Ah yes, I'm processing it and all the links around it, because I did all that manually this morning so I can have a pretty good opinion about what the bot comes up with.

neat!  i presume you meant you’re digesting that zerohedge article, not that you wrote it (unless you’re Tyler Durden)

interesting tools to shield yourself against clickbait world we now seem resigned to. 

seems like there might be a market for some sort of community notes like tool that inspects articles for BS provides ratings for relevance etc 🤔 

I mostly do research to create new tooling, but 

 with ChatGPT, because I am convinced that that (and all hosted chatbots with it) is a massive trojan horse, like FB. So I only use local implementations and I since a month and a bit I have a little set-up where I queue up large LLM requests and run it on rented cloud hardware.

How do you use AI ?

I mostly do research to create new tooling, but **explicitly not** with ChatGPT, because I am convinced that that (and all hosted chatbots with it) is a massive trojan horse, like FB. So I only use local implementations and I since a month and a bit I have a little set-up where I queue up large LLM requests and run it on rented cloud hardware.

Under (early) development:

- A "dream" mode where idle GPU/NPU time is being used on a RAG knowledge base to re-prioritize knowledge and update importance of stored knowledge. I'm not sure if this is going to work. For now I'm re-tooling it completely because I started in a wrong direction with it initially.

In "testing":

- A case builder that judges relevance of additional information to be added to a case. I.e. I can feed it a link, I'm currently working on [this article](https://www.zerohedge.com/technology/house-committee-leaders-urge-fbi-halt-certifying-chinese-biometric-devices) from this morning. I programmatically extract all links and fetch all these, convert videos, podcasts and pdfs to text, which are added to a RAG knowledge base powered by [`all-MiniLM-L6-v2`](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) as a sentence transformer. Then 2nd layer links are each examined for relevance, and so on. This is currently eating most my NPU on my laptop, especially since I'm messing with PDF transposition with [`minerU`](https://github.com/opendatalab/MinerU) which takes for. ev. er.

In "production":

- News article aggregation against clickbait titles (https://stacker.news/items/1006160/r/optimism?commentId=1006868) I use this daily as an alternative RSS reader. I need to rework it though, I'm thinking to maybe use [Airflow](https://github.com/apache/airflow) for orchestration, the prototype suxx because of `NATS` not being stable when ran locally (which was why I wanted that and not `kafka`).
- I noticed that I use Brave's AI-enhanced search results more often than the algo results.