Closed chillatom closed 3 months ago
I’m sure that @toolmantim has better ideas but ever since we chatted about surfacing the differences of models, I've been thinking of this as a 3 point scale across two dimensions and have envisioned a UI like this:
I took a stab at a first draft of this here: https://www.figma.com/file/jEsnzgsf0hNuJbqB9pKaGm/VS-Code---Cody---Commands-up-front-and-ChatGPT%2B%2B?type=design&node-id=2270-46492&mode=design&t=PP9VQP2xJha34ErH-4
Click on one of the model-selector frames and press shift-space to prototype the hover/scroll. See the comments for notes.
Let me know what you think!
Nice @philipp-spiess — love this sketch! I ended up going with a simpler grouping… but you know what, we could actually make the selector an actual 2-axis chart that you use to select the model. As long as you make it clear which ones are the recommended ones.
Not sure if we've got enough data/info to be that granular to compare model quality/speed though, do we? And it adds a bit of a bar for adding a model… vs just having an "Experimental" box to throw stuff into w/o clear data.
OR going back to something simpler… we just add some little indicators/icons to the right of each row. Only if we feel it's really needed though.
@toolmantim Ha love the idea but the route you're going in the Figma is probably enough for now and already a good improvement (and much easier to visually parse).
I love the grouping idea!
Personally think it'd be helpful to show what tasks each model is good at to help users to decide which model to use! Something like: Claude Opus: Great for answering large codebase questions, or Claude Haiku for code formatting etc
Ref: GPT marks ChatGPT3.5 as Great for everyday tasks
and ChatGPT4 as the smallest and most capable 😆
A + button to add new modeld would be cool too like they do in Opera:
Love the plus button idea! Could link to some page explaining ollama etc :) Not sure about the text from ChatGPT. I think this UI only works if you have very few models. Imagine presenting 6 models like this. You'd have to scroll pretty quickly. Plus: How would you describe GPT4 vs Opus in these terms? The differences become blurry quickly.
Model selection should be for free users, too, but we can still gate certain expensive models.
Updated design in Slack: https://sourcegraph.slack.com/archives/C04MSD3DP5L/p1713858613785279
Problems
Desired state