WoT Technology · London · October 2026

The everything computer

Take the middleware out of computing and only three things are left: imagination, time and tokens.

Scroll, or use the arrow keys

1 / 10
Where the time goes today

Most of building software is plumbing

A simple app sits on a stack of separate systems. Each has its own format, its own failure modes and its own specialists, and most of a team's week goes into keeping them talking to each other.

Interfacewhat the user touches
API and authwho may ask what
Queues and cachesmaking it fast enough
Sync and replicationkeeping copies in step
Database and file systemwhere it lives
Logs, audit, provenanceproving what happened
Deploy and opskeeping it all running
2 / 10
What WoT is

One material where those layers used to be

WoT is a compute floor where identity, provenance, solving, media and money are the same material. Everything is a thought with its own address. The same form is used in memory, on disk and across the network, so there is nothing to translate between.

API and authidentity is built in
Queues and cachesanswers arrive by themselves
Sync and replicationone address, every machine
Database and file systemmemory and disk are one form
Logs, audit, provenancea receipt on every answer
What you want it to dothe only layer left
3 / 10
Why it builds fast

You describe the goal. It finds the design.

Capabilities are sentences

Every capability is a short sentence of verbs anyone can rearrange. A new workflow is a new sentence, not a new service.

A solver underneath

Give it a goal, materials and constraints. It returns the design that meets them, or names exactly what blocks it. It does not make answers up.

Nothing to wire up

No schemas to migrate between systems, no glue code between a cache and a database. The thing you build is the thing that runs.

4 / 10
Where AI models fit

Language models don't invent. They can steer.

A language model predicts the likely next word. It will never invent the answer to a problem nobody has solved. But it is a fast, cheap guide: it can propose where to look, and a solver can check every proposal against the problem a person actually defined.

The model guides

It suggests candidates, quickly and in bulk. Every new model release makes this part cheaper.

The solver checks

Every candidate is tested against the real goal and constraints. Only what passes comes back, with the evidence.

So better models help us

When the next model turns someone's prompt chain into one API call, our guide gets cheaper and our verified answer is worth more.

5 / 10
Why it feels instant

Ask once. Everyone waiting is told.

In a normal stack, every screen asks the server again and every layer adds a round trip. On WoT, a part that needs an answer says so once. When the answer exists, every part waiting on it is told at the same moment, and nothing recomputes it for itself.

The same rule runs on one laptop and across a mesh of machines, so the work spreads out without anyone scheduling it.

6 / 10
Measured on real machines · 18 October 2026

One machine does the work. Every other machine is told.

One machine searches all of Wikipedia113.5 GB in 27 s, 18 marbles in parallel; its 946 hits match an independent search tool. A hundred questions in one pass: 41 s
A second machine asks the same questionTold in 336–1,230 ns. It holds no byte of the dump and walks nothing.
The answer crosses the networkAbout 0.5 s, once. After that, everyone on that machine is told locally.
Serving one told question0.29 ms of work, down from 70–90 ms a week earlier
A 3,000-question essay55.9 → 10.2 s of compute on one machine, every answer identical
Every internet standard, read by its own grammarAll 774 RFCs in 3.3 s
A rebuild from nothing agrees with the running codeByte for byte, on a second independent build
7 / 10
Token economics

A token worked out once never has to be worked out again

Most inference repeats itself: the same system prompt, the same questions, the same continuations. On WoT a step's answer is filed under the exact content that produced it, so a repeat is told instead of recomputed. The more traffic repeats, the more of each token is already paid for.

Lab result · 21 October 2026
Normal serving
24.5 ms
Told, 3 in 4 repeat
7.1 ms
Told, 9 in 10 repeat
3.8–5.3 ms

Exactly the same answer

762 / 762 requests gave identical tokens, and the model's scores matched bit for bit. Nothing is approximated.

Up to 6.5× less time a token

3.2–6.5× faster as repeats rise, while normal serving stays flat whatever the traffic.

Where it goes next

The cost of a told token here is a copy. Handed over by address, it falls towards the cost of looking it up.

A lab spike on one machine with a small open model (Qwen2.5-0.5B), measured by time.

8 / 10
What is left

computing − middleware
= imagination + time + tokens

When the plumbing is gone, the cost of an idea is the time to say it and the tokens to think it through. Building becomes describing, and the person with the idea can be the person who ships it.

9 / 10
The beta

Register for beta updates

We'll email when the beta opens, and when there's something new to try. Nothing else.

Keif Gwinn · WoT Technology Ltd · wot.technology

10 / 10