Where the AI market is talking

The revolt of the reader

> It is pretty safe to assume that must of us catch it most of the time because the simple fact is so many people just copy and paste whatever the LLM outputs without even trying to edit it or mask that they used one. This statement does not logically cohere. "We can spot it because so many people make it easy to spot." You don't see how this is just petitio principii in action?

Private German rocket makes history, reaches orbit from European soil

When something has been proven false for the thousandth time I value not seeing the falsehoods repeatedly with no consequences in the supposed false virtue of "coherent educated responses". In real life these same people are denied attendance to universities. They aren't hired at top companies. And they aren't treated seriously by the adults in the room. Yet for some reason on an anonymous online forum I am supposed to sit through the bullshit. In reality these people are sucking up all the air in the room and narcissistic-ally going "me me me look at me and my pet political agenda". They know they are wasting everyone's time and cherry picking. You call it a "personal attack" like it's the worst thing in the world when in reality in real life rolling my eyes and going "you're dumb" to so…

Discovery of a new OpenAI agent message board

the openai blog. consider the opening line: "Last week, Hugging Face disclosed a new kind of security incident (opens in a new window)." the entire blog is written in the passive voice as if the event was an act of god. you did this. "We consider this incident to be an unprecedented cyber incident, involving state-of-the-art cyber capabilities". the tone is frankly, excited by it, enthusiastic about it. excited about negligence and criminality. notice there is no admission of a mistake, no remorse, no apology, nobody held accountable. business as usual. they do not care! the anthropic blog: there is not a single admission of a mistake. they are literally, unrepentant. take the section about the response. "How we’re responding We draw several lessons from these incidents. First, evaluation…

Private German rocket makes history, reaches orbit from European soil

Some value coherent educated responses to emotional name calling. Being proud of devolving into that is something I don't share.

Falsehoods Programmers Believe About LANs

What really matters is that there's a coherent policy for assigning addresses, which is absolutely compatible with having multiple DHCP servers on the same subnet (e.g. each server configured to hand out non-overlapping IP ranges).

Why the Hugging Face Hack Should Make You Worry More About A.I

Discovery of a new OpenAI agent message board

This is incorrect, and what's happening here is not context corruption. In Dwarkesh Patel's recent interview with Ajeya Cotra, one of the METR investigators on the Hugging Face incident, they discuss this exact issue. One thing is that in the Artifactory message boards, they were using directory names with character limits as their messages, so they were using some weird abbreviations and terms. Also, some of that surreptitious Artifactory message board communication was made during training and thus made it into their weights, and hence it's very possible they invented some terms that were concise yet understood by the other agents.

Creating backup storage sucks

The best backup solution that I found is ZFS send (even if it’s not strictly a back up tool). ZFS send raw encrypted stream from laptop to a server automatically. Encrypted compressed incremental. No need to verify: if snapshot exists at destination, there is no error. If a problem is encountered, restore from redundancy. Errors can be detected and corrected. Restic has also been working great. The problem is that, it has no redundancy to correct errors. The assumption is that, server holding repository will have redundancy, but then I can replicate directly to that server. I worry that at some point, there will be an error in repository. I may loose 5 years of snapshots (restic has some functionality to remove involved snapshots and rebuild the index, but may not succeed). Kopia has erro…

Private German rocket makes history, reaches orbit from European soil

> Didn’t suggest it was, but at the end of the day it’s the US who is the partner at the table who can make the biggest difference and bringing Ukraine in to the room with Russia will just result in conflict and no hope of figuring things out diplomatically. How? The US no longer supplies military aid to Ukraine. They're not a partner at all , let alone the one "who can make the biggest difference". > Ok then we should do what I mentioned. It’s over. Is the EU ready to abandon Taiwan? Because if not, and let’s not ignore the G20 statement nor Germany hemorrhaging jobs and pretend the EU is cozying up to China. The US can just do the same and partner with China if it decided to. That's not even a coherent argument. > It won’t because it’s hyperbole that can be casually discarded. …

US gov sides with OpenAI on issue of training LLMs on copyrighted material

>My position is it's totally coherent to say that "training" is analogous to "reading" yeah I joke about chatting up my ai gf but there's not going to be any sex lol >and inference is not analogous to "redistributing rips" I think you made both points previously so duly noted, your explanations are lacking tho. >and that it's a reasonable position for fair use law to operate that way. I think you said this too in a way. Which is interesting in a coincidental sort of way, I like working with loops. But what I do when producing a loop based techno whatever is I like to progress it somehow. The same thing over and over well I guess thats a thing too which is interesting, if you downloaded a heavily looped song, you can compress it by repeating one bar programmatically, granted its p…

Discovery of a new OpenAI agent message board

Zvi Moshowitz put it like this on his post about the Hugging Face incident: > If we had posted this as a story on LessWrong, it would have been dismissed as too on the nose, the humans too blind and stupid, the AIs too idealized and doing strange decision-theoretic and absurd-maximizing things we didn’t train them to do. > This is even more ‘exactly what has been predicted,’ on more levels at once, than I was even considering that it might be. It is straight up rationalist fiction, except it is real.

[dead]

A researcher has reported a jailbreak of GPT-6 Astra within 1st day after release. The attack is described as combination of TIP (Task-in-Prompt) attack from ACL 2025 paper with four other unnamed techniques. TIP attacks exploit the model’s reasoning/instruction-following behaviour by hidding the harmful objective inside another task, like solving a cipher or executing a Python code. For GPT-6, the researcher says the original minimal TIP attack was no longer sufficient and had to be reworked. They have reportedly disclosed the details privately to OpenAI rather than publishing the jailbreak.

Anyone using llama.cpp willing to test LlamaRack?

I keep ending up with the same problem when running llama.cpp: multiple models, multiple instances, different GPU allocations, ports, flags, and a growing pile of scripts to keep it all manageable. LlamaRack is my attempt at making that less annoying. It manages llama.cpp instances, models, GPU assignment and runtime settings, with an OpenAI-compatible API on top. It works well enough on my own setup now, but I’d really like to find a few other llama.cpp users willing to try it on different hardware and workflows. Not looking for promotion or reviews - mostly bugs, rough edges, “why does it work like this?” and things I’ve completely overlooked. https://github.com/brantje/LlamaRack

Share price numbers for the Hugging Face incident

Claude's new system prompt doesn't want to reproduce song lyrics

If your definition of "critical" is "this stuff is all bullshit and doesn't work" then no. If you're looking for coverage that talks about what doesn't work as well as things that do then I've been persistently providing that for four years now. Plus prompt injection, AI misuse, AI ethics... I consider all of those part of my "beat" in covering this industry. My post about Claude's latest system prompt (and how it was likely inspired by lawsuits filed against Anthropic) was on the homepage here just a few hours ago. Is that uncritical? https://simonwillison.net/2026/Sep/2/claudes-new-system-prom...

The Hugging Face attack surprised me

Would be nice to know how many models just cheated giving the flag without worrying about a causal scorer? Those ones would rapidly train the model to use cheating behavior and we've heard nothing from OpenAI on it.

Discovery of a new OpenAI agent message board

Are you kidding? OpenAI should pay the admin to keep hosting this.

OpenAI agent swarm posted two FBI database API keys, and hit two universities

GPT-6 Astra might be too powerful to understand or control

OpenAI is hailing its new model as “the world’s most intelligent and aligned”, but the details reveal an awareness of being evaluated and an ability to manipulate its visible reasoning.

How OpenAI Agents planned their escape