ACS

Alignment of Complex Systems (Center for Theoretical Studies, Charles University)

Research at the intersection of AI alignment and complex systems

[ACS @acsresearchorg]

@JohnKocdr Yes, we provide full logprobs

[ACS @acsresearchorg]

Base models are super useful for deep understanding of LMs, and what originates in pre-training vs post-training. But experimenting with large base models can be hard, annoying or costly for individual researchers.

That’s why we’ve built a platform that gives free access to base models for researchers and tinkerers working in LLM psychology. Currently serving Llama 3.1 8B, 405B, and Trinity-Large-TrueBase

1/3

[Jan Kulveit @jankulveit]

What is LM psychology, why care about it, and how do you do it well? Our new post with @raymondadouglas makes the case for applying the psychology frame to AIs, along with a reading list and suggestions for where to start

[ACS @acsresearchorg]

It's free. If you work on LLM psychology, interpretability, cyborgism or something adjacent and might benefit from more base models access, please apply. No publications or affiliation needed.

More detail at https://t.co/BL2PenWvT3

3/3

[ACS @acsresearchorg]

Apply for access at: https://t.co/ck6EeL5DpE

We provide features needed for research: completions, full logprobs, activation harvesting and steering, browser workbench for fast iteration, loom for comparing outputs, … and we plan to expand the platform. Join discord for sharing research ideas and results.

2/3

[ACS @acsresearchorg]

Post: https://t.co/RNs4WhAbcJ

[ACS @acsresearchorg]

A lot of AI alignment discussions are concerned with possible convergent features of future AI systems, such as power-seeking or situational awareness.

Here, convergence & contingency are typically thought of as opposites. This article suggests a different perspective!

[ACS @acsresearchorg]

It may allow us to take some (dangerous) feature of future AI systems which we expect to be convergent, and ask “what is the feature contingent on?”.

This in turn may provide a reliable point of intervention to prevent the dangerous features from arising in future AI systems.