My Python program could be worth open sourcing. It's pedagogical and a decent answer to the question of "how can a computer play chess?"
The Java program is a mess though, and there are a lot of chess programs out there, both of those programs are based on knowledge I got from
https://www.chessprogramming.org/Main_Page
I've made some really simple demos of VR rendering that play well on my MQ3 when it is attached to a computer which is running the web browser -- they tend to choke running in standalone mode which I think is the main market so the product I want is going to be one tuned up for memory use. I think the problem is that the MQ3 just doesn't have a lot of RAM and a lot of of it taken up by the OS and the standard UI. A single DSLR photograph would be like 6000x4000x3 = 72MB as a texture and it would have to be unpacked to view it. I've seen some good demos that run in WebXR on the MQ3 so I know it's possible but I'll have to really tune it.
As for text classification I have an RSS reader that uses text classification for a recommender, here it is running on the MQ3:
https://mastodon.social/@UP8/114910543438621522
It downloads maybe 20,000 RSS feed items, runs them through
https://sbert.net/
to turn them into vectors (easy!) and then clusters them with k-means clustering to divide them into 20 "topics"
https://scikit-learn.org/stable/modules/generated/sklearn.cl...
it picks the 10 highest scoring article out of each cluster and adds another 100 randomly chosen articles to show me 300 articles. I give a thumbs up or thumbs down judgement of each article and use the vectors as X and the judgements as y for
https://scikit-learn.org/stable/modules/svm.html
with the probability option to compute the scores.
This system is super-reliable and fast, in three minutes it trains something like 20 models and picks the best.
My approach based on SBERT gets a good sense of the "gist" of something but doesn't really understand the order of words, can't handle negation, is not so good for sentiment analysis and more complex kinds of tasks -- but my recommender doesn't really need a highly accurate model because my judgements are not accurate, I might judge the same article up or down depending on how I feel that day.
People who hold court on the forums on huggingface tend to advocate "fine tuned BERT models" for classification, I think they for the birds. I see a lot of arXiv papers where people copy a training recipe from another paper, I haven't seen a recipe that consistently makes good models -- I don't want to write papers, I want a system that a person who just has text and judgements can push a button and get a good classifier for a wide range of problems.
I've worked on LSTM trainers in the past and found I could develop reliable training procedures for them, the literature tends to show these often beat the "fine tuned BERT" by a bit, so I am really interested in making one that "just works", like you give it documents as your X and your judgements as y and it will train a bunch of models and give you the best.
https://scikit-learn.org/stable/model_selection.html