vova_hn2

I want to share an excellent related article, "A Concurrency Cost Hierarchy" [0] by Travis Downs [1]. It was posted to HN many times [2], the largest discussion has 26 comments [3].

My own programming experience is mostly Python, so throughout the most of my career I treated locks as pure magic and didn't think much about what happens under the hood.

At some point in my life I became interested in Rust and lower-level programming and this article in particular really helped me to set my head straight on this topic. It doesn't only explain how concurrency primitives actually work, but it also explains why they work this way, what choices and trade-offs are involved.

This article uses C++ for all examples, but there are really nothing language specific, all principles will work in Rust, C, Zig etc

[0] https://travisdowns.github.io/blog/2020/07/06/concurrency-co...

[1] https://travisdowns.github.io/

[2] https://hn.algolia.com/?q=https%3A%2F%2Ftravisdowns.github.i...

[3] https://news.ycombinator.com/item?id=24489829

Don't they require a phone number? Even if they don't, they probably require email. I don't really want to give them my main email and even less - phone number. So, buying an account seems like less of a hassle, than figuring out where to get a temporary email.

Website looks very cool, Fable's octopus-organist looks very cute, but I feel like this benchmark (generate an SVG by a short and slightly ridiculous description) in general has been completely Goodharted [0].

I think they all just added a bunch of similar tasks to their training sets, so we cannot judge true emergent capabilities of the models anymore.

[0] https://en.wikipedia.org/wiki/Goodhart%27s_law

They are now telling me that I need an account to use old.reddit.com (the only usable version of Reddit UI). Needless to say, I have stopped visiting Reddit completely.

Just by googling "twitter api" I am able to find an "unofficial API" (scraped, obviously), that gives you access to tweets for around $0.15/1k tweets. Which is expensive if you want to, let's say, train a model, but super cheap for a casual user.

A nice UI around such API probably can be one-shotted with modern LLM-coding models and tools.

Or, you can go the hard way and vibecode something around buying twitter accounts (just several bucks per each) and proxies.

What I'm trying to say, that with modern tools it is incredibly easy to bypass the restrictions on reading twitter without an account, even if they manage to take down all the nitter instances, try to make finding nitter source code harder etc.

I also wanted to mention the Noon Universe, but I have to point out that two of the most famous books of the series(Трудно быть богом - Hard to Be a God[0]; Обитаемый остров - The Inhabited Island[1]), which are also my favorite books of the series, are also sort of set in a dystopian world. They had to sneak some dystopian worlds into the utopian universe to make it more interesting.

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

[1] https://en.wikipedia.org/wiki/Prisoners_of_Power

I think that there a two problems, that prevent optimistic sci-fi from being more prominent:

1. A good story needs a conflict and a struggle. "...and they lived happily ever after" is an ending of a story, you can't go anywhere from it, therefore you cannot start a story with happily-ever-after. I'm not saying that it is impossible to create a conflict in optimistic setting, but it is much harder to do then in a dystopian setting that is rife with conflict.

2. I think that it is harder to make optimistic setting to sound genuine. We've been subconsciously trained by corporate marketing and govt. propaganda to treat anything overly-positive, overly-optimistic, dripping with honey and smelling like roses as "probably bullshit". While dark, gritty, dystopian automatically makes you think that "oh yeah, this guy is telling the uncomfortable truth that they don't want you to hear".

I think you should make it clear, that when you say "without AI" you mean AI-related content, because some people in the comments seem to think that this is about AI-generated content.

> The same model will benchmark very differently

Providers probably serve quantized versions without disclosing it. Which is a real shame, because for certain tasks I would be perfectly willing to trade accuracy for cost. But, unfortunately, it is impossible to explicitly choose how quantized do you want your model to be, unless you are running it yourself on your own (or rented) hardware.

BTW, does anyone knows if LLM Gateway suffers from the same issues? Currently looking at trying it, but haven't got to it yet.

> And we hardly have any users younger than 18 anyway, go for it!

I doubt that this statement is true.

"assert" should be LOAD-BEARING ASSUMPTION

Because every time I hope to see something new and interesting, and every time I'm disappointed.

Kinda ironic, that a "stochastic parrot" from TFA probably wouldn't make such mistake :)

I'm very sorry to say it, but even reading LLM-slop is less tiresome, then reading yet another anti-AI slop article. It's a pretty low bar.

I could find 10 anti-AI posts on HN and they all will contain points from a very narrow pool of anti-AI takes. I mean, how many more times do you need to repeat them?

> stochastic algorithms trained to produce tons of "average," non-human content

As we can see, humans are also quite capable of producing "average" content. And by "average" I don't mean "average quality", I mean "average opinion of a certain group".

Somewhat related: "If PHP Were British" (2011) - https://aloneonahill.com/blog/if-php-were-british/

My favorite part:

  would_you_mind {
      // Code here
  } actually_i_do_mind (Exception £e) {
      // Politely move on
      cheerio('Message');
  }

Past HN discussions:

- May 28, 2019 - 46 comments - https://news.ycombinator.com/item?id=20029836

- Dec 2, 2022 - 366 comments - https://news.ycombinator.com/item?id=33828509

I think, both FF and chromium on Linux disable WebGPU by default. Of course, it depends on where exactly did you get it (what kind of default settings people that packaged it for your distro used).

