• Subscribe
  • What are your biggest learnings when building with GPT?

    Christophe Pasquier
    5 replies
    I was thinking about this when reflecting on our latest launch (basically chatGPT for your knowledge base, check it out), all the builders out there will face with exactly the same issues. So I listed my 3 top learning/overcome challenges working with it. Sharing them here, I'd love to hear lessons from others!

    Replies

    Christophe Pasquier
    Here are mine! #1 You can't just feed everything to GPT and expect an answer. Whatever your use case, these models are craaaazy expensive, and have all a limit of input. An average team on our platform has 5000 docs, 1.5M words, 3M tokens. The limit is 4,000 tokens... ๐Ÿ˜… -> You need a way to cut your input ๐Ÿคซ #2 Keep cost low with other models ๐Ÿ’ธ Even if you restrict your GPT usage - and you should - it is crazy costly. GPT is not great for everything. You have much better models for classifications, embedding and so on. ๐Ÿ‘‰Be smart, and as much as possible keep GPT for the last mile of your interface. #3 Generate *accurate* results. In our case the result is an answer. Now, let's say you write in a doc "Van Gogh painted Mona Lisa". If you ask Slite who did.., it will answer VanGogh ๐Ÿ˜… Your docs, your answers ๐Ÿคท It's our secret sauce but we manage to give the sources the answer went from, and while its just a UX tweak - we'll still say Van Gogh ^^ - it changes everything for your user. -- That's it, here are my learnings! Share yours and if you're curious about what we build, support and check it out ๐Ÿ™Œ https://www.producthunt.com/post...
    Uday Patel
    Really true and we have faced similar issues.
    Uday Patel
    What did you do to cut the cost? Specifically in fine tunning model?
    Christophe Pasquier
    @uday_patel4 aha thatโ€™s the secret sauce ;) but as said the big part is about not trusting GPT to be made to do everything, while it can feel quite smart, itโ€™s just a text generation model and other models are much more suited and cheaper for certain tasks
    Uday Patel
    @christophepas which other model do you suggest we have used davinchi 3 and ada