RE: https://sauropods.win/@futurebird/117181914667887463

I boosted this because wanting to talk about ants in a video is something that seems entirely normal to me now.

Look what this place has done to me!

Just hypothetically, what would happen if Canada put a 100,000% export tax on toilet paper being shipped to the USA four weeks before the midterms?

@rysiek @futurebird

Yup, that's a legitimate concern. I wouldn't be surprised if they've been doing this for a while anyway to avoid training on their own output, but releasing an API to query gives them a lot of control.

And, of course, you can't detect slop, you can detect slop generated by a specific tool.

@dee

If I understand correctly, you are now required to undertake a quest to destroy the ring by throwing it into a volcano.

@futurebird

From a CS perspective I can't think of any way to have a watermark that couldn't be easily defeated through additional processing

The way that the current ones work is that they tweak the probabilities to generate specific patterns in the output. Where there's an equal probability of two words following another in the raw weights, they'll tune it so that there's a higher probability of one than the other.

If you know the weights and know the biassing factor, you can look at each word pair and see what the probability would be of the model generating that.

This means that the watermark is smeared all over the output. And it's not a binary thing though, each pair of word contributes something to the probability of matching the watermark and looking at the whole thing will give you a probability at the end.

Changing every other word should give you close to a 0% of matching the watermark but, at that point, why bother using the LLM at all? If you're going to change half of the words, you may as well just write them yourself. And that's a problem for people who want to share low-effort slop and pretend to be creative.

@futurebird

for decades no one could be bothered to fix it because the small vendors aren't a powerful political lobby.

I don't have specific knowledge of New York in this regard, but it's quite common that big business is lobbying for this kind of regulation. Entirely sensible-sounding regulations add barriers to entry and the costs of compliance are often easier to amortise over a large business.

Investors: Facebook needs to stop jumping on bandwagons and get back to its roots.

Facebook: *releases a device to enable wannabe frat bros to perv at women*

Investors: We walked into that one, didn’t we?

@robinwhittleton @isotopp

Additionally: The problem he identifies is only that they may be open to lawsuits and not that they're filtering out a load of qualified and experienced candidates?

Also that the only issue was ageism (I'd be shocked if this approach didn't also discount most female candidates, though the 'youthful energy' bit of the prompt will almost certainly filter out anyone who has age-related signals in their CV).

It appears that this is the original tweet (xcancel because I don't want to send traffic to the Nazi Bar, remove the 'cancel' from the URL to avoid the mirror).

The post appears to be trying to drive traffic to his legal AI newsletter. So I suspect it's a made-up story by someone trying to sell legal-adjacent advice.

@viq @beeftacos

And, to me, that is very different from an LLM patch. You have put in the effort. As with everyone, there is a limit to what your effort will achieve that is bracketed by your knowledge, but you’re asking for help after trying, and you’re explicit about where the edges of your knowledge are.

I love getting patches like that because they come from someone who has identified a problem and the kinds of things that they get wrong are predictable and so review is fairly easy. They generally fall into two categories: either they’re nearly right and a small tweak is all that’s needed to land them, or there’s some fundamental misunderstanding that means the code is useless as is, but serves as a very detailed bug report for something people have actually encountered as a problem in real use. And these are both things that have value to me and so make me want to put in the remaining effort to make sure that we have something that benefits us both.

And, if someone who doesn’t think they’re a programmer (you wrote some code and sent it to a real project, you’re a programmer, sorry to be the one to break this to you) put in the effort, that’s also a strong signal to me that this is a fix that actually matters to users.

If you hadn’t bothered to try to understand the codebase at all and just sent an LLM patch, it’s more work to review because LLMs have very different (and far more complex) failure modes than inexperienced programmers. The effort to say ‘Claude, do the thing’ is so low that I don’t know if the person who did it even uses the project or if they were just looking for something to pad their CV, so I have no idea if this is a change anyone wants. And if you don’t understand what the change you’re making does, you’re putting 100% of the cognitive load on me.

Some of the most useful contributions I’ve had to various F/OSS projects have not even been code. Detailed bug reports with test cases that let me reproduce them are incredibly valuable.

@Robotistry I had not considered any of that, but it's a fascinating theory.

@miblo

I spent some time about 20 years ago talking to someone who wanted to contribute to one of my projects who didn't understand why he should be polite. He viewed politeness as a set of arbitrary rules that made no sense to him. This forced me to think of the underlying commonality between things that I considered polite. That was what I came up with and I've found it a useful framing since. It's also helped me discard some notions of polite behaviour that really are just cultural conditioning and don't benefit anyone else if I follow them.

I was trying to work out yesterday why LLM users annoy me so much and it came back to my basic notion of rudeness: impolite behaviour stems from assuming that your time is worth more than other people's.

If someone doesn't bother to write coherent words in a message to you so that they can type faster at the cost of your comprehension speed, they're trading your time for theirs and that's rude.

If someone uses an LLM to create more words from fewer ideas, that saves them time in careful expression at the cost of more work for you. That's rude.

If someone uses an LLM to write code that doesn't work and expects other people to help debugging it, then they're saying that it isn't worth their time to learn but it is worth your time to help them. That's rude.

If someone uses an LLM to write some code that looks like it works and expects you to put effort into reviewing a PR, again, they're saving work for themselves at the expense of your time. That's rude.

Almost all LLM uses are like this: they save the user time (at least, in the short term) by making more work for someone else. They're machines optimised for creating rudeness.

There are a bunch of places where you do trade different people's time. In particular, if tasks A and B need doing, but person 1 can do either and person 2 can do only task B, then it makes sense for person 2 to do task B so that person 1 can do task A. That's not because person 1's time is more valuable, it's because person 1 doing task A is more valuable. It's not that your time is more valuable than the cleaner's time, but the cleaner doing their job means that you can spend my time doing more valuable tasks.

@logaldeveloper @GossiTheDog

1K more users, 1K fewer servers, and you can both be happy.

@lproven Mmm, salted supply-chain attacks.

I wish people wouldn’t have screaming arguments at 6:45am while I’m trying to sleep.

And, by ‘people’, I mean ‘cats and herring gulls’.

@db @TauPan

I view the existence of boilerplate as code smell for missing abstractions. If writing boilerplate is more than a tiny trivial amount of your work, something has gone badly wrong somewhere.

@TauPan I cringe so much when I see people say 'I use it for repetitive things, like tests!'. Because writing good tests requires understanding which bits in the code are corner cases and writing tests that exercise those. I've seen LLM-generated tests with hundreds of tests for the happy path and none that trigger any of the error-handling.

I can maybe imagine a loop with an LLM and a coverage tool to get proper coverage, but then you'd need to review the tests generated by the LLMs for the corner cases to make sure you weren't just generating tests that enforce bugs in the implementation.

@kirtai Exactly. Imagine doing code review where every PR was an entry into the underhanded C contest.

@spitfire I think that’s part of why people trust it for code. You know it was invalid because the parser rejected it. If you have to compile the code, run it, and pass tests, then you have some confidence that it is not completely wrong. The problem is that you have a tool that is especially good at producing subtly wrong output.

@spitfire

It’s something you can predict if you’ve read LLM-generated prose too. They generate entirely plausible-sounding text. Spotting the errors is hard. If you’d written the text, you wouldn’t introduce errors from completely fabricated data and citations. If someone who wasn’t actively trying to trick you wrote the text, they wouldn’t include made-up numbers and citations. But somehow people expect code to be magical and special.