Details:

Eliezer Yudkowsky has a very similar sounding question on this platform. However, I believe this question is distinct. Whether or not Eliezer's market would pay out for any scenario which this market will pay out for is a question you'd have to ask him.

This question is whether by the end of 2025 a technique will be developed which can take as an input the weights of an LLM (>50b parameters) and translate some algorithm from those weights into a program in a procedural programming language (like BASIC, C, or Java) or an object-oriented programming language (like C++ or Python.)

The program need not be functional in the same way as it is functional within the LLM, nor does it need to be interpretable by modern computer science. It must be able to run on an available computer with the same or lower specifications as can run the model it was extracted from.

This question is not "Will an LLM write a computer program by being used to do so?" It is about whether novel programs will be able to be "distilled" or "translated" from the transformer weights into a programming language in either of the categories mentioned above.

"I made a new way to represent all the weights of an LLM by automatically generating [parameter count # of variables] in a Python program which then behaves identically to the LLM!" does not count either.

I am concerned that this type of technique will be an additional unexpected (to the people releasing them) source of civilization-scale distribution from "open-weights" LLMs, and predict it would open up new options for capability improvements & model hybridization - so I am posting this question to check the temperature in the sections of the field participating in these markets, or others who are confident they can predict the near-term feasibility of this specific kind of technique.