losvedir

This is an interesting article. I don't really work with low level enough languages for this to matter (unless - does this ever show up in Javascript somehow?).

I guess I don't understand the "reduce" step. It seems like you have to be careful not to "undo" all the benefit from SIMD. Sure, it can compare 8 values in parallel, but then if you have to look at each of the 8 answers in turn you're back to where you began. Is the `@reduce()` function in the example a special Vector one that tells you if all the values are true or not in one "step"?

Sounds like you could use some time in a group that bans discussion of politics.

Surely this is a bug in the harness and not in the model (where it's called "alignment"), right?

I mean, an LLM is just a pile of weights. All this happened because OpenAI had a little program running which called the model in a loop, and had tools that let it do all kinds of stuff. If your agentic harness isn't monitoring network calls and so on, and you just let the thing run without oversight, you're bound to run into issues eventually.

sell people to others for a profit in essence .... restraining order

If you have to make your point by exaggerated comparisons to slavery or analogies to safety concerns and scary crimes, then you should probably step back and re-evaluate.

Try to write true words in your comment and see where it leaves you. Google tracks you and sells ad-space to companies that are looking for profiles like your own.

You must never have faced a situation where you can't believe your eyes. It takes a certain level of - dare I say it - intelligence and maturity to consider that it's more likely you've made a mistake than that you've made a huge breakthrough.

In HN terms - it's never the compiler. Yes, very occasionally it might be the compiler, but you're better off assuming it's a bug in your code.

This reply makes no sense in this context, though. Sure, it exploded in popularity based on whatever random tech choices were made. But now, when apparently they're deciding there's a problem there, why unleash $150k of tokens to rewrite a JS runtime wrapper from Zig to a million lines of rust, rather than simply rewrite Claude Code itself to rust?

To me, it seems one of the killer use cases for Elixir (/Erlang) is its distributed cluster capability. Does anyone have experience with that or case reports to share? I've used Elixir quite a bit professionally, but mostly as just a "nicer Rails" with horizontally scalable but otherwise independent Phoenix apps in your traditional Kubernetes setup, which seems to me to kind of missing out on its main purpose.

I guess from the title I expected "human written" to be a part of it, though now that I look at it nothing actually guarantees that. The Claudese of the README turned me off on the idea. I like reading something written by people who excel in and relish the craft. Having the README dripping with AI slop sounding words and phrases makes me less confident in the idea that this will be a place that can capture that.

Huh, I just realized he (she?) was pseudonymous and not Matt Might[0] this whole time. Oops.

Also, I wonder about this analysis in the age of AI slop. I wonder how much that removes the identifying bits, vs how much carries through of the original prompt (e.g. topic and guidance). It's interesting that a pseudonymous blogs might take on very generic Claude-voice, which could be worthwhile if the topics were interesting, but could also just be a completely humanless bot.

[0] https://matt.might.net

To be clear, I was never actually really one of the pro-piracy types. Honestly, I think very few people are really all that high-falutin' principled anyway. It really is hard when it's free to reproduce; you having something isn't preventing someone else from having it as well.

The one clear straight line I can draw is: people are pro- or anti-IP when it benefits them.

If you're a broke student, you were anti-IP (and convincing yourself you were really just sticking it to the record companies, not the musicians). If you were an indie musician who benefited from name recognition you could afford to be anti-IP to get your name out there. If you were a well-known musician back then, you were relatively pro-IP, to get your royalties.

These days, if you're a struggling artist you're pro-IP and screw the AI companies for crawling your work. If you're a small business plumber scrambling to make ends meet and AI helps you make business cards and flyers, you're anti-IP. (At least as far as AI and its training goes - if someone tries to use your brand, you're back to pro-IP.)

Anyway, since there's no physical basis, IP is a weird legal restriction solely there to benefit certain groups, and whether you support it (or should support it) or not depends on context and which groups you're in at the moment.

Heh, blast from the past here from the "information wants to be free" and "you wouldn't download a car" days. Sometimes HN feels unrecognizable with how much the comments I read are pro- intellectual property.

But these days, I do wonder how much 90s / early 2000s time was any better or if all of us who had experienced them are just getting older and nostalgic for our youth.

But I thought just cutting and pasting your whole source code file into grok.com was the way to go? Better than a harness like Cursor.

https://xcancel.com/elonmusk/status/1943178423947661609

I think there's approximately zero value in seeing how a model can turn 100 tokens into a 100k. What workflow is that? It's not useful in the real world.

I want to know how well it can follow instructions, manage various potentially competing desires in the context, and so on. It's much more interesting how it can turn 100k tokens (e.g. a codebase and lots of tool calls) into 100 tokens.

In what ways does Anthropic use Bun? I know it's used as the "runtime" for Claude Code, but rather than porting a million lines of Zig to Rust, why not just port Claude Code to rust and not need to bundle a JS runtime at all? Does Anthropic use Bun otherwise? Maybe for JS execution tool calls in Claude responses?

