Almost nobody can describe what they actually want.
Every travel app opens with the same question — pick your interests, choose a vibe, tell us who you are. It is a reasonable question and it has a hard ceiling: the answer is only ever as good as your ability to describe yourself to a form. Byway asks it once, then stops asking and starts watching the road. The model that builds is pointed at one traveler, trained on nobody else, and we measured it before we asked you to believe in it.
You drive
Nothing to fill in. Arriving at a stop, keeping one, rating one, skipping one — every signal the model reads comes from something you actually did on the road.
The model moves
Those signals become small, signed adjustments across your thirty-two dimensions. A stop you kept pulls the model toward it; one you skipped pushes away. One drive tilts the model. It never overwrites it, and nothing anyone can say will rewrite what your driving has taught it.
The next shortlist shifts
The engine reads your model when it decides which verified stops to offer along your route. Whether that actually helps is not something we ask you to take on faith over a season of driving. It is measured below.
Not likes nature.
These are the axes the model actually carries — the same list the app writes to and the engine scores candidates against. Your model holds a weight on every one at once, which is why two travelers asking for the same road get different stops. A few of them are practical constraints rather than tastes — what a stop costs, how hard it is to walk, whether a dog is welcome — and the model carries those the same way, because they decide whether a place suits you just as firmly as scenery does.
- scenic
- viewpoint
- coast
- beach
- mountain
- nature
- waterfall
- historic
- cultural
- food
- local gems
- hidden gems
- hidden towns
- family
- wildlife
- adventure
- art
- photography
- sunset
- sunrise
- quiet
- popular
- free
- paid
- easy
- strenuous
- pet friendly
- accessible
- remote
- urban
- educational
- romantic
A brand-new traveler starts with all thirty-two equal, and equal means silent. Until your driving gives the model something to say, it says nothing at all and the road and the record carry the plan — you are never handed a stranger's stereotype while the model waits to meet you.
You are never overruled by a model.
Three things have a claim on which stop you are offered, and the question that matters is what happens when they disagree. They resolve in a fixed order, and the order is the promise.
- 01
History settles first
Among everything the model has learned, what you have actually done carries the most weight. It does not overrule the interests you ticked — those keep more weight than anything learned, which is why a traveler who states a preference is never talked out of it by their own history — but your driving is what moves the model underneath.
- 02
Recent signals season it
Deliberate signals — the stop you called the best of the drive, the place you lingered at rather than drove past — reach the model faster than the nightly pass. They are capped by design, because the freshest signal is also the thinnest one.
- 03
Today's choice always leads
Whatever you ask for on this drive is applied last and outweighs everything learned, put together. A model that has watched you for a year still loses to the thing you just said you wanted. That is the product, not a tuning choice.
And a local's word is not a fourth opinion about you
It answers a different question. Taste asks whether a place suits you; quality asks whether it is worth anyone's time at all. They stay separate scores, combined only at the end, which is how a local signal can lift a place for every traveler without ever pretending to know your taste.
A model of one traveler. Trained on no one else.
Personalization is the one claim on this site you are entitled to be suspicious of, so here is the whole shape of it in plain terms — enough that an engineer reading this page can tell whether we know what we built.
- The model
- A preference model carried per traveler across thirty-two dimensions of taste, one weight on each. Yours moves as you drive; nobody else's moves it. It is a classical, well-understood family of model — relevance feedback — chosen because its behaviour can be explained to you in a sentence, which matters more here than novelty.
- What it reads
- Behaviour, not forms. Arriving somewhere, keeping a stop, rating one, skipping one. No questionnaire beyond the interests you tick yourself, no data bought from a broker, and nothing we measure anywhere else on the site ever reaches it.
- How it ranks
- Vector similarity between your model and each candidate place, computed in the database at query time, across every candidate the engine is considering for your route. Personalization happens inside the plan, not as a filter bolted on afterwards.
