News for hackers. Stuff that matters.
Re:Ask HN: What makes your AI special that can't be replicat... by PaulHoule on Monday July 31, 2023 @12:23PM (#36944991)
It's trained on my data which is perfectly suited to my problem, which my competitors don't have.
Re:Ask HN: What makes your AI special that can't be replicat... by saliagato on Monday July 31, 2023 @12:28PM (#36945073)
Cool. Do you mind me asking what do you mean by “trained”? Fine-tuned on your data?
Re:Ask HN: What makes your AI special that can't be replicat... by PaulHoule on Monday July 31, 2023 @12:51PM (#36945398)
I think the specificity of data is general, but I have a number of models that predict preferences and engagement for myself and communities. The production models either use classical ML with bag-of-words and other features or classical ML on BERT-like embeddings. The data is a set of judgements about individual documents.

I tried fancier “fine-tuning” methods and the noisy nature of the data means they won’t perform vastly better than my current models and have training times more like 30 minutes instead of 30 seconds but it is worse than that because if I want good results I might need to fine-tune 8 models and pick the best.

So I stick with my system because I can retrain as much as I like, run thousands of experiments a day if I want, etc.

Re:Ask HN: What makes your AI special that can't be replicat... by schappim on Monday July 31, 2023 @05:50PM (#36949359)
This is the answer. Very smart people [1] are disregarding the worth of proprietary data.

[1] https://twitter.com/nikiscevak/status/1650295590809636864

Re:Ask HN: What makes your AI special that can't be replicat... by rpastuszak on Tuesday August 01, 2023 @08:48AM (#36955286)
+1 this

I can't wait till the tech becomes even more easy to use, so one could work on ML projects with the same ease I had in the late 90s when learning from inspecting HTML source of pages I liked.

(also, curiously most of the AI services I'm paying for are also trained on my data!)