It seems to write in the generic "style" of GPT, instead of in the style I would recognise as a HN poster. Is that because of something baked into how the training process works? It lacks a sort of casualness or air of superiority ;)
There was no training process, this is just running GPT with relevant HN comments as part of the prompt.
If he wanted it to replicate that classic HN feel he would either have to extend the prompt with additional examples or, better yet, use finetuning.
I guess he could also just randomly sprinkle in some terms like 'stochastic parrot' and find a way to shoehorn Tesla FSD into every conversation about AI.
> “AskHN” is a GPT-3 bot I trained on a corpus of over 6.5 million Hacker News comments to represent the collective wisdom of the HN community in a single bot.
First sentence of the first paragraph on OP's page
EDIT: it's a bit misleading, further down they describe what looks like a semantic-search approach
I agree, that language could be very improved. This is not a GPT-like LLM whose training corpus is HN comments, which I found to be an extremely interesting idea. Instead, it looks like it's finds relevant HN threads and tells GPT-3 (the existing model) to summarize them.
To be clear, I think this is still very cool, just misleading.
Soon we will see language style transfer vectors, akin to the image style transfer at the peak of the ML craze 5-10 years ago -- so you will be able to take a HN snark vector and apply it to regular text, you heard it here first ;)
Joking aside, that does seem like it would be very useful. Kind of reminds me of the analogies that were common in initial semantic vector research. The whole “king - man + woman = queen” thing. Presumably that sort of vector arithmetic is still valid on these new LLM embeddings? Although it still would only be finding the closest vector embedding in your dataset, it wouldn’t be generating text guided by the target embedding vector. I wonder if that would be possible somehow?
Last year (pre the chatGPT bonanza) I was using GPT-3 to generate some content about attribution bias and the responses got much spicier once the prompt started including the typical HN poster lingo, like "10x developer":
To truly capture the HN experience, the user should provide a parameter for the number of "well actually"'s they want to receive.
So initial response should demonstrate clear expertise and make a great concise point in response to question, and then start the cascade of silly nitpicking.
I wish the results were reversed, so I could "well actually" your comment, but 'site:news.ycombinator.com "well actually"' gives ca. 4k results in Google and 'site:news.ycombinator.com "I think you'll find"' gives close to 17k results, so you appear to be right.
