So, once again, I neglected to write a blog post last week. I don't want to make this a habit, but I guess it's been so busy these past two weeks that I didn't want to bother writing a whole post last week.
We've been focusing on getting our v0 (aka alpha version) out so that investors and such can start testing the app.
To generally cover what we've gotten done:
- iOS app is out on Test Flight
- revamped prompts
- speech to text
- two new architectures/prompts beyond what we had before
- a chat history sidebar
- selection of the architectures
- message limits
- a bunch of other small improvements/bugfixes
Generally, I've been working with my fellow prompt engineer and the others in the product team to get the prompts to do what we want, as has been the goal for the past 3 months. I believe I mentioned last time that we were transitioning from using fine-tune data to using just prompts for now. We need more resources to do fine-tune.
That being said, I also designed a script to get two AI models to talk to each other in a loop until the conversation naturally finishes. This was mostly to test that our prompts are adequate, but I can also take them and adapt them to fine-tune data. But the major issue is that it costs a lot and I'm actually not sure if it's generally better in terms of quality of the conversations.
Anyway, another thing we're considering is the fact that there's no chain-of-thought reasoning possible for our app. We are designing it so that the bot gives the shortest, most concise lines of questioning, which disallows for it to reason for the previous context. This causes the bot to follow the prompt and goals that we laid out a little too rigidly instead of following a natural flow. So to mitigate this, I'm thinking of designing a two-step process where it will take a long-winded response with chain-of-thought reasoning, then reduce that down to a singular question. This way, it can follow its own logic better.
So that's what's happened so far. Hopefully it turns out for the better. Until next time, cheers.
Top comments (0)