That's why I have some private benchmarks and I'm sorry to say that the transition from GTP4 to o1 wasn't unambiguously a step forward (in some tasks yes, in some not).

On the other hand, private benchmarks are even less useful to the general public than the public ones, so we have to deal with what we have - but many of us just treat it as noise and don't give it much significance. Ultimately, the models should defend themselves by performing the tasks individual users want them to do.

I'd guess it's doing natural language procedural synthesis, the same way a human might (i.e. figuring the sequence of steps to effect the transformation), but it may well be doing (sub-)solution verification by using the procedural description to generate code whose output can then be compared to the provided examples.

While OpenAI haven't said exactly what the architecture of o1/o3 are, the gist of it is pretty clear - basically adding "tree" search and iteration on top of the underlying LLM, driven by some RL-based post-training that imparts generic problem solving biases to the model. Maybe there is a separate model orchestrating the search and solution evaluation.

I think there are many tasks that are easy enough for humans but hard/impossible for these models - the ultimate one in terms of commercial value would be to take an "off the shelf model" and treat it as an intern/apprentice and teach it to become competent in a entire job it was never trained on. Have it participate in team meetings and communications, and become a drop-in replacement for a human performing that job (any job that an be performed remotely without a physical presence).

Agreed.

> And nobody else has come up with a comparable benchmark that is so easy for humans and so hard for LLMs.

? There's plenty.