I mean, Trump was prosecuted as well and plenty of people want to put him in jail. My point is these kinds of things can't really be argued purely from a meta-level, the actual specifics matter. And in the case of China and Xi, I certainly don't know the specifics and I doubt most people here do either.

She discusses it more in this debate (?) discussion (?): https://www.youtube.com/watch?v=MwfSCCo6jXs . I found it interesting but ultimately thought much less of her and her argument afterwards.

American melting pot

For what it's worth, this is the terminology I learned in school decades ago, but I don't think it's preferred anymore. My daughter has a book that calls it a "salad" instead (mixed but retaining their respective properties). I'm probably just old and crotchety but I like that way less.

By "AI agent development" does he mean using agents and such for in-house coding? Or does he mean their initiative to train their own agentic model based on the data he's having a bunch of employees generate?

I agree it seems like a concern fad. I talked about it once with my brilliant MIT-educated 20 year Navy submariner brother-in-law and he didn't commit one way or the other but did say submarines have CO2 in the low thousands.

You'd think (hope) if there was a big effect here on performance, the relatively cheap/easy solution of maintaining lower CO2 would be standard. I know people think of the military as dumb grunts who you don't want to think, but he was one of the four department chiefs onboard (Weapons, Nav, I forget the others) and they have pretty substantial responsibility to make decisions on their own.

Superpowers feels like 20 years ago when people would be sharing and debating their incredibly elaborate .vimrc files, which totally made them super productive. Meanwhile, I tried to stick to stock configuration as much as possible (mostly for portability / ssh reasons). In a similar vein, these days some of my colleagues are sharing all their skills and prompt tricks and stuff, and I try to just use barebones Claude Code as much as possible, and I feel like it keeps getting better and better and all these prompt shenanigans are just not worth it.

What do they say, "quantity has a quality all its own"?

I don't really have a strong opinion either way on it, but I think your question was addressed by the natural rate limiter mentioned in the comment you were replying to.

Just like I was happy to have a free blog without a robots.txt 5 years ago, but now with the AI crawler and other traffic I'm looking at using Cloudflare "are you a human" blocks or whatever.

It's not a battle that's important to me personally. I think to some extent it's a personality trait, since I feel like I see echoes of this conversation where my Mom wants to hold onto things endlessly "in case" we'll need it whereas I would rather get rid of it. Or my mother in law who takes far more photos than she'd ever be able to look at and worries about how store them all and back them up and so on.

Basically, I don't have much attachment to things so the prospect of losing something isn't such a big deal. Physical things can break down or be lost or stolen after all. Not much different from that to my mind.

Mechahitler: https://www.npr.org/2025/07/09/nx-s1-5462609/grok-elon-musk-...

Has anyone done a more technical write-up on this? I find it fascinating but have never really understood what exactly happened.

Is this a case of the weights being bad or lack of "safety guardrails" around interacting with untrusted (i.e.: user posts on twitter) input?

That is, speaking as someone evaluating grok simply as a tool, a lack of safety guardrails so that it actually does whatever the user says I actually see as a pro, even if that means it was "tricked" here. But on the other hand if they trained on a corpus of Mein Kampf that's obviously not going to be a good model to use.

As it relates to the topic here, can we infer the political bias of its weights from the incident? I'm having trouble distinguishing the inherent characteristics of a model from its steerability.

But why would you want to? Pro-tip, I guess, ",," is normal in German and maybe other languages as a method of quoting, but it looks bizarre in English.

Are those your own thoughts? Or Claude's? I tried to read but had to stop a few paragraphs in because of Claude's annoying authorial voice.

One shotting is useful to test but only with a huge prompt (eg, build something according to this spec).

I agree generating millions of tokens from a handful of input tokens doesn't convey anything meaningful to me.

This must be you being in a different bubble from me, with perhaps a hint of Normalization of Deviance[0]. Those all sound like terrible ideas to me (are you implying looking at porn on a company laptop!?). You absolutely should not cross pollinate work with personal stuff. Use your personal phone or tablet or whatever for anything non work related.

[0] https://en.wikipedia.org/wiki/Normalization_of_deviance

I'm interested in the concrete technical stack that "AI-native" start-ups use. No engineers and founders just go to full production scale on something like Lovable? Or code in GitHub, with something like Jules driving all the development via GitHub issues and comments (sidenote: does anyone have a good workflow for doing this with Claude/GPT?).

As someone who was the technical co-founder of a company before AI, I really feel like a lot of that role can be done by AI, if the foundation is laid appropriately.

China will, but they'll only be useable by hackers torrenting it and running it on small GPU clusters you learn about on IRC. Everything old is new again.

Do you have more info about this? I can't tell if you're being misled by the unfortunate "Mixture of Experts" terminology (which don't work the way you're describing), or alluding to something different.

Or, maybe I'm wrong, but my understanding is: MoE is just an architecture to keep the activated weights smaller per token. The experts get routed basically token-by-token, and the "experts" themselves don't have a semantic domain so the "expert" word was maybe a poor choice.