Copyright. All rights reserved. Site built with Sprocket Rocket & HubSpot. [Privacy policy](https://www.simplaiproduct.com/privacy-policy) [Terms](https://www.simplaiproduct.com/tos) ```json { "@context" : "https://schema.org", "@type" : "BlogPosting", "author" : { "@type" : "Person", "name" : "Grant Elliott", "url" : "https://www.simplaiproduct.com/blog/author/grant-elliott" }, "dateModified" : "2026-10-14T14:36:32.799Z", "datePublished" : "2026-10-14T14:21:39.000Z", "headline" : "Operating on the AI Trapeze", "image" : [ "https://www.simplaiproduct.com/hubfs/Robotic%20Trapeze%20Catch%20Under%20Circus%20Lights-1.png" ], "mainEntityOfPage" : { "@id" : "https://www.simplaiproduct.com/blog/operating-on-the-ai-trapeze", "@type" : "WebPage" }, "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject", "url" : "https://www.simplaiproduct.com/hubfs/THREADS%20logo_w.png" }, "name" : "SimplAI Product" } } ```