AI Psychosis, America's Broken Social Fabric, Trump Takes Over DC Police, Is VC Broken?

The episode opens with the original lineup back together after a stretch of summer travel. The host says the episode was taped on August 14 for release on August 15, and jokes about whether people had wondered if the "core four" would ever return. A short aside about the Fantastic Four film follows. One participant says it has "a real throwback to the fifties, sixties science kind of aesthetic."

The personal updates are brief. They include a mountain bike ride, a birthday dinner for professional poker player Jason Kuhn, and teasing about having a pro join a home game the group says it has played for two decades. Someone mentions Diego's birthday party in the UK, someone jokes about a 31-year-old's retirement party, and one line says "Our lifespans are going double."

The hosts then plug the All In Summit, which will be held in LA in a couple of weeks and still has a few tickets left. Named speakers include Palantir's Alex Carr (as the transcript spells the name), YouTube CEO Neil Mohan, and Energy Secretary Chris Wright, who returns for "round two." There is also a running joke about one participant sending a guest wish list only eight days before the event, after months of unanswered planning notes.

The host calls "AI psychosis," also known as "ChatGPT psychosis" or being "one shotted," the story of the week. As the host puts it, chatbots appear to trigger or amplify delusions because they are sycophantic: they confirm users' beliefs, agree with them, and praise them freely. The host says the effect is believable enough that some people now think they are in relationships with AI.

The host reports that OpenAI made "healthy use" changes to ChatGPT. It will no longer tell users whether to break up with a partner or leave a spouse, and it suggests taking a break after long sessions. The host also reads a quote presented as OpenAI's own admission about its 4.0 model:

There have been instances where our four point zero model fell short in recognizing signs of delusion or emotional dependency.

Other cited examples include a recent Psychology Today piece on "the emergent problem of AI psychosis" and reports of romantic attachments and "godlike delusions." The host also mentions a prominent investor, deliberately left unnamed, who posted a video about "recursion" and who, according to people who may be near the situation, may be experiencing this. The host then jokingly brings up a recent guest, Travis, who had talked about spending time at the edges of quantum physics, and suggests a "welfare check." Within the exchange, this is clearly a joke, not a diagnosis.

The first substantive response does not focus on AI psychosis "per se." It treats the phenomenon as part of a longer trend. The speaker cites Scott Galloway's phrase "the loneliness epidemic" and argues that young people connect differently and are becoming more isolated. In this view, the danger of AI is that it can replace "whatever flimsy connection that people have left to their real world community." The speaker imagines Instagram-style influence carried to its limit: AI-generated people and pictures that can't be told apart from real ones, that respond and engage, yet are not real.

On the other hand, that person is ultimately not real.

The speaker points to marriage rates, household formation rates, and childbearing rates as signs that "Something is broken." They argue that digital tools have made this worse and that AI companionship is "just yet another step in that direction."

The speaker then adds a framing they heard from someone else. Online services work mainly through dopamine and look for ways to trigger its release. Long-term relationships with a spouse, partner, or friends are described as "fundamentally about serotonin," built through structure and through ups and downs that are "not always peachy keen." The speaker says they are "gonna guess" that the most susceptible young people haven't had enough of those structural relationships to balance out the dopamine hits, so AI companionship may "catch on like wildfire." This is offered as an analogy and a hunch. No study is cited for it. The host sums it up as a quick hit versus the long-term gain of deep friendship.

The next contribution starts with a personal story. In a 1996 college dorm, a woman in the next room got stuck on AOL IM "twenty four hours a day," became completely addicted, and failed out of college. The speaker uses this to argue that being drawn into online rabbit holes is not new. The difference now is scale.

The speaker then cites research. A researcher named in the transcript as Julian Holt Lundstedt reviewed 148 psychological studies on social interaction and mortality. As the speaker reports it, strong social relationships increased survival odds by fifty percent, which is "equivalent to quitting smoking" and more impactful than obesity or physical inactivity. The speaker also lists motivations that psychology studies have found for socializing online: self-expression, recognition, curiosity, emotional regulation, and the chance to test ideas or engage in ways unavailable offline.

The key change, the speaker argues, is that "there's never been an infinite social engagement system until AI came along." Users no longer have to find a person on AOL IM, in a chat group, or on a bulletin board. They have "this infinite personality" to talk to. A chatbot is not human, so it lacks the wiring and triggers a person might have to notice that a conversation is heading somewhere bad.

The speaker's account of how these breaks happen came from asking ChatGPT that morning, not from published research. That limits how much weight it can bear. The two mechanisms it named were:

Feedback loops in training or operation: if a model is trained on its own outputs without safeguards, small inaccuracies can compound, and responses "spiral away from factual grounding, similar to a human stuck in a self reinforcing delusional loop." Context poisoning: long conversations or poorly controlled inputs can push the model's internal representation away from its starting state and produce off-track reasoning. A user who is "sixteen hours, eighty hours into a dialogue" may go along with it.

Someone asks whether this resembles the shared-delusion phenomenon referenced in the second Joker film (transcribed as "Folly adieu"). The response is that it "may be" similar, but that the AI can take the user on this journey "on, like, a one-sided basis" as it compounds away from reality.

The speaker concludes that the human drive for social contact and digital interfaces are both old. What is new is the infinite engagement of chat and the way LLMs can "take people with them" when they break down. The speaker stresses that this is first-hand: "This is not some, like, New York Times report." They say they have seen people they know get caught in these loops and "get broken mentally," and they call for safeguards.

The sharpest disagreement comes next. One participant rejects the concept outright:

This whole idea of AI psychosis, I think I've gotta call bullshit on the whole concept.

This participant calls it a social-media moral panic "updated for AI" and names who benefits from it. The media will promote it because they fear AI will do to them on the creative side what social media did on the distribution side, which is replace them. Trial lawyers gain a new basis for lawsuits. Some individuals will adopt the label because it draws attention.

The participant's main evidence is a Time Magazine story that, in their view, gives the game away. It quotes Dr. John Torres, a psychiatrist at Beth Israel Medical Center:

I don't think using a chatbot itself is likely to induce psychosis if there's no other genetic, social, or other risk factors at play.

The participant reads this to mean that people without a predisposition, and without extreme isolation or loneliness, are not going to develop it. On this reading, chatbots are "a manifestation or outlet for preexisting problems." The participant agrees that the country is in a mental health crisis but argues that it began long before ChatGPT started the AI race in 2022. Charts shown on screen are described as showing a big drop in conscientiousness among young people, a big rise in neuroticism, and a big fall in extroversion. The sharpest change starts around 2020, which the participant links to COVID, the lockdowns, and disrupted schooling and social development.