Recently, AI Models have become good at long horizon tasks. In addition, they seem especially good at the types of long horizon tasks that allow for a quick and short feedback loop. For example, on MirrorCode, models were able to generate tens of thousands of lines of code, a task that would have taken a human several days to complete. This was possible because of the number of tests provided to the model, such that the model was able to write code that would pass many tests.
In no way is this discrediting the capabilities of the models, but instead showing that they excel on tasks that allow for the models to see when they’re making progress.
For more evidence, Anthropic published Automated Weak-to-Strong Researcher, which had models hill-climb an alignment task in an autoresearch loop, outperforming human researchers (though I think that, if humans had run the same number of experiments, they could have outperformed the model, the model in the autoresearch loop could run more experiments).
Altogether, this presents the case that models are good at long horizon tasks that are attached to some metric when being asked to push that metric to some extrema.
I think the right framing for automating alignment research is to find tasks that a) are either directly alignment tasks that you’d like to improve or b) find tasks that when improved will lead to better alignment indirectly. Anthropic did the former, but there are likely many tasks where the latter is also possible.
You could do this quite feasibly across many fields of alignment; controls, monitorability, weak-to-strong generalization, and so on. This leads to forms of reward-hacking really easily, and this encourages reward hacking if you do not do your environment design carefully. However, you could keep your validation set in a place the model couldn’t access, add honeypots, and so on as a way to detect or prevent reward hacking.
I think people are probably underestimating the amount of work that automated alignment research can do because this capability is quite new, and there are often small changes that allow you to get useful information from hillclimbing a task. For example, you can optimize a data filtering experiment by using a metric along the lines of reduction in misalignment for compute / number of datapoints removed, and this would yield a viable way to get useful information from hillclimbing a task. Another example is a task a friend is working on, figuring out whether certain RL environments lead to emergent misalignment: while you should probably have a human put a few hours into this task, having a model burn a billion tokens on crafting environments costs around the same as the human here, and makes the case for your argument much stronger, in either direction. I write this post in the hopes that more people will consider converting alignment tasks into this hillclimbable form as a first attempt at doing their alignment research.