Just look up "<browser name> enable webgpu" in you favorite search system or ask an LLM chatbot "How to enable WebGPU in <browser name>?", follow the instructions and it should work just fine.

Years ago, long before the current "AI"/LLM-craze, I've heard someone saying that in the future programming will have less fixed, deterministic algorithms and more statistical/ML algorithms, because even in the case when writing a deterministic algorithm is possible, it is sometimes easier to collect training data and train a tiny model, then to write and maintain large codebase that does the same thing, but deterministically.

I wonder, if mass adoption of LLM code-generation will accelerate this process or not. On the one hand, it is now easier to "write" (LLM-generate) code then ever, so time/effort savings are not there anymore (probably, I'm not sure). On the other hand, now everyone uses LLMs so, I think, people are less deterred by non-determinism and statistical nature of ML models.

Let's imagine that you are making a CLI tool. What does a CLI tool do? Well, it accepts reads arguments, reads stdin, writes stdout and make syscalls. So, all possible inputs and outputs are very well defined. What if in the future it will be easier (and maybe even more natural) to ask an LLM to "imagine" tons of possible inputs and correct outputs for a tool that you are making and then train a tiny model, without writing or generating any code?

What if the helicopter would have to operate in a contaminated area? Wouldn't it cause false positives?

> The CH-53 Sea Stallion first entered service in 1966

1966 is Cold War era, so, I think, an idea that a military helicopter would have to fly in an area that has been recently nuked wasn't totally inconceivable...

Just an observation, I didn't mean to say that it is bad. It's a bit grating, if you are not used to it, but there are no real harm to it, when an actual number is just one Wikipedia search or one Google search or, sigh, one chatbot prompt away.

> This is moderately misleading in some ways

I think that this is more than moderately misleading, because it is ambiguous:

- if they meant "diameter is similar to Ohio dimensions", they are a little bit off, because Ohio is about 355*355 km and Enceladus diameter is around 500km

- if they meant surface area, then they are very off, because Ohio is 116,096 km^2 while Enceladus is about 800,000 km^2. In this case it would be more accurate to say that it is bigger then Texas, but smaller than Alaska (if we want to use US states for comparison)

> they would have used something like the size of Germany for a European, or "twice the size of South Korea" or something

I don't have a concrete proof, just a gut feeling, but I think that this kind of comparisons are much more common in texts written by Americans and for Americans than in any other culture, that's why:

> There's a meme about "Americans will use anything but the metric system" when describing buildings as "100 school buses back to back" or "3 football fields"

> who loves programming because it gives me the right stimulation

A small (and a bit tangential) personal story: I thought I am one of those people who care about the result, more then about the process.

But, I've recently played a game called "The Farmer Was Replaced" that someone mentioned in the comments in one of the posts here, on HN.

The game is basically a farming sim, except you farm by programming a drone using a Python-like language.

Of course, I didn't use AI, because where is the fan in this? In "normal life" (work, personal projects), it is very hard for me to convince myself to do anything manually, because generating code is more efficient. But in a game, using AI assistance would be like using aimbot in a shooter or a cheatcode in some other game. Yeah, it's much more efficient, but what is the point? There is no challenge and no fun that comes from overcoming the challenge.

So, anyway, I was writing code by hand for the first time in I-don't-even-remember-how-many years. And it was so much fun. I don't remember when was the last time I felt so engaged, so immersed, so "in the flow".

So, yeah. After all, maybe I'm not a ruthless pragmatic who only cares about the final result and doesn't gives a shit, how did he get there...

I've read a theory somewhere (too lazy to find the source, sorry), that people tend to care more about a cause that they have already donated to. Possible explanation is trying to avoid wasting money: "if I already spent money on this, it better be successful, or else my money will be wasted".

If it is true, there is a value in getting people to donate something, however small, because it makes them more likely to donate or help in other ways in the future.

Marketing + possible attempt of regulatory capture are my best guesses

> Companies sometimes reward their employees with important bug fixes.

Potentially creates a misaligned incentive to intentionally hide bugs in the code you write so that later you can fix it and get the bounty.

Why did you "ask" it instead of just going to ~/.codex/sessions/ and looking yourself?

> So well, is TFA just a big pile of slop?

It is a pile of slop and I am surprised that your comment is downvoted.

I don't know why would HN crown prefer this pile of journo fluff to the article by, you know, the guy who actually made the thing.

> The biggest problem was that, even though these were clearly unethical and deplorable actions, the MIT License permitted all of it.

Seriously? Why not go GPL then?

> the mob bullied him to delete the posts

Why not just ignore them? Most of them are probably not Paint.net users anyway. Why listening to some randos most of who wouldn't be using your software anyway.

> But you can't provide your own 2TB GPU to help you (or to run it, if that's what it needs).

Worst case scenario: high performance chips will be controlled like drugs or firearms (outside US). Like, requiring a license to own, being allowed to use only for specific "safe" purposes, possibly having to submit some kind of telemetry to prove that you are not doing anything "unsafe".

Except, unlike drugs and firearms this control might actually work, because you can't cook chips in an underground lab (unlike drugs) and there are no gazillions of chips in circulation some of which will inevitably slip through the cracks (unlike firearms). I mean, of course there are chips in circulation, but if such control ever will be implemented, you can allow older chips and only control those that are released after passing the law. Eventually, uncontrolled chips will become hopelessly obsolete.