- Where it lives
- Scoped to your account in the database itself, enforced by the database rather than by our good intentions. Nothing writes to it but your own driving. Byway AI can read it back and explain what it believes about you — it cannot change it, and there is no dial here for us to turn on you either.
What we will not do
No collaborative filtering. Nothing learned from other travelers can outrank something learned from you — a bound on the arithmetic itself. What travelers share is the record, never each other: a locked gate or a view grown over is reported, dated, and carried with the place, because that is a judgement about the place and not about you. Even there the rule is narrow — one account's report can degrade a stop and never erase it, and only a verified channel closes one. Your own driving decides what you are shown, and anything we ever learn from the crowd will sit under that bound rather than beside it. We would rather be slow to learn you than quick to assume you — the hard thing to copy was never the model anyway, it is the corpus underneath it. We do not sell personal data.
We measured it before we asked you to believe it.
Byway is in private beta, and a model this young has not logged enough road yet to grade itself on real trips. So we graded the mechanism instead: 9,600 simulated travelers, each with a taste they never state out loud, choosing stops along real recorded routes. The model sees only what they kept. Every figure is the average of 12 independent runs, so no single lucky run is doing the talking. The large travel platforms run evaluations like this one too. What almost none of them do is publish one where a traveler can read it.
- Byway learns it
- asking you
| how well you can describe your own taste | the model | your own answers |
|---|---|---|
| can't at all | 28.5% | 15.5% |
| barely | 28.1% | 18.9% |
| roughly | 29.0% | 20.8% |
| well | 27.6% | 24.2% |
| very well | 27.7% | 26.5% |
| perfectly | 27.6% | 27.0% |
The finding, and it is the whole argument
At or below 40% self-description the model wins outright — across 12 runs the two do not overlap at all, which is a harder test than statistical significance. Above that they converge, and we will not call that a win: for a traveler who can name their own taste perfectly, a questionnaire does the job just as well. That is the result, not a caveat. The claim was never that a model is cleverer than a form — it is that it does not need you to be good at forms, and nobody planning their first drive through a place they have never been is good at forms.
And how quickly it learns you
- ceiling
- notability
- your answers
- learned
How it was run
The one thing worth checking on a simulated result is whether the imaginary traveler thinks like the model being graded. If they do, the evaluation measures nothing but itself. So here is the whole setup, printed from the run that produced the numbers above.
- Travelers
- 9,600 across 12 runs
- Places
- 772 real recorded places
- Shortlist per query
- 40 candidates
- Held out
- the next stop chosen, never shown to the learner
- How the traveler chooses
- softmax over the pool at temperature 0.35
- What the traveler wants
- the weight the traveler puts on the place's primary category, multiplied by how far they will detour for it
- What the model scores by
- cosine similarity between model and place — mirroring SQL 1 - (feature_vec <=> taste) — multiplied by the same detour curve the engine applies
- Why that is not circular
- the traveler decides by a single-category weight; the ranker sees only a 32-dimension vector and never reads that weight. Distance is shared deliberately: it is not hidden from either side
One metric is a cut, not a result, so here is a second one. Mean reciprocal rank — which rewards putting the right stop high rather than merely inside five — reads 0.186 for the model against 0.164 for the questionnaire, with the oracle at 0.236. The ordering is the same one the headline reports.
How to read these numbers
These are simulated travelers, not customers: the figures grade the mechanism, and the ceiling above shows exactly how much room is left in it. Personalization in this category is asserted to travelers and measured in private, which is the part we are breaking with. Method and every figure below, including the ones that do not flatter us. Last run 2026-08-20.
The next number is already committed to. Once one hundred travelers have driven five or more routes, the same evaluation runs against real held-out drives, and that result replaces this one on this page — better or worse. Writing the threshold down before seeing the data is the only thing that makes the eventual number worth anything.
One traveller, everywhere.
Your model does not live inside one screen. It is stored against your account, not against the thing you happened to open, so the same understanding of what you like is there whether you are planning at a desk or asking an assistant for a detour. You do not start from zero when the interface changes.
- Byway on the web
- Byway AI
- ChatGPT & Claude
- The phone app
- In the car
One per traveller. Carried across every surface above it.
- Plan
- Travel
- Learn
- Plan again
What you have done before gives Byway context. What you ask for today still wins — the order in Composition above does not change because the surface did.
The first three read one model today. The phone app and the car are not yet places you can use it, and they are marked as such: the architecture is what makes adding them a wiring problem rather than a second model.
It never invents a place
Personalization chooses among verified places; it cannot conjure one. Every stop still traces to the public record behind it, whatever your model says.
Nothing outranks what you did
There is no path into your taste model from anyone else's. Byway will not pool travelers to predict travelers, and if we ever learn anything from the crowd it enters capped and underneath you — able to break a tie you have not broken yourself, never to reverse one you have. What travelers do share is the record: a locked gate, an overgrown view, a stop that is no longer what it was. That is the atlas correcting itself, and no single account can erase a place on its own.
It should never flatten you
Where the model is thin, the engine leans on the road and the record rather than guessing at you. An untouched model expresses no opinion at all — it waits until your driving has given it one, instead of filling the silence with an assumption.
Where we are taking it.
The loop above runs today and the measurement above is real. Everything below is what the evaluation says to do next — written as intent, not as a shipping promise.
Everything above is simulation. The threshold is already written down: one hundred travelers, five drives each, the same harness against real held-out sessions. That number replaces this one whichever way it lands.
Recent signals should count for more when they confirm the longer record and less when they contradict it, rather than gaining weight on volume alone. Agreement is the better measure of confidence, and the evaluation is how we will know it worked.
Thirty-two dimensions against a handful of stops per drive is a thin signal, and the honest name for it is sparsity. The fix is not to pool people, it is to pool the SHAPE: which kinds of places tend to go together, learned in aggregate and used to spread a few signals across the dimensions they imply. That is a fact about taste, not a guess about you — and it enters under the bound above, able to break a tie you have not broken yourself and never to reverse one you have. The cap and its decay to nothing are the part we would publish and test.
The same traveler wants a different road in February than in July, alone than with kids in the car, on two hours than on two days. The model should carry the context, not just the preference.
See what the model believes about you, in plain words. Correct it where it is wrong. Forget a stretch of road you would rather it did not remember.
The web, the assistant and an outside AI already resolve one model per traveller. The dash does not: turn-by-turn is its own surface with its own constraints, and the work is to reach it without the model becoming a second copy that drifts from the first.
The most personal signals — where you slowed, where you turned around — should be able to shape your model without ever leaving your device. It is the natural end of a model that already belongs to one person.
Drive it once. Then drive it again.
The model has nothing to say about you until you go somewhere. Plan a drive, keep the stops worth keeping, and it starts to move — and unlike a questionnaire, it never needs you to explain yourself.
Questions about the model
Is this really machine learning?
Yes, in the textbook sense: parameters fitted from data by a named algorithm, evaluated against baselines. It is also nothing like what the phrase tends to imply — no deep network, nothing trained on a crawl of the internet, and no model of anyone but you. We would rather describe it accurately and publish an evaluation that does not flatter it than let two initials do the work.
Why not learn from other travelers?
We do learn from them about places — a reported locked gate or a view grown over is carried with the stop, dated, for everyone. What we will not do is let anything learned from them outrank what we learned from you. Your model is yours permanently, not a stage we are passing through. The cost is real and worth saying plainly: learning across everybody is what makes most recommenders work, and holding it under a bound means a smaller model that leans harder on the road and the record. It also means nothing you do on Byway is ever used to profile anyone else.
Can I see what it learned about me?
Yes — ask Byway AI about your taste profile and it reads back your strongest dimensions and the categories you picked yourself. It is a read-only view: the assistant can explain your model, never rewrite it.
What happens before it knows anything about me?
The interests you tick carry the plan, and they outweigh everything learned anyway. The model itself stays quiet until your driving gives it something to say — deliberately so, after the evaluation showed that an untouched model was not as neutral as we assumed.