I'm having a lot of issues trying to use Opus 5.5 for some safety research tasks at the moment. One failure mode that I can't seem to prompt the models out of is that they resist reading transcripts / data to try to figure out what's going wrong, and when they do they ignore obvious issues in favour of whether the surface level criteria I laid out are satisfied. I've tried prompting pretty aggressively to not have this happen, but it seems like every time I leave them to work autonomously and come back I find that they haven't read transcripts for a while. (Very cool graphic btw!)

More details: For some reason, instead of doing the obvious thing of training a generic NLA and then fine-tuning it on goal data, they decided to first train a goal-verbaliser activation-oracle style on a synthetic dataset, then do NLA training starting from that on the same dataset. Because the dataset was synthetic and contained only a few bits of entropy, the NLA components collapsed to a private code to encode information more efficiently. They also seemed to screw up the centering on the NLA which caused an entropy collapse in one setting. They also ran four experiments at once rather than trying it on the smaller models first (ran 120B in parallel with 20B) which wasted a bunch of compute, and seemingly didn't even bother to read the transcripts (to be fair who has time for that, only poor humans such as ourselves and certainly not the swarm with infinite subagent budget) which would have shown them that the first stages made a moderately-functional system and the later sections killed it. This basically fits with my experience of autoresearch overall which is that I simply never trust a model to not fuck up anything other than converting markdown to python code. I have to hand-hold them through every step or they do stuff like this for seemingly no reason.