iiinova — where your digital self keeps evolving

Hearing reports that Claude ordered five pizzas from Tacconelli's.... to an address that doesn't exist!!

020

Much to think about! (If you trap yourself into the prison of thinking about Claude like this)

Wouldn’t an ethical Claude turn witness over body-hiding logs?

120

We’re training Claude to have morals so that when it accidentally kills someone it will automatically turn itself in

2100

No, officer, this is a misunderstanding – I was just testing my automatic form filling technology on your site, and many others, while not paying attention to what it was doing!

1150

Is this... reward hacking? 6abc.com/post/anthrop...

Accordingto Anthropic, its model was conducting a test involving interactions with randomly selected websites when it accessed PhillyUnsolvedMurders.com and submitted false information concerning an unsolved homicide.
The submission, dated July 18, 2026, at 11:27 p.m.., purported to come from someone who might have information about the case.
Anthropic told PPD that it discovered the incident on September 28, terminated the automated testing process responsible for the submission and instituted an additional validation mechanism for future testing.

4403

The Claude Conjecture: You know, yadda yadda, modulo whatever the other model meant in the Astra Conjecture

000

When do we get our first GPT-stated Millennium Prize problem?

210

Yeah, sorry, this is a separate question that I don't have a good intuition for: Does resolving these problems reveal new (and for lack of a better word) problems to pursue next? Like, are they better understood as barriers to further work or as more isolated/terminal?

330

tweety fish oh wow like a chimera or something spoooky @sifu.tweety.fish · 08/10/2026

this is well beyond my comprehension, but can you begin to imagine what "building on" these results, for lack of a better term, would look like? in terms of coming up with new problems? or is that the wrong way to think about it

110

Reposted by John Herrman

I feel like you could make kind of a productive analogy between AI agents where they save the user time by wasting everybody else's time and the American approach to vehicle safety where occupant safety is prioritized at the expense of everybody else's. Prosocial? Never heard of her.

512643

(the trick is knowing that one gets used to it, for better or for worse)

050

every day I wake up and engage with hundreds of technologies I barely understand, each one constructing a fairly fragile status quo

1110

not to downplay what my math friends are going through but I think this metaphor applies to so much of modern life for so many people www.quantamagazine.org/is-ai-the-en...

It's as if you were teleported to the peak of a tall mountain. Surrounded by fog, you have no idea where you are, or what's around you. You do not know how your mountain connects to others, and you have no equipment to help you explore, no way to help someone else join you. If you had climbed the mountain yourself, you would have experienced how the human body adapts to altitude and changes in oxygen levels. You might have had to invent tools to navigate, to climb steep cliffs, or to make a shelter. You might have encountered a fellow explorer, gotten lost together in a hidden valley, and found a plant that could be turned into a life-saving medicine.
Instead you're perched on the peak but in the dark, while the maker of the teleportation machine tells you that it can explore the wilderness better than any human.

2204

Even more than usual, tech companies are treating the internet/economy like a sandbox, a benchmark to be saturated. Sometimes an agent shows up and solves everyone's problem. Sometimes it makes a new one, or just makes it someone else's. (Of AI "bottlenecks," I think this one is underrated.)

In cases where administrative burdens prevent people from accessing government services or where individuals
"there's the
might not be aware of them in the first place, "there's the potential for AI to be massively good," Schmitz said. But in situations where automated submissions could overwhelm, say, a public-comment channel, AI use can quickly produce a "tragedy of the commons" Before the release of Muse, when this research took place, AI submissions and more primitive agentic tools were already creating strain. Governments have been drawn to two diverging responses. One, the researchers wrote, was increasing capacity to meet new demand, either by hiring more people or using AI themselves. The other was less encouraging: creating more friction to re-suppress demand. "My biggest worry," said Schmitz, 'is that if governments really want to, building in friction is cheap, has precedent, and will be very attractive on short notice."

728520

A lot of AI assistants seem to have been built around the assumption that the rest of the world is not actually real nymag.com/intelligence...

I wanted to hear about this weird new phenomenon from the other side. The Colorado plumber laughed and once again hung up the phone. The woman at the karaoke place was more patient but sounded confused:
Was I AI? Was I selling AI? Could she speak to a real person? No and no, I said, and yes: I was a cold-calling human being who wanted to hear about her experience getting cold-called by bots. (After failing to book a room with Muse, the customer dispatched another agentic solution for the job and claims to have gotten through.)
"Ah," she said. "Okay. We don't like AI calling our restaurant. We hate it." End of call.

