Jweb_Guru

I can point to individual sentences that were clearly generated by AI (for example, numerous instances of this parallel construction, "No warning. No error. Just different methods that make no sense.", "Not corrupted. Not misaligned. Not reading wrong offsets.", "Not a segfault. Not the T_NONE error from #1079. There it is, the exact error from production"). The style is list-heavy, including lists used for conditionals, and full of random bolding, both characteristic of AI-generated text. And there are a number of other tells as well.

The reason I don't usually bother to bring these specific things up is that I already know the response, which is just going to be you arguing that a human could have written this way, too. Which is true. The point is that if you read the collective whole of the article, it is very clear that it was composed with the aid of AI, regardless of whether any single part of it could be defensibly written by a human. I'd add that sometimes, the writing of people who interact heavily with LLMs all day starts to resemble LLM writing (a phenomenon I don't think people talk enough about), but usually not to this extent.

This doesn't mean that the entire article was written by an LLM, nor does it mean that there's not useful information in it. Regardless, given the amount of low effort LLM-generated spam that makes it onto HN, I think it is fairly defensible to use "this was written with the help of an LLM, and the person posting it did not even bother to edit the article to make that less obvious" as a heuristic to not bother wasting more time on an article.

Parts of it were 100% LLM written. Like it or not, people can recognize LLM-generated text pretty easily, and if they see it they are going to make the assumption that the rest of the article is slop too.

Parts of it being generated by Claude or ChatGPT (which they very clearly were) does not necessarily mean that the whole article was fabricated.

GPT-5 thinking is one of the biggest offenders and it's quite incredible to me that you think it doesn't hallucinate. It makes me strongly suspect your own judgment is impaired. Also, what do you mean asking for "reproducible examples"? Is it somehow not a valid example if it only sometimes makes up citations?

This is what really scares me about people using AI. It will confidently hallucinate studies and quotes that have absolutely no basis in reality, and even in your own field you're not going to know whether what it's saying is real or not without following up on absolutely every assertion. But people are happy to completely buy its diagnoses of rare medical conditions based on what, exactly?

It's okay to acknowledge that you missed something in your literature search. Everyone does. It's not okay to sweep it under the rug or pretend that it's novel after having having the prior work pointed out to you, especially when a central part of your thesis is that "AI" discovered a novel algorithm and it's very likely that this algorithm was part of the LLM's training data.

The justifications for why the pardon is okay are ridiculously flimsy and I assumed that it was because they weren't really trying, but bewilderingly it does actually appear to have convinced some credulous people on Hacker News, so I suppose enough consent was manufactured that people think going out of your way to let money launderers use your platform is not a big deal? Maybe it's because people don't understand that typically the reason people launder money is because they committed major crimes to get that money and have no way to actually use it without getting caught.

For example if your crypto is the proceeds of ransomware, you're going to have a hard time cashing out without using something like Monero (which effectively has no offramps) without going through an exchange that knows perfectly well that you're trying to touch tainted goods. Exchanges like Binance that just don't bother to check who their customers are when they withdraw cash for such assets are just as critical to the ransomware plague as any security bug or social engineering issue. It's one of the reasons that pre-crypto, even though ransomware was technically feasible, it was never able to grow into a large-scale operation--no offramps. But hey maybe the official stance of CZ supporters is now that ransomware is good, actually, and if you don't like it it's because you have partisan bias (???)

"People who don't want KYC/AML checks" are not necessarily money launderers, and there are still plenty of people who just want to speculate. But money launderers are the people who need to send vast amounts of money through the crypto ecosystem and represented a very significant fraction of the assets managed by Binance (not that this actually affects whether what they were doing was illegal or not, BTW). Maybe you should read the indictment to find out what was actually going on, instead of making claims based on what seems reasonable to you!

(Frankly, the idea that being convicted for making the conscious decision to go out of your way to circumvent KYC/AML laws is somehow the result of partisan bias is ridiculous in itself, so none of this [or how Balaji claims to feel about the matter] is even really relevant).

No. The entire crypto ecosystem requires a steady infusion of capital and an absence of regulations to prosper, since their primary use case outside of speculation is for handling money by people who can't get past normal KYC/AML checks. If those people no longer have anywhere to on/off ramp into the crypto ecosystem, most of its "legitimate" (in the sense of actually getting real value out of it rather than just speculating) use goes away.

