New Alexandria.

A free gateway to open models, and a library of open data that can't burn.

ProposalNot open yet

Use open-weight models for free. If you choose to, give a conversation to the commons: personal data is removed, the result is refined into open datasets, and those datasets train the next open models, which come back to everyone, free.

The old library

Every library begins with a first scroll.

Ancient accounts say that in Alexandria, books arriving on ships in the harbour were copied for the library. For a few centuries, a large share of what people knew was gathered in one place. Then, over time, it was lost.

Today, what people learn by talking to machines is gathering again, in a few private places. Each question, correction and answer teaches the next model, and almost none of it comes back to the people who gave it.

New Alexandria is a library of the conversations people choose to share, cleaned of what is personal, copied everywhere so it can't burn, and open to anyone who wants to learn from it.

How it works

Ask. Share. Clean. Refine. Publish. Train.

One loop. What people choose to give back becomes the material for the next open models, which are served free again.

Ask, for free
Open-weight models through one gateway, at no cost, running on donated and sponsored compute.
Choose to share
Nothing you say is used unless you give that conversation to the commons. You decide, one conversation at a time.
Clean
Names, contacts, identifiers and other personal details are removed before anything is stored, and every removal is logged so it can be checked.
Refine
Shared conversations are sorted, deduplicated and labelled into datasets, each with a datasheet and a manifest of exactly what went in.
Publish
Datasets are released openly and mirrored in many places, so no single server or company can take them away.
Train
Anyone can train new open models on them, and the gateway serves the best ones back to everyone, free.

Two lanes

Private when you want it. Shared when you choose.

Proposal

Private lane

Just use it

Free access to the same models. Your conversations are not kept for datasets or training.

Contribute lane

Give it to the commons

Choose a conversation to share. It's cleaned, reviewed and, if it passes, refined into an open dataset. You can withdraw it at any point before it's published.

Honest about privacy

What cleaning can do, and what it can't.

Removal is logged: Every dataset ships with a manifest: which records went in, what was removed from each, and the consent each one carries. Anyone can check it.
Unclear means excluded: A record whose consent is unclear is left out, not "cleaned" and kept.
Other people stay out: Details about people who didn't consent, the friend you mention, the colleague in a screenshot, are excluded by default.
Not a promise of anonymity: Removing personal details is necessary, but it doesn't prove nobody could ever be recognised. That's why sharing is a choice, review comes before publishing, and you can withdraw until then.
Published means copied: