I'd like to start off with some good news: I already found a new job. And once again without doing a technical interview! However, I can't really talk much about the product because it's in stealth mode and I signed an NDA. However, it's another AI-based thing so I'm glad that I can continue in that space.
Unfortunately, and this is the bad news, the compensation is not nearly as good as Comigo, so I think I'll stay working on this part time while I find a more sustainable full time position. I did have some interviews earlier in the week but they were all dead-ends, unfortunately. I did get the CSO from Comigo to join this new company though which is great.
As for the Safari extension, I didn't really get a chance to fully work on it in between everything. I did realize that there's an extension handler .swift file that you're actually supposed to put the code in, and I found this out via Claude 3 Opus, which I got access to finally over the weekend. I haven't noticed much of an improvement from GPT-4 yet but I think it's pretty decent so far.
Regardless, I tried getting code into the .swift file but had trouble compiling it, and that's where I left it before I got really into my new work.
I am currently on a trial week kind of like I was at Comigo, so I'm just trying to prove my worth at the moment. I am current hired officially as a "prompt engineer" and since there are already lead, backend, and frontend engineers, I can finally stay specialized and focused on what I care about. However, perhaps the team is pigeonholing me too much considering that they aren't even letting me access the frontend code, which is a little strange but I understand.
They wanted me to revamp their core prompt so that it actually follows the instructions. I noticed right off the bat that it was extremely complex and not really organized properly, so I got to work immediately creating the same sort of class structure that I'd theorized and planned out at Comigo but never got the time to actually implement. It came out pretty well, I'd say. I have a base prompt, a persona, and goals for the AI to follow. You can plug all of them in and it should be able to construct a well-defined string.
I also noted that the bot was using GPT-3.5 instead of GPT-4 which is contributing to its inadequacy in following the prompt set out. I also adjusted the temperature since the use case doesn't require much variation. In addition, I noticed that there wasn't an actual message history being recorded as a proper Langchain so I'm working on that now. Lastly, I suggested some text streaming to improve UX.
Overall, since it's such a new project, there's a lot to improve on and there's also a lot for me to contribute to, which is great and an opportunity for me to continue to grow and develop. I only wish that the compensation were more aligned with what would work for myself, but alas. It should still be a nice passion project.
Regardless, until next time, cheers.
Top comments (0)