Here's a number that shouldn't still be true. The trick that makes a Plus Code short is the same trick that makes it fragile. Drop four digits off the front, and you've saved six characters at the cost of a dependency you cannot control.
And that trade is entirely invisible to most people using them.
Daniel's question picks up where we left the last one. He's been on a hunt, in his words, for the ideal open geographic addressing system. Something that spits out a short, human-readable, globally unique identifier for a precise spot, generated off a phone's GPS, decodable offline by whoever receives it, with no locality name, no proprietary database, no geocoding service standing in the middle.
And the target resolution he names is fourteen metres by fourteen metres, though he says the exact precision is negotiable.
He also splits it into two questions, and the first one is sharper than it sounds. Has Google's consumer implementation obscured the real potential of Google's own open standard? Because his complaint isn't about the standard. It's about the interface. The shortened code they put in front of everyone.
And the second question is whether anyone else has built the technically purer version of the idea. Whether the ideal already exists, just badly presented, or whether there's a real gap.
So let's start with what Google actually shows people.
The full-length code is the whole point of the design. It's self-contained. Google's own README says a full code can be converted back to a location completely offline, without any data tables to look up and without any online services required.
Which is what you'd want in a flood.
Right. And then the same README describes the shortened form. Those can be used regionally, or in combination with a reference location that all users of the short code need to be aware of. And then the sentence that does all the damage. If the reference location is given in the form of a location name, use of a geocoding service might be necessary to recover the original location.
So the short code isn't wrong. It's just not finished.
It's incomplete by design. And the README draws the line itself. Geocoding services are not a part of the Open Location Code technology. They're a part of the product built on top of it. The locality dependency is introduced by the interface.
Which means the standard's defenders are correct and the users are still stuck.
So the arc for today is, why shortened codes exist at all, what the independent ecosystem actually offers, how the competing systems compare, and what all of it means for the one case where the answer really has to be right.
The emergency case. Walk me through the shortening, then. Mechanically.
You drop four or more leading digits from the front of a full code. What's left works on its own within roughly fifty kilometres of the place it belongs to. If you supply a city or a locality, it resolves globally, because the missing prefix can be reconstructed from the reference point. And the short code always keeps the plus sign, so the receiver can see how much is missing and compute the rest.
The plus is the seam.
And the precision numbers are worth having. A ten-digit code is one eight-thousandth of a degree, which at the equator is about thirteen point nine metres by thirteen point nine. That's Daniel's fourteen by fourteen. Add a digit and you get roughly three point four metres by two point seven. Add another and you're under a square metre.
So a person could hand over a code smaller than a doorway.
A person could. And the reason Google shortened them at all is usability. Nine seven M F plus two seven, Mumbai, India, is a thing you can say out loud and a stranger can write down. The full ten-digit string is not. The locality name is doing the work of memory.
It's doing the work of memory and borrowing against trust.
That's the trade. And the standard's own design paper is explicit that this trade was never supposed to be made. The requirements list reads like a manifesto. Complete, it says. The codes must have enough information on their own. Extra information such as street number, locality or country name could be helpful, but should not be required.
Should not be required. Somebody wrote that sentence and then somebody shipped Mumbai, India.
Same paper, three more. Functions offline. Easy to implement, and it should not depend on a single provider. Free, and it should not require a licence fee.
So the standard promised self-contained and the flagship interface leads with dependent.
And Google's own documentation concedes the failure mode. There's a worked example in the README about Zurich. Although one service may place Zurich close to the Google office, another may move it by a hundred metres or more, and this could be enough to prevent the original code being recovered.
A hundred metres. That's a different street.
It's a different building, in a case where the point of the code was to say exactly which one. And the README's advice is telling. Use a smaller neighbourhood feature rather than a large city name.
Which is good advice and not the one the interface follows.
The interface optimizes for a tourist typing a place they've heard of.
The character set is the part I find elegant. Tell me about that, because it's the bit people miss.
The alphabet is two three four five six seven eight nine C F G H J M P Q R V W X. Base twenty. No vowels, so you can't spell anything. And the confusable letters are gone. No I, no L, no O, no S, no Z, no zero and no one to fight with them. They picked it by generating every possible twenty-character combination, over eight billion of them, and scoring each one against ten thousand words from more than thirty languages. Anything that spelled a real word got thrown out.
They ran the world's languages through it and kept the set that says nothing.
Deliberately. Transcribing something out loud over a bad radio, you cannot accidentally read a word that isn't there. There's no word to read.
And then the imperfections, because there always are.
The character set is Latin-only, so you get the Cyrillic N that looks exactly like a Latin H, and that ambiguity is real the moment a code crosses a keyboard. There are discontinuities at the poles and at longitude one eighty. Area distortion grows at high latitude. And the scheme can't exactly represent latitude ninety.
The top of the planet is undefined.
The top of the planet is undefined, which matters to approximately nobody, and I still want it fixed.
Daniel also cites a discussion on the Google Maps subreddit from twenty twenty-three about how hard it is to get a full-length code out of Google's own interface.
And I want to be careful there. We could not verify that thread. Daniel cites it, and I don't doubt he read it, but I haven't independently confirmed the content, so treat that as his claim and not as something I checked.
Fair. And notice what it claims, though. Not that the full code is unavailable. That it's undiscoverable.
Which is a different kind of problem. The capability ships. The path to it doesn't.
So: has Google obscured its own standard? First pass, yes, but not by hiding anything. By leading with the version that needs them.
That's a fair first pass. Though I'd push on the word obscured.
Push.
Shortened codes exist because the full code is hard to communicate by voice. That's not a marketing decision, that's a real constraint. Somebody in a call centre reading nine seven M F plus two seven plus a city gets it right. Somebody reading the full string gets it wrong twice.
So the shortened form is a concession to humans, not a lock-in.
It's a concession to humans that happens to require the vendor.
Both can be true. It's good for the caller and bad for the flood.
Yes. And here's the thing I keep circling. The dependency isn't the shortening, it's the name. You could shorten a code against another code. Against a known landmark set offline. The standard supports a reference location that isn't a name at all. Google chose the name.
Because the name is what people have. So now the second question. Who else built the pure version.
The independent ecosystem is real, and it's broader than people assume. The reference repo is multi-language. Java is about twenty-two percent of it, JavaScript fourteen and a half, Go nine, then PL/pgSQL, VBA, C++, Dart, C, Python, Rust, Ruby. There's even a Visual Basic for Applications implementation, which means somebody is generating Plus Codes inside Excel.
Of course there is.
And the third-party ports are the interesting layer. There's a JS and TypeScript library by janne, two to fifteen digit precision, zero dependencies, and it advertises that it matches Google's reference output. There's a Python package on PyPI, version twenty twenty-two point one point three, Apache licensed. There's a pure Kotlin Multiplatform port. And then there's the one I'd actually point a listener at, which is a Garmin Connect IQ datafield. It displays the Plus Code for your current location, and it doesn't use any network.
A watch. Offline. On a wrist.
A watch, offline, on a wrist, doing the thing Daniel asked about, and it isn't Google Maps.
So the libraries are there. Now ask the hard version of his question. An ordinary person with an ordinary Android phone, no Play Services, no Google Maps, no Google APIs. Can they do it?
This is where I have to be honest with you, and with Daniel, because he asked specifically for actual applications and not demonstrations.
Go on.
I could not confirm a standalone, mature, F-Droid-distributed Plus Codes app.
Say that again plainly.
There is no Plus Codes app in the F-Droid catalogue that I can point to and say, install this. The libraries exist in eleven languages. The Garmin datafield exists and is offline and shipping. But the thing Daniel actually asked for, a mature consumer application that generates and shares a full-length code with no Google components, I could not confirm exists. That's a genuine gap and I'm not going to paper over it.
So the algorithm is solved and the app is missing.
The algorithm is solved and the app is missing. Which is a stranger result than either of us expected going in.
What about the repo's own health? Because that tells you whether anyone's still home.
Four thousand three hundred and thirty-five stars, five hundred and twenty-nine forks, seventy contributors, fifty-six open issues. The latest tagged release is one point zero point four, from July twenty nineteen. But the main branch was pushed in March of this year.
So the standard is frozen and the plumbing is maintained.
The standard is stable, which is what you want in a standard. The release cadence is slow, which is what you'd expect from something Google treats as finished.
Now the field. Because if OLC isn't the answer, something else might be, and Daniel wants to know about the systems that tried.
Start with Geohash, because everybody's used it and nobody's checked it. Thirty-two character set, vowels included, so it can spell things. Not just profanity, though that's the famous one. The one that should worry you is that it can produce almost-words. There's a documented example where a Geohash looks like Z U R one C H. That's Zurich with a one in it.
Which a tired person reads as Zurich.
It's variable length. And it has five discontinuities, at the equator, longitude zero, longitude one eighty, and both poles. Those are coordinates where adjacent squares in the real world sit at opposite ends of the string.
Adjacent geography, distant codes.
That's the thing to remember. Two neighbours whose codes look nothing alike.
Geohash thirty-six.
Drops the vowels but gains mixed case, so now you have a capital O and a zero and a lower-case L in the same string. It can be one character shorter than a full Plus Code at similar accuracy, and it carries an optional altitude field and a checksum.
The checksum is the one good idea in the room.
The checksum is the one good idea in the room, and Plus Codes don't have one. There's an optional check digit in the specification but it isn't in the codes people actually use.
Keep going. MapCode.
MapCode produces global codes about the same length as Plus Codes, which is respectable. The territory codes are shorter, but they need a maintained data file, and they're keyed to ISO three one six six country codes. Which works beautifully until a country is disputed, and then the address system is taking a side.
The map decides the border.
And one location can carry several codes at once. Schiphol Airport has a global MapCode, and then at least three territory forms. Four valid answers to where is the airport.
Open Post Code.
Global codes and country codes, and the two sets are completely dissimilar. Which means a global code typed into a country field will be valid, and will decode somewhere else entirely. It won't fail. It'll succeed, incorrectly.
That's worse than no answer.
That's worse than no answer, and it's a design choice rather than a bug.
Natural Area Code.
Proprietary, licence required, whitespace significant, and no truncation. Three of those are fixable and the licence isn't.
Maidenhead, then. Because that's the one that's been quietly right for decades.
That's the one that's been quietly right for decades. It's area-based, you can truncate it, it lands at similar accuracy and similar length to a Plus Code. Radio operators have used it forever. Its limit is that it's only formally defined to eight characters, so the precision stops where the standard stops.
And What3Words, which is what most people have actually heard of.
Three-by-three metre squares, proprietary algorithm, and the part that matters for this conversation. The offline SDK was announced late twenty fourteen, but it doesn't support the OneWords people actually bought. So the thing you licensed isn't the thing that works on the plane.
Three words is easier to say than a code. That's the whole pitch.
Three words is easier to say, and it's dramatically easier to say wrongly. Banana rabbit monkey is in Argentina. Banana monkey rabbit is in Russia. Same three words, order reversed, wrong hemisphere.
And this is where the published research comes in, because Daniel specifically wants the engineering literature and not vibes.
Rudy Arthur at Exeter, in PLOS ONE, October twenty twenty-three. A critical analysis of the What3Words geocoding algorithm. Two findings, and both are simulation-based. First, most simulated addresses have one or more word triples they could be confused with. Only about a third of triples have no possible confusion at all, under typographic and homophone error modes.
So two thirds have a twin.
Two thirds have a twin somewhere. And around a quarter of addresses have more than three confusions, which matters because What3Words' own autosuggest offers three suggestions. Four confusions and the correct answer isn't on the list.
That's the number that should end the argument for emergency use.
That's the number that should slow it down, certainly. The second finding is about proximity. Confusable pairs near each other show up far more often than the one-in-two-and-a-half-million figure What3Words quotes. The cause is the linear-congruence assignment, and in a simulated area about fourteen and a half kilometres by four and a half, the peak of confusable pairs sat about ten kilometres apart. Two places you could walk between, whose codes are one letter apart.
So the errors aren't random. They cluster in exactly the geography where a mistake costs you the response.
That's the finding, stated plainly.
And then Arthur praises Plus Codes in the same paper.
He does, and it's the reason the citation is honest rather than an attack. He writes that it's obvious the Plus Codes P F Q nine plus six P seven, plus six P eight, plus six P nine are close together, while it's far from obvious that washed dot expand dot logs, poems dot joke dot urban, punks dot nails dot villa are neighbours.
Adjacent codes look adjacent.
That's the property the whole system is built on, and Arthur confirms it while demolishing the alternative.
But he also notices Google's own habit.
He does. He points out that Plus Codes are integrated into Google Maps and made more user-friendly by substituting the top-level information with standard place names and keeping the lower level for fine detail. His example is the Exeter computer science department, which comes out as UK, Exeter, P F Q nine plus six P seven.
So the researcher who defends the standard also documents the substitution.
Both things are in the paper. Nothing is being hidden.
And his recommendation for the lost walker?
The OS Grid Reference. Open source, widely understood, and critically, printed on most maps.
Printed on paper.
And that's the dimension I underweighted going in. Every phone-based identifier has the same failure: the battery dies, the phone goes in the river, and the code goes with it. A grid reference printed on a paper map in a rucksack survives all of that. The identifier and the medium are two separate design problems and we've been treating them as one.
Arthur also says something about the token set that we should quote, because it's the strongest argument against words.
Alphanumeric codes, he writes, could be much easier for non-native speakers and are less likely to be affected by accent, since the sender and receiver are only using tokens from a very restricted set.
Restricted set. That's the whole philosophy of the Plus Code alphabet in one phrase.
It's the same argument the designers made when they scored eight billion combinations against thirty languages. Arthur arrives at it from error modelling and lands in the same place.
And his closing caution.
That until such an evaluation has been carried out, he would caution against the widespread adoption of What3Words, especially in applications with a noisy communication channel between sender and receiver.
Noisy channel. Which is a radio in a storm, or a phone with one bar.
Which is every emergency.
Daniel also raises Uber's H3.
H3 is a spatial indexing system, and that's the right way to frame it. It's for databases and analysis, hexagons at multiple resolutions, and it's excellent at that job. But it's not designed to be said out loud by a person who is frightened. Nobody's reading a hexagon index over a radio.
Man walks into a shop with a hexagon index.
Man walks into a shop and the shopkeeper asks him to repeat it.
Back to Plus Codes, then. Because they do have real deployments, and we should say so.
CEP Digital runs across fifty-five Brazilian favelas. Favela Brasil Xpress uses it to address a hundred and twenty thousand residents. Shelter Associates addressed one thousand seven hundred families in Rajendra Nagar. Those aren't pilots. Those are people receiving post.
In places where there was no address to receive post at.
Where there was no address at all. Which is the strongest argument for the system that exists. It's not a thought experiment about emergencies. It's the mechanism by which somebody gets a delivery.
And there's a comment from the Hacker News thread on the original announcement that's worth reading, because it's the commercial version of Daniel's objection.
Doctor underscore Fegg, in twenty fifteen. Neat idea, shame the data and algorithm is closed. In the Rio example, a postal service or courier would have to pay What3Words to license the words-to-latitude-longitude lookup before they could deliver anything.
Read that back. Before they could deliver anything.
That's the sentence. The address system sits between the courier and the door, and it charges rent.
That's the one I'd want a listener to remember from the whole episode. An address is infrastructure. Infrastructure that can deny service is a toll road wearing a map.
And the other comment, from aw3c2 in twenty fourteen. I find them all really questionable as they try to push non-free, centralised setups.
Which is the same complaint, ten years earlier, before any of this was settled.
Before any of this was settled, and it held up.
So let's do the ledger. Six properties, and you can't have all of them.
Global uniqueness, compactness, memorability, transcription reliability, precision, and offline independence. Shorten for memorability and you buy a dependency. Use words for memorability and you buy a dictionary, a language, and a confusion set. Go finer on precision and you lengthen the string, and every added character is another chance to drop one.
And the one people forget. A globally unique identifier doesn't mean the location is right.
That's the distinction Daniel flags and it's the one that gets lost. Uniqueness is a property of the encoding. Accuracy is a property of the fix. You can have a flawless, globally unique, offline-decodable code for a position that is forty metres wrong, because the phone was under a bridge.
Perfect address, wrong door.
Perfect address, wrong door, and the code did nothing wrong.
So where does that leave the two questions?
First question, yes, the consumer implementation has obscured the standard, and the obscuring is presentational rather than malicious. The standard's own requirements say the codes should be complete on their own. The flagship interface leads with an incomplete one. Google's own README concedes the recovery can fail. Arthur's paper praises the design while documenting Google's substitution of place names. Every source agrees on the shape of the thing.
And the second question?
Plus Codes already meet almost every requirement Daniel listed. Complete, offline, open, no licence, no single provider. The system he describes mostly exists, and the piece that doesn't exist is the app.
Which is the strangest answer available and therefore probably the true one.
It's not thirteen point nine metres.
At the latitude you'd actually be standing at, it's narrower. The grid's fixed in degrees, so the ground distance shrinks as you go north. Thirteen point nine is the equator number and everyone quotes it like it's the number.
That's right. It's a degree-fraction grid, so the east-west spacing tightens with latitude. At Jerusalem's latitude the cell is noticeably narrower on the ground than the equator figure.
You've been sitting on that one.
I've been sitting on a lot of things. My brother-in-law sells surveying equipment. Has for years. And he will not use any of these.
Why not?
He says a code is a promise that someone else's software still exists. A mark on the ground is a promise that the ground is still there.
That's better than anything we said.
He's also not to be trusted. He once told a customer a boundary line ran through a shed, charged him for the survey, and was right. The customer had built the shed four metres onto the neighbour's side. The neighbour had known for eleven years and said nothing, waiting.
Waiting for what?
A reason. When the survey landed, the neighbour produced a folder. Photographs, dated, going back eleven years. He'd been documenting a shed that didn't exist yet when he started.
That's not possible.
The foundation was poured first. It sat there bare for two years while the neighbour photographed it. That's what the dates show.
So the neighbour documented a hole for two years.
He documented a hole. Then a slab. Then a shed. Then he waited eight more years for someone to ask. My brother-in-law was the first person to formally agree with him.
What happened to the shed?
It's still there. The neighbour decided he preferred having it. He'd spent eleven years documenting it and didn't want to start over on something else. He asked for a written apology instead, which he got, and which he then framed.
He framed the apology.
He framed the apology and hung it inside the shed. My brother-in-law says it's the only boundary dispute he's ever closed where both parties ended up with a document they liked.
What's the code for the shed?
I've never looked it up. I know the code for the pub, though. Nine characters, I've had it memorized for years. Couldn't tell you my own phone number.
Why the pub?
It's where the folder got handed over. Seemed worth keeping.
So the ground was there before the code and outlasted the argument.
That's about the size of it. Though the survey still cost four hundred.
Hilbert, thank you, and thank you to Hilbert Flumingtop, our producer.
Before we go. The thing on the cutting-room floor, because it doesn't fit anywhere and it's the detail I keep thinking about. The Plus Code alphabet was chosen by generating every possible twenty-character combination and scoring each one against ten thousand words from more than thirty languages, specifically to make sure the codes never spell anything.
They ran the world's languages through it so it would say nothing.
That the only thing a Plus Code can ever mean is a place.
Here's where I land. The technology for a universal, open, human-readable geographic address already exists, and the reason it isn't a public utility isn't technical. It's that nobody's built the boring app around it.
The boring app is the whole thing. A library that matches the reference implementation in eleven languages is not a utility. A watch datafield that works offline is closer, but it's still not the thing a person reaches for.
The invention happened. The adoption didn't.
Which means Daniel's quest has an answer, and the answer is a to-do list rather than a research programme.
If you enjoyed this one, rate and review the show. It helps other listeners find us.
We'll be back soon.
If you want more of this, try episode fifty-nine twenty-three, Plus Codes; episode forty-three seventy-three, Why Jaffa Street Has Three Names; and episode twenty-four thirty-seven, The False Precision of GPS Coordinates. This has been My Weird Prompts.
