Situational awareness benchmarking shows increasing performance with newer LLMs, but not on this one: ANTI-IMITATION tasks challenge LLMs that naively imitates training distribution. To succeed, an LLM must use details of the LLM itself and its particular non-human capabilities.

LLM-robo girlfriend

1/ Excited to finally tweet about our paper “Implicit meta-learning may lead LLMs to trust more reliable sources”, to appear at ICML 2024. Our results suggest that during training, LLMs better internalize text that appears useful for predicting other text (e.g. seems reliable).

Great opening to a philosophy book review. "Barrister" is roughly "trial lawyer"

"The Annunciation", Oleksandr Murashko. 1909, National Art Museum of Ukraine Saw this in Bratislava (Slovakia) on loan from Ukraine. Not much info online about this painting.

Jan Leike on two AI alignment threat models

I'm at #ICML2024 in Vienna this week.

I so enjoyed talking with Dwarkesh Patel about my essay series “Otherness and control in the age of AGI.” He engaged so deeply, asked such great questions, and aimed so directly at the core of the issues at stake. It’s a pleasure to be a part of conversations like this.

I've interviewed Owain Evans about his work on Al situational awareness and out-of-context reasoning in LLMs Owain has been publishing some of the most surprising and important Alignment papers in the past year, and I'm proud to be his first longform podcast ever

Great interview: 80000hours.org/podcast/episod…

Susan Sontag and John Berger, podcast, circa 1983 youtube.com/watch?v=MoHCR8…

Useful thread on proposed California AI bill (SB 1047)