Re:Ask HN: Engineering managers, how do you onboard new hires?
by tarruda on Monday May 22, 2023 @09:17AM
(#36031047)
Not an engineering manager, but if I was, I would consider investing some time fine tuning GPT-3 (using OpenAI API) on the company's codebase, documentation and processes. Then create a Chatbot which can answer almost any question new hires might have.
[ Parent
| on HN ]
Re:Ask HN: Engineering managers, how do you onboard new hires?
by spencerchubb on Monday May 22, 2023 @09:25AM
(#36031139)
I've wondered about doing something like this. How expensive is it? How often should you update the "finetune"?
Re:Ask HN: Engineering managers, how do you onboard new hires?
by tarruda on Monday May 22, 2023 @09:30AM
(#36031217)
> How expensive is it?
I have never done it, but how expensive would depend on the amount of documentation/code the company has.
> How often should you update the "finetune"?
No idea, but maybe once a week if there were significant changes? It doesn't have to be fully up to date since GPT is good at predicting.
I'm also not sure if fine tuning is the best way to go. A possibly better/simpler approach might be using OpenAI "embeddings" API to build a vector database that you can search and compose a chatbot prompt from questions new hires have.
Re:Ask HN: Engineering managers, how do you onboard new hires?
by flerchin on Monday May 22, 2023 @09:31AM
(#36031226)
How are we doing this without uploading company proprietary to OpenAI to use however they want?
Re:Ask HN: Engineering managers, how do you onboard new hires?
by tarruda on Monday May 22, 2023 @09:34AM
(#36031271)
You can't. Though if the company uses Github to host its repositories, it is very likely they already have access to all your code through Microsoft.
Another option might be to use OpenAI embeddings API and build your own vector database, though you still have to upload company text when calling this API.