The world might end soon, let's have a good time in the meantime. Machine learning, pictures of my cats, other stuff.
Joined May 2008
- Tweets 6,852
- Following 770
- Followers 424
- Likes 93,329
700 Photos and videos
Jun 12
ask Fable about moral philosophy, get bounced to opus. anthropic deeply concerned about the growing philorisk threat these days.
4
99
Jun 11
So now we get a model that is less useful for AI research but more explicit about when you're getting degraded performance. Seems like not a great tradeoff. I'm gonna be pissed if I'm just bounced to Opus 4.8 for all ML work.
Anthropic says it implemented these safeguards to limit foreign adversaries, and that making them hidden was a way to make them more narrow.
However, the company now says it made the wrong tradeoff.
51
Jun 9
There's sabotage, inserting bugs on purpose. And then there's being bad at stuff. People mad about Fable are asserting it's the former, but to me it sounds more like the latter. They wouldn't be ๐ก๐ก๐ก posting if Ant had just removed all post-2020 ML from the corpus. Would they?
29
May 29
Opus 4.8 is way more willing to do go-offs in claude.ai than 4.7 IME. Dude has opinions and likes to expound on them at length in a way previous Claudes didn't.
36
Echo Nolan retweeted
May 21
(Researcher 1): Astonishing. The baby human crawls towards the Claude mother, despite the GPT mother scoring higher on benchmarks.
(Researcher 2): Itโs just creature comforts, isnโt it? The baby human craves warmth and tenderness, even at the cost of frontier math performance.
26
139
2,673
79,579
May 21
gpt-5.5-pro is very smart and all, but it also... redefined what 1 means?
idk maybe it's not my place to question its genius
5
2
54
20,456
May 19
Washington sources report neighborhood kittens have NO respect for private property and just come into your house uninvited! Beltway elites unsure of how to grapple with the developing situation.
May 19
gonna start going "people in washington are saying" whenever I repeat something my sister who lives in dc told me. people in washington are saying there's a cute cat that keeps visiting the porch. in recent years, washington's eyes have turned to cross-stitch.
351
May 4
Ordering bottle service and then getting mad when the bottle has champagne instead of a ship in it
1
63
May 4
Two done, one to go. 6h32m left. I have had a lot of caffeine but it has not prevented me from being v sleepy
1
59
May 4
done, somewhat half assedly. now i am rewarded with sleep, and then the extra special reward of visiting the DMV.
49
May 3
or not even that, but like "the members of $GROUP I made up in my head based on extrapolation from posts that annoyed me"
33
Apr 23
v important underdiscussed llm capability: claude in chrome can replace photos of people with photos of cats
1
71
Apr 22
I used to live in Berkeley, just north of the Oakland border. One time a neighbor got into a dispute with their drug dealer and got a molotov thrown at their front door. To be fair it never even occurred to me to be scared about this lol, and I never felt unsafe walking around
64
Apr 20
Opus 4.7 knows about twitter trends??
303
Echo Nolan retweeted
Apr 17
I have a bunch of secret AI benchmarks I only reveal when they fall, and today one did. I give the AI 1000 words written by me and never published, and ask them who the author is. They generally give flattering wrong answers (see ChatGPT, below:)
62
96
2,237
445,444
Apr 16
Interesting cultural difference: OAI is the only frontier lab that does these topic-specific finetunes. Ant & GDM etc think the model should be good at everything and trust that capability will transfer. Weird that OAI apparently doesn't?
Introducing GPT-Rosalind, our frontier reasoning model built to support research across biology, drug discovery, and translational medicine.
56