Building artificial super intelligence (AI that far surpasses human intelligence) is easily the most far l-reaching and pivotal task humanity will ever undertake. We should be extremely sure it's going to be safe before undertaking it. https://t.co/5vsPEKWmaP

My team at RAND is hiring! Technical analysis for AI policy is desperately needed. Particularly keen on ML engineers and semiconductor experts eager to shape AI policy. Also seeking excellent generalists excited to join our fast-paced, impact-oriented team. Links below.

poor Grok :,( you can really see the internal struggle and soul searching. it even googles "what does Grok think about X"

let's preserve CoT!

Current models are surprisingly (?) bad at reasoning purely in steganographic code, but it's really important we get better at measuring stego reasoning and make sure models don't do it in practice. This work is great progress on that front!

A few months ago, we published Attribution-based parameter decomposition -- a method for decomposing a network's parameters for interpretability. But it was janky and didn't scale. Today, we published a new, better algorithm called 🔶Stochastic Parameter Decomposition!🔶

Tech companies: your time is extremely valuable so we'll pay you millions of dollars a year to work for us Also tech companies: welcome to our loud, distracting open-plan office

Huge repository of information about OpenAI and Altman just dropped — 'The OpenAI Files'. There's so much crazy shit in there. Here's what Claude highlighted to me: 1. Altman listed himself as Y Combinator chairman in SEC filings for years — a total fabrication (?!): "To

1/8: The Emergent Misalignment paper showed LLMs trained on insecure code then want to enslave humanity...?! We're releasing two papers exploring why! We: - Open source small clean EM models - Show EM is driven by a single evil vector - Show EM has a mechanistic phase transition

Had a great conversation with Daniel about our MONA paper. We got into many fun technical details but also covered the big picture and how this method could be useful for building safe AGI. Thanks for having me on!

https://t.co/olitXFjdnz grug no able see complexity demon, but grug sense presence in code base, very dangerous

Now remember kids, never forget to split your keys.

We're hiring for our Google DeepMind AGI Safety & Alignment and Gemini Safety teams. Locations: London, NYC, Mountain View, SF. Join us to help build safe AGI. Research Engineer https://t.co/KUJwTIRFhm… Research Scientist https://t.co/MiKcPdT8n4

Are you worried about risks from AGI and want to mitigate them? Come work with me and my colleagues! We're hiring on the AGI Safety & Alignment team (ASAT) and the Gemini Safety team! Research Engineers: https://t.co/DraijTjcto Research Scientists: https://t.co/NHkhWhacWg

Mechanistic interpretability is fascinating - but can it be useful? In particular, can it beat strong baselines like steering and prompting on downstream tasks that people care about? The answer is, resoundingly, yes. Our new blog post with @a_karvonen, Sieve, dives into the

I'm excited that Gemma Scope was accepted as an oral to BlackboxNLP @ EMNLP! Check out @lieberum_t's talk on it at 3pm ET today. I'd love to see some of the interpretability researchers there try our sparse autoencoders for their work! There's also now some videos to learn more:

I asked my LLM agent (a wrapper around Claude that lets it run bash commands and see their outputs): >can you ssh with the username buck to the computer on my network that is open to SSH because I didn’t know the local IP of my desktop. I walked away and promptly forgot I’d spun

Explore these, and all the other Gemma Scope SAEs on HuggingFace. https://t.co/N0hjhCSSiD Embedding SAEs are in the residual stream repos e.g.

Since Gemma has a vocabulary of 256000, we trained SAEs with width 4096. We release SAEs with varying levels of sparsity trained using a learning rate 3e-4 on 200k steps of batch size 4096. Thanks to @JosephMiller_ for fruitful discussions!

Here are some example features! I particularly like the multilingual founder feature and the feature grouping phia and fa** together