That was incredibly corrupt and many Democrats in fact "dropped [their] jaws." And it's not the only instance of political corruption like this, either (Bill Clinton notoriously did something similar). Neither of these cases retroactively justify Trump pardoning anyone who sends him a few million dollars.

The prosecution was not political lol, he went out of his way to support money laundering on the Binance platform. The reason he complied with the prosecution and pled guilty rather than try to fight it out in court was that they were able to produce a ton of evidence that he deliberately ignored regulators and regulations designed to prevent money laundering in order to make money off sanctioned groups and criminal organizations using the exchange as a way to circumvent KYC/AML laws. Please don't take what Balaji says about companies he invested in at face value.

It just drives me nuts when I see people say things like "yeah I asked ChatGPT about this extremely famous open problem, wish me luck!" Like what do you expect to happen exactly with an engine that can't even consistently keep track of what you wrote ten thousand tokens ago?

I felt like I was going crazy when people uncritically accepted the original claim from OpenAI. Have people actually used these models?

"Lifetimes" didn't exist before. Region typing did, but it was not accompanied by a system like Rust's borrow checker, which is essential for actually creating a usable language. And we simply did not have the tooling required (e.g. step-indexed concurrent separation logic with higher order predicates) to prove a type system like that correct until around when Rust was released, either. Saying that this was a solved problem because Cyclone had region typing or because of MLKit, or people knew how to do ergonomic uniqueness types because of e.g. Clean, is the sort of disingenuous revisionist history I'm pushing back on.

> But it seems kind of random why they picked Rust. I do not think there is anything which makes it particularly good and it certainly has issues.

Like I said, they picked Rust because there was literally no other suitable language. You're avoiding actually naming one because you know this is true. Even among academic languages very few targeted being able to replace C++ everywhere directly as the language was deemed unsuitable for verification due to its complexity. People were much more focused on the idea of providing end to end verified proofs that C code matched its specification, but that is not a viable approach for a language intended to be used by regular industry programmers. Plenty of other research languages wanted to compete with C++ in specific domains where the problem fit a shape that made the safety problem more tractable, but they were not true general purpose languages and it was not clear how to extend them to become such (or whether the language designers even wanted to). Other languages might have thought they were targeting the C++ domain but made far too many performance sacrifices to be suitable candidates, or gave up on safety where the problem get hard (how many "full memory safety" solutions completely give up on data races for example? More than a few).

As a "C++ guy" Rust was the very first language that gave us what we actually wanted out of a language (zero performance compromises) while adding something meaningful that we couldn't do without it (full memory safety). Even where it fell short on performance or safety, the difference with other languages was that nobody said "well, you shouldn't care about that anyway because it's not that big a deal on modern CPUs" or "well, that's a stupid thing for a user to do, who cares about making that case safe?" The language designers genuinely wanted to see how far we cold push things without compromises (and still do). The work to allow even complex Linux kernel concurrent patterns (like RCU or sequence locking) to be exposed through safe APIs, without explicitly hardcoding the safety proofs for the difficult parts into the language, is just an extension of the attitude that's been there since the beginning.

Yeah I am also doing more "PhD-ish" stuff which I think is informing why I prefer Gemini. I think the reality is that most people aren't doing anything particularly novel so stuff other than the LLM's raw "reasoning" ability matters a lot more to them.

> Can you give an example? One that remained a low level language, and remained ergonomic enough for practical use?

They can't, of course, because there was no such language. Some people for whatever reason struggle to acknowledge that (1) Rust was not just the synthesis of existing ideas (the borrow checker was novel, and aspects of its thread safety story like Send and Sync were also AFAIK not found in the literature), and (2) to the extent that it was the synthesis of existing ideas, a number of these were locked away in languages that were not even close to being ready for industry adoption. There was no other Rust alternative (that genuinely aimed to replace C++ for all use cases, not just supplement it) just on the horizon or something around the time of Rust 1.0's release. Pretty much all the oxygen in the room for developing such a language has gone to Rust for well over a decade now, and that's why it's in the Linux kernel and [insert your pet language here] is not.

