Is there a list (something analagous to the excellent list of examples of specification gaming by Victoria Krakovna) where examples of impressive capabilities of frontier models are compiled?  And if not, can you provide your own best examples in response to this post?

Context:

I want to give a presentation explaining AI Safety to AI university students.

I want to mention the debate regarding whether current AI systems are "stochastic parrots", whether they "lack any real understanding", etc. As part of this, I want to clearly share examples of peak capabilities of current frontier models (e.g. something very impressive that GPT4 can consistently do), so that the audience has more information with which to explore this debate further themselves.

I think examples where it is impossible / very difficult to argue that "this is (likely) just because this was in GPT-4's training data" are best.

Currently, I am thinking to draw from the Sparks of AGI paper, the GPT-4 technical report and to make reference to 'SmartGPT' as a way of showing that greater "reasoning" abilities can be drawn from models like GPT-4, given the right prompt.