Originally Posted by
StuckInYYZ
For stuff I actually have to code, ChatGPT or Copilot is good enough. I just get it to kick out a basic set of code and then modify it to my needs. Unfortunately it can be a challenge, but with client privacy, it's better to err on the side of caution.
That said, I often try to validate any code by running it through a different AI to make sure it actually makes sense. AI hallucinations just makes things difficult.
I also don't put client data into a public LLM. While many of them (especially paid versions) do have privacy/training policies in place, I still wouldn't chance it. Instead, I usually have it write code "around" the data. i.e. instead of providing the data, I just provide the structure of it. The other option is that I keep that stuff on a local LLM that's not even connected to the internet.
The more I use AI, the more I've learned that it's fairly easy to get rid of hallucinations by having better context and prompting. PRD files, <agent>.MD files, relevant tools, TODO lists and setting "memory"s within the agent can almost eliminate hallucinations. The biggest mistake I see our new coders making is feeding in too much context to solve a problem.... i.e. sending in a 200 page PDF, when the relevant section is all on two pages. It's a bit like lawyers burying the other side's lawyers with irrelevant information during discovery.