Rajashree Agrawal
@___rajashree___
Followers
317
Following
3K
Media
3
Statuses
140
We automatically generated the unit test that would’ve caught @AnthropicAI’s top-K compiler bug without relying on their bug reproducer code. Most testing pipelines never hit rare bugs until they fail in production. Ours do.
3
9
52
An interesting concept! I'm curious to see where it goes
Theorem (@theoremlabs) is an AI-coding IDE for mission-critical software. They're making program verification 10,000 times faster, so your systems code does exactly what you asked for. Congrats on the launch, @diagram_chaser and @___rajashree___! https://t.co/hqZJINedRj
1
3
44
If Andrew says so, it must be true 😁
This just might be the greatest Y Combinator company of all time
0
0
6
Theorem (@theoremlabs) is an AI-coding IDE for mission-critical software. They're making program verification 10,000 times faster, so your systems code does exactly what you asked for. Congrats on the launch, @diagram_chaser and @___rajashree___! https://t.co/hqZJINedRj
9
23
168
Mechanistic interpretability gives us rich explanations of models. But can we convert these explanations into formal proofs? Surprisingly, yes! Mech interp helps write short proofs of generalization bounds — and, shorter proofs provide more mechanistic understanding. 🧵
1
34
178
I believe much more of our bandwidth should be spent getting ready for the next generations of models, on security, monitoring, preparedness, safety, adversarial robustness, (super)alignment, confidentiality, societal impact, and related topics.
30
248
3K
MSJ is an order of magnitude more effective at jailbreaking Claude than SOTA attacks!!! The paper is a cool step in thinking about *model capabilities* as *attack surfaces*. Congrats to @cem__anil on knocking it out of the park! Super glad to have been a part of it!
New Anthropic research paper: Many-shot jailbreaking. We study a long-context jailbreaking technique that is effective on most large language models, including those developed by Anthropic and many of our peers. Read our blog post and the paper here: https://t.co/6F03M8AgcA
0
2
13
There's no clearly winning score on OCEAN, but there is a losing score. Smh. Just like BMI.
0
0
6
Obvious but surprising fact from analysing my self-reported daily productivity metrics from 2023: using AI tools was anti-correlated with deep work. I would like to experiment with this more, but don't have any bright ideas for it.
0
0
4
Are you afraid to value bet unless you’re almost positive your hand is good? It’s not uncommon, but you’re missing out on a massive amount of value and it’s costing you $$$!
18
2
129
games i have experienced strong tetris effect for so far: 1. wordle (+variants): thinking in 5 letter words 2. set: 2 colorful images popping up, find the third one 3. kakuro: quiz me on how many ways there are to make 5 unique single digit integers add up to 28
0
0
3
it is interesting how much search results change when googling symptoms vs symptoms + "woman"
0
0
0
"substantive" economics and mathematics discussions in academic settings are only bearable to me when i go in with the intention to make notes on how people will miscommunicate. i do this about once a week, also the frequency of my desire to start a new teaching program.
0
0
0
misplaced attempts at structure can overwhelm. dependency on structure can squeeze out serendipity.
0
0
0
finding structure can give mobility, add efficiency, illuminate beauty.
1
0
1
being obsessed with finding underlying structure is a gift and a straitjacket
1
0
2