bakugo

Fun fact about K3's distillation:

As of a couple months ago, when using Claude to write adult content through the API, sometimes it will silently inject a system prompt giving the model a bunch of guidelines on exactly what kind of adult content it's allowed to write, steering it away from anything "questionable" ("Claude will not write etc etc").

Moonshot distilled Claude so hard recently, they actually ended up distilling this prompt injection, too. Using K3 to write adult content results in it randomly hallucinating the injected Claude prompt during thinking, and it will quote parts of that prompt, complete with the name "Claude".

Not that I think distillation is a bad thing, just thought this was funny.

It's not really the same thing.

"Normal" streamers are watched for a variety of reasons: maybe their personality is entertaining and charismatic, maybe they're really good at the game(s) they play (speedrunners in particular fall into this camp), and so on. There will always be viewers watching purely for parasocial reasons, but it's usually not the primary reason to watch them.

With vtubers, though, the parasocial element is essentially the full extent of the appeal. They're not very charismatic (vtuber humor is infamous for consisting mostly of loud childish noises and sex jokes), they're usually not very good at games and rarely dedicate themselves to one game or one genre. They are watched because the viewers are lonely and want to feel like they have a (girl)friend, that's about it.

People will make up a million different reasons, but at the end of the day, the vast majority of vtuber viewers are single men who like to pretend the streamer is their (girl)friend.

I've heard the "if you hate vibe coding you must also hate compilers" strawman so many times, I'm sure it's been said by someone with a "name and reputation" at least once.

The latest and greatest method of exploiting and monetizing the male loneliness epidemic.

Correct.

Would anyone pay a human to create an SVG of a pelican riding a bike?

How can building an application feel like a reward when you're not the one who built it?

And before you post the obvious response, no, if you're truly "100x faster", you're not reading or even thinking about anything the AI is outputting. The time math doesn't add up.

The AIs will adapt to this

I don't think this is true, at least not right now, and in a way I'm actually thankful for it.

The frantic rush to chase the only potentially profitable use case for LLMs found so far (writing code) and the resulting focus on coding RLHF means models are actively becoming worse at sounding like humans.

This is my favorite example, and it's already relatively outdated: https://progress.openai.com/?prompt=10

If something like this was implemented, the benefit of the doubt would have to be given in ambiguous cases, but I don't think it's that hard to tell most of the time.

The latest codemaxxed models all tend to write in very distinct, instantly recognizable ways unless carefully instructed otherwise (honestly a good thing if you want to avoid wasting time reading AI text). A great example is this submission that's currently #1 on the front page (which is also just a thinly veiled advertisement): https://news.ycombinator.com/item?id=48884853

If this becomes common, it's trivially detectable with good ol' toString().

As far as I understand, "flagging" is intended for things that break the guidelines. AI-generated content is explicitly banned, but only in comments, not submissions.

I think it would be great if AI-generated submissions were outright banned as they fundamentally break the balance of effort that HN was built upon, but as was already stated by another user, YC is heavily invested in AI so there's a conflict of interest there.

Part of it might feel like ad hominem, but I think the bluntness is justified in this case. It provides some pretty important context on the situation.

As an outside observer who knew little about Zig and Bun, the impression I got is that the rewrite came out of nowhere, and that the Zig community was surprised by it. The post makes it quite clear that this wasn't the case, and describes how the actions and overall mentality of Bun's creator already suggested such a thing was likely to happen, straight from the mouth of someone who worked with him and isn't afraid to voice his thoughts. Not everything needs to be PR speak.

Vibe code was never meant to be reviewed.

These rewrites are just test-driven development taken to the absolute extreme. Created under the hope that the existing tests are exhaustive and cover every relevant use case, such that if they all pass, the rewrite must be at least as good as the original. So just go with the vibes and burn tokens until they pass, and your job is done.

In practice, this is never true for any codebase above a certain level of complexity, especially not one as mature and widely used as Postgres. But reality doesn't seem to be an obstacle for vibe coders.

I don't really understand how "written by AI" and "for learning purposes" can ever be compatible. What exactly does one learn from typing "Rewrite this in Rust, make no mistakes" into a terminal?

But I read here every day that agents can't code. And that "real developers" spend more time fixing AI bugs than producing code, and it slows them down.

This is all correct, though. I haven't tried this, but I can guarantee it's a buggy, incoherent mess, same as every other vibe coded app I've ever tried, no exceptions.

I would rather use a proven reliable system that’s been used and tested by millions of other users

You're describing ZFS.

It wasn't, this is just another free marketing piece for Anthropic.

This article seems quite drawn out for what is essentially an ad hominem attack on the personal views of the creator of the language.

In a sane world, this would be a clear cut case of false advertisement, and the real estate agents would be held liable for fraud. Sadly, we don't live in a sane world.

This specific form of steganography was not present when the leak happened, as far as I can tell.

I've heard that it was possible to trigger really obvious output poisoning on Fable with something as basic as asking the model to think outside of its built-in hidden thinking delimiters.

This watermark may trigger a similar mechanism.

steal the models or illegally distill them

Oh no, they're trying to steal the models that were trained on stolen data? That's horrible, I feel so bad for Anthropic.

Output poisoning and/or eventual account bans, if I had to guess.

It's not just the price, but also how "legitimate" the registrars are and how well they deal with abuse.

Here's an anecdote: I know someone who insisted on using a .tk domain for legitimate business purposes for many years. When I heard of this, I immediately asked "isn't that the TLD managed by a shady company that gives domains away for free and then steals them back if they become popular?" He insisted this did not affect him, as he was a legitimate customer who had been paying for the domain for over a decade.

Fast forward a few years, the company behind the TLD (Freenom/OpenTLD) went under due to their shady business practices, he lost the domain, and was told he had to register it again at a new registrar for a much higher price to recover it.

Look closer at the capitalization of the words in the quoted sentence.

To put this in perspective: If the AI had recommended Black and Asian candidates at the same rate as it recommended the most-favored group (typically white applicants)

Some people just can't help but put their biases on display at every opportunity, even when it comes to the most minute details.

"Vibe coding" implies minimal to no human involvement. It doesn't matter how good of an engineer the person who typed the prompt was, they were not involved in writing or reviewing the code, so the end result will not reflect their skill. The whole point of vibe coding is making software engineers irrelevant.

People like to go on about how "good engineers review their AI code" but that's just not what's happening in reality. Not only is reviewing large amounts of AI generated code unpleasant and mentally taxing, it also negates most of the perceived productivity boost, so people are simply not doing it.

Proper testing

There is no formal testing that would be expected to catch an issue like this. It can barely be classified as a bug, the logging is working as intended, just with negative side effects that weren't accounted for.

The only real way to proactively prevent an issue like this is for a human programmer to stop and think about this code as they're writing it and go "hmm, we're logging large amounts of data to disk at a fast pace here, this may be a bad idea". Without human involvement, this is just going to keep happening. All vibe coded software is bloated and unstable, I have yet to see a single counter-example.

I like how the page is actually struggling to load due to the sheer amount of bot activity on the PR.

On a completely unrelated note, I wonder why Github is always down. Real mystery there.

No human has ever read or will ever read the PR description.

No human has read or will ever read any of the code, nor was any human thought involved in its creation.

Everything is performative now. As long as you just keep your eyes closed and believe it all works, that's all that matters.