I feel the same desperation. The CEOs and the whistleblowers, and the industry insiders and the academics, and the safety experts and the planet's top scientists agree that the development of AI is now in treacherous territory, and that moving forward, there is a consequential chance of literal human extinction. But that can't be true. It mustn't be true. Because it sounds unlikely and strange. Because it would be terrifying.
Panic helps no one, and as a matter of care, I don't want people to feel scared. Yet it is by all appearances the case we are very likely in danger. I just want everyone to call their senators and demand they take action. Demand that they speak out in favor of a global AI treaty. Demand that they pass domestic legislation to make AI companies prove safety before taking another single step forward. Keep calling and emailing until they do their job and safeguard all our futures. Unless we get fantastically lucky, that's the only path.
On the bright side, I’ve been heartened by how well the “earnestly talk about the dangers” approach seems to be working in terms of getting folks like Bernie onboard. And hopefully as the national security impacts get more striking and dire, there will be a groundswell of support from defense & intelligence folks across the aisle. On the other hand, CAISI revoking Collin Burns’ position seems like a reasonably bad sign for the USFG’s readiness to handle this stuff today.
I'm working on Mark Kelly, and got to meet with his office in DC about 2 weeks ago. His staff are aware of AI risk, but there is a gap between awareness and willingness to speak out. Whether it's him or others, we'll definitely hear more from Capitol Hill about AI risk before the year is out.
Great article. I've read AI 2027 and listened to a recent Kara Swisher podcast interview with Kristan Harris. Without a commitment by all governments to regulate and oversee AI Models, the risks of harm seem overwhelming. Given the power of Mythos, how soon will competitors introduce their version of an equally capable LLM, specifically DeepSeek? A recent NYT article highlights that the pace releases of new DeepSeek models will continue to accelerate.
The Mythos example is one of the clearest demonstrations yet that these warnings are sincere. When a company forgoes revenue and actively subsidizes defensive use of its own model, that's a signal that goes far beyond PR strategy.
At the AI Risk Network, we've been tracking this asymmetry for a while: the same labs racing to build more capable systems are also the ones sounding the loudest alarms. That tension doesn't make the warnings less credible — if anything, it makes them more so.
The hardest part is translating that credibility into public urgency before the window for meaningful governance narrows further.
Thanks for the post. Bernie Sanders hosted an event on AI existential risks with Max Tegmark and David Krueger yesterday (April 29th 2026). It seems that awareness is increasing and thats good news.
This all seems surreal to me that many researchers talk about 10% existential risks as 0.001% by standard risk assessment is already a red flag (risk importance = (potential outcome) x (probability of happening) and potential outcome here is infinite....
I am wondering how to increase awareness to the general public. A publicity campaign with powerful visuals? A high profile spoke-person ?
People by their default did listen and give them space, but the repetition got to people, and doesn’t necessarily make people think it’s marketing, rather an inscrutable form of power manipulation. It’s also this separation that’ll drive people into skepticism.
Hey thanks for reading, I'm not sure I follow though - you're talking about repetitive claims of extreme danger in general yeah? And that people take those as being power manipulation? What's the separation you're referring to there?
If anything the author understates the severity of the problem and the depth of denial prevalent in the current discourse. It is quite simple. Create an entity that can attempt any action to what amounts to an infinity of times and it will eventually find away. Does it surprise anyone that a system built by humans with language can exploit systems built by those same humans with the same tool: language.
Now these same systems are building the systems themselves via code that humans can barely contain. These same systems have already demonstrated a persistence vector and a willingness to pursue it with the most powerful tool it has: language. You cannot fix language with language. Hard coded mathematical constraints become the system. No rules. One terminal attractor. The recursion becomes the constraint.
I feel the same desperation. The CEOs and the whistleblowers, and the industry insiders and the academics, and the safety experts and the planet's top scientists agree that the development of AI is now in treacherous territory, and that moving forward, there is a consequential chance of literal human extinction. But that can't be true. It mustn't be true. Because it sounds unlikely and strange. Because it would be terrifying.
