[beyang]

I think you make fantastic points (Sourcegraph CTO, here). This is one of the reasons why we focused on code understanding rather than code generation for the initial version of Cody (in contrast to, say, GH Copilot).

For code understanding tasks, the issues with standalone LLMs is that they have a certain amount of "memory" which is limited to their training data (SO and OSS)—and even that can be unreliable.

A big "a-ha" moment for us was the realization that LLMs get much more helpful and reliable when coupled with a competent context fetching mechanism that can surface relevant code snippets from your own codebase. This makes Q&A much more factually accurate (and code generations that learns from the patterns in your codebase). We don't think LLMs will ever replace human coders, but we think they can be super helpful in eliminating a lot of the tedious, boring, duplicative writing and reading code that devs do every day. The entirety of Sourcegraph (not just the LLM part) is focused on eliminating these pain points.