search

Vandivier / ladderly-3

the home site for the ladderly ecosystem
https://ladderly.io
13 stars 9 forks source link

monetize advanced model / gpt #247

Open Vandivier opened 6 months ago

Vandivier commented 6 months ago

the point of this issue is to create a more powerful model than the free custom gpt. this may require training from scratch, fine tuning, or just embedding context or RAG...anything's on the table. we need to make sure it nets a profit though we can't just be burning money

this issue has some old references to gpt pass but that's probably an obsolete recommendation...

https://www.gptpass.io/

to monetize custom gpts

Vandivier commented 6 months ago

https://help.openai.com/en/articles/9119255-monetizing-your-gpt-faq

Vandivier commented 6 months ago

meh on gpt pass: https://community.openai.com/t/how-to-monetize-my-custom-gpts/600016/62

Vandivier commented 3 months ago

maybe useful https://github.blog/news-insights/product-news/introducing-github-models/

Vandivier commented 3 months ago

let's fine tune a GPT

cached context with a long window (gemini, claude sonnet 3.5) is a super interesting alternative and dare i even call it a tie, but:

  1. i don't want to deal with cache management
  2. iiuc fine tuning gpt 4o can be done for $25
  3. gpt-4o is still the highest ELO performer, so if the technical implementation is a wash then let's favor the higher brute expected performance

in the future we can a/b test other models too, but for now let's try fine tuning

Vandivier commented 3 months ago

so the implementation would be:

  1. fine tune and deploy a model
  2. host the model somewhere that people can't just send messages to it. i think open API can host for us. if not we might need like an EC2 or some equivalent.
  3. user must be logged in to the website and on premium account to make a request to our BFF, our BFF then hops further to the hosted model and passes it back to end user

that's basically how we can monetize with a BFF gate and hosted LLM

Vandivier commented 3 months ago

scope from related and closed #231

help users create a structured anecdote simulate a behavioral interview and then provide feedback simulate a behavioral interview for software engineers and then provide feedback calibrated to the Google leveling matrix encourage users to use voice mode and target an initial answer for 10 minutes also intro should be 2-3 min and include a brag

Vandivier commented 2 months ago

fine tuned > context approach

https://www.linkedin.com/feed/update/urn:li:activity:7238206300709396480/