Hacker News Reader

Comment by serendipityrecs | Hacker News Reader

I built a recommendation engine on top of GPT-3: http://serendipityrecs.com/

I think other applications exist, you just have to figure out how to extract what you want from the prompt/response interface.

View on HN

I tried the recommendation engine with some prompts that are fairly simple, but would require a deeper understanding than just a keyword search to get right (e.g. "books where magic is secretly technology"). The recommendation engine does not seem to be doing significantly better than a keyword search would. I would probably have more success with raw Google.

This is the sort of thing that dampens the hype for me a bit. I keep assuming that the demonstrations are not cherrypicking examples, but it's kind of hard to believe I'm just uniquely good at picking problem cases.

Actually I like the results a lot. How does the logic behind it work, is it just the output of GPT-3? Do you need to parse it somehow? If it is just GPT-3 its kind of amazing how good the results are, since it was never trained for that. A bit scary to be honest.

It's mostly just GPT-3. You do need to parse it, as far as GPT-3 is concerned it's just doing text in/text out. I'm doing some ranking logic behind the scenes to make the results more consistently reliable, but you could directly ask the model for recommendations. GPT-3 has a lot of knowledge of the real world from having read so much of the internet.

Wow kind of amazing. Don't think a human would give me much better results. It would be interesting to know how fast OpenAi is able to retrain / update the model based on new data.

Reply on news.ycombinator.com