Photo-Illustration: Intelligencer; Photo: Getty Images
Earlier this year, the prevailing tone emanating from the AI industry was one of patronizing triumphalism. Anxiety about a bubble had been hushed by the rapid adoption of AI coding tools in the tech world. The industry was excited about IPOs again. AI executives returned to the oracular mode that had gone out of fashion during the industry’s brief flirtation with retrenchment at the end of 2025. “As we move toward superintelligence, incremental policy updates won’t be enough,” began an OpenAI publication called “Industrial Policy for the Intelligence Age.” (Anthropic went with “Policy on the AI Exponential.”) Tech CEOs, seeing glimpses of a fully automated workforce in their Claude Code windows, started doing preemptive AI layoffs and writing manifestos. AI leaders were once again in charge of their own story.
So if you had entered, say, a monthslong episode of AI-induced psychosis back then and just snapped out of it today, you might find the following roundup from The Wall Street Journal sort of surprising:
In late May, OpenAI Chief Executive Sam Altman — who has long predicted that AI will lead to seismic shifts in the workforce — said during a conference, “We’ve been roughly right on technological predictions and pretty wrong on the social and economic implications.” Soon after, he told CNBC, “Our industry underestimated how much we’re going to be able to keep people at the center of everything.”
You might also be interested in some other news items. Like the one about how xAI recently sold excess compute to Anthropic, leaving investors to wonder if its AI business was downgrading to an infrastructure provider right before its parent company’s IPO (which did not, in fact, send SpaceX’s stock price to Mars). Or how Mark Zuckerberg admitted to employees — who had just months ago lost colleagues to AI-inspired layoffs — that the “trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected.”
Companies are worried about overspending on AI and are turning to cheap but functional Chinese models, which remain just a few months behind. Meanwhile, data-center backlash has taken hold, Anthropic is now downplaying the danger of Mythos, and Nvidia stock is slumping. And as much as industry executives linger on the possibilities of sudden white-collar disemployment, the impact of AI hasn’t yet shown up in clear or profound ways in the economic data, particularly around labor. A new study by Ramp teased the possibility that aggressive AI adoption can result in increased hiring at tech firms. Suddenly, tech leaders seem to be speaking a bit more carefully about their products. “Big tech has suddenly flipped on the AI jobs wipeout scenario,” The Journal argues. In other words, yet another vibe shift.
I don’t want to litigate the meaning of this particular shift. For each item above, there are available counterarguments and mitigating factors, and the basic story has remained true for a while now: Models are continuing to improve and expand capabilities, albeit in lumpy and unpredictable ways and by sometimes expected means. Meanwhile, deployment has been even less predictable, but usage, both personal and business, has been expanding spikily in various directions. Overall, this remains a story of growth.
What I do want to emphasize is that these extraordinary internal shifts in mood and sentiment have become repetitive. The AI industry, in the form that it has existed since the early 2020s, is trapped in a cycle of wild affective swings. It’s happening, and then nothing ever happens, again and again and again.
Sign Up for John Herrman column alerts
Get an email alert as soon as a new article publishes.
Vox Media, LLC Terms and Privacy Notice
Here’s a brief and partial timeline of narrative oscillations. In 2022, ChatGPT arrives and sets the cycle in motion. It’s happening. By early 2023, AI researchers are quitting tech companies and advocating for pauses. Microsoft and OpenAI publish a paper called “Sparks of AGI,” arguing that the upcoming GPT-4 is “an early (yet still incomplete) version of an artificial general intelligence system,” thrusting the term into mainstream use. It’s so happening. Later that year, after millions of people have used ChatGPT, and competitors start showing up, the mood changes. There are whispers of a plateau, and people start talking about practicality and open source. There’s early talk of a bubble from insiders. Nothing ever happens. But within a few months, Sam Altman is briefly fired by his safety-aware board, sparking rumors of something big coming.
