brian @briancloutier.com · 24/01/2025the three kinds of gaming streamer: - entertaining: come and laugh with me - educational: learn how to get better at this game - competent: watch someone who's incredible at this game 030
brian @briancloutier.com · 21/01/2025I'm very curious about the rationale, LSM seems like it should work pretty fine here. The circular bufferarity means every insert also inserts a tombstone so there's some unfortunate write amplification, but it doesn't seem outrageous to add a max keyrange length while sticking to the LSM model 130
brian @briancloutier.com · 17/01/2025a "small world" custom feed which throws you into a cohort with ~20 other small accounts. the cohorts slowly change over time as it learns from who you interact well with and tries to match you with the perfect little community key engagement metric: how many users eventually become irl friends? 040
brian @briancloutier.com · 17/01/2025the paradigm of running `sudo apt install calibre` seems wrong every tool in the apt repositories should be "installed"; it is downloaded into a local cache when I first try to use it and automatically removed from the cache if I haven't used it in a while. nix, uv, bun, docker approximate this 030
brian @briancloutier.com · 13/01/2025a cool visualization from the washington DOT, the optimal amount of traffic is a lot higher than you might prefer it to be (ht www.construction-physics.com/p/reading-li...) 030
brian @briancloutier.com · 12/01/2025some proposals for passwordless wifi. All of these are possible, and have been possible for many years, but standards are hard to bootstrap: 110
brian @briancloutier.com · 08/01/2025love it, - did:plc, you trust bsky not to censor your account, otherwise have full control - did:web (typically), you trust your registrar to control your identity - did:web (this monstrosity), you trust your ISP and their dynamic DNS, famously very trustworthy and stable (!), to host your identity 070
brian @briancloutier.com · 02/01/2025this must be the mathematician's version of finding a bug in your compiler all your derivations are correct, it's the published result which is wrong! 230
brian @briancloutier.com · 31/12/2024a single-payer blockchain: the network is subsidized by one entity which directly pays out block proposers this makes sense by the same economic logic which powers franchise restaurants 020
brian @briancloutier.com · 30/12/2024love this! some more nostr's which nature has grown: - historically: books are sent to many libraries - academic papers: submitted to both journals and arxiv - music: submitted to both spotify and soundcloud - burning man: if it ever fails there are many other events which could pick up the slack 121
brian @briancloutier.com · 30/12/2024I always assumed PLC uses some kind of merkle-powered append-only ledger a la certificate transparency, call me a crypto bro but that's the right way to do what PLC attempts to do appalled to find out today PLC uses the nostr approach but with a single trusted relay which can censor anything 😮 230
brian @briancloutier.com · 27/12/2024does anyone run an alternative atproto relay? everyone seems to read from "the firehose" but surely there are others somewhere out there? 120
brian @briancloutier.com · 27/12/2024after reading every NIP the nostr protocol also seems very cool, it's a shame it is currently being used exclusively to talk about bitcoingithub.comGitHub - nostr-protocol/nips: Nostr Implementation PossibilitiesNostr Implementation Possibilities. Contribute to nostr-protocol/nips development by creating an account on GitHub. 2102
brian @briancloutier.com · 26/12/2024I love the idea of an atproto-powered wiki! With atproto we can have an opinionated wiki which shows you your favorite version of each page, ignores all contributions by anybody you've muted / blocked. We already have a wiki for consensus knowledge, atproto can give you a wiki without edit wars 120
brian @briancloutier.com · 26/12/2024a lexicon which defines the rules for a custom feed, and a universal custom feed which reads those rules from your repo and serves you exactly what you asked for 130
brian @briancloutier.com · 25/12/2024I am continually surprised this doesn't cause more problems git merge mostly just works without introducing any bugs but it has this seemingly huge flaw! looking forward to an LLM-aided git merge which considers the intent & semantics of each commit while merging 020
brian @briancloutier.com · 23/12/2024there's an opportunity to create the netlify of bluesky, put all your records into a git repo and this service serves them for you - a bluesky account is a bag of tuples, all a PDS "really is" is an interface relays use to fetch your tuples why not decouple the data storage from the serving? 040
brian @briancloutier.com · 22/12/2024how would one go about automatically finding the closest renaissance (in this case baroque) painting for some image? mb similarity search over an embedding db but this match is so perfect because it's a crop of the original image, some kind of registration algorithm has to happen during the search? 020
brian @briancloutier.com · 14/12/2024my first custom feed: Top Articles It shows you the links the people you follow have most shared in the last 24 hours 031
brian @briancloutier.com · 11/12/2024New atproto applications require a webserver for hosting the interface to the application. Your PDS just holds the data, not the application itself. A new lexicon for defining interfaces/applications would enable atproto to become beautifully self-hosting 040
brian @briancloutier.com · 09/12/2024Don't think I'll be doing a firehose post every day but really enjoying watching new services launch. In the last 24 hours a few more new atproto applications launched and saw some adoption: - @pinksea.art, a drawing BBS - pastesphere.link, a pastebin - @woosh.link, a linktreepastesphere.linkPastesphere 030
brian @briancloutier.com · 08/12/2024Yesterday the firehose averaged over 700 events per second. Two million people liked at least one post and one million people made at least one post, an incredibly high ratio. New today: 11 people placed a bunch of pixels @place.blue 071
brian @briancloutier.com · 07/12/2024this is incomplete data (still working on stability) but I've started drinking from the firehose, over one million people sent at least one like in the last 24 hours! I'm enjoying browsing this long tail of experiments & typos 010
brian @briancloutier.com · 28/11/2024Going the other direction: people keep trying to give every webpage a comment section (hypothes.is) but it's really hard to get a critical mass of users But now you can bootstrap using the bluesky network: a browser extension which adds a sidebar of (filtered) bluesky comments to _every_ page 2102
brian @briancloutier.com · 27/11/2024the bluesky architecture enables overlay feeds: it pulls from your Discover BUT also syncs with mochi and occasionally throws in some of your flashcards. It syncs with readwise and throws in some longforms you might prefer to read 010
brian @briancloutier.com · 25/11/2024The custom feeds API is incredibly flexible and could be co-opted to expose some fun features: - a choose your own adventure where you react to a post from a limited menu of posts then are shown the next menu - voting: you're shown two SoME math videos and asked to react to the one you like most 060
brian @briancloutier.com · 25/11/2024The game theory of medieval monks bestowing curses upon those who betray them: academic.oup.com/jleo/article...academic.oup.com“God Damn”: The Law and Economics of Monastic MaledictionAbstract. Today monks are known for turning the other cheek, honoring saints, and blessing humanity with brotherly love. But for centuries they were known 010
brian @briancloutier.com · 25/11/2024The existence of undetectable backdoors seems to suggest some kind of impossibility result for alignment: you have no idea what your optimizer might have planted in there: arxiv.org/abs/2204.06974arxiv.orgPlanting Undetectable Backdoors in Machine Learning ModelsGiven the computational cost and technical expertise required to train machine learning models, users may delegate the task of learning to a service provider. We show how a malicious learner can plant... 010