Sept 2023, https://blog.sbensu.com/
For more like this one, see Friction logs
Summary
- The onboarding didn’t help me to use the model effectively.
- The model rarely hallucinated which was great (AFAICT!).
- The UI was clean and well-done but with a few small inconsistencies and room for improvement.
- It set up the wrong expectations regarding the free plan, how long I could use it for, and how I would be able to upgrade. I was making good progress and now I am very frustrated that I am locked out.
Suggestions:
- Redo the “Meet Claude” first screen to show:
- More diverse prompts regarding what I could do with the model
- Put the full prompt in each option, so that it is clear what information will go into it
- The input field gets really complicated when you add prompting, search, and files, all in the first use-case. Simplify with the most popular suggestion that guides that user to use the model well: “I want to …”.
- Highlighting parts of the answers should yield a menú with:
- “Expand on this”
- “Explain this”
- “This is wrong”
- CTRL+UP should put the last submitted prompt into the input field, so that I can rapidly tweak and resubmit.
- Copy is much less useful when the model is so deferential “You are right…”. I want to copy the information, not the pleasantries-boilerplate.
- When you tell it the previous answer was wrong via the prompt “That was wrong because…”, it could possibly downvote the previous answer and feed that into the RLHF system.
- The generated title of the conversation should probably be re-generated every 10 prompts or so (if it wasn’t edited by the user) or simply generated a few prompts in, as opposed to in the first prompt.
- It is unclear when it presents the option to “Retry” and when it doesn’t
- When I use Retry, it doesn’t show the previous answers so that I can compare them.
Stream of thought