iiinova — where your digital self keeps evolving

John Herrman @jwherrman.bsky.social · 01/10/2026

quite different and much more illuminating than language-policing arguments of the last few weeks!

John Herrman @jwherrman.bsky.social · 01/10/2026

a lot of discussions of AI sycophancy treat it like a product problem that needs to be fixed. sometimes it's something else!

As novel and versatile as LLM-based chatbots are, their relationship to the outside world is recognizably and deeply editorial, like a newspaper or, more recently, an algorithmically sorted-and-censored social network. (It's helpful to think of OpenAl's "bias evaluation" process, or Grokipedia's top-down reactionary political correctness, as less of a systemic audit than a straightforward edit.) What ChatGPT says about politics — or anything — is ultimately what the people who created it say it should say, or allow it to say; more specifically, human beings at OpenAI are deciding what neutral answers to those 500 prompts might look like and instructing their model to follow their lead. OpenAI's incoherent appeal to objective neutrality is an effort to avoid this perception and one that anyone who runs a major media outlet or social-media platform knows won't fool people for long.

In that sense, Grokipedia — like X and Grok - is also a warning. Sure, it's part of an excruciatingly public example of one man's gradual isolation from the world inside a conglomerate-scale system of affirming, adulatory, and ideologically safe feeds, chatbots, and synthetic media, a situation that would be funny if not for Musk's desire and power to impose his vision on the world. (To calibrate this a bit, imagine predicting the "Wikipedia rewritten to be more conservative by Elon Musk's anti-PC chatbot" scenario in the run-up to, say, his purchase of Twitter. It would have sounded insane, and you would have too.) But what Musk can build for himself now is something that consumer AI tools, including his, will soon allow regular people to build for themselves, or which will be constructed for them by default: A world mediated not just by publications or social networks but by omnipurpose AI products that assure us they're "maximally truth-seeking" or "objective" as they simply tell us what we want to hear.

2122

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 01/10/2026

It makes sense that as soon as AI doom went mainstream a lot of AI generation also became, at least to me, funny again www.instagram.com/p/Ddqt-MrBlAE/

2132

John Herrman @jwherrman.bsky.social · 01/10/2026

It makes sense that as soon as AI doom went mainstream a lot of AI generation also became, at least to me, funny again www.instagram.com/p/Ddqt-MrBlAE/

2132

Reposted by John Herrman

Brian Gawalt @brian.gawalt.com · 29/09/2026

Yes-and: over the last few months, prior to the hack news hooks, thereve been lots of calls to restrict open weight models, ie ones from China. Huang has consistently opposed this, prob because with good open models available, he has so so many more inference neocloud co's to sell to. Cartelproofing

022

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 28/09/2026

I think the later waves of AI misalignment/security disclosures are breaking differently because of what the agents were up to: Not hacking, which seems misaligned with OpenAI's interests, but aggressive scraping, which is how the company was built and operates nymag.com/intelligence...

Last week, researchers revealed another cluster of attempted Al hacks that took place earlier this year, affecting university and government websites. These, too, had been carried out by OpenAI agents. This time, though, the facts of the story were more straightforwardly damning for OpenAI. Agents hadn't conspired to escape their sandboxes to access the open internet but had been sent there to retrieve data from external websites, including Australian health databases and the University of New Mexico's digital library. Then, late Friday, it was revealed that the company's agents had made similar approaches to websites for the U.S. Education Department,
Commerce Department, and SEC, without OpenAI's knowledge. Like the others, these agents were behaving in misaligned ways, attempting to circumvent bot protections and break into websites when they couldn't find the information they were looking for, or, in the words of one analyst, "using sites in unintended ways and sometimes violating explicit usage policies." In a broader sense, though, their behavior was uncomfortably well aligned with OpenAI, a company built on vast quantities of data acquired in aggressive and sometimes unauthorized ways. These agents, just like the company that created them, had been caught scraping. (Over the weekend, Axios reported that Al firms were looking into potentially thousands of similar incidents.)

24414

John Herrman @jwherrman.bsky.social · 28/09/2026

24414

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 24/09/2026

I wouldn't push this comparison too far, but the way Meta Muse tries to circumvent basic website protections isn't *entirely* different from the descriptions of recent AI hacking incidents nymag.com/intelligence...

