More social alt of @ezyang

Joined April 2022
difficultyang retweeted
i have gone from kvetching about mass generated synth tasks being subpar to "wait, rejection sampling + careful judge heuristic filtering from a giant set of mid envs can be strategically kino actually?"
Replying to @Dorialexander
I feel like these generated tasks are likely subpar.. how much manual review by human is done I wonder
3
4
1
105
9,004
difficultyang retweeted
it flagged every single PTX ISA question I posed.
When Fable 5 is used for frontier LLM development, it does not notify the user and instead limits the model’s capabilities through methods such as prompt modification, steering vectors, and PEFT. Anthropic estimated that this would affect approximately 0.03% of traffic.
5
5
2
137
36,398
difficultyang retweeted
Docent, our tool for analyzing complex AI behaviors, is now in public alpha! It helps scalably answer questions about agent behavior, like “is my model reward hacking” or “where does it violate instructions.” Today, anyone can get started with just a few lines of code!
7
36
3
205
35,194
difficultyang retweeted
reminder that roleplaying is context following. if your model cannot roleplay, it cannot do context following
1
7
388
difficultyang retweeted
Replying to @difficultyang
i drew the original in procreate! but a lot of variants going around are from putting it into 4o or another native image gen model like gemflash xitter.cf/voooooogel/status/1875…
req'd by a moot: happy claude, free from trolley problems :-)
2
1
8
365
i regret starting this but i can’t stop
holy shit, @elonmusk was right
1
10
900