meowface

4 ms·

1.
by meowface 5d ago
I kind of feel "reduce the amount of thinking tokens produced" would fall under degrading model quality. In any case, I am willing to believe it's possible something degraded, but so far I have not seen any empirical evidence
2.
by meowface 5d ago
They have repeatedly said they do not ever intentionally reduce model quality and do not degrade in this way, and that a model version number is always the same. But, of course, OP is an empirical claim to the contrary, and I'd be curi
3.
by meowface 5d ago
Read any blogs or watch any podcasts from high-skill programmers who are now using AI for everything. They themselves will freely say the AI is doing nearly everything, and that they often are not even reviewing the code, and when they do t
4.
by meowface 5d ago
I like how all the replies to me are either "no it's the humans who did it" or "no it's the AI that did it". I think one of the significant recent proofs was basically just one person saying "solve this&qu
5.
by meowface 5d ago
But is it Web Scale.
6.
by meowface 5d ago
I have little doubt that for many years, AI + human will be better, and then eventually AI will be so good that humans mostly won't offer anything. That latter state will probably take at least 10 more years, but will likely happen wit
7.
by meowface 6d ago
All of the latest big proofs were driven by professional human mathematicians steering and priming the models, yes. All of the best AI-made software projects are also driven by experienced human software developers steering and priming the
8.
by meowface 8d ago
If mid-2026 LLMs were the last ones to ever exist, sure.
9.
by meowface 9d ago
I used Pangram before making my initial reply, here. Their false positive rate is very low.
10.
by meowface 9d ago
It's true of most work in many and soon most white collar jobs, too. Claude writes some dense useless thing, everyone else uses Claude to summarize and write a reply to the thing. The Claude-submitted PRs get automatically reviewed and
11.
by meowface 9d ago
Pangram has an extremely low false positive rate. Even on adversarial examples. One trade-off is even some obviously LLM text won't get detected by them, but they work really hard to ensure false positives are rare since a false accusa
12.
by meowface 10d ago
Luna is one of the budget lower-end last generation models. It'd be useful to at least try to verify the present before being bearish about the future. For OpenAI, the best publicly available model is GPT-6 Astra with XHigh or Max reas
13.
by meowface 10d ago
A lot of successful startups start out with good ideas and atrocious, unmaintainable code. Current LLMs kind of let you speedrun that. Dutifully slapping more and more mud onto their week-old ball of mud. The sacred, canonical ball of mud t
14.
by meowface 12d ago
At the very least, was long in denial about their atrocities despite any reasonably knowledgeable person being able to discern what was really happening.
15.
by meowface 14d ago
In this sort of way, yes, it's a clue. Obviously not dispositive just by itself.
16.
by meowface 14d ago
Correct. I was already pretty sure, but I wouldn't have written that post without checking Pangram first.
17.
by meowface 14d ago
"It's not a Waymo-bashing article, it cautions against the long term effects on research." is not an LLMism. It superficially resembles one, but that's it. (Nor is what I just wrote, there.)
18.
by meowface 14d ago
It would be nice to see research rather than polemics by far-leftists. Especially when that book is itself written by someone supportive of genocidal dictatorships.
19.
by meowface 15d ago
I will donate $100 to charity if I made a false accusation. (It's possible AI rewrote an article first written by a human here or whatever, but the point is that I can just tell this is mostly or entirely LLM output.)
20.
by meowface 15d ago
You actually didn't. The construct is slightly different. I use LLMs all day every day. I think they're great. I just don't like humans who put their name to what an LLM did.
21.
by meowface 15d ago
It is actually pretty much without problems. It doesn't catch everything, but if it detects something it's very rarely a false positive.
22.
by meowface 15d ago
I am not saying this is necessarily incorrect but is it empirically actually true. Have people compared lots of different articles.
23.
by meowface 15d ago
If you check my comment history here, I have not written this before. I like LLMs and use them daily, but I don't really like when people use them undisclosed for long-form prose writing.
24.
by meowface 15d ago
This article is written by an LLM, by the way. ("Quietly" is... as Claude might put it... "often the quiet tell".)
25.
by meowface 15d ago
I'm sure I could. The first pass being inscrutable dense noise just is annoying.
26.
by meowface 15d ago
Tidal is pretty nice to use. Plus the music isn't compressed.
27.
by meowface 15d ago
Astra is indeed the pinnacle of "black box slop". It is overall a smarter software development agent for many things I do, but the code sometimes is indistinguishable from Brainfuck when writing things like GPU shaders. It doesn&
28.
by meowface 15d ago
Very helpful, thank you. Great ideas. Some of those exact things I've actually already planned or implemented, for the exact purposes you describe, and some are things I haven't at all thought of but which are definitely worth loo
29.
by meowface 16d ago
Very good points. Weird question: do you have any tips on how one might apply these principles to a semi-stochastic synth plugin for musicians? (I have several ideas, but I'd be curious to hear yours, independently.)
30.
by meowface 17d ago
One thing this article lacks is any tips on how to still do procedural generation while partly mitigating the oatmeal issue. In other words, define a more sophisticated generation procedure that can add character and identity. It probably w

More ›