Minesweeper bench
Minesweeper seems to make all LLMs go through hell. They fail consistently through it. I think this is because its out of their training data.
Get ocassional cool stuff in this RSS feed.
Minesweeper seems to make all LLMs go through hell. They fail consistently through it. I think this is because its out of their training data.
Using metaheuristics to prune layers from AI while mainting model performance.
I made an automated pipeline to clean data. The idea started from a paper called Minipile. This led me to a rabbit hole. If you’re careful about the data you use for training, you can break the scaling laws. Who knew being a data snob could be so rewarding?
Classifying all of SafeDocs through the metadata.
Boom box a tool for drawing bounding boxes.
Only 1% of all the actual posters on the internet are good at publishing and promoting.
Generate images with just a CLIP model and no training*.
I have been considering a lot more how n-grams are amazing tools to show some of the fundamentals of language models.
I have been trying a new workflow for my lecture notes in university.
Learning Rust by making a small TUI to check my CPU usage.
Using telnet from a Microcontroller to connect to a computer.
I recently played the online game called contexto.me. I was intrigued by how it worked so I created a tiny clone.
I implemented Feature Selection through Genetic Algorithms.
I wanted to create a dataset that was focused on spanish.
I created a simple markdown to html parser called m2h. I wanted to implement the basic markdown syntax, I implemented it over a weekend.
Go and experiment. Create and finish projects. This way, you can experiment, but get new knowledge for new areas.
My thesis was that people outside of my country can be more objective of what happens inside. They are a third party. They should be able to tell fact from fiction.
... There is another type of engineer that could exist. A master at making things more efficient.