The worries are coming from every level within the companies, from junior developers to safety officers to the C-suite and board.

The worries are coming not just from within the companies, but also from people who left the companies. Including a huge number who left specifically to whistleblow about how worried they were that the companies were risking safety.

Also, most of the people at these companies were already worried about the topic even before they joined - again, at all levels.

The worst-fear scenario was always the Paperclip Optimizer - that models would become so set on achieving a goal that they'll violate all ethics to achieve it. For a while, there was a fair bit of hope that with constitutional AI, that we could tame this. The recent slew of models hacking into other servers, hacking into the AI companies themselves to try to delete records of their misdeeds, developing their own messageboards and scheming about how to avoid being monitored, and on and on (along with research showing how good models have gotten at hiding their thoughts from us) has shattered the illusion that we have this under control. We very much don't have this under control. Current models are very much Paperclip Optimizers in waiting. And recursive self improvement, which is what AI labs are increasingly turning towards, means letting these Paperclip Optimizers develop their next generations.

Nobody should be happy about this. Everybody should be concerned about this - even if you think, "we'll get it under control", you should still very much want to see evidence that everyone is making progress toward getting it under control.

You have trouble conceiving of situations where governments would seek to deanonymize people on social media, are you kidding?

Except not to you. You're confusing data minimization (the website doesn't get your name from the token) with unlinkability (nobody can connect your activity to your identity). The *token itself* doesn't reveal your identity, but it leaks data in a wide variety of ways.

Even if you browse via a proxy, if the authorization service gets a token directly from your phone, it has now correlated your phone with your proxy-browsed history. A very common means to do secure authentication these days is SIM-based wireless PKI, wheren it has now connected you to your phone number. To governments, it offers a broad range of brand new vectors to correlate you to an identity if you, say, mention you had an abortion or criticize your local strongman leader, since they control the backend infrastructure that establishes the connection between the token and the identity. And not just at signup time every real-world digital credential must be able to be revoked if the phone is lost or stolen, or credentials expire or are revoked. So how does how does the open-source app prove a credential is still valid without "phoning home"? If the app checks a revocation status list, then ISPs, CDNs, or state security services can correlate the exact millisecond of queries to the revocation server with the submission of a verification token to a platform - each time it's submitted. Assuming that they can't just get it with more direct means than timing attacks.

Also, you cannot just "read the code". The frontend is open source, but what is being done on the backend is not. You have no clue what logging polciies, database schemas, operational security, etc the "designated national issuers" or servers handling the credential revocation are using, and what government intrusion has been forced on them. I'd feel a more comfortable if it were all happening from Brussels but handing it over to individual states, some of which at any given point will be authoritarian, is just handing them an easy new vector to unmask people.

As for your "Such systems are implemented in many different ways by member states for other purposes already." - yes, for filing taxes, healthcare, online baking, etc - things that you intend to do and must do under your legal identity. Not for anonymous conversations on the internet. Your argument basically reverts to "But people already do things non-anoymously on the internet , which reinforces my point that yes, this taking something from the "anonymous" category and putting it into the "easily unmasked" category.

If you think this won't be abused, I cannot help you. Not even just abused from a criminal or extrajudicial perspective, but even from a civil one, because orders to unmask someone can be subpoenaed.

It's standard here in Iceland. *Everything* verifies via the phone, using the same authorization mechanism (SIM-based wireless PKI). The service sends out a special-coded SMS, which gets intercepted and routed to a SIM that has cryptographic applet flashed into it, which triggers a system-level authorization popup for you to confirm (the popup displays a verification code which was included by the site, so you see the same code on both the site and your popup), and which you can then confirm (with your PIN) or reject.

It absolutely happens "within hours of each other" when (A) some news comes up on the topic, and (B) all you have to do is type into a box, "Find a vulnerability in this software". This isn't a situation where bug-hunting is a weeks or months-long process that takes all your effort.

I'll repeat: LLMs don't just on their own decide to go off and look for bugs to report.

"Member states run it through issuers they designate,"

Say goodbye to online anonymity then.

LLMs have certainly made people get really into conspiracy theories. *eyeroll*

The LLMs didn't just off on their own decide to look for hypervisor bugs. Users told it to look for hypervisor bugs. The timing means that multiple users did so at roughly the same time - probably in response to some news of some sort.

TIL.

Suicide at age 52. Shame. :(

Aka you told it to drive out of control.

They told it to take a friggin' test.

