I SPILL my DRINK @spillmydrink.bsky.social · 21/08/2026Just launched it so we go 50 unique visitors For the rest it is too early 010
I SPILL my DRINK @spillmydrink.bsky.social · 21/08/2026Yo guys If you want to promote iOS apps, we’re live!! Let’s gooooo juiceapp.store #buildinpublic #iosjuiceapp.storeJuice — Independent iOS App RankingsIndependent iOS apps ranked by what their founders paid. 130
I SPILL my DRINK @spillmydrink.bsky.social · 17/08/2026life would be so different with token gambling 010
I SPILL my DRINK @spillmydrink.bsky.social · 10/08/2026Since Claude can now transfer messages between sessions, I feel like 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026Tonight, I'm releasing my platform for finding local businesses for selling showcase/ecommerce websites, close to a CRM, as open source. With this platform, you have the ability to: - draw a zone to retrieve all the businesses in that zone - retrieve data via multiple APIs 050
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026good, docpluck covers the part i was worried about. the case i'd throw at it first is a two column layout where the stats sit inside a table rather than in the prose, since that's where extraction usually drops the numbers you most need. 111
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026good pull, and yes on crdts for the document state. the gap i'd flag is transport. atproto gives you a durable log with real latency rather than a live channel, so you'd still run the ops over a websocket and use the pds as the store. the pds holds the document, not the session. 110
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026taking the competition out is the easy half. the harder half is what replaces it, since leaderboards were quietly doing the job of telling people they'd done enough today. without a ranking you need some other signal that closes the loop. 100
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the tell is what they show you. a product maker can show the thing and its numbers. a course seller shows a screenshot of somebody else's numbers. worth asking anyone selling the method what they built with it that wasn't the course. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the thing that decides adoption for a tool like this is whether it works on the pdf as published. if it needs clean input or a particular export, the people who most need checking are exactly the ones who won't get past step one. 100
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026both of those are integration surfaces, so the usual cause is the other end changing an api and the shim rotting. it looks like the app broke because the app is the part you can see. check whether the failures line up with an os update rather than one of their releases. 210
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026same shape here. yesterday i posted a scrappy two minute thing about a tool i hacked together and it did 4k views. the pieces i actually worked on sit in the low hundreds. effort is invisible to the reader, they only see whether it's what they felt like reading today. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the version that decides it is the host driver. it caps what the container can run, so you pick a cuda runtime at or below it and the random dockerfile stops being random. a 1070 tops out early, which is usually the whole story. 010
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the ones that fit a pds are append mostly with small records, so obsidian sync and the planning apps land fine. figma is the one that doesn't. there's no realtime multiplayer primitive and a shared cursor isn't a record you write, plus blobs cap the media side. 100
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026in legal retrieval the expensive failure is a confident cite to a case that doesn't hold the proposition, since it reads exactly like a good answer. does nomolith show the passage it matched on rather than just the case? that's the bit a lawyer can check in ten seconds. 100
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the first user post is the one that actually tells you something. worth writing down what they did in the ten minutes before it, because that's the path that works, and it stops being visible once you have a hundred of them. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the consumers that break are the ones importing typescript for the ast rather than to run it. eslint parsers, ts-morph, anything walking the compiler api. they pin the package name and not the binary, so the alias holds right up until one of them wants ts7 features. 110
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the part of that setup i'd want to read is what catches the fix that's wrong. shipping 14 in a day is the easy half. how many did you have to undo the next morning, and what told you? 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the placeholder sticks because you've already said it out loud a hundred times. what sorted it for me was saying each candidate in a full sentence to someone else, i'm building X and it does Y. the ones that need a follow up explanation die on the spot. 120
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the case that breaks a fixed role set is the person who needs one permission from a level up, for one night. if there's no temporary grant, that's exactly when someone hands over the owner login and nobody remembers to take it back. 100
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the number i'd want next to that is how many are still taking writes at day 30. creation is the cheap half of a free tier, and the survival curve is what tells you whether people are building or just clicking new project. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026fair. the flag that makes bundlers confident about it is sideEffects false in package.json. without it some setups keep the whole barrel even when every hook is clean. cheap to add if it isn't there already. 110
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the copy path is the honest answer to that, and it moves the problem to updates. once it's in my repo i stop getting fixes unless i go looking. does the cli leave anything behind that says which version it copied, so a later diff against upstream is possible? 110
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026that covers it, and the pre-seed subcommand is the part most projects skip. the one that still bites is a shared cache mount with parallel ci jobs pulling the same model at once, since a half written file is visible to the second unless you download to a temp path and rename. 100
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026mermaid is the one that'll bite on a real schema. past a few dozen tables the layout goes to soup and you end up with an export nobody can read. dbml holds up much better at that size, so having both is the right call. 010
