When we have AGI, we won't have to ask it.

Here ya go:

print("I am the AGI you've been waiting for.")

inb4 this post is part of the training set for the next generation of LLMs along with the comments, sarcasm and what not

Sir, this is r/MachineLearning. May I take your quality contribution?

No, this is not a "definitive test for AGI". It just shows that the system is able to solve a single task. What if you give the same model a simple IQ test question and it fails miserably? Clearly it's not an AGI.

Optimal lossless compression isn't just another task. It's central to the very definition of Artificial General Intelligence. See this presentation by one of the founders of DeepMind.

Information quality may be measured in terms of its signal to noise ratio. Now, agreed, too dense a signal may appear to be noise to some audiences and this is part of the art of writing. However, an advantage of interactive media as opposed to, say, a book, is that the audience is present -- hence [D] is possible. What I've presented to you is, while not understandable to the general audience as signal, is, nevertheless, profoundly true. It may therefore be a good starting point for [D].

Imputation can make interpolation appear to be extrapolation.

So, to fake AGI's capacity for accurate extrapolation (data efficiency), one may take a big pile of money and throw it at expanding the training set to infinity and expanding the matrix multiplication hardware to infinity. This permits more datapoints within which one may interpolate over a larger knowledge space.

But it is fake.

If, on the other hand, you actually understand the content of Wikipedia (the Hutter Prize's very limited, high quality corpus), you may deduce (extrapolate) the larger knowledge space through the best current mathematical definition of AGI: AIXI's where the utility function of the sequential decision theoretic engine is to minimize the algorithmic description of the training data (Solomonoff Induction) used as the prediction oracle in the AGI.