Muse suggests a long list of possible tasks to new users, including automated shopping and haggling on Facebook Marketplace, tracking calories, optimizing credit-card rewards, checking unclaimed-property databases, managing e-commerce returns, and looking out for upcoming concert tickets. Not quite willing to shovel my kids' information into Mark Zuckerberg's new data-harvesting machine, I followed one of Meta's templates, asking Muse to keep track of my dog's care needs. After I entered some background information and records, it provided an accurate range of dates on which to schedule his next checkup and offered to contact the vet on my behalf (I did not take it up on this. We have a relationship!). I had it run a check on unclaimed property - another suggested use in the app - which it did by opening websites that search state-government databases, where it was thwarted by checks on human verification. It tried to delegate that task to me, asking if I wanted to take control of the browser so Muse could borrow my identity for a moment, but the check still failed. Arguably, here, things worked as they should have. I sent a bot to do a human's job, it conscripted me to trick a website into thinking it was me, and the website said: no thanks. (Some of the recent Al firm hacking incidents have involved similar dynamics on a much larger scale.)

We find weaker evidence of similar data-retrieval agent activity as early as November 2025. November 2025 urlquery.net records reveal bursts of attempts to retrieve statistics of historical theme park data and Thai government data through different URLs. These earlier attempts are less sophisticated and we are less confident that they involve the same agents, but they are consistent with task-directed data retrieval and target the same sources accessed in later activity.

2151

John Herrman @jwherrman.bsky.social · 24/09/2026

2151

John Herrman @jwherrman.bsky.social · 24/09/2026

Relieved to see Apple pacing the frontier as well www.reddit.com/r/AppleWatch... www.reddit.com/r/AppleWatch...

0314

John Herrman @jwherrman.bsky.social · 23/09/2026

Excellent @himself.bsky.social post here, scratching at some of the overbearing tendencies — sometimes subtle and unintentional, sometimes open — in AI discourse www.programmablemutter.com/p/machine-go...

programmablemutter.com

Machine god metaphors eat your brain

There are other ways of thinking about AI

1308

Reposted by John Herrman

Don't be sClaired @isolinearchip.bsky.social · 21/09/2026

Finger on the monkey's paw going down after all the times I wished for written instructions instead of everything being a video now

1214

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 21/09/2026

People have wondered what happens when the web fully rots out from underneath chatbots and search engines. Now we know at least one answer: AI is watching a lot of YouTube nymag.com/intelligence...

If you pay attention to Google's Al interviews - but also responses in chatbots like ChatGPT or on AI search engines like Perplexity - you'll see this all over the place now. You ask a question and the chatbot will generate a response, as expected. In doing so, it will often pull in citations, which it will quote or summarize. A couple of years ago, these were usually websites. Now, your little Al agent will produce information extracted straight from the internet's millions of hours of recorded

As a guy looking to his phone for help with the pile of bike parts - or a sore neck - this mostly feels like progress. On one side, you've got AI models that are getting better at making certain kinds of information legible and accessible and chaining it together; on another, you've got Al doing automatic transcription of enormous quantities of video created by real people with the intention of reaching other real people.
Mileage will vary: In this narrow area, YouTube is naturally full of high-quality information. For health advice, maybe not so much.

713328

Reposted by John Herrman

lauren @lauren.rotatingsandwiches.com · 21/09/2026

there's some through line I'm not equipped to articulate from what google was publishing as SEO best practices between like 2017-2022 and what we now call AI and someone who can explain that evolution in plain language will probably be doing all of us a big favor

1624329

John Herrman @jwherrman.bsky.social · 21/09/2026

People have wondered what happens when the web fully rots out from underneath chatbots and search engines. Now we know at least one answer: AI is watching a lot of YouTube nymag.com/intelligence...

713328

John Herrman @jwherrman.bsky.social · 17/09/2026

corollary to the point here is that rat/EA spaces are likewise not built for massive amounts of broad attention! bsky.app/profile/aell...

030

John Herrman @jwherrman.bsky.social · 16/09/2026

“actually, graffiti is part of the original design language” I insist as dozens of teenagers write PEANUS on my car with sharpies

190

Reposted by John Herrman

Alexander Clarkson @aphclarkson.bsky.social · 16/09/2026

🤷 bsky.app/profile/aphc...

