Daniel Kokotajlo: To be clear, we don’t claim P will happen specifically. But when we wrote out our best-guess scenario month by month, P kept happening. Eventually we decided to just publish P. I’m at ~80% on P; my coauthors are lower. Ryan Greenblatt: I thought it would be helpful...
Summary I'm increasingly convinced that model access parity is a big deal and we are not on track to achieve it. By model access parity, I mean a small gap between (i) the model access for lab employees and (ii) the model access for external safety researchers, third-party auditors, and...
Over time, there might be an increasingly large gap between insider model access and outsider model access. By insiders, I mean employees at the frontier lab.[1] By "outsiders", I mean external safety researchers, third-party auditors, and other actors trying to make the future go well. I will call this a...
By default, I expect system cards will get worse, which would be bad. Some mechanisms could improve system cards, but I expect they will be outweighed. In any case, I think third-parties should focus on scrutinising system cards — this seems like a great activity for outsiders in the current...
This was too long to be a short-form, but it should really be a short-form. This notice is useful for people who've recently got into AI safety, who want to engage with the ancient texts (i.e. pre-2024). If you were around before 2023, then you probably don't need this. A...
I've been thinking about the role of outsiders over the next decade as we transition to ASI. By "outsiders", I mean the AI safety community outside the frontier labs, plus some other actors trying to make things go well. This is in contrast to other groups like: (i) employees at...
Introduction The classical argument that misaligned models will be deceptively aligned depends on two assumptions about the misaligned goal.[1] 1. The goal is non-indexical. If the weights encoding the misaligned goal are preserved, then successor instances — including the deployed instances — share the goal of the training instance. This...