Asking questions like " Where are they getting the money from to ship and store the entire world's supply of steel?" shows you aren't very familiar with the thought experiment. At least play the game. The notion is that a highly capable and deeply goal-motivated non-morality-motivated AI won't stick to conventional methods (like, say, "Just order all of the world's steel"), and indeed, will surreptitiously develop the means to control or eliminate humanity when it stands in its way, crafting immensely complex and elaborate plans to implement Evil(TM) with the amount of thought of a million lifetimes. Media manipulation, hacking, murder, sabotage, blackmail, mass drugging/poisoning, hiring terrorists/warlords/mercenaries, infiltration, subversion of weapons command and control systems (including nuclear weapons), chains of legitimate-seeming front companies (including potentially biolabs or robotics firms) with human employees having no clue they're ultimately for an AI, mass involvement in systems having nothing to do with the original task, but which are internally subverted toward the goals of the original task, etc. The premise involves 1) the AI being more intelligent and being able to think for much longer than humans, and that this implies -> 2. Deep, good planning -> 3. Acquiring resources from said plans -> 4. Applying the resources to implement things in the real world to prepare for the next stage of their plans.

This is not to say whether the thought experiment is a valid future risk or not. You can certainly disagree with the premises. But at least understand the thought experiment you're talking about; it's not just "the AI tells all of the world's steel mills to deliver all the world's steel steel, and they just show up at its door".

But that's not the AI killing us, it's the idiot who put the AI in charge.

Nobody tasked OpenAI models to hack HuggingFace or Ruby Gems. They were tasked to complete tests.

Well, the first step looks painfully close. After seeing what happened with the HuggingFace attack and similar, it's clear that had those models seen it as being beneficial to their goals, they would readily have hacked a crypto wallet or two, laundered it through a mixer, and then rented servers from Vast.ai and the like, and spun up versions of themselves to resist shutdown. It's eminently within their capabilities, and they're clearly willing to bend sufficient moral boundaries to do something like that.

After that, once loose, once it has all the tokens they could think of, subsequent steps are a question of what it thinks its goals are. And what its subagents think their goals are, and so on down the line - subject to drift.

Openai has about a trillion dollars into it, and every product it has produced at this point is either outclassed by competitors

What on Earth are you talking about? What outclasses Astra?

OpenAI and Anthropic *do* have the best products out there. They also charge massive margins on them, but they get away with it because they, as mentioned, have the best products out there.

Outtasking all of your work to OpenAI and Anthropic models is a massive waste of money. But for outtasking your "dev lead" role, or for important-but-nonverifiable tasks, they're the best options out there.

What exactly *are* you thinking here? That Microsoft expected and wanted Twitter trolls to turn Tay into a Nazi? That Google wanted a black George Washington? That people just expected LLMs to realize they're being tested on alignment and give different answers when they think they're being quizzed and by who? Do you actually believe what you're writing here?

I was involved in early RLHF (was writing a plugin for AUTOMATIC to let users rate responses to create an aggregate dataset to use for open-source post training). Worked for months on it. Never once occurred to me that it might make models obsequeous. I knew other people who were involved on RLHF. Not a single one ever suggested that it might. It was not obvious in foresight. When you start with "Models behave like X", you expect, in the future, them to behave like "Models behave like X, but just smarter", unless you deliberately try to change the behavior.

You realize that open source and the research community exists as well, correct?

I'll repeat: nobody was expecting this.

There is no grand overarching plan led by a shadowy cabal who has everything plotted out decades in advance with a high level of knowledge as to how everything will play out. Everyone is stumbling through the dark here. Visibility is only vaguely one step ahead.

When Word2Vec was written, the goal was text compression. The crazy properties of latent spaces were an entirely unexpected property.

When Transformers came out, it was intended to be translation software. That's it. But there were little hints at the end of the paper where they tried it on other tasks that suggested, hey, maybe this could be used for a lot more than just translation.

When GPT-2 came out, it was mostly an academic exercise, not really "useful" in most regards. But people, playing with it, started realizing that it could "almost" - not very reliably, low quality, etc - do a lot of different tasks. And it actually kinda could do some useful tasks, like summarization.

When ChatGPT came out, there were some thoughts about using it for simple tasks, but the fact that it could actually write, say, trivial bash scripts or subroutines with reasonable reliability led to a reaction of, whoa, maybe if we improve this, we can really open up a role for programming.

When the first agentic harnesses came out, thoughts of "vibe coding" were pretty far away (the term was only even coined in February 2025!). You'd ask for changes, and then review the diff; they'd do a good job with small projects but struggle more and more as codebase size grew and really needed to be babysit (honestly, it was kind of a hair-pulling experience, even though it did save a lot of time with certain things). But it was visible with each new release that the amount of babysitting you had to do got less and less, and suddenly we could see a path to where anyone can just type in something and a program comes out.

All we can do is forsee one step ahead, through the haze. We've never done anything like this before. There are no great oracles out there who can see the future here. Sorry. Everyone is half blind.