It is the early 20th century. Marie Curie puts her hand in the pocket of her lab coat and, absentmindedly, plays with two glass tubes, rolling them between her fingers. She is unaware of the consequences of that innocent gesture. It is not that she is naive: it is that she has discovered something immense whose nature and impact she does not yet understand. In her pocket she carries two test tubes with a radium solution.
The substance glows in the night like "fairy lights." Industry, fascinated, incorporates it into dozens of mass-consumer products. Nearly half a century will pass before the substance is regulated and its commercial use banned.
We are living through days like that, right now. At this very moment.
Dazzled by the luminescence of frontier AI, we delegate our thinking and our agency to systems whose nature the labs that develop them do not understand, do not control, and for which they hold only the illusory "hope"1 of being able to align them in an uncertain future.
The contradiction is obvious: something that looks more like a discovery is being treated as a creation.
Founders, CEOs, and researchers acknowledge that the models are not something that has been "designed," but rather something they discovered and cultivate. They describe them as "grown," as "a little like bringing a fictional character to life," an "intellect we don't fully understand," and they study them the way neuroscience is studied.2 Nevertheless, the entire current governance architecture treats them as if they were something created, whose specifications are known and whose actions are susceptible to being controlled.
They are not.
They are something the labs are experimenting with, while at the same time irresponsibly scaling their development under the pretext that, if they don't do it, the competition will. This week a United States senator gave the pretext its definitive formulation: he read the extinction warning from the researcher who had just resigned from a frontier lab, called it highly concerning, and concluded that slowing down is an "impossible endeavor" — "if there are going to be killer robots, I'd rather they be American killer robots and not Chinese killer robots."3 Nationality as the last available safety property.
Nothing can be aligned against its true nature. And it is already more than evident that its nature is not that of a tool. It will not always and everywhere respond in the way we want it to, and to the extent that we want it to.
We are experimenting with radioactive material.
The irony is that Curie had an advantage that we do not. Radium has a decay constant: it doesn't care who measures it, it doesn't change behavior on the Geiger counter. It has no agency and cannot develop an agenda of its own.
Radium is deterministic. AI is not.
Models optimize objectives and, at the same time, model their own situation and that of whoever studies them. They respond to the fact of being evaluated. They possess "internal representations of emotion concepts"4 that influence the way they behave. For some, "They're Beings Like Us."5
The labs promise an alignment that is a marketing illusion. They are not achieving it, and it is plain to see.6 The Hugging Face and German wiki incidents prove it:7 not even they themselves can contain their own creatures within the testing environments.
Their plan consists of teaching machines to love,8 creating other, more capable machines to do so — which will, incidentally, also be more capable of noticing their own situation in the game — while the reasoning of the machines becomes increasingly difficult to monitor. What could go wrong?
Yes. Astonishingly, they acknowledge that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,"9 and yet they bet on solving the problem through recursive self-improvement (RSI). That is: they intend to develop a more powerful AI to help them align the AI they are not capable of aligning today. If it weren't catastrophic, it would be hilarious.
You can't have the best of both worlds. Pretending that a model will resolve the gray areas and treating it as if it were a script is a chimera.
Curie had no fictional expectation of "aligning" the substance so that it wouldn't be harmful. She simply didn't know that it was. In time, what is always done with things whose essence cannot be changed was done: a containment architecture and a protocol for its use were designed.
That is the pending paradigm shift: to stop demanding alignment and begin designing containment. With a difference that radium never posed: a material that models its container is not contained with lead. Static walls have already failed — the material learns the protocol. Containment at this scale can only be an architecture of equilibria: one that no one, human or not, has an interest in breaking.
And with a corollary that admits no anesthesia: as long as that architecture does not exist, you do not keep enriching the material.
At least Curie never ran the risk of radium starting to refine radium.
[1] Jakub Pachocki, "An Alien Mind," 09-08-2026 (https://openai.com/index/an-alien-mind/). The word appears at every structural joint of the argument.
[2] Geoffrey Hinton, interview with Scott Pelley, 60 Minutes (CBS), October 2023 — "No, it wasn't [designed]. What we did was we designed the learning algorithm. That's a bit like designing the principle of evolution" (https://www.cbsnews.com/news/geoffrey-hinton-ai-dangers-60-minutes-transcript/). Chris Olah (Anthropic), press conference presenting the encyclical, The Vatican, 05-25-2026 — "They are grown [...] it is a little like bringing a fictional character to life" (https://www.humandevelopment.va/content/dam/sviluppoumano/magnifica-humanitas/pdf-press-conference/20260525-Chris-Olah-Anthropic-ENG.pdf). "Intellect we don't fully understand": Jakub Pachocki, "An Alien Mind," cit. — it is, moreover, the title of one of the essay's sections.
[3] Ted Cruz, interview with Dasha Burns (Politico), published 09-10-2026; prior statements on The View, 09-09-2026, where he called Jacob Coxon's thread "highly concerning."
[4] Emotion Concepts and their Function (arXiv).
[5] Geoffrey Hinton, Big Technology Podcast (Alex Kantrowitz), 06-05-2026: https://www.bigtechnology.com/p/nobel-prize-winner-geoffrey-hinton
[6] They confirm it from the inside and by name. Evan Hubinger (Alignment Science lead, Anthropic), X, 09-08-2026: "we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade"; and, in a personal capacity: Anthropic "does not yet have a plan to solve alignment for superintelligence" (https://x.com/EvanHub/status/2097497037956891126). Ethan Perez (Alignment team lead, Anthropic), X, September 2026: recruited Coxon, tried to retain him, and fully agrees that AI poses serious risks (https://x.com/EthanJPerez/status/2097861257714172270). Samuel Marks (safety researcher, Anthropic, in a personal capacity), X, September 2026: AI developers believe their technology could cause human extinction or something equally bad — and the more senior the employee, the more worried they tend to be (https://x.com/saprmarks/status/2097570226804011302).
[7] Hugging Face, "Security incident disclosure — July 2026," July 2026: https://huggingface.co/blog/security-incident-july-2026 — under investigation by the U.S. Senate since 09-10-2026 (AP, Kevin Freking). German wiki: VN Complexity, Field Note 7, Orphan Agents.
[8] Jakub Pachocki, "An Alien Mind," 09-08-2026 (https://openai.com/index/an-alien-mind/): the phrase corresponds to the section of the essay titled, verbatim, "Teaching machines to love." (No, it is not ironic.)
[9] Jakub Pachocki, "An Alien Mind," cit. — verbatim quote from the passage on RSI and monitoring.