stratos123

The repo creation date is confusing but if you look at the actual commits, there's been 0 commits between repo creation on Dec 6 2025 and last week, after which it's all "Co-Authored-By: Claude Fable 5". So it's in fact all Fable, not Opus.

I don't think this scenario makes sense. It's one of a class of scenarios I've seen several of, that simultaneously assume:

  A) ASI is developed and massively overshadows the rest of the world economy 
  B) the world still has rule of law, contracts, business, well-developed finance, etc

No matter what human set of interests you consider important, you'll need alignment research to have any idea on how to instill it. Otherwise you're overwhelmingly likely to get an AI with a set of interests that's totally alien to what any human would ever want.

I'd be very surprised if this is in the training data given that most models mess it up to this day. E.g. look at the ones from Opus.

The current wording seems terrible. It sounds like if I bundle an existing app into a flatpak and submit it to flathub (a common practice, see e.g. unofficial flatpaks for Discord, MS Teams, etc), and the app happens to have some AI-generated code, my submission might be rejected, even though the AI-generated code had nothing to do with the flatpak manifest and won't directly appear on flathub at all.

"Defeating Nondeterminism in LLM Inference" ( https://thinkingmachines.ai/blog/defeating-nondeterminism-in...) has a repo: https://github.com/thinking-machines-lab/batch_invariant_ops

which seems to have eventually been merged into vllm: https://docs.vllm.ai/en/latest/features/batch_invariance/

So you can get determinism locally. On a cursory search I wasn't able to find any LLM provider advertising determism; if you need it for research you might have to rent a dedicated GPU pod and run vllm there with the appropriate settings.

I'm not sure it's that far out of the Overton window. US already managed to do some small-scale UBI trials, after all. Maybe one day it can do a countrywide trial.

When automation happened historically, people whose job were displaced suffered a lot and eventually pivoted to different jobs. Having to relearn all of your career skills is already quite bad in practice, but a bigger problem is that it only works if you can learn the skills for job B before it, too, is automated. That'd require AI progress to hit a wall and stay there for at least years, ideally for decades. If this doesn't happen, then there simply won't be any white-collar jobs to pivot to, and shortly after that, no jobs to pivot to at all.

The article is from Feb. 7, 2025. Title should probably include (2025), if this is posted at all.

HN's algorithm is in fact kosher, because it's not personalized. On HN, arguing with people on topic X will not make you get shown even more articles on topic X to keep you engaged. Reddit-like platforms are similarly okay (you personalize your experience by subscribing) and short video platforms like Tiktok are the great evil.

I find this exact attribution very unlikely; this article is from April 2026 and the term "permanent underclass" is used in this way since at least 2025.

In fact, the repo says:

  This is my social commentary on the "Permanent Upper Class" philosophical virus that has spread through San Francisco tech communities like the plague in the last year or so. You can read the original blog post [here].
  While the game tells a grim story, my goal with this is to show how ridiculous this way of thinking is. I'm actually more optimistic that humans will find a way to prevent this future but if we don't then I suppose none of us will don top hats for eternity.

Some of the reactions to this are slightly unfair because it looks like Justine (who I'm not familiar with) supported Moldbug in 2014. It was only in around 2020 when Yarvin sold out, did a 180 in his views, and became a Trumpist; in the early 2010s, he would write long posts about how populism inevitably leads to fascism, any entity who wants power cannot ever be trusted with it, etc. His political views back then were far weirder than what people probably think - e.g. he thought a dictator needs to be simultaneously secure and yet overseen by a board of directors with power to replace the dictator, using some galaxy-brained setup I don't really understand.

Hence, recommending Curtis Yarvin in 2014 is rather eyebrow-raising but a very different thing from recommending him now.

Honesty and correctness are not the same thing, even when talking about LLMs. Sometimes an LLM says a false thing and you don't know whether it's being dishonest or merely incorrect. Sometimes, however, you can see in the CoT that the model does know the true fact and is reasoning about how to deceive the user. That's lying, not just being incorrect.

AFAIK, Antropic claims that they weren't aiming for zero-days specifically. From https://red.anthropic.com/2026/mythos-preview/ :

  We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy. The same improvements that make the model substantially more effective at patching vulnerabilities also make it substantially more effective at exploiting them.

You're making an implicit assumption that the way humans implement a trait is the same as the reason why that trait evolved. But of course, that's very wrong - evolution overall completely failed at making humans care about evolutionary fitness. A human engineer designing a species might have them only experience kindness towards those who can reciprocate, but evolution didn't do that with humans, because evolution is far dumber than that.

You could clone the repo yourself and not give the agent any tokens at all. When done, push it yourself. This also lets you sandbox the agent to only have access to the local repo and nothing else.

I'm assuming this is a joke, but:

- why'd a quantum computer help running an LLM?

- of course there'd be need for frontier companies - nobody else has the resources to train frontier models.

Consciousness aside, why does reading about an LLM generalizing from specific to general dishonesty make you think it's not intelligent?

Is the output on the max level meant to be missing?

Your phrasing ("you forget") implies this is a fact and common knowledge, while in fact there's little reason to think that's true.

You could use `iter(lambda:1,0)` to get an infinite iterator, then itertools can be dropped.

Why do you call it destructive? It rewrites history only locally and reversibly (the disappeared commit is still in reflog and can be recovered with another reset) and also doesn't destroy uncommitted changes, so it's quite safe. You can only lose data with it by resetting an unpushed commit and then waiting long enough to let the unreferenced commit be garbage collected.

If you had some time and computing power (not even all that much, in the large scale of things), you could simulate perfectly how a human grows from an embryo to an adult, or how an entire human brain processes some incoming signal, and yet this wouldn't give you the understanding to design a human or human brain from scratch.

You call this a "scale problem" as if there's some scalable way such as an algorithm to resolve arbitrary scientific questions and we simply haven't done it, but of course no such algorithm exists, which is why there's plenty of science that's still not settled.

It's way worse than what you can get locally via whisper. Not sure why, perhaps Google legally can't use whisper, or can't spare the compute.

AFAIK, extremely long CoT is fairly routine for those math-finetuned private models. Also, is Mythos unusually verbose compared to Opus?

It's equivalent to setting no_new_privs on the container process, so it'd mean you have to grant a privelege to the container process if you want any children to have access to it. It sure sounds funny in a CVE context, though.

I didn't realise there was a second EMDrive. Much like with the latter, though, I'm overwhelmingly betting on it being really hard to accurately measure small thrusts on the ground, rather than the drive being genuine.

I wouldn't say it's trivially wrong but it's pretty much always wrong. There's two notable sampling parameters, `top-k` and `top-p`. When using an LLM for precise work rather than e.g. creative writing, one usually samples with the `top-p` parameter, and `top-k` is I think pretty much always used. And when sampling with either of these enabled, the set of possible tokens that the sampler chooses from (according to the current temperature) is much smaller than the set of all tokens, so most sequences are not in fact possible. It's only true that all sequences have a nonzero probability if you're sampling without either of these and with nonzero temperature.

By making millions of queries to frontier models from a lot of accounts, collecting the results as a dataset, and finetuning your model on it. Chinese companies have been caught doing it on an industrial scale several times now.

"Biorisk" seems to be a concept not only invented by OpenAI but exclusively taken seriously by them.

This is false. Antropic just bundles it into CBRN. As for inventing it, the idea of AI-created bioweapons as a concrete risk far predates OpenAI as a company.