Lex Fridman on programming with AI agents by recording long audio notes

1. 1. Lex Fridman prefers typing over voice interaction with computers, despite trying voice tools.
2. 2. The guest, an introvert, finds voice-based interaction powerful and uses a Plaud device to record long audio notes for AI processing.
3. 3. The guest records long audio notes (10-20 minutes) that capture stream-of-consciousness reasoning, which he finds powerful for early design stages.
4. 4. The guest uses an LLM to process the recorded audio, with transcription and cleanup steps that require a dictionary and awareness of the codebase.
5. 5. Lex Fridman has not tried the approach of talking continuously for 20 minutes, but he sees its potential for stream-of-consciousness input to an LLM.
6. 6. The guest emphasizes that speech-to-text errors are a problem that needs solving, especially for dictating emails where mistakes are unacceptable.
7. 7. The guest is willing to wait 5-10 seconds for processing, arguing that if a problem has no current solution, users will tolerate latency.
8. 8. The guest advises not letting perfection be the enemy of good, especially when enabling new capabilities.
9. 9. Lex mentions that Box Type, an open-source speech-to-text tool, works well for short commands and is built into some systems.