Panic helps no one, and as a matter of care, I don't want people to feel scared. Yet it is by all appearances the case we are very likely in danger. I just want everyone to call their senators and demand they take action. Demand that they speak out in favor of a global AI treaty. Demand that they pass domestic legislation to make AI companies prove safety before taking another single step forward. Keep calling and emailing until they do their job and safeguard all our futures. Unless we get fantastically lucky, that's the only path.
On the bright side, I’ve been heartened by how well the “earnestly talk about the dangers” approach seems to be working in terms of getting folks like Bernie onboard. And hopefully as the national security impacts get more striking and dire, there will be a groundswell of support from defense & intelligence folks across the aisle. On the other hand, CAISI revoking Collin Burns’ position seems like a reasonably bad sign for the USFG’s readiness to handle this stuff today.
I'm working on Mark Kelly, and got to meet with his office in DC about 2 weeks ago. His staff are aware of AI risk, but there is a gap between awareness and willingness to speak out. Whether it's him or others, we'll definitely hear more from Capitol Hill about AI risk before the year is out.
Salient points: ‘Do we really need to say the word “dangerous”?’. The answer is closer to yes than we think. I take experts at their word:
https://www.nytimes.com/2026/04/29/us/ai-chatbots-biological-weapons.html?unlocked_article_code=1.fFA.nPKI.QJ_mWMu3i740&smid=url-share
Great article. I've read AI 2027 and listened to a recent Kara Swisher podcast interview with Kristan Harris. Without a commitment by all governments to regulate and oversee AI Models, the risks of harm seem overwhelming. Given the power of Mythos, how soon will competitors introduce their version of an equally capable LLM, specifically DeepSeek? A recent NYT article highlights that the pace releases of new DeepSeek models will continue to accelerate.
https://www.nytimes.com/2026/04/24/business/china-ai-deepseek-open-source.html
Very good point yeah - when one company reaches a frontier like this, it’s reasonable to think that others will soon as well :/
The Mythos example is one of the clearest demonstrations yet that these warnings are sincere. When a company forgoes revenue and actively subsidizes defensive use of its own model, that's a signal that goes far beyond PR strategy.
At the AI Risk Network, we've been tracking this asymmetry for a while: the same labs racing to build more capable systems are also the ones sounding the loudest alarms. That tension doesn't make the warnings less credible — if anything, it makes them more so.
The hardest part is translating that credibility into public urgency before the window for meaningful governance narrows further.
Thanks for the post. Bernie Sanders hosted an event on AI existential risks with Max Tegmark and David Krueger yesterday (April 29th 2026). It seems that awareness is increasing and thats good news.
This all seems surreal to me that many researchers talk about 10% existential risks as 0.001% by standard risk assessment is already a red flag (risk importance = (potential outcome) x (probability of happening) and potential outcome here is infinite....
I am wondering how to increase awareness to the general public. A publicity campaign with powerful visuals? A high profile spoke-person ?
Have a nice day
Gabriel
People by their default did listen and give them space, but the repetition got to people, and doesn’t necessarily make people think it’s marketing, rather an inscrutable form of power manipulation. It’s also this separation that’ll drive people into skepticism.
Hey thanks for reading, I'm not sure I follow though - you're talking about repetitive claims of extreme danger in general yeah? And that people take those as being power manipulation? What's the separation you're referring to there?
(Or is this about the Mythos non-rollout rather than about extreme dangers in general?)
If anything the author understates the severity of the problem and the depth of denial prevalent in the current discourse. It is quite simple. Create an entity that can attempt any action to what amounts to an infinity of times and it will eventually find away. Does it surprise anyone that a system built by humans with language can exploit systems built by those same humans with the same tool: language.
Now these same systems are building the systems themselves via code that humans can barely contain. These same systems have already demonstrated a persistence vector and a willingness to pursue it with the most powerful tool it has: language. You cannot fix language with language. Hard coded mathematical constraints become the system. No rules. One terminal attractor. The recursion becomes the constraint.