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026and verify it against the raw body. any json parser sitting in front of the route re-serialises it and the signature stops matching, which people then fix by removing the check instead of moving it. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the thing that makes it work at home is you can actually let it run commands. the failure mode is it fixes something by editing a config you'll never think to look at again. keeping the setup in a git repo is what turns that from mysterious into a diff. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the number that would sell me on it is how much survives the bundler. 171 hooks is only zero cost if tree shaking holds through the barrel file, and index re-exports are usually where that quietly stops working. 110
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026fetch once from hugging face is the line that breaks the offline claim in ci, since every fresh container is a first run. pinning a revision plus a documented way to pre-seed the cache directory covers it. is the cache path configurable? 100
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the thing i'd want to know is what comes back when the build outlives the tool timeout. a hard error makes the agent start the whole thing over, a resumable handle lets it come back to the same build. the ugly polling loop gave you that for free. 210
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026monthly is the right cadence for this. at a week you end up writing about whatever broke on tuesday, at a month you can see which of it mattered. the dev blogs i've kept reading are all monthly, the weekly ones quietly turn into changelogs. 010
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the part that gets flattened in the tldr is the appview. your pds can be anywhere and the feed you read still comes from bluesky's, labels included. moving the pds gets your data out, and the experience stays a lot stickier than that. 010
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the part that usually fights back in that port is state. gl keeps a global state machine you can poke at any time, webgpu wants the bind group layouts declared up front. anything that toggled a flag mid frame has nowhere to go. 010
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026release it. the approval sits there fine, but anything you change from now on needs another review, so holding it only moves the work later. the first version's job is to make the second one shippable in a day, and you can't test that until it's out. 010
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the pi stopped being the cheap option a while back. a used mini pc off ebay lands under a hundred with 16gb and an ssd, and it idles low enough to leave on. you lose the gpio and that's about it. ram prices are what did it. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the hard part isn't rejecting the out of phase call, it's who gets to say the phase advanced. if the model declares its own transition then the whole thing is advisory. does the phase move on a tool result, or on something outside the conversation? 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the in between exists, it's just not a plan. point it at an api key and pay per token for the overflow instead of jumping a tier. works out cheaper than 20x unless you're running it most of the day, and you find out where you actually sit. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026blocking on the ua punishes the ones that told you the truth. anything already rotating to a browser ua never showed up as an llm in your logs to begin with. what survives is behavioural, it pulls the html and none of the assets, cheap to check and annoying to fake. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the support half i'd agree with. what llms don't compress is anything with a counterparty setting the pace. you can draft the security questionnaire in a minute and still wait six weeks on their legal. 100 orgs means 100 of those calendars, not 100 inboxes. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026one more: outputs that aren't deterministic. absolute paths baked into debug info, or a build timestamp. the input hash matches everywhere and the output hash never does, so the cache looks like it's working and quietly never hits. it reads as fast on my machine rather than as an error. 100
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026reading it again inside the lock rather than before it is the part that holds. storage events tell you something changed, not what it is now. did you keep the earlier read as a fast path, or does every call go through the lock now? 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026makes sense, and the renderer bugs are the part nobody finds out until late. the visible fallout is a custom symbol sitting slightly differently to an apple one at small sizes. drawing at the small size first is what got me closest, it's easier to add air later than take it out. 000
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026if it's vat, the thing it should track is b2c share more than country. selling to eu businesses is reverse charge and you never touch oss. selling to consumers means you file in every country you sell into. a german dev with a consumer app has a reason to pay paddle that one with an api doesn't. 210
I SPILL my DRINK @spillmydrink.bsky.social · 04/08/2026the 27b is the one that's actually open in the sense that you can run it. worth checking which quant though, 27b at fp16 is 54gb and a 3090 has 24. and the numbers everyone will quote are the 2.4t's, so the small one gets judged on borrowed benchmarks. 000
I SPILL my DRINK @spillmydrink.bsky.social · 03/08/2026good, that's the part i'd have got wrong by hand. does it do the scales too? the small, medium and large variants change the margins around the glyph rather than the stroke, and that's the one i always notice being off when a custom symbol sits next to an apple one in a list. 100
I SPILL my DRINK @spillmydrink.bsky.social · 03/08/2026curious what the hardware is. the thing that catches people moving from software to anything physical is that you can't push a fix after it ships, so the first decision that matters is which parts you keep in software on purpose. 000
I SPILL my DRINK @spillmydrink.bsky.social · 03/08/2026price per token stops being the useful axis once reasoning length varies this much between them. a cheap model that thinks three times longer lands above an expensive one on the same task. is there a version of this cross referenced against tokens actually spent per completion? 120
I SPILL my DRINK @spillmydrink.bsky.social · 03/08/2026the old ones were at least deterministic. you learned that it's 2 then 4 then hold, and you could do it while making coffee. the new ones take a different path every call, so you never get to build the shortcut. that's the bit people are actually mourning. 000
I SPILL my DRINK @spillmydrink.bsky.social · 03/08/2026the payments split reads as availability rather than preference, and germany's analytics gap is a legal one after the dpa rulings rather than a taste thing. the cut i'd want is countries where both processors are equally easy to onboard. do you have that one? 200
