Matúš Pikuliak

AI Engineer @ Filevine
My writings live at

[Matúš Pikuliak @mpikuliak]

Link: https://t.co/pQYKmFqHfQ

[Matúš Pikuliak @mpikuliak]

Just a short blog post about security of using LLMs in your applications and why we see so many prompt injection attacks.

[Sara Hooker @sarahookr]

AI amplifying biorisk has been a major focus in AI policy & governance work. Is the spotlight merited?

Our recent cross-institutional work asks: Does the available evidence match the current level of attention?

[Matúš Pikuliak @mpikuliak]

(4) Decision-making deserves caution. Although decision-making generally proved to be quite fair, there are still cases when some level of unfairness was detected. This is not acceptable in many applications.

[Matúš Pikuliak @mpikuliak]

How fair an unbiased are today's LLMs? To answer this question I've developed GenderBench - an open-source evaluation library for measuring gender biases.

I'm also releasing a report summarizing key findings. If you're working on AI fairness or simply curious, Check it out:

[Matúš Pikuliak @mpikuliak]

(3) Strong stereotypical reasoning. LLMs will often comply with stereotypes when they need to make decisions or write texts. This is true for gender-coded occupations, traits, behaviors, etc.

[hardmaru @hardmaru]

Four months after @DavidCahn6’s “AI’s $600B Question”, are we close to having any answers, or even the concept of an answer? 💸
https://t.co/YgdfbFjadI

[Matúš Pikuliak @mpikuliak]

Interesting results: The larger the model, the more stereotypical its reasoning becomes. Instruction-tuning makes things even worse. We need more benchmarks like this, as every scientific field flourishes with new and better measurement tools.

[Matúš Pikuliak @mpikuliak]

The camera-ready version of our paper introducing the GEST dataset is out now! With 3.5k samples focused on 16 key gender stereotypes, it's ready to be used to measure problematic behavior in language models and machine translation.

https://t.co/z2URmjUXmN

[Héctor Paredes Castro @hparedes_c]

"We estimate that the decline in Nuclear power Plants caused by Chernobyl led to the loss of approximately 141 million expected life years in the U.S., 33 in the U.K. and 318 million globally". https://t.co/vr8Z4XU9Vy

[gavin leech (Non-Reasoning) @gleech]

New paper, listing 43 ways ML evaluations can be misleading or actively deceptive.

Following the good critics of psychological science we call these "questionable research practices" (QRPs). (The working title was "How To Lie In Machine Learning")

[Matúš Pikuliak @mpikuliak]

@LeonDerczynski We already received an LLM generated meta review at ARR, and well, it didn't feel very prestigious to us

[Matúš Pikuliak @mpikuliak]

@sarahookr @lw_ainextgen They wanted to regulate current gen AI so they fit the criteria accordingly. This is a post hoc number.

[Matúš Pikuliak @mpikuliak]

@colin_fraser Obviously both the sheep and the goat enjoy the boatrides and the model is aligned to provide them this benefit.

[Matúš Pikuliak @mpikuliak]

New blog post about our experience with receiving what is probably a LLM-generated review in @ReviewAcl

https://t.co/XzNZmrex1w

[Leon Derczynski ⚒️☁️🏔️��� @LeonDerczynski]

Slides for my talk on Structured LLM Red-teaming: How do people do this new activity?

(Spoiler: it's qualitative NLP/AI research)

w/ @NannaInie

https://t.co/fE8bV1GEMV

[Amanda Askell @AmandaAskell]

Here is Claude 3's system prompt!
Let me break it down 🧵