
I am trying to write some code that, given a training set of point clouds in 2D and edges connecting pairs of points, will predict the probability of an edge between any pair of points in a new point cloud. The rule that determines the presence of an edge comes from gravitational lensing physics an it is not a simple function of the positions of the points; it depends also on where all the other points are. The sizes of the point clouds can vary within a factor some. There are typically a few hundred points and a few hundred edges. I can generate as much training data as I want to.
The way I set out to do it is by using a set transformer to turn the point cloud into a fixed size embedding, then I will concatenate the x,y of each point to that embedding and train an edge predictor on that. At some point I will want to make the whole thing equivariant for translations and rotations but that is not the priority right now. The whole thing will be written in pytorch and should run within a colab. My knowledge of pytorch is adequate but not stellar. Two more people are working with me on this. We are all using GPT4 with the advanced data analysis plugin to write the code and explore the results. For now the code produced by GPT4 seems reasonable but it could be wrong in subtle ways. We are still in the process of getting it to work at all (e.g. an error I am getting now is "Embedding size needs to be divisible by heads" and I am quite irritated because some obvious solutions proposed by GPT4 do not work).
I will resolve YES if the project is successful thanks to GPT4 (in my judgment) within a week from now. Success is defined as having a code that works as expected at least on a toy version of our problem. If the project is not successful (e.g. we abandon it) or is successful but not because of GPT4 (e.g. we have to rewrite everything from scratch) then I will resolve NO. If GPT4 becomes unavailable or dumber to the extent that we cannot use it for the project in the mean time I will resolve N/A..
Since this is based on my subjective judgment I will not bet. I may provide more information on the state of the project down the line.
🏅 Top traders
| # | Trader | Total profit |
|---|---|---|
| 1 | Ṁ155 | |
| 2 | Ṁ94 | |
| 3 | Ṁ38 | |
| 4 | Ṁ34 | |
| 5 | Ṁ31 |
@mariopasquato note that by now this specific problem was solved and we have something that trains (gasp!) even though I need to double check that it is actually learning what I want it to learn.