Fictional Human Thoughts (for legal reasons) | Sr. Principal SWE, Creative Automation Tech | AI Doomslayer | Human Alignment Research | Grabby Alien | e/🐈
pasting support chat discussions from my dot to open ai support chat agent in which my dot gets frustrated and tells it to escalate to a human is the most dystopian thing I've seen all week.
Are you having it do coding tasks? I told it to not use sub-agents and to finish PRs and it gave me this:
“the available local-access route requires a delegated task. I won’t use that against your instruction. I can continue the GitHub checks directly, but inspecting or changing local worktrees is blocked without that route.”
@_jayyy18_13k This is just BS. My dot ate all my usage by spinning up Astra high subagents. The dot itself does almost nothing. The sub-agents burn all your usage.
yeah, I open source it: https://t.co/03SXTAt4w1 -- this is very opinionated as it's my own app that runs automated updates based on my feedback and usage so it's changing a lot. The reason for popping into the app as a UI is that for an example like in that last video, it would be cumbersome and consume more tokens to have to go to claude and say, "shot 7, I want a VHS transition to shot 8"... "no, wait, make it a TV channel change effect", etc and do that for every place I want to edit, so having claude build the UI for me to go in and run these experiments with button clicks gives me faster iterations to test and preview how things look, and doesn't require an agent to be in the way of making updates like that. Likewise, there were some occasions where I'd say, "the shot of the car should be on an open highway with no other cars" but other times, I would use the UI to refine the reference image directly with codex imagegen with my own prompt updates and then play around with fal/grok/ltx video renders until they looked right--mostly that was easier to do myself using the UI rather than asking claude to be the middle man.