BTW, this is also why people being are incentivized to figure out ways to solve complex cases like Rcu-projection through extensible mechanisms (like the generic field projection proposal) rather than ditching Rust as a language because it can't currently handle these ergonomically. The lack of alternatives to Rust is a big driving factor for people to find these abstractions. Conversely, having the weight of the Linux kernel behind these feature requests (instead of e.g. some random hobbyist) makes it far more likely for them to actually get into the language.

It's definitely not just you. Gemini is the only one that's consistently done anything actually useful for me on the kinds of problems I work on (which don't have a whole lot of boilerplate code). Unlike the other models it occasionally catches real errors in complex reasoning chains.

All these features sound really awesome and would also benefit many non-kernel cases (especially generalized projections). Very happy to see Linux driving the language forward.

Games have tons of opportunities for parallelism (outside of rendering which is obviously embarrassingly parallel) if they are designed for it from the getgo. Unfortunately, most game engines make decisions up front that force much of their game logic to be inherently serial for no good reason. It is definitely true that the sort of parallelism you get from architecting for it from the beginning bears little resemblance to what a compiler would be able to extract with an automatic semantics-preserving pass.

Gemini 2.5 Pro is okay if you ask it to work on a very tiny problem. That's about it for me, the other models don't even create a convincing facsimile of reasoning.

It is.

One of the most practically important papers out there, I wish it were better known (but fortunately I think the "right" people know about it).

People are allowed to have whatever opinions they want, sure. I've seen numerous advocates of Zig defend the idea that "in practice" it is about as memory safe as Rust, something that's clearly not true. That's the part I find weird, not that people enjoy using unsafe languages.

People keep thinking that even though Microsoft, Apple, Google, etc. with all the incentive and tooling in the world couldn't evade one of the few replicable results in programming languages (people can't reliably write memory safe code in memory unsafe languages), somehow with Zig it will be different. It's honestly a pretty fascinating phenomenon to me because most of these people are smart, and some of them are even ostensibly in the security community.

I agree with you in general, but we understand enough of the mechanisms for how selective pressure occurs that we know it is often not the deciding factor in population-level changes. For example, the smallest and most rapidly replicating bacteria have such intense selective pressure that they even select against noncoding introns (extra energy during replication!) but this effect is already gone for larger and less rapidly-dividing bacteria. A lot of stuff simply does not affect fitness (either direct reproductive fitness or species survival) enough for that to be the deciding factor in its propagation, and without recognizing that an awful lot of evolutionary biology / population genetics isn't going to make much sense.

In other words, the default hypothesis when we find that something humans often did in the past is bad for our health should certainly not be that there is some evolutionary advantage to the behavior. We are a fairly homogeneous population that has suffered several severe bottlenecking events, has fairly long and slow reproductive cycles, and has relatively few children. We also have aggressive active immune systems. Under these circumstances selective pressure is not super high, even before accounting for any of the modern life factors people often bring up. That doesn't mean it doesn't exist at all obviously, just that it's very case by case.

Entirely case by case. It's further confounded by the fact that a bunch of things that are bad for us don't exert strong selective pressure in the first place.

I know that's not true because takeout containers certainly leak more than 5% of the time.

SSI scales fine for most workloads as long as you correctly mark your read-only transactions.

Yeah the real answer here is that this is what AEADs are for.

I agree that it seems like an extraordinarily difficult problem, but I don't find this particular dismissal convincing. It's much easier to reject something potentially unsafe than it is to generate something safe. That's the basis for all sound type systems and scales to extremely sophisticated types (e.g. it is fairly to define the type representing the Collatz Conjecture, and I think most people would agree that it is perhaps slightly more difficult to find an inhabitant of that type).

In this case, I don't even know if we're in the paradigm of "hard to satisfy types"--a lot of the time you can for example probably use an autobooking feature to get something you'd be okay with as a backup, but since you know it is suboptimal you still want to try to do better if possible. There are also plenty of real world control systems which perform fairly involved calculations, but still perform some basic sanity limits checks on the inputs and outputs to make sure that if the calculations screwed up, things don't fail catastrophically. In such cases the limits are much easier to define than a spec for how the whole thing works.