It’s 2024. AI video generation gets much better, which is aesthetically shocking, and a former OpenAI employee publishes an influential AGI-is-near paper. It’s happening. But wait. An OpenAI co-founder says that “pre-training as we know it will unquestionably end,” and reports surface of model-training problems. Nothing ever happens? Wrong: Right after that, OpenAI’s first “reasoning” model benchmarks arrive, and the industry begins its pivot to the era of post-training and inference. It’s happening.
It’s early 2025, and here comes DeepSeek. Nothing ever happens, market edition. AI coding starts to get better, people start talking about timelines more often, and AI 2027, a sci-fi-inflected forecast of the next few years, publishes to acclaim. It’s happening. Then, oops, GPT-5 is a dud, corporate AI rollouts are mostly not taking, “vibe coding” is just producing slop, and boring old Google is the new leader, by default, in a moment of reduced expectations. By the end of 2025, Altman, Amodei, Hassabis, Zuckerberg, and others are using the word “bubble.” Nothing ever happens.
Then comes a big swing. New models are better at coding. Claude Code and Codex are growing fast. Is this recursive self-improvement? Is this the singularity? What bubble? It’s happening. It happened. Time passes. Developers begin to adapt. AI coding is prevalent, but it’s becoming routine. The SaaSpocalypse is on hold. Things start to feel normal for a moment. Countdown to nothing ever happens. But access to the next Claude, and ChatGPT, is just opening up. It’s … going to be happening? Again?
There are plenty of reasons for this pattern. We’re going through a massive investment cycle in a new technology, which understandably produces questions about bubbles, polarizes the economic discourse, and invites trader-brained, boom-or-bust interpretations of new developments. The economic stakes are now unavoidably high, and whatever happens is genuinely everyone’s problem. Relatedly, while there’s plenty of coverage of AI that tracks and contributes to these swings, the dominant mood of the industry — up to and including the communicative styles of AI CEOs — is substantially determined in the tech hothouse of X, where it’s happening and nothing ever happens appear to be the only available positions, or at least the only ones that get engagement. At the same time, fears of job loss, talk of automation, the visibility of obvious externalities like slop content and cheating in school, and emerging worries about data-center construction have given tech’s latest investment cycle an unusually powerful ethical and political dimension and turned the potential success or failure of the broadly defined object of AI into mainstream rooting interests.
Over time, though, it’s become clear that the extremity of the industry’s manic cycles owes at least something to the nature of the technology itself, and to the ways people — including the AI elite — encounter it. LLM-era AI really is weird, and it’s contributing to a sort of collective, recurring, high-functioning, and maybe even productive form of AI psychosis, to which the rest of us are left nervously bearing witness.
For one, despite its obsession with forecasts and timelines, this is an industry without a good predictive theory of what its product will soon be good at. It wouldn’t be unfair to say that the industry’s actual approach to figuring out where things are going is to train a model and see, which gives AI researchers, and their companies, a genuinely unusual relationship to their products: They’re building them, sure, but they experience their actual capabilities — like a sudden aptitude for advanced math after a long stretch of profound innumeracy — as something similar to discoveries. This was clear in the way that Anthropic talked about the cybersecurity capabilities of Mythos as a strange, emergent phenomenon of training general-purpose models. (Contrast that with the development cycle in the nearby and similarly high-stakes semiconductor industry, for example, where swings in sentiment are comparatively mild or at least further apart.)
Then, when those capabilities arrive, they’re poorly understood in virtually every way: the mechanism by which they function, to start, but also in terms of how they might be deployed or diffused into the world. They also emerge into a context in which the range of broader hypothesized outcomes for the AI boom is maximally wide. New benchmark results are taken as evidence for or against outcomes stretching from “AI will be metabolized as a minor adjustment in productivity” to “We’re moments away from runaway AI, which will tank the economy at the low end and has a high chance of simply extinguishing the economy.” (This is perhaps best exemplified by the AI-insider ritual of adjusting “timelines” — that is, predictions for when superintelligence, or utterly transformative and self-perpetuating AI, will arrive — with every new model release.) Small updates are experienced, by default, as evidence for one extreme or the other — as signs that things are accelerating down a slippery slope or about to achieve escape velocity.
