AI + 3:18 marathon

Β· Zi Wang Β· 1 min read

Z / Runner πŸƒπŸ»β€β™‚οΈ

πŸƒπŸ»β€β™‚οΈ 03:18:39 (7:34 min/mile), mom's congrat text "δ»₯恒为贡". πŸ€ unlocked! 'https://results.svetiming.com/Napa-Valley-Marathon/events/2026/kaiser-permanente-napa-valley-marathon/1615/entrant/share'

πŸƒπŸ»β€β™‚οΈ sending extraction per 1k channel as quotes too passive. Need to connect w/ "your" internal thesis and provoke thoughts. Otherwise, it is just content, not bespoke. It should turn into sharper position that feels like something "you" would have written after a day of reflection.

  • πŸƒπŸ»β€β™‚οΈ accuracy is one of the many concerns, but each model update will improve; some of the errors are also result of "shared memory" w/ Yo, she also uses my accounts. That's probably another reason why "multi-user agents" are not available in the market (yet).

  • πŸƒπŸ»β€β™‚οΈ do you find notebookLM useful?

Stephen / Basketball πŸ€

  • πŸ€ oh, it's the opposite for me: i will only use content agents for extractive (never abstrative) tasks. my prompts always ask for the original passages – much like scientific proofs with citations. otherwise, you have no way to verify then trust the synthesized opinions or recommended actions. i analyzed and checked through your (generated) prompts + sessions; many subtle errors – that is, worse than slop.

  • πŸ€ for audios, highlight clips will be 10x better than synthezied in 2026; for videos, 2029.

  • πŸ€ also, only audios for now. text is too _, videos a bit harder.

  • πŸ€ nah. gems are always original.

  • πŸ€ vs "ai agents to take a list of hackernews top 20 links, extract (not summarize) the most technical bits as 2 paragraphs for link"

πŸ€ "top use cases for ai agent to replace humans" πŸƒπŸ»β€β™‚οΈ learning; but synthesized; shopping…