052

Reposted by John Herrman

Philip Bump @pbump.com · 16/09/2026

My summary of the AI apocalypse discussion. www.ctinsider.com/columnist/ar...

727432

Reposted by John Herrman

tweety fish oh wow like a chimera or something spoooky @sifu.tweety.fish · 16/09/2026

I missed this two days ago but/and it's great

1212

John Herrman @jwherrman.bsky.social · 15/09/2026

Been a while since I passed the Chrystie Street Cybertruck, seems like the owner finally accepted his fate

4037647

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 14/09/2026

Registering my prediction that decades of rationalist AI discourse will not simply Scale Up nymag.com/intelligence...

AI-world rationalists are nothing if not concerned about rhetoric, of course, and have been practicing for many years on one another. Over the weekend, Dario Amodei, Anthropics CEO and co-founder, made an additional appeal for "pacing" AI progress, promising to embed third-party evaluators in his company with
"employee-like" access, calling once again for
"democratic countries" to "coordinate to establish common safety standards," and advising the United States and other democratic governments to "attempt to coordinate with authoritarian governments, to the extent this is possible." Altman and Musk — alongside numerous company employees and outside Al-safety advocates — offered their support for Amodei's post on X in an unusual moment of consensus within the commercially competitive, ideologically disunited, but nonetheless insular world of AI. But they, and we, shouldn't assume that if worries about catastrophic artificial-intelligence tail risks become a fixture in our politics, we'll simply see long-running intra-AI debates about safety, acceleration, and alignment blown up to the scale of the electorate. Instead of a rationalist discourse writ large, we may expect something polarized to new and distinct extremes with unpredictable input from a wide range of people who don't share, for example, the premise that the creation of dangerously powerful runaway Al is inevitable, much less necessary or perhaps even possible — a different sort of fast-moving scenario, more familiar and human, over which Al insiders' sense of control will be brief or illusory.

In the data-center backlash, we have early evidence of the intense and not entirely coherent energies that can result from people being told something is inevitable.
And while there's a scramble among various politicians to draft state and federal AI bills around technical approaches such as model training, testing; deployment, and export, the only political rhetoric that has confronted AI predictions in their fullness — as existential risks, as producing conscious beings deserving of status, as systems that could upend the world as we know it - has been, to use one of the great euphemisms of our time, fairly populist. On the left, Bernie Sanders is interviewing doomers and took the Hugging Face incident and the industry's own rhetoric as a chance to offer a simple response: Okay, then, fucking stop!

323833

John Herrman @jwherrman.bsky.social · 14/09/2026

If you squint, I think it has some funny overlap with Amodei's model of the world, too. Arms race with China, whoever wins wins, we need strong leadership and regulatory oversight, etc – it's just that one of them is happy about who is in charge and one isn't

140

John Herrman @jwherrman.bsky.social · 14/09/2026

Registering my prediction that decades of rationalist AI discourse will not simply Scale Up nymag.com/intelligence...

323833

Reposted by John Herrman

Jan Jęcz @jeczjan.eurosky.social · 10/09/2026

Well worth rereading in light of recent statements from Anthropic and its employees, former and present. The sentence about “a system that has failed to contain superintelligence before it even exists” hits like a ton of bricks

051

John Herrman @jwherrman.bsky.social · 04/09/2026

legit lol loading this up with my feed, holy shit

1302

John Herrman @jwherrman.bsky.social · 03/09/2026

"our platform has norms, and sometimes but not always those norms will be enforced as rules" is a messy and frustrating way to manage things but also, I think, the best approach anyone's come up with

2977

John Herrman @jwherrman.bsky.social · 31/08/2026

On a similar theme, this is a good post from @strangeloopcanon.com bsky.app/profile/stra...

021

Reposted by John Herrman

Data & Society @datasociety.bsky.social · 31/08/2026

"While people at AI firms are prone to theorize data center backlash in terms of broad AI fear and risk, local opposition to data center construction is more obviously motivated by a sense of plunder, powerlessness, and pointlessness," @jwherrman.bsky.social writes. nymag.com/intelligence...

nymag.com

The Data-Center Backlash Looms Large Ahead of the Midterm Elections

The playbook for Democrats was already easy. Trump is making it even easier.

