3 ms·

Re: #1 chess - this might be worth open source-ing? I know that's the standard vibe here but if it truly is a dead-end for you it might be a perfect fit for som

Re: #1 chess - this might be worth open source-ing? I know that's the standard vibe here but if it truly is a dead-end for you it might be a perfect fit for someone else. Just put it on not-GitHub and no one will ever hold it against you in a future job interview! Re: #2 VR - there are a number of one-and-done demos rendering 2D or 3D images/models for a VR gallery on github, though none seem too concerned about performance. i'm not sure if you'd get more traction perfecting that or just focusing on your stereo camera (which might be valuable just as a page to view one image but ultra niche). Sorry I don't know enough about text classification so instead I'd ask if you could share the best resource you used to get started working them? It seems to be necessary to have the best foundational understanding as LLMs take over.

PaulHoule 1y ago
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 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 https://mastodon.social/@UP8/114910543438621522 It downloads maybe 20,000 RSS feed items, runs them through https://sbert.net/ 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.cluster.KMeans.html 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 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 https://scikit-learn.org/stable/model_selection.html
- password4321 1y ago
  Thanks for the details on text classification! > I want a system that a person who just has text and judgements can push a button and get a good classifier This seems very practical. These systems could add a lot of additional value filtering sites like HN to follow the trends for technologies someone is interested in, or drilling down into popular trends as a tool to find what to generate content about when building an audience.

PaulHoule 1y ago

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 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 https://mastodon.social/@UP8/114910543438621522 It downloads maybe 20,000 RSS feed items, runs them through https://sbert.net/ 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.cluster.KMeans.html 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 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 https://scikit-learn.org/stable/model_selection.html

password4321 1y ago
Thanks for the details on text classification! > I want a system that a person who just has text and judgements can push a button and get a good classifier This seems very practical. These systems could add a lot of additional value filtering sites like HN to follow the trends for technologies someone is interested in, or drilling down into popular trends as a tool to find what to generate content about when building an audience.