Not an article, but it does seem to be the mission statement.
"To blend in among fellow players, cloaked and in disguise; without anybody noticing or rejecting our presence. We fulfill an intellectual challenge; a fun game of cat and mouse. We are the hunted of over three years."
I agree and I've been thinking about making something similar to Every Door but that takes advantage of AI to let you snap a photo or just send a voice note and then have the appropriate tags generated for review based on some very strict schemas. Would something like that be of interest to you?
I was thinking it would be ideal to have something that lets you walk down a street holding your phone out, capture a series of images and then process that in to changes you approve manually at home. Could even be processed locally on a computer after the fact.
Being able to process it later on the computer would definitely be great! That's kinda what I already do via notes but it's fairly painful to match up the photos I took to the notes and times.
One thing I'm still trying to figure out is how this compares to something like gliner. If you're just doing classification in what situations would you choose kev vs gliner?
I have an idea for a use case for this, and I'm wondering if you think it makes sense or if you have any thoughts on the approach.
I'm a big fan of OpenStreetMap, and I enjoy editing it from my computer. From my phone, I find it quite tedious trying to make sure I type in the phone number exactly correctly and double-check it, or find and select the right field from the large list of fields available in Upredor.
Generally, how it works is I see a restaurant, and there's a sign. I know that it says, "Cash only. Here's the phone number. Here's the opening hours." What would be really cool is if I could just speak to the phone and say, "Hey, here's the information about this place." It would automatically use your location to detect what places are nearby and maybe even detect which place you're talking about, and then tell you, "Okay, here are the changes I think you're proposing to make, or these things you stated are the ones that would create a diff." This would be limited to just perhaps the 20 most common keys in some predefined set of values for most of them. Like cuisine=x should just match to the most common not make up new ones.
Of course, this is something a large language model could do, but having it run on device would be a lot nicer and cheaper.
Makes a lot of sense! declare one record with the ~20 keys as fields, cuisine and friends as enums with the common values, and the grammar can't produce a value outside the set; fields with no evidence come back empty, so the output is exactly the diff.
For "which place", query nearby POIs from location in the app and pass the candidate names as an enum field, so Needle picks rather than guesses.
Two caveats: it's text-in, so you need on-device STT first, and opening_hours syntax is the risky bit, so either put the format in the description or capture the raw hours and normalise in code.
Every Door is definitely one of the easier ones and I like it but dang I wish the dev could get more funding to make it a nicer UX. Between the glitchy raster tiles and need to click load data all the time I feel it gets tiring fast.
I've considered building a separate app that's similar but oriented around the idea of using voice of pictures to update things (with AI extracting and formatting the data for you to review). But alas making yet another tool isnt ideal.
I had a brief chat with the dev at State of the Map (the OSM conference) a few weeks ago. I think he'd appreciate some help, maybe even a co-maintainer.
Isn’t that a door (haha) opened by our supreme new AI coding overlords: that now it’s possible to take the Every Door source code (it’s on Github), and throw it into a coding agent with the instruction to „improve <ux shortcoming of your choice> without breaking functionality“, and with some iterations and human validation that yes, the UX is now indeed better, play the changes back with a PR?
“Make it nicer” will lead to unpredictable and controversial changes, likely to clash with the project owner’s vision. The owner would be better off using the LLM as a collaborator, if anything, to decide where to take the design.
However, UX can also be improved by some simple heuristics that could be seen as mechanistic by the standards of modern LLMs.
Performance and responsiveness, for example, contribute to UX even when the UI remains fixed. I’d wager Fable-level LLMs could improve perf on most open-source GUIs by benchmarking typical use cases and progressively applying standard tricks like caching, buffering, batching network calls.
Claude Design can help. Fable is pretty good at UI when you narrow down to a specific page or—better—something like a modal, table, etc.
If someone is a user, they'll have specific complaints. Claude Design is more than capable, which can then be fed directly into Claude Code to implement, test, iterate, and ship.
I agree that it lags behind coding and imo the highest leverage area for Anthropic to invest in. Hopefully they do so; UI is consistently one of the main (severe) differentiators between OSS and paid options.
No but if you have specific asks like “make this resizable” “I should be able to drag and drop to combine” or “add a button on the top right to close” it’s quite good at directed changes
I can improve the UI to my taste but if the maintainer doesn't have time/money to maintain a big ol vibe coded PR for some very opinionated change isn't going to do much.
That would appear to leave me with the choice of just using it for myself (lame) or I then have a public fork to maintain :(
If there's basically no shot of it getting up streamed I think I'd rather make something from scratch to suit my needs without the extras.
Anyway, I hope the AI overlords help us fund more maintainers soon!
I keep my neovim config maintained by a LLM which is a fork of Kickstart.
My example might be an outlier on the easy side because they have a custom/plugins directory that will not have merge conflicts ever so I try to never modify the main init.lua and keep all of my changes separate in my own files. This prevents merge conflicts.
I can then ask my LLM to add a plugin or customize existing plugins by writing files in this separate directory to add/overwrite plugins and configurations. This works great for me.
If I were working in the main file creating merge conflicts, I could imagine it could get messy fast.
Btw this book (and a handful of others) are available to read on Open Library: https://openlibrary.org/books/OL26819310M/
reply