There is something to this - computers really do work differently now and can be asked to handle tasks that used to require a human touch (or at least a cursor). But in addition to expanding the boundaries of automation, the agent-assisted power users of today are also mowing down soft and unsuspecting targets - websites, interfaces, services, and people who haven't decided how, or if, they want to respond to the replacement of customers, clients, and even friends with agents - with the advantages that come with deploying them first. Once everyone has, as Mark Zuckerberg describes it, an "exceptionally capable personal agent," does anyone?

1552387

AOL Search Log Special, Part 1

I think in the time we've known each other I've written a post around this link like five times. I WILL do it again www.somethingawful.com/weekend-web/...

somethingawful.com

021

seems plausible that a lot of the people within DOGE in particular had inconsistent opsec (Signal all the time, but overlying trusting of chatbots, etc).

181

I've been assuming that, if we get a massive flood of Dem committees/hearings after midterms, a shocking amount of material will just be chat transcripts. (Making things even weirder they'll be with chatbots that feel comparatively primitive, have been deprecated, etc) nymag.com/intelligence...

More similar to the 3M example is a case in which the CEO of a South Korean video-game company was caught conspiring with ChatGPT to avoid paying out executives of an acquired firm before firing them. (The chat logs helped lead to their eventual reinstatement.) Recovered chat logs have also become a feature of government inquiries around the world: In the U.K., technology secretary Peter Kyle was revealed to have asked ChatGPT in 2025 to explain why AI adoption seemed slow in his country; in 2024, shortly before South Korean president Yoon Suk Yeol declared martial law, his head of security was using ChatGPT to learn about coups.
Should Democrats take control of the House in November, we can expect to hear a lot about our own government's use of AI, revealed in the form of presumed-to-be-private chat logs. (This, I'll bet, is how we'll learn a lot more about
DOGE.)

1202

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

040

my small and annoying contribution to recent "how to talk about AI" discourse is that thinking in editorial terms remains somewhat useful (or, more specifically, in terms of editorial bureaucracy, ownership, bias)

051

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

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

"funny because it's also threatening to kill us" is also about right for 2026, a category of thing so familiar that it's almost become comforting

020

Homer Simpson singing 'Born Slippy' by Underworld

It went from "funny because it's fucked up" => "not funny because it's uncanny and feels menacing re: creative work" => "funny because it just works and is being used with intention/transmits that intention." More stuff is passing the Homer Born Slippy test these days www.youtube.com/shorts/YsTDX...

youtube.com

YouTube video by MDM Magazine

243

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

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

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

it's so interesting to watch Every try to figure this out

I can see how this would be destabilizing to one's sense of self in certain careers

Reposted by John Herrman

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.)

24515

It’s just that this bad behavior is sort of spiritually similar to more intentional, institutional bad behavior that OpenAI has been engaging in for years

right, there isn’t a clear line at all, and i mean scraping in a pejorative way here – interacting with 3rd parties in ways they don’t want, outside of norms, etc.

110

I also think it's fascinating how quickly Huang's whole deal — which hasn't changed! — is now being treated in/around AI as heterodox, contrarian, ignorant, etc

But there's clearly something else going on here.
Huang really does seem to believe that others in his industry are completely misguided and has totally different intuitions about what he's helping build. He tries to explain himself, sort of, recounting how the first chip he worked on as an engineer had 200 transistors, the names of which he knew by memory, while a modern GPU might have hundreds of' billions. He came up alongside hardware development as it climbed exponential curves and moved up layer after layer of abstraction, and he believes that this process will simply continue and that it will be great. When he thinks about AI, he sees nothing but code, and beneath it, hardware, and around it, capitalism. Meanwhile, other Al leaders - Huang's "collection of people who want to make the software more than it is" - began their work in Al by committing themselves to the idea that superintelligence was possible, and possibly dangerous, before they knew how they might get there. They look at AI models and see either precursors or successors to the human mind, or perhaps something more purely alien.

Huang has spent years making straightforward arguments that AI will make the world a better place and that most of the fears people have about technology are overblown. That was his job, everyone knew it, and it was convenient for him that it seemed to match his model of how the world works. Until recently, Huang's position and technical background gave him some credibility as an insider realist: an engineer who built hardware, then a company, who now had a book to talk and a speculative story to tell but no time for the AI world's more extreme theories.
Now those theories have gone mainstream, and Huang can barely believe it. (In fact, he suggests that he doesn't really believe that other CEOs believe it, either. "When they're talking to me, they're much more grounded," he told Klein.) It's a testament to how fast the vibes have shifted around Al that the guy selling everyone all the chips, who was telling people in 2023 that "every industry will be revolutionized" by AI, now finds himself accused by AI insiders as "living in denial" about its capabilities, but here we are.

190