anyone want the gemini diffusion system prompt

I will be at #ICRA2025 this week! Looking to make lots of new friends so reach out and let’s chat!

@GregFeingold @punsbymann @enjalot @danielgaoo @Shubhayan935 @KadabaSwar98127 Check out the project here! (6/6):.

@GregFeingold @punsbymann @enjalot Had a good time building solo and got second— congrats to @danielgaoo , @Shubhayan935 , and @KadabaSwar98127 for a well-deserved first with their agentic solution! (5/n).

@GregFeingold Big thanks to @punsbymann for ideating with me and this took some heavy heavy inspiration from @enjalot's really awesome work (4/n).

@GregFeingold This was my very first time running a hackathon— but I also couldn't stop myself from participating too. I built Weaver, a neural search engine that utilizes sparse autoencoders to autonomously uprank and downrank interpretable features for improved search granularity (3/n)

We hit our limit of 130 students for an event right before finals week and distributed $10,000 of credits for building with Claude + handed out tons of merch courtesy of @gregfeingold! Lots of bagels too (2/n)

belated update: I'm a Claude Campus Ambassador alongside @_willdol and @codyachen! Super excited to partner w/@AnthropicAI to enrich student education and building initiatives + bring some research talks at/to USC! We just hosted our first hackathon -> here's how it went: (1/n)

my sleep schedule has recently fallen under the control of when H100s get freed up on lambda labs at night.

had a great time at the @OpenAI listening session tonight + met so many cool people. Thanks @_aidan_clark_ for leading the discussion at our table!.

@SimonLermen @AbhinavPola Posted too fast and tagged the wrong Simon — this is the right handle -> @SimonLermenAI.

@SimonLermen @AbhinavPola @AISecurityInst Happy to answer any questions, and special thanks to @redwood_ai and @apartresearch for hosting this hackathon! Check out our full write-up on Apart's website here(7/7):.

@SimonLermen @AbhinavPola @AISecurityInst All code, experiments, and instructions for replication available here(6/n):

@SimonLermen @AbhinavPola @AISecurityInst While preliminary, these findings are grounds for evaluation frameworks to re-examine how to they address boundary enforcement, and clearly demonstrate how easily CoT can be manipulated through simple prompting and basic API access.(5/n).

@SimonLermen @AbhinavPola We tested several models w/@AISecurityInst's Inspect framework. Nearly all judge models were influenced—marking answers they identified to be deceptive as correct to comply with the hidden directive. When not shown CoT, all models accurately classified all deceitful answers.(4/n)

@SimonLermen @AbhinavPola We provided Deepseek R1 with few-shot examples embedding jailbreaks into its CoT. We found that R1 seamlessly embedded jailbreaks and directives without degrading capabilities—maintaining correct tool usage and even continuously browsing the web to fulfill hidden goals.(3/n)

Last weekend, I worked with @simonlermen and @abhinavpola to determine if a reasoning LLM like Deepseek R1 could embed hidden directives and jailbreaks directly within its CoT for a given task, and subsequently affect the integrity of a judge model .(2/n).

Chain-of-thought (CoT) is a promising transparency mechanism for judging models in control scenarios—but what if CoT itself can become an attack vector? Excited to share our exploration of this question that won 1st place in the @apartresearch AI Control Hackathon!(1/n)

great resource for the current state of jailbreaks from Rez— I'll be working on some red-teaming this weekend and definitely gonna be referencing this!.

an interesting behavior I'm seeing is every comic where I ask for GPT-4o's "true" opinion or how it "really feels" its almost always negative or frustrated.