I have never written an OBS plugin before.
With the release of Claude 3.5 Sonnet just in the nick of time for an event tomorrow, I decided to see if I could create a plugin that would generate a slide with a countdown clock. Instead, it generated this:
Rather than just write what I asked, it output the initial code and then continued to make suggestions to improve the program. The code contained docstrings, type hints, and helpful comments. My limiting factor was that IntelliJ does not have a feature where you it can search for the function you want to paste over from the clipboard, and it took very long to constantly search with find and replace.
Claude 3.5 Sonnet eventually arrived at a 715 line program, which took 146 minutes to write, debug, test, and achieve stability sufficient for a critical event attended by 100 people. The plugin also displays green, yellow, and red timing cards at specified intervals, saves configuration to disk and does automatic crash recovery, loads JSON from an agenda.json and automatically sets the timer to keep every event on track for the day, and contains a UI to change all the configuration options.
It is unquestionable that no human, regardless of experience with OBS and Python, has ever written a working, user-friendly, bug-tested 715-line plugin with this quality of code in 146 minutes.
In a previous poll:
/SteveSokolowski/did-openai-achieve-weak-agi-with-gp
I voted that GPT-4o had achieved weak AGI. In this poll, I will vote that Claude 3.5 Sonnet has achieved weak superintelligence. Do you agree?
1,000I also feel hampered by the slowness of copy and pasting things in between LLMs and code editors. I have copilot set up, and it's good for small local changes, but to do any kind of reasoning about how the whole codebase works, I always need to upload all the files to a web client and then copy and paste the suggestions back in locally. It's very tedious.