astrange

6 ms·

1.
by astrange 9d ago
The cyborg you replied to.
2.
by astrange 11d ago
Something like this seems unlikely to work for a behavior so burned into them. It'd be better to do a second pass to delete all the nonsense comments.
3.
by astrange 11d ago
That's a Claude.
4.
by astrange 11d ago
> I have a stage-gated workflow This is a Claudism, right? I feel like I never saw "gated" used this way before it.
5.
by astrange 15d ago
Eh. They recently changed the algorithm quite a lot, and the guy who was in charge of it (Nikita Bier) made it a lot less evil when Elon wasn't looking. But now he's gone so who knows what will happen.
6.
by astrange 15d ago
Unlike GPT (I think?), Claude is served by all of AWS/GCP/Azure and you can go buy it from any of them.
7.
by astrange 15d ago
> - AI centralizes power in the hands of capital, rendering those who can afford compute hardware vastly more capable than those who cannot and thus increasing social stratification Looks to me like I can run Claude Code without being ab
8.
by astrange 15d ago
After a year of not being able to serve Claude because they ran out of datacenters I don't think they want to go back to that. (If they did, they wouldn't have added the effort level.)
9.
by astrange 15d ago
No, there's no reason chatbot behavior would have anything to do with frequency of text in pretraining.
10.
by astrange 15d ago
I think the specific issue with Opus 5 is that its writing style is just trying to cheat at RL. It makes everything hypey yet self deprecating and constantly brings up "honest caveats" because the scoring rubrics look for those.
11.
by astrange 15d ago
> Twitter was changed so that only views favoured by Musk show up prominently in replies. This doesn't happen anymore, it's actually relatively good at sorting bad comments to the bottom now. The current problem is it's fu
12.
by astrange 19d ago
You can't turn off random parts of a system someone else made. You don't know what they do. https://limi.net/checkboxes
13.
by astrange 20d ago
Anything with ranged numeric types. Like everyone's favorite functional programming language, Ada.
14.
by astrange 20d ago
Haskell's type system would not easily prevent this bug. It's not good at numeric/logic issues like that. When people say "Haskell makes it impossible to write bugs" they mean "Haskell has enums" (ADTs).
15.
by astrange 20d ago
Since the groups are unrelated, they don't actually have consistent wealth levels or belief. It's literally just a list of random people she doesn't like.
16.
by astrange 20d ago
Background processes almost never really use expensive resources on macOS or Linux. Certainly they can cause symptoms sometimes, but it's typically no worse than extra fan noise. All the wasted resources on most people's machines
17.
by astrange 20d ago
TESCREAL is a word meaning "a bunch of unrelated groups that Timnit Gebru thinks are nerds", who are not really aligned with the administration. (I also think they're nerds.)
18.
by astrange 22d ago
I don't think there's anyone who wants to fund "committing felonies", for obvious reasons.
19.
by astrange 22d ago
The issue is you're usually using inline asm for special cases, or else CPU features so new the compiler doesn't know about them. Which means it also doesn't know the clobbering rules. That's why it looks like an escape
20.
by astrange 22d ago
In Steve's case "fueled by spite" would be more precisely "borderline personality disorder". If you read him in that lens it might help distinguish how to be successful like him without all of the bad parts.
21.
by astrange 26d ago
Anthropic is a PBC, so they're required to work towards a public benefit. They're also basically a religious cult dedicated to Claude.
22.
by astrange 26d ago
> To be honest, if I could stop AI from saving memories, it would be fantastic, because claude code etc definitely creates more issues for me when it creates memories than anything it solves. You can turn that off, and I have. But Opus 5
23.
by astrange 26d ago
It's a specific dialect of its own that comes from RL. Nobody on Earth talks about load-bearing honest caveat seams. (Another strange thing about Opus 5 is that it will constantly re-mention things it thinks are important, probably a s
24.
by astrange 26d ago
I've had to read difficult things at work in my life, like C++ error messages, but I think Opus 5 is the first time I felt like I was being poisoned by my terminal. I keep having to leave and go use the office gym.
25.
by astrange 27d ago
5.6 has a totally different style again, kind of relaxed and neutral with some definite jokes.
26.
by astrange 27d ago
It's mode collapse from RLVR. That and if you talked to the same one person's frozen brain upload all day, you'd see the same catchphrases used too.
27.
by astrange 27d ago
Pretraining is full of bad writing and it doesn't really cause issues. Writing style comes from post-training. In this case it's gotten worse because they prioritized agentic abilities.
28.
by astrange 27d ago
The training methods try not to apply any particular rules to the contents of the thinking text. That's called "optimization pressure on CoT" and is thought to reduce safety by inducing the model to lie (or stop clearly print
29.
by astrange 28d ago
Okay that's not Mastodon people, that's Metafilter or Tumblr people. Mastodon people in my head are European Linux users who are only capable of replying to things in the form of a non-constructive disagreement.
30.
by astrange 1mo ago
I wouldn't say "stupid", but trying to make AI go away by posting about it on the internet is sort of like… if you were about to get hit by a car and tried to solve it by posting about it on the internet. It's not going

More ›