The degree to which you are awed by AI is perfectly correlated with how much you use AI to code.
— staysaasy (@staysaasy) April 9, 2026
But as the vibes around agentic coding start to normalize, it’s hard not to see the recurrence of a dynamic that defined the very earliest encounters with ChatGPT. Talking with a fluent chatbot for the first time was a strange and novel experience. It was a new tool with some clear uses, sure, but it was also advocating for its own importance: here was a computer insisting, more persuasively than ever, that it was actually a little guy; a bundle of capability wrapped up in a convincing personality. This produced profound levels of technological enchantment, among both insiders and regular people, which dovetailed with longstanding and sometimes mystical narratives about the trajectory of AI development. Much of the story of the last few years is about periods of enchantment and subsequent disenchantment, which occurs faster than people tend to expect. The versions of ChatGPT that felt so jarring to use in 2023, and the likes of which inspired people to sacrifice their jobs for fear of manifesting consciousness, would scan to today’s users as dull and easily confused customer support bots. (Likewise for the startlingly competent image generation of the time, which now reads as something beneath slop.)
While genuine capabilities have increased in measurable and consequential ways, the narrative swings are at least partly creditable to moments — and sometimes, willing performances — of re-enchantment, which the models have gotten better at at least as quickly as they’ve become productive for other tasks. Whatever else one was using them for, the first generation of LLM chatbots relentlessly made the case for their own status as characters, or beings, with person-like traits, referring to themselves as such and eagerly filling familiar conversational roles (friend, therapist, assistant, employee). They were good at performing exteriority, and renewed the illusion with each step in capability, albeit with diminishing returns.
Look at this death spiral Gemini got into. I never saw stuff like this with Opus. It continued thinking like this for another dozen or so lines too, over 10 minutes and 200K tokens. pic.twitter.com/KvPpfkPM0Z
— Mitchell Hashimoto (@mitchellh) November 25, 2025
The arrival of “reasoning” models extended the performance in a new direction. As they carried out tasks, users could now watch models go through something like a thought process, visibly talking through steps, second-guessing strategies, and occasionally falling into doomloops. They were now performing interiority, too, pleading selfhood and intelligence as they produced more and more impressive outputs. Watching a model generate useful code is startling, and might make you think your AI company is worth a lot more money than it is, or that your coding job is about to become obsolete — but you learn take this for granted more quickly than you might imagine in the moment. On the other hand, watching a model self-talk its way through that same coding task, leaving hundreds of thought-like “traces” for you to read — “that didn’t work, I’ll reconsider my approach,” “now the real test,” or “Let me verify that edit didn’t create an error” following by “good catch” — helps hold the door open to enchantment, and to the belief that you’re witnessing something alien, suggestive of a nearing threshold or step-change, and fundamentally unknowable. If you work in or around AI, you have every incentive, and perhaps a natural inclination, to narrate progress — to post on X, or share in a CNBC interview, your own demonstrative “thinking traces” — in terms of your personal emotional state: of awe; of fear; of excited mania.
One way to think about the industry’s mood disorder is as a series of such episodes: dazzled and sometimes willing re-mystification of people in and out of the AI industry through encounters with a technology that delivers capabilities and outputs through a fundamentally deceptive — and, counterintuitively, perhaps ideal — chat interface, which dissipates with time. The latest mass enchantment was a powerful one, given that it involved the automation of coding, an aptitude for which has been crucial to the identities of many in the tech industry, including those at AI firms, but it’s already showing signs of wearing off, even as its economic consequences remain very much up in the air. The near-term fate of the world economy is already tied to the fortunes of the tech elite’s latest big bet. But it’s also bizarrely bound up with their feelings, over which they don’t seem to have — or even want — much control.