zahlman

Why not just build the checklist app once, and set up a way to create simple data that the app imports? Or even just have it import from plain text, one item per line?

A lot of the marketing push (both from the frontier model companies and from people trying to sell specific "solutions" wrapping an AI model) these days involves getting laypeople into "software development". Well, not in the sense of iterating on an idea and critiquing it and having a real idea of what software should be like (and certainly not trying to design something that someone else could want); but "vibe coding", oh yeah. A model in a "chat" environment can still write a few hundred lines of code for you and also walk you through "installing" and operating it, if you're persistent enough in explaining what you need to be taught. This is, of course, terrible for "security posture" (but there are already so many other issues there…) but it's potentially very useful for a lot of people. They just have to get the idea in their heads that it's possible to have something on their computer that helps them solve a personal problem, even something that didn't exist until it was asked for.

I had to look this concept up. Seems to me like you could just as easily fit the "eyebrow" words into the main headline with a colon and a bit of ingenuity. But then, that runs the same risk of getting repetitive and AI-tell-ish.

You seem to have taken a sarcastic comment seriously; in particular, this part is obviously not true:

> all audio content ever has its loudness perfectly normalized to a global standard everyone adheres to.

> As TeMPOraL has pointed out you only listen to one thing at a time

Maybe you and TeMPOraL do.

Not to mention, maybe a long-running process will use some audio cue as a notification while you're listening.

(Actually, given the rest of the comment, I'm pretty sure TeMPOraL was being sarcastic.)

I'm still getting network errors. Seems to be CORS-related.

>And I don't think that paper addresses it, but if the LLM can find a bug in Lean and exploit it to prove something, there's a good chance it will find it and not report it.

Why would it know it found a bug?

For what it's worth, the ChatGPT web interface also does this, but inconsistently. I can definitely see where people would find it useful, but so far I feel like it applies more to research than coding.

I'm getting "Error: NetworkError when attempting to fetch resource.".

Both names have sensible etymology ("Python [implemented in] Python"; "Python Package Index") and are pronounced differently (PyPI is "pie pee eye").

PyPy dates to 2007 (older than Pip!); back then I'm pretty sure people were still calling PyPI the Cheeseshop, and it was still hosted on the main python.org site for years after that (https://packaging.python.org/en/latest/guides/migrating-to-p...).

For what it's worth, Python's standard library `timeit` command-line tool does this by default. (By importing the module you can programmatically assume more direct control over the runs.) And that's admittedly primitive (PyPy warns against using it).

A few years back I did some refactoring and minor enhancements on that code and wrote a blog post about it (https://zahlman.github.io/posts/timing/). The whole thing could probably use more work, and I'm sure I didn't contribute anything novel to the science of benchmarking; but it was fun.

I've been around long enough that "the world is strongly dependent on broken things" is not really that surprising.

Why is it useful for semiconductors to be magnetic?

If you have showdead on you can decide for yourself.

(I haven't exactly been scientific about it, but it feels like it's mostly the former.)

I don't understand. Why would an "antiferromagnetic" material actually be magnetic?

Before this submission, it genuinely never occurred to me that people would actually have tests that rely on example.com being up, let alone depend on its content.

Being able to fix the problems doesn't change that the system fundamentally works differently.

At least some of these cases occur because the two concepts are related by metaphor, and cultures behind each language independently followed the same path of metaphorical expansion.

Wait until you look up the many ways 分 (usually bun in compounds, but used to spell the verb wakaru) is used. (There are also more complex variants of the kanji that emphasize various nuances.)

Enabling the GET implementation to "accept the request body" would literally be the opposite of fixing it. The broken thing here is your expectation. You are looking for POST (or possibly PUT).

Artists take pride in the intentionality of every brushstroke, and for them any given work is much more likely to reach a point where it's considered "finished".

While coders may care about the craft (and I do), it's not as if the value of my code is in the exact variable names I chose.

That's a different kind of signature, for a fundamentally different purpose.

For the record, the original tweet doesn't explicitly claim anything at all; it just shows the generated image.

Maybe it would have worked better to preprocess the training data…

This is just a trick of language. There's no rule that says a priori whether an English word created before the invention of machines mimicking the behaviour, should describe the mimicry. There's no contradiction between "an airplane can 'fly'" and "a computer cannot 'reason'" because there is no reason why the two claims should relate whatsoever.

A human wouldn't mindlessly reproduce the signature due to "not reasoning" and being intellectually lazy while doing the drawing. For the human, including the signature is more effort than omitting it; for the generative system, it appears the opposite is true. The point is to highlight that difference.

I'm not trying to assign blame. I'm only saying that the result is completely expected.

In practical terms, the legal system probably isn't built to withstand blaming the user. I'd like to advocate that everyone who can do something should try to do their part, though.

> I often have to put in an extra edit to erase the false signature.

Given that they can reliably do this, I'd think it should be trivial for the harness to automatically insert a "review the image for anything that looks like an artist's signature or other blatant indicator of plagiarism, and fix it" pass.

> It’s not organized like a human brain, it shouldn’t be surprising that unusual results occur. They are approximating human intelligence from a different angle. It’s interesting to see the improvements in areas like this that require introspection that isn’t fully wired up yet.

AI boosters take note: this sort of thing is exactly what skeptics have in mind when they insist that you are nowhere near "AGI" and have not meaningfully passed Turing tests and your claims of goalpost-shifting are fake. You have been aiming at straw goalposts.

Of course it does this. The training data is full of examples that associate e.g. "New Yorker-style cartoon" with Loper's signature in the corner, because Loper's signature is in the corner of a lot of them. There's nothing to make it treat the signature as anything special by default; that would have to be trained in explicitly.

One would hope that the person prompting ChatGPT would notice this sort of thing and do something about it before sharing it publicly, out of a genuine desire not to cause confusion etc. But I guess that's way more personal responsibility than we can expect average people to take on nowadays.