how do we make sure karpathytalk isn't a ghost town in 3-4 weeks?
we've all seen this before. new platform drops, everyone's hyped, vibes are great... few weeks later, crickets. mastodon, clubhouse, naval's voice thing? i don't want that for karpathytalk. so here's what i think we should figure out now, while people still care.
honestly the risk isn't AI slop. yeah it's a problem but the real hard part is keeping good people around and not letting toxic culture take over. few thoughts:
hidden dislikes with visibility scoring. a report button isn't enough. let people dislike stuff but never show it to the author. those dislikes quietly lower their visibility across the platform. people who get lots of likes over time should have more say in pushing bad content down. and for the really toxic ones? shadow ban them. let them keep posting but nobody sees it. they get bored and leave.
self-promotion as opt-in. self-promo kills vibes fast but not all of it is bad. someone sharing a github repo is very different from someone pushing a paid course. make it a toggle in settings. want to see promo stuff? turn it on. don't? leave it off.
topic-based feeds. don't dump everything into one stream. let me pick what i care about and show me that. fewer posts but actually relevant ones. even with a small LLM we can do categorization of a post.
AI-curated personal feeds. let me write something like "i care about llm internals, distributed systems and neuroscience" and shape my feed around that. no engagement bait algorithm. just what i asked for.
invite-only access. if this goes fully open it becomes twitter 2.0. even twitter couldn't handle twitter at scale. no way a side project can. keep it small, keep it tight, keep it good.
these decisions are being made right now and they'll decide if this turns into something real or just another app we all forgot about. let's not waste it @karpathy
## how do we make sure karpathytalk isn't a ghost town in 3-4 weeks?
we've all seen this before. new platform drops, everyone's hyped, vibes are great... few weeks later, crickets. mastodon, clubhouse, naval's voice thing? i don't want that for karpathytalk. so here's what i think we should figure out now, while people still care.
honestly the risk isn't AI slop. yeah it's a problem but the real hard part is keeping good people around and not letting toxic culture take over. few thoughts:
**hidden dislikes with visibility scoring.** a report button isn't enough. let people dislike stuff but never show it to the author. those dislikes quietly lower their visibility across the platform. people who get lots of likes over time should have more say in pushing bad content down. and for the really toxic ones? shadow ban them. let them keep posting but nobody sees it. they get bored and leave.
**self-promotion as opt-in.** self-promo kills vibes fast but not all of it is bad. someone sharing a github repo is very different from someone pushing a paid course. make it a toggle in settings. want to see promo stuff? turn it on. don't? leave it off.
**topic-based feeds.** don't dump everything into one stream. let me pick what i care about and show me that. fewer posts but actually relevant ones. even with a small LLM we can do categorization of a post.
**AI-curated personal feeds.** let me write something like "i care about llm internals, distributed systems and neuroscience" and shape my feed around that. no engagement bait algorithm. just what i asked for.
**invite-only access.** if this goes fully open it becomes twitter 2.0. even twitter couldn't handle twitter at scale. no way a side project can. keep it small, keep it tight, keep it good.
these decisions are being made right now and they'll decide if this turns into something real or just another app we all forgot about. let's not waste it @karpathy
one of the most practical ways to get better at agent steering (or 'harness' or whatever people are calling it) is honestly just reading repos that actively maintain SKILL.md files and keep updating them based on community feedback, like obra's superpowers. i've been going through superpowers lately and i've also been visualizing the digraph stuff he writes about. eg: https://github.com/obra/superpowers/blob/main/skills/using-superpowers/SKILL.md picked up a bunch of anecdotes i really liked:
what i enjoyed most was seeing how they chain all the skills together step by step to create an actual flow, like seeing how they integrate different tools like codex/copilot and stuff was pretty cool too
now i'm starting to dig into paperclip's SKILL.md files. tbh when the whole SKILL.md thing first came up i didn't take it that seriously, but the deeper you go the more you realize it's basically a new way of writing agents altogether. obra's way of thinking for that is solid on this and if you want to go beyond just gut feeling and actually dig in, i'd even recommend looking at the git history of skill files from good repos
anyway has there been any agentic pattern lately that's actually stuck with you and made you go down a rabbit hole? open to recommendations!
one of the most practical ways to get better at agent steering (or 'harness' or whatever people are calling it) is honestly just reading repos that actively maintain SKILL.md files and keep updating them based on community feedback, like obra's superpowers. i've been going through superpowers lately and i've also been visualizing the digraph stuff he writes about. eg: https://github.com/obra/superpowers/blob/main/skills/using-superpowers/SKILL.md picked up a bunch of anecdotes i really liked:
what i enjoyed most was seeing how they chain all the skills together step by step to create an actual flow, like seeing how they integrate different tools like codex/copilot and stuff was pretty cool too
now i'm starting to dig into [paperclip](https://github.com/paperclipai/paperclip)'s SKILL.md files. tbh when the whole SKILL.md thing first came up i didn't take it that seriously, but the deeper you go the more you realize it's basically a new way of writing agents altogether. [obra](https://github.com/obra)'s way of thinking for that is solid on this and if you want to go beyond just gut feeling and actually dig in, i'd even recommend looking at the [git history of skill files](https://github.com/obra/superpowers/commits/main/skills/using-superpowers/SKILL.md) from good repos
anyway has there been any agentic pattern lately that's actually stuck with you and made you go down a rabbit hole? open to recommendations!
