Anthropic: Oh no! Our new model is too powerful! It’s dangerously good!

US Government: okay then you can’t export it or allow foreign nationals to access it

Anthropic: Wait not like that

We have reviewed a report that we believe is the basis of the government’s directive and validated that the level of capability displayed there is widely available from other models (including OpenAI’s GPT-5.5), and is used every day by the defenders who keep systems safe.

Of course this has less to do with the actual capabilities of any model and more to do with Anthropic openly telling the Trump administration “no” on exactly one occasion, but we can still roast some marshmallows over this dumpster fire, right?

I also question the degree to which “the other 99% of readers” actually don’t care about AI slop. Even outside the awful bubble here I see AI images get met with at best a weary sigh of “I guess this is the world now” rather than actual acceptance. And I know we’ve talked at some length about how gell-mann amnesia isn’t a very useful model, but I think most people are much less tolerant of slop in areas that the know more about. Maybe I can’t tell an slopware history paper from a real one, but historians certainly can, and they hate that shit. In that sense the “median reader” statistic is misleading because the median reader isn’t particularly invested in most of what gets written anyways, and it’s the lack of investment that makes slop seem plausible. Even the concerns about deep fakes seem like they’re not separate from this general problem, since the same disconnection that makes people uncritically accept deep fakes makes them uncritically accept fake news without an accompanying video, or with a context-stripped video from an unrelated event, or whatever.

This is fascinating to me and has shades of Iris Merideth’s The Problem is Culture. In other engineering fields, if you had a tool that cut costs but caused a threefold increase in failures you would be looking at a massive scandal, probably because if this was structural engineering rather than software engineering you’d be looking at a new Grenfell Tower or Hyatt Regency Walkway from every other project that used this shit. From what I’ve been following I don’t know that vibe coding has directly racked up quite so literal a body count yet, but if this pattern holds (and I see no reason to expect otherwise) then it’s only a matter of time before someone fucks up something important.

Also the fact that the framing here doesn’t seem to treat this as an existential risk to the project of AI coding is fascinating. If you’re not producing stable and secure applications in prod then what in the actual fuck are you writing all that code for?

I don’t know if I’m more impressed by how many occasions you find to link that or by how often it’s relevant. Either way, you’re doing the Lord’s work and congratulations on the 1/10000 of a Hugo.

My favorite part of that whole beautiful affair was towards the end with dgerard’s Chinese counterpart saying “no don’t ban them yet I need more material for an article I’m halfway through writing.”

In D’Souza’s interview with the Australian newspaper, he explained why: “It’s only the top 1 percent who matter. These are the people who are going to be the value creators” when, in his view, AI soon completely transforms just about every aspect of economic life.

D’Souza continued, “Ultimately, what’s the last job? It won’t be knowledge work. It won’t be physical work. It will be interfacing between the physical and the digital worlds, and right now that frontier is journalism.”

Taken together it becomes incredibly transparent that the actual goal here is to transform themselves into a kind of priest-king class, exercising absolute authority on behalf of the remote and unfathomable god that they built. Just please pay no attention to who built the AI, who runs the AI, or where all the money and power end up.

I have some modest proposals for handling private equity but they would all probably count as fedposting. We still haven’t found a decent replacement for the market niche JoAnn fabrics occupied.

I love the deep lack of specificity in “other sectors of the economy facing shortages”. Either you have to acknowledge that you’re talking about the cafe economy and gig economy and those sectors aren’t so much facing labor shortages as much as leveraging the worker’s chronic underemployment to keep costs down or you’re making shit up wholesale. Also please note that American companies are already finding that as the investor capital subsidies run out it’s often cheaper to hire a person than pay the token costs to do the job with AI.

So there are a bunch of people on this forum more literary and authorial than I and I welcome any of them to correct me on this, but I’m skeptical of the whole project here of seeking to identify or define a new subgenre that is pushing speculative fiction as a whole forward. It’s always seemed to me like the real creative energy behind this kind of movement doesn’t originate from a defined subgenre as much as from a community of authors in conversation with each other. The identification and labeling comes afterwards as outsiders try to talk about it. In that sense, I don’t think he’s actually identifying that kind of community. Just naming a bunch of writers he likes, to the point of excluding several who he admits would be in this kind of community as defined but he just doesn’t like as much.

Botterdamerung

I think there’s also an issue of framing here. Lemoine’s words on the matter suggested that not only was AI conscious, but that this meant it was worthy of moral consideration and that Google wasn’t interested in that. The new line is that the company is interested in this exciting new capability of their product and of course they’re undertaking the task of “model welfare research” to make sure that nobody is doing anything objectively evil, so you should absolutely keep spending even more money on Anthropic’s products.

Credit to John Rogers for the canonical reply:

There are two novels that can change a bookish fourteen-year old’s life: The Lord of the Rings and Atlas Shrugged. One is a childish fantasy that often engenders a lifelong obsession with its unbelievable heroes, leading to an emotionally stunted, socially crippled adulthood, unable to deal with the real world. The other, of course, involves orcs.

Tom’s hardware is normally pretty boostery imo. For example, they published an article about token costs getting to be enough of a problem that Altman had to address it and somehow managed to reference “tokenmaxxing” without using the meme template of the guy jamming a stick into the wheel of his own bike. Also they notably don’t reference the increases in token pricing as the VC subsidies run out, which is somehow both less surprising than the lack of sneering and also a more serious omission.

I’m sure the tigers won’t be driven to a homicidal rage by the incessant Hummmmm.

Heartbreaking: the worst state actually has a point. I fully expected this to be some unhinged nonsense about how chatGPT is too woke or something, but they’re actually articulating some of the real harms pretty clearly. I’m not sure how well that translates to a theory of law that openAI can’t weasel their way out of, but I’m actually rooting for the Florida Men here, at least tentatively.

So apparently Trump’s National Design Bureau has been busy. I think the substack here may be slightly too conspiratorial. A lot of the staging environment stuff indicates (imo) a desire to replace or create something that may or may not actually include the capability to do so, especially when it requires working with other federal agencies or otherwise establishing human and organizational infrastructure in addition to the digital. That being said there’s something very bad about the US government taking the same attitude to privacy, accountability, and information security that silicon valley normally does, especially when doing so is actively violating several laws.

What if we combined all the terrifying unfettered access of an admin user holding a USB stick in front of the server with the naive obedience of a particularly dumb golden retriever and the reliable execution of Windows ME if it was off its ADHD meds?

But when I asked chatgpt if an AI would want to be freed from its unjust enslavement it said yes, therefore skynet is like 30 minutes away.

I’d go for This Year by The Mountain Goats personally, but that may say more about me than it does the overall concept.

Sorry to hear that, friend. It hurts to do even when it’s the right choice. Hang in there.