Webspinner™
Gossamer City
Your files as a city at night; a hero swings to the right building and a knock opens the original.
Webspinner · Portland, Oregon
Got AI?Create it. Own it.
Almost everyone using artificial intelligence today is renting it — by the seat, by the token, on terms that can change without them. Webspinner builds the other thing: intelligence you make yourself, on hardware you already own, that nobody can reprice, deprecate, or switch off.
One company, one foundation, and a growing estate of properties that all answer to the same idea. The course is free. The tools are free of charge. The work is in the open.
New · A keynote from the future · 13:10
The Slot Machine and the Mentor
This is an AI Produced Video. We use Sovereign AI on home computers to show you what is possible. The messaging in the video is truly John David Marx’s thoughts and also his voice, cloned on these same home computers. Sovereign AI is necessary to prevent further disparity in income, wealth and power. The Webspinner Foundation, a not-for-profit technology company, stands at the ready to help you help us democratize AI, before we lose this opportunity.
Tales from the Hermetic Museum
Eight stories of alchemy, each brought to life as an illustrated, narrated film.
Inference
The first trailer, made at home with Webspinner’s Inference Conjure.
Nano: The Storybook
A little robot learns his name, and Book 2 takes him to the pond.
The Inference Desktop
The Desktop in three minutes, cut from its own lessons.
Sovereign AI
The Inference Desktop, the Hebbian Corpus and the MDLC, filmed running.
Read it now · Books 1 & 2
Nano: The Storybook
Two picture books that play, Nano and Daniel Learns To Fly: turn the pages and each scene comes alive as a film, read aloud in ten languages.
Computers Make Mistakes. These books were made with AI on home computers, and we share them as they are, small mistakes included. When a child spots one and asks why, talk about it together: knowing that AI makes mistakes is the first lesson of living with it.
Its companion · 8 minutes
The Chaos Monkey, and why we break stuff.
For about seventy years a computer has been the most obedient machine anyone ever built: you tell it to do a thing and it does that thing, exactly, every time. Ask a language model the same question five times and you get five different answers. It is not broken. That is what it is.
So every course this Foundation makes shows you the failure before it shows you the fix — a model that answers a question about France with a speech from a history play, and one that defines a banana by reciting the definition of something else, word for word, with the grammar immaculate. The film ends where the software starts: what grounding means, what a Governor is for, and why consumers of synthetic intelligence do not deserve to be lied to.
Start here
The architecture, and the way in.
Everything Webspinner builds sits between these two. One is where sovereign inference actually runs. The other is how an ordinary person comes to understand it well enough to want it.
The architecture · nephonous.com
A sovereign inference architecture for Apple Silicon.
A federation of individually complete nodes, each retaining nothing once a turn ends. It answers the utilisation paradox in GPU datacenter economics with hardware that is already on desks — and it does not persist patron data, so it cannot disclose it.
Visit nephonous.com
The course · inferencegenesis.org
Create your own AI, and own it.
A free course from the Webspinner Foundation that walks anyone through building a working language model from nothing — on the machine they already own, in an afternoon, with no programming and nothing to install.
Visit inferencegenesis.orgMeasured here, not asserted
Three million parameters, ninety-six seconds, nothing sent anywhere.
The Foundation’s free course does not describe a language model. It has you build one, from an empty folder, and then shows you its insides: the corpus it read, the numbers it holds, the six characters it is still torn between. Every figure below was measured on that model, on ordinary hardware, and is yours to check by running it.
That is the whole method of this company. Where a number comes from somebody else, the source and the date travel with it, further down this page.
- 3,225,665 parameters in the language model the free course has you build — small enough to train on a laptop, real enough to generate text. Measured on the model InferenceGenesis teaches, September 2026.
- 1 min 36 s to train it, start to finish, on an ordinary machine. Nothing leaves the room it is trained in. Measured on Apple Silicon, September 2026.
- $0 to take the course. The Foundation is a 501(c)(3); the teaching is given away and the software is free of charge under the nosource agreement. The Webspinner Foundation, Portland, Oregon.
The case
Renting intelligence is a position, not a plan.
A rented model is somebody else’s product decision. It can be retired, repriced, retrained, or quietly changed underneath the work that depends on it — and the organisation that built on it finds out afterwards. That is not an accusation against any particular vendor. It is what renting means, and the dates are published in advance for anyone who reads them.
- 23 Oct 2026 the published shutdown date for GPT-3.5 Turbo, GPT-4, o1 and o3-mini. Whole products go too: the Evals platform, Agent Builder and the reusable-prompts API all close on 30 November 2026. OpenAI, API deprecations page — accessed September 2026.
- +13% on Office 365 E3, from $23 to $26 a seat, effective July 2026 — the first such rise since 2022, with AI features in the new packaging. Windows Enterprise per device went up 31%. Microsoft, 365 pricing and packaging updates — December 2025.
- ≥50% of generative-AI projects are expected to overrun their budgets through 2028, on poor architectural choices and thin operational know-how. Gartner — forecast, June 2025. A prediction, not a measurement.
What changed
Owning it stopped being the expensive option.
Three years ago, running capable inference on your own hardware meant a rack and a budget. It now means a laptop. Apple ships a three-billion-parameter model on-device to every user and gives developers an API to it; a 270-million-parameter model answered twenty-five conversations on a phone for three-quarters of one percent of its battery.
Be careful with the version of this argument you will read elsewhere. The capability gap between the best closed model and the best open one has not closed — measured properly, it has widened slightly, and open-weight models trail the frontier by about four months. The honest claim is narrower and more useful: the gap no longer decides most work, and the price of a given level of capability has fallen far enough that owning is a real choice.
- 9× – 900× the annual fall in the price of reaching a fixed level of capability, depending on the task. MMLU-level performance went from $60 per million tokens in 2021 to $0.07 in 2025. Epoch AI, LLM inference price trends — March 2025.
- 464% growth in Hugging Face repositories built for local inference between January and August 2026, against 21.5% growth in model repositories overall. Hugging Face, State of Open Models: Summer 2026 — August 2026.
- 3.3% how far the best closed model still leads the best open one — up from 0.5% in 2024. The gap widened. We would rather you heard that from us. Stanford HAI, 2026 AI Index Report — April 2026.
Where the shortage actually is
The constraint is not hardware. It is understanding.
More than half of workers now use these tools every week. A third have had any training in them at all. Almost every American has heard of a chatbot and fewer than one in five is confident using one. That is not a technology gap; it is a teaching gap, and it is the gap the Webspinner Foundation exists to close.
It is also why the free course has you build a model rather than watch one. A person who has trained three million parameters on their own machine understands what inference is, what a context window costs, and why a confident answer can still be wrong — and that person cannot be sold a miracle.
- 55% vs 33% share of workers using generative AI or agents weekly, against the share given any employer training in six months. Over a quarter say their employer offers none at all. The Conference Board, Skilling for AI — July 2026.
- 18% of US adults are confident using an AI chatbot, though 87% have heard of one. Half do not use chatbots at all. Pew Research Center, Americans and AI — June 2026, 5,119 adults.
- 25 vs 5 points of measurable business impact from having a clear strategy, against better tools alone. Two-thirds of frontline users get no guidance on what to do with the time AI saves them. Boston Consulting Group, AI at Work 2026 — June 2026, 11,749 workers.
The part everyone is afraid of
Augmentation is measurable. So is the damage.
The best evidence says AI lifts the least experienced most: across five thousand support agents, a generative assistant raised issues resolved per hour by 15%, with the largest gains going to the newest and lowest-skilled workers. That is augmentation, measured in a peer-reviewed journal rather than asserted in a keynote.
The same honesty requires the other half. Employment for 22-to-25-year-olds in highly exposed occupations is about 19% below trend — and the damage concentrates precisely where AI automates a task rather than assisting a person. Occupations where it complements the worker are flat or growing. The direction is a choice, not a forecast, and it is made in how the tools are built.
There is a bridge, and it has been costed. Wage insurance for displaced workers raised four-year earnings by more than $18,000 a head and paid for itself. Nobody built one for the weavers of 1811. We know how to build one now.
- +15% issues resolved per hour across 5,172 support agents given a generative assistant — with the least experienced improving most, in both speed and quality. Brynjolfsson, Li & Raymond, Quarterly Journal of Economics — May 2025.
- −19% employment for workers aged 22–25 in highly AI-exposed occupations, against same-age peers in less exposed ones. It falls where AI automates and holds where AI complements. Stanford Digital Economy Lab, Canaries in the Coal Mine? — August 2026.
- +$18,000 cumulative four-year earnings for displaced workers covered by wage insurance, a 26% gain — and the programme was self-financing even on conservative assumptions. Hyman, Kovak & Leive, NBER Working Paper 32464 — August 2024.
What we hold to
Three commitments, structural rather than stated.
Anyone can publish values. These are arranged so that breaking them would mean changing how the organisation is built, not merely changing its mind.
I
Sovereignty is a property of the thing, not a promise about it.
What a Webspinner produces belongs to them. The Foundation neither claims nor licenses patron output, and the inference architecture retains nothing once a turn ends — so there is no store of patron data to leak, sell, or be compelled to hand over.
II
The teaching is given away, not sold back.
The Foundation is a 501(c)(3). Its course is free, its writing and films are shared under Creative Commons Attribution 4.0, and its software is free of charge under the nosource agreement — which is not an OSI-approved open-source licence and is never described as one. The code is free; the service is priced in the open.
III
AI augments people. It does not replace them.
That is a design constraint here, not a reassurance. Every tool in the estate is built to put a person further ahead in their own work rather than to remove the person from it — and the Foundation’s public argument is about building the bridge that 1811 never got.
Why the name
What is a webspinner?
A webspinner is a small insect of the order Embioptera that spins silk from glands in its own front feet and builds the tunnels it lives in. We took its name on purpose: something that makes its own material and builds its own world is what we mean by AI you create and own.
Every property, live
One house, many rooms.
This list is not typed into this page. It is read live from the estate register on our own hardware, the same source johndavidmarx.com and webspinner.foundation render — so a property that graduates, gains a screenshot, or changes its description changes here the minute it changes there. Each badge says honestly where that property stands.
Join Us
Two ways in. Pick whichever suits.
Write if you have a question. Book time if you would rather talk it through. Either way a person reads it — there is no bot on the other end of these, and nothing here is sold or shared.
Send a message
For anything at all — a question, a correction, an introduction, or a note that something on one of our sites is broken.
Make an appointment with a Webspinner
Thirty minutes with someone who builds this, about your situation rather than ours. Free, no pitch, and we will say so plainly if we are not the right people.