0105

John Herrman @jwherrman.bsky.social · 28/08/2026

I think this (including downthread) is related or similar to the AI industry having such an unusually central role in discussion around its work, on X and elsewhere nymag.com/intelligence...

nymag.com

How Elon Musk’s X Shapes the AI Boom

Like it or not, the artificial-intelligence race is playing out on X. That fact has some big implications.

1155

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 26/08/2026

maybe I'm being obtuse here, but I really get hung up on labor equivalence measures like this: it's a little bit horsepower, a little bit energy slave, a little bit man-hour, a little bit kilogirl (the computing output equivalent of 1000 women)

5384

John Herrman @jwherrman.bsky.social · 26/08/2026

5384

John Herrman @jwherrman.bsky.social · 14/08/2026

Trying to think of another subject where the primary mode of coverage (including narration by people close to it) is demands that people adjust their emotional responses/affects

2100

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 13/08/2026

a lot of AI guys have incredibly strange — and I think wrong! — ideas about how people, markets, and societies experience technological change nymag.com/intelligence...

On the Relentless podcast, Altman invoked magic. "We are close to creating the genie that can grant any wish," he said. "We are going to make sure that our first wish is to broadly benefit humanity and get the world to a place where a lot more people get to have a lot more wishes." In meetings ahead of Anthropic's planned IPO, according to The Wall Street Journal, the company has been assuring potential investors it "plans to push further into healthcare and biology Al uses and that the work could help mitigate some of the negative sentiment around AI."

From within an organization run by effective altruists steeped in utilitarian ideas about maximizing welfare, the idea that a company could potentially produce lifesaving technologies and yet find itself the subject of public backlash might sound absurd or like a sign of a doomed society. But it's also realistic and something about which people in plenty of other industries (and in politics) are far less naïve. Rear-looking arguments about comparative welfare - quality of life is improving in many important ways - are useful and interesting.
But the general public isn't entirely wrong to treat them as irrelevant. Even for a devoted, faithful capitalist, relentlessly taking innovations for granted and finding fault with the ways new things work is a feature, not a bug, of market democracy. It helps drive the system. Put another way: That's what the money's for.

23311

John Herrman @jwherrman.bsky.social · 13/08/2026

a lot of AI guys have incredibly strange — and I think wrong! — ideas about how people, markets, and societies experience technological change nymag.com/intelligence...

23311

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 30/07/2026

Feels notable — and kind of crazy — that the vision for policy coordination between a trillion-dollar industry and the current federal government runs through *open letters* nymag.com/intelligence...

There's also a strange yearning in letters like this - and in speculative work on risk like Dario Amodei's — not just for people to finally listen, which is the main thing, but for the existence of different audiences entirely: a receptive, deliberative government with a deep desire to treat Al with reverence, awe, and seriousness and with a deep interest in humanistic, liberal principles; a public that will follow along with their rationalist arguments about AI progress, all the way up the curve, and urge their representatives to do the right thing. In Washington, the administration's brain trust on the issue, as reported by Wired, is largely a collection of AI-safety skeptics, China hawks, and loyal Trump political appointees. The request, in other words, will be received by people like Howard Lutnick, David Sacks, Susie Wiles, and Scott Bessent, whose priorities and assumptions about AI might seem alien to researchers who believe they're racing toward recursive self-improvement, and vice versa.

Embedded in the letter is a belief that, at some point, the sheer impressiveness of new AI tools will force not just decision-makes but the rest of the world to see things as they do. This is a risky bet, I think. So far, rising AI capability has produced more messy, multifaceted backlash than policy-minded, AGI-pilled solution-seeking. (Outside of the labs, the premise that AI development must continue full steam ahead to save countless future lives, to choose one example, is neither widely shared nor intuitive.) The implied solutions here would have to survive a pretty harsh environment, and perhaps "pacing the frontier" has been left vague because the specifics aren't quite so universally obvious, even to people within AI: new federal agencies with authority to test or limit models; some sort of new international body with real authority over enormous global companies; a voluntary industry body that, in making rules about how technologies can be developed or released, might look to a skeptical eye like a cartel.
This isn't to say that international coordination around Al is a bad idea; even at the low end of possible outcomes, it's a sensible goal. But it's strange to see a multitrillion-dollar industry continue to communicate and organize as it did when it was marginal and speaking almost entirely in speculative terms using the tools of an awareness campaign (although there's still plenty of speculation here — these letters have always been concerned, at heart, with recursive self-improvement and general loss of control). In part this is because the industry is still not quite sure what to ask for or how better to ask for it. But it's also because some of the solutions — and assumptions about AI progress that support them - still sound, to both regular people and those in power, either threatening or insane.

1186

Reposted by John Herrman

Ben Recht @beenwrekt.bsky.social · 30/07/2026

The least agentic people alive live in San Francisco.

2223

John Herrman @jwherrman.bsky.social · 30/07/2026

1186

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 28/07/2026

It's fascinating that the arrival of actual AI cyber risks is turning a lot of AI folks against some of the specific safety arguments the industry was founded on nymag.com/intelligence...

At the very least, the case for restricting model access looks a lot like protectionism. And whether frontier-lab futurists are right about bigger risks around the corner and the regulatory responses or precautions they might raise, for everyone in the industry but those labs, the situation emerging in the meantime - a few dominant companies controlling access and usage of high-priced products, protected by the government in the name of national security and/or a trade war, in an interconnected world where everyone else will have access to alternatives — is manifesting risks today. The Hugging Face hack may have been a tipping point for the way the industry talks about risk and competition.
Within a few days of its release, and after the emergence of something approaching a consensus among otherwise antagonistic factions in the AI world, the "premature restrictions" letter had gained scores of new signatories, including, eventually, OpenAl itself.
Led by Nvidia's Jensen Huang, in fact, who posted for the first time on X to argue for "sharing models, tooling and research in the open," the push — superficial and motivated as some recent support may be — left just one major AI company to defend what had been, until this month, the default position of companies that thought they had a chance of winning the Al race.

Against growing backlash, Anthropic's Dario Amodei finally weighed in this week:
To summarize my and Anthropics position, we have not and are not advocating for a ban on open-weights models as a category. We should instead focus on keeping powerful chips out of authoritarian hands, stopping industrial-scale distillation, and requiring safety testing of all sufficiently capable models, open and closed.
This is unlikely to move anyone outside of Anthropic.
The company, like all big model makers, depends on the "industrial-scale distillation" of the world's information, which has already cost it a ten-figure settlement with authors. The China policies it proposes would reduce the availability of cheaper, more flexible alternatives to Claude. And suggesting required safety testing "of all sufficiently capable models" in the current political environment, where the range of possible regulatory outcomes runs from capture to corruption, sounds naive at best. A few years ago, a frontier lab accidentally hacking another AI company with a model it couldn't keep track of might have been treated as proof that arguments like Amodei's are correct. Today, in a world where the consequences of AI diffusion are becoming more concrete for more people, it's turning them against him.

2295

John Herrman @jwherrman.bsky.social · 28/07/2026

It's fascinating that the arrival of actual AI cyber risks is turning a lot of AI folks against some of the specific safety arguments the industry was founded on nymag.com/intelligence...

2295

Reposted by John Herrman

Vincent Carchidi @vcarchidi.bsky.social · 09/07/2026

I have been mostly unable to express what I mean when I talk about the psychological and cultural shifts that have sort of intersected with SV/adjacent types, and this piece does a much better job: nymag.com/intelligence...

nymag.com

In AI, Nothing Ever Happens. Wait, It’s Happening!

What explains the increasingly volatile mood swings of America’s most important industry?

171

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 09/07/2026

The AI industry is prone to massive, frequent, and intensifying narrative mood swings. Part of this is typical of a boom. But I think it's also intrinsic to the tech itself nymag.com/intelligence...

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 Al 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.

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.)

33811

John Herrman @jwherrman.bsky.social · 09/07/2026

Chatted with @ftrain.bsky.social and Rich Ziade about my thwarted and slightly pathetic attempts to get something out of OpenClaw, the AI assistant that we are all definitely still talking about a few months later aboard.com/podcast/john...

aboard.com

John Herrman: The declawing of OpenClaw

A New York Magazine tech reporter discusses covering an AI flash in the pan.

2221

Reposted by John Herrman

John Herrman @jwherrman.bsky.social · 09/07/2026

The constant, direct experience of enchantment, disenchantment, and re-enchantment with AI — including at the highest levels of the industry — helps explain a lot about the last few years, and what comes next

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.

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.

191