I'd consider using something besides 'list all files' as first example in the gif. I'd think anyone who opens a terminal knows how to do that so listing files in a directory and hitting a spinner isn't very inspiring.
The second curl example is better since plenty of people won't know that off the top of their head.
I loved Warp for it's speed, but the keybindings and lack of configuration didn't work for me. For example, it doesn't support "ctrl+x ctrl+e" which I use daily for editing long prompts in vim.
I recently switched to Kitty (https://sw.kovidgoyal.net/kitty/) and I don't think I'm ever going back. It's not flashy and doesn't have any AI features. It's just a donation-funded, wickedly fast, highly configurable, dotfiles-friendly, modern terminal emulator.
The only thing holding me back from loving Kitty was its behavior when ssh’ing, even if ssh was aliased to kitty +kitten ssh. Namely, that if you sudo’d to another user, the TERM settings were lost and so backspace became space (among other things).
Today, I finally figured out to export XTERM before sudo, which fixes it. Hallelujah.
I was surprised that Whiz could use tools like ffmpeg even though it tells ChatGPT to only use available shell commands. I asked it to "convert demo.mov to an mp4" and "cut the first five seconds of demo.mp4" and it came back with the correct commands. I guess that was enough to allow it to assume ffmpeg is probably installed. Pretty cool!
Looking at the code, the system prompt says "You MUST NOT use functions that are not available." but apart from the OS info, it's not giving GPT any info on what commands are/aren't available, so I'd imagine it's a bit of a crapshoot.
Seems like a perfect use case for local models. Not sure I want to be sending my .bash_profile or .bash_history (or local env vars...) to OpenAI. And I can't imagine doing anything in the terminal that llama2-code-7b couldn't make sense of. That can trivially run on an M1 with 8GB.
We need good locally installed LLMs (and cheap hardware to run it). I hope there can be some kind of breakthrough for this similar to what Stable Diffusion made for image-generators. I tried to generate some simple code using a few of the llama-models small enough to run on my computer and it did surprisingly well, but still far from good enough to be useful.