small fun project ideas i hope someone steals from me

• all non-ai podcasts on {china,bio,near-history,far-history,misc}, ordered by ?date? ?popularity?
• all podcasts on AI sorted/categorized… somehow?
• selected ‘best’ or ‘favorites’ from dwarkesh?

i love dwarkesh’s podcast. for the last few years, it’s been one of few podcasts i regularly listen to.

[cont description]

really something like this but for dwarkesh

tldr: micromorts.rip but really-really good.

features:

• add cruxes

[add good description]

LLM flow that processes news articles from NYT, WSJ, CNN, MSNBC, and a few other top major news sources, then generates overlays of each article with places where context is missing/things are straightforwardly false.

would probably look something like a combination of community notes and the ‘is this true?’ chrome extension, with overlays that look something like Curius comments. could be generated once per article, then cached and shown every time a user with the extension opens that article. stretch goal is for people to vote up/down on the overlays about whether they were helpful, accurate, etc.

main problem i haven’t fixed is incentives, though i have some ideas for potential solutions

current vision is something like:

• when you start, you spend 2-3 mins selecting the sorts of historical bits of news you’d like to be notified about
• then, it synthesizes everything that’s happening — across prediction markets, forecasting aggregation platforms, MSM, probably even some twitter accounts, maybe something like the setup nuno sempere has at sentinel — and sends you notifications at a frequency you want about news that you want.
• each news ‘article’ is a very-brief articulation of what’s happening, alongside the best relevant external links (where you specify what sorts of external links you want)
• super easy mechanism for providing feedback on any given notification — e.g. tap and hold and click “see more of this” or “see less of this” — that actually goes into the end-of-the-line decision of which notifications to send you

i think one possible way of improving incentives is to:

• not have ads at all;
• give users a 3 month free trial (so that they experience at least 1-2 events that are particularly important)
• then, a 6-month recurring, reasonably expensive ($100?) subscription

means that people are thinking in terms of what they want over a longer time horizon, which means that you aren’t optimizing for short time horizons by seeing their attention. (same goal re: not having ads.)

a freemium model doesn’t seem crazy, but the free version should be pretty great for most people’s needs, and the premium version should be way way better for a very specific kind of person, and also quite expensive.

update march 2026: i tried for ~an hour to make this with claude code / opus 4.6, and couldn’t build a good workflow — but i’d give 60% that i could build a reasonable workflow in another ~3 hours that costs ~$10 in march-2026 token costs per book.

there’s a lot of great writing that’s only ever been published online. i sometimes want to read it in a physical book:

1. i mostly read things online, but i sometimes prefer paper (e.g. reading-before-bed, or by-the-lake).
2. many people basically don’t read (long) things online. many of those people would still get quite a bit of value out of the writing-which-is-currently-only-available-online.

however, it currently takes quite a lot of hours or dollars to typeset a book. for me, an amateur who’s done it only once before as of this writing, i’d guess it’d take me ~10 hours to do a mediocre/sufficient job. an expert charging $50/h (for example) and taking ~5 hours would cost $250 — which doesn’t include the cost to actually print the books.

i’d bet that it’s possible to build some scaffolding for an LLM to do this reasonably well — something like ‘cursor but for indesign’ would be a dream, but more feasibly something like ‘have an LLM do the first 2/3, then a human does the last 1/3’ would still be a huge improvement on the status quo.

GitHub lets you track contributions via a graph, where: each day is a square; each square that’s unfilled indicates that you didn’t commit anything that day; squares that are filled with lighter shades of green indicate fewer contributions, and darker shades indicate more contributions.

There’s a popular Anki add-on that does the same with quantity of cards reviewed. I find this to be a surprisingly strong motivator!

What if you had this for other habits? Examples:

• Number of miles ran per day
• XP on Math Academy or PhysicsGraph per day
• Number of hours studying $TOPIC per day

in my current view, one of the most significant limitations of domain-general memory systems (e.g. Anki, SuperMemo) is the complete lack of modeling the dependencies/prerequisites that much of knowledge has. Math Academy and PhysicsGraph model dependencies/prerequisites really well, but aren’t domain-general — you can’t use them for history.

i’d love to be able to mark dependencies in my Anki collection among cards.

one important kind of dependency is: wait to show me Z card after I’ve mastered cards X and Y. another kind of dependency (less important IMO) is: Z card relies on knowledge in X and Y, so if I get Z wrong, you should lower the probability that i can recall X and Y.

relevant reddit thread:

give some website a list of Twitter accounts; get emails every day of the summary of posts from them are.

For LW: each month, the most-upvoted, curated, etc posts from that month get printed & sent to your doorstep.

For newsletter: idk, maybe manual curation from the user? Or maybe just auto sent based on analytics?

then just send me a beautiful, physical magazine, maybe quarterly or monthly, with typeset essays. maybe with images?

maybe you get an email saying “here’s the auto-created proof of your upcoming magazine; you have until MM/DD to tell us there’s something wrong with it”

Occasionally, I’ll add an Anki card to my deck that ends up being false/misleading/poorly framed. So I’ve started to just chuck a prospective card into Claude/ChatGPT and ask “is this right?” before clicking ‘add,’ but this is a kind of inefficient/annoying workflow. Ideally, this would all be automated/happen in the background. A few concrete ideas for how a plugin could do this (these aren’t mutually exclusive!):

• A button on the card creation modal (alongside the text formatting buttons) that, when clicked, sends the cards that would be created to an LLM, which returns either “Correct” or “False/Misleading/Poorly Framed: [explanation].”
• A cron job which checks through one’s whole collection (or specified decks) for cards that are false/misleading/poorly framed, and sends you a notification or an email or something for cards that ring alarm bells
• A utility that, upon starting the process in the add-on, checks through one’s whole collection (or specified decks) for cards that are false/misleading/poorly framed, and returns a browser list of cards that are false/misleading/poorly framed

Conway’s game of life is “a cellular automaton … its evolution is determined by its initial state, requiring no further input. One interacts with the Game of Life by creating an initial configuration and observing how it evolves in discrete time steps called generations. There is no limit to the number of generations (computational resources permitting) or win condition” (from Wikipedia).

A physical version of Conway’s game of life would be so, so cool — I’d love to hang one in my living room. My vision is a 4ft by 4ft filled with little squares that have one side that’s black and the other that’s white, with small servo motors turning them back & forth in association with an infinite Conway’s game of life.

Pretty straightforward — a simple, canonical website that clearly & neatly displays model cards for every frontier model released by every frontier lab since GPT-2’s model card was released in 2019. An LLM should be able to update this automatically as models get released, though this might need to be done manually until agentic models get better (than they are 2026-03-19).

i started building this just for myself, but i could imagine a wider-scale version of it that just makes sending link posts super easy

built this: saulconsuming.substack.com

should send you things like:

• are you keeping pace with your budget? are you over (or under!) spending in some category?
• do your investments need attention (e.g. some options that are about to expire)?
• etc

on a regular basis that the user sets

something like a 2-5 day bootcamp for people like me, rachel weinberg, or arden berg, or noah birnbaum, or carolanne jiang — none of us have any sort of background in ML, and spending a few days intensely learning the technical details of ML systems seems plausibly super high impact and quite valuable.

i’d imagine this would look like a long weekend at an airbnb with 5-10 people or something and ~2 domain experts, such that there are quick and constant social & technical feedback loops.