The path to M1

We are preparing to release M1, our next visual intelligence model. We are describing it before it ships rather than after, because the thing that makes it useful is the same thing that makes it worth being careful with, and we would rather set that out in public while it is still a choice than explain it later under pressure.

This post says what M1 changes, what it does not change, what we are not going to claim, and how an organisation can ask for access.

One model, three questions

M1 reads photographs. Where a picture was taken is one of the questions it answers, and it is the one people ask about first, but it is not the whole of what the model does.

M1 is one model, not three. It does three jobs, each a mode that answers a different question and fails differently. Geo-Estimation says roughly where on Earth. Street-Match says this exact place. Property-Verify says and here is the proof, or I abstain. You choose the mode that fits the work, and the evidence the model gathers carries across all three rather than being rebuilt for each.

What M1 changes

Our current models reason toward an answer. They work through what is in the frame, the vegetation, the hardware, the signage, the road furniture, the quality of the light, and narrow the world down until the evidence stops supporting a smaller answer. That process is general. It works anywhere, including places nobody has ever photographed for a map.

M1 adds a second thing it can do. Where a place already sits inside reference coverage we have surveyed, M1 resolves the scene against that coverage directly. It is the difference between working out where you probably are and recognising the street. When the place is known to us, the answer arrives faster, it is more consistent between runs, and it is easier to check, because the claim is no longer an inference but a correspondence to a specific surveyed place.

Outside that coverage, M1 behaves as our models do now. It reasons. That fallback is not a weakness bolted on. It is most of the world, and it is the case our existing work was built for.

What it is good at

Speed, where coverage exists. Work that took an analyst an afternoon of manual cross referencing comes back in seconds, and it comes back the same way the second time.

Evidence, not just an answer. Every result carries the observations behind it: what the model read in the frame, which candidates it weighed, and why it preferred the one it chose. That is the part that survives somebody questioning it a month later.

Reach beyond the map. Geo-Estimation works on imagery nobody has surveyed, which is most of the world. Coverage sharpens an answer where it exists. It is not a precondition for getting one.

Repeatability. The same image returns the same answer. That sounds like a small thing until you are the person signing your name under it.

On accuracy

We measure M1 against the standard public evaluation sets continuously, and we will publish those numbers with the method printed alongside them. A figure without a method is marketing, and in a category where nearly every published accuracy claim is self reported and almost none come with a method, the method is the part that matters.

A public figure still only tells you how a model performs on somebody else’s fixed set. It does not tell you how it performs on your imagery, inside your workflow, on the kind of material your people actually receive. That second number is the one worth deciding on, and it is one we can measure with you during an evaluation, on material you choose. If you are weighing M1 seriously, ask us for it.

Knowing when it knows

The more interesting consequence is not the speed. It is that M1 can distinguish between two failures our current models cannot easily tell apart: a place it does not recognise, and a place that is not in coverage at all.

A model that only infers will always produce something. Asked about a stretch of road it has no purchase on, it returns its best guess with a number attached, and the number is doing work it has not earned. A model that can check its coverage first is able to say that this particular question is outside what it can answer well. For anyone whose job involves writing down a conclusion and then being asked how they know, that distinction is worth more than a faster answer.

When the evidence is thinner, M1 still answers. It returns the strongest thing it actually holds and labels which kind of answer that is, so the claim widens honestly instead of keeping the same confident shape all the way down.

Why we are saying this before shipping

Reading photographs for what they can be made to tell you is genuinely dual use, and pretending otherwise would be the least trustworthy thing we could do. A system like this is useful to a claims investigator testing whether damage happened where a policyholder says it did. It is useful to a newsroom deciding whether footage is what it purports to be. The same capability, pointed the other way, helps somebody work out where a person spends their time.

Making that capability faster, more consistent and available at volume does not create the problem, but it does change its shape. Something that took an analyst an afternoon and could only be done occasionally becomes something that can be done continuously. Capability that scales needs safeguards that scale with it, and those are easier to build before a launch than after one.

What we are doing about it

M1 answers where and whether, not who. It reads scenes, places and properties. It is not an identity system, it does not do face recognition, and we do not permit it to be used to locate a person. That is written into our terms rather than left to good intentions.

It declines rather than guesses. Precision stays bound to evidence. When the evidence supports a region, M1 names a region. When corroboration is not available, the confidence comes down to reflect that, even when a more specific answer would look better. A number that is only sometimes meant is not a number. Property-Verify is built to abstain outright, and that is the behaviour we want rather than a gap in it.

Access is granted to organisations, not to anonymous accounts. We verify who we are dealing with, what they intend to use it for, and under what jurisdiction, before an evaluation begins.

Retention is set per organisation and written down. Customer imagery is used to perform the requested analysis and is not used to train our models. Where an organisation carries obligations from its own clients, we set retention explicitly as a term rather than leaving it to a platform default.

A person stays in the loop. M1 produces evidence for a human decision. It is not built to be the decision, and evaluation access is granted on that basis.

What we are not claiming

Coverage is finite and always will be. Recognition depends on reference coverage we have surveyed, and that coverage will never be everywhere. Where it does not reach, Geo-Estimation is the mode that serves you, and it is the one our existing work was built around. If you are not sure which of those describes your material, that is worth establishing in an evaluation rather than guessing at from a blog post.

We are not giving a date. M1 is in preparation. Evaluation access will open before general availability, and a date in a post like this one is a promise the work has not made yet.

Requesting access

Evaluation access is staged and granted to named organisations. We are prioritising work where being wrong has a cost somebody bears: verification desks, fraud and claims investigation, trust and safety teams, and research groups.

Tell us who you are, what you would use it for, the jurisdiction you operate under, and roughly what volume you expect. That is enough for us to say yes, no, or not yet, and to tell you which of those it is quickly.