Very good point! So right now what I have is a pipeline that looks at the real news and generates (using an LLM) a mixture of fake headlines that are inspired from the recent events and some fake headlines that are not conditioned on the current events (I, of course, do a final pass and ranking). I do agree that this isn't what most people call "fake news" but maybe more funny, cards against humanity abstract headlines.
Do you think that maybe I should source from snopes.com (or another fact-checker) or create another pipeline that tracks down more what people call fake news? How would you improve this project?
I've thought about going that route but defaulted to just making something fun
Hi HN! You can get a sense of each article's full text by reading the summary first. If you want to read the full article later, just save it to your profile. The main idea of the website is to allow readers to get a sense of the broader news faster!
I am mainly looking for any feedback and also any kind of feature that would fit well with the current implementation.
Some tech details:
The website is made with NodeJS and Supabase and runs on Vercel. There is also a separate Python service that runs an LLM (on a local or remote GPU).
I experimented with vibe coding as much as I can so I decided to pick the web stack based on what LLMs were suggesting, assuming that would be the most time and effort saving stack. I think I vibe coded about 80% of this in 6 days. It would have been difficult to get through some stuck points without actually having a background in programming though (so I recommend vibe coding only if you are willing to pause every so often and dig into the code to debug what an LLM agent can't). I am thinking of making an engineering blog to discuss my experience because this approach to personal project development felt new and engaging.
It loops through some news sources and asks the LLM to summarize it. Then there is a second pass to rate the quality. Finally there is a materialized view of ordered articles that the front page reads from.
Currently, I am just utilizing my own 4080 GPU and running a distilled version of DeepSeek
Chrome extension: Gives Ctrl+F like find results which include non-exact (fuzzy) matches using string edit-distance and GloVe/Word2Vec. Also searches by regular expressions.
