Ask HN: My company doesn't let me use Copilot or Tab9 – any non-hosted options?

2 points by fapi1974 2 years ago | 7 comments

As is the case with many companies, our team has decided that the risk of using hosted AI solutions is too high. Are there any other solutions out there that would allow us to train on our own codebase without exposing it?

rektide 2 years ago |

Just started trying to figure out how to get StarCoder helping me today. Runs on my MacBook & can generate a fibonnaci unsurprisingly but I have so little idea how to use this thing to help myself, so far.

Repl.it and Salesforce also have code LLM, ReplitLM and CodeTF.

Different topic but how does one train / tune to their existing codebases? StarCode for example also ships a python tuned model. Do any of the services the submitter mentioned actually do any tuning to your company's codebases in any way?

Jimmc414 2 years ago |

The Falcon 40B param model may have potential, but at this point, I would be surprised if it were ready for primetime code completion tasks. It might be useful for answering questions about the code from a local vector store. I'd imagine that not using AI for code completion tasks in 2023 also carries some degree of risk. What are the primary concerns of your team? OpenAI using your IP against you or something else?

version_five 2 years ago |

Look at "starcoder". It's not strictly open source, is has a condescending "responsible AI" license, but you could easily download it and run it on premise.

32gbsd 2 years ago |

Code has become really complicated nowadays. I don't know if AI training will work only on a single code base. Unless you are only looking forward to auto complete

fapi1974 2 years ago | |

I should have been more specific - fine tune on our own code.

throwaway2990 2 years ago |

I would look for a new job.

beardyw 2 years ago | |

Some folks work behind an air gap, hand in their phones at the door, and still turn out good code. Not saying this person does but not everyone's environment can be the same.