Notes

The embedding matrices used for instance in ChatGPT-3 are a vector space representation of co-occurrence frequency matrices for a given context-window size. This matrix can also be seen as a complete graph with edges labelled by probability values (we can assign values also to polyhedra).  It is important to study the properties of these matrices.

Locally integrable functions (fundamental to distributions and weak solutions to PDEs) are a kind of sheaf-theoretic completion of the $L^p$ spaces. On the other hand from a finitist point of view simple functions are the basic kind of function and these are 'dense' in many fundamental spaces in an appropriate sense. What is the meaning of a PDE in a scientific context ? A positing of certain recursive numerical algorithms and approximation ideals.  How curious that numerical conditions for stability involve the pairing of space and time.  And fascinating is that smooth initial conditions for PDEs as simple as $u_t - g(u)u_x = 0$ may generate shock waves. And what is, from a computational finitist point of view, a weak solution ? A good philosophical goal: to gain a deeper understanding of distributions (cf. Sato's hyperfunctions).

How could we prove that a physical system is computing beyond the Turing limit ? We could of course produce experimental evidence to the contrary, producing a certain algorithm that agrees with know observations. There are irrational numbers whose sequence of digits are not computable. How are we to view scientific theories about such numbers (which could represent measurements of some fundamental physical constant), calculations and approximations of such numbers, and confrontation with experimental evidence ? Obviously such a question is only interesting from a non-finitist perspective.

The interest in solving the P=NP problem hinges on the complexity of the algorithm for transforming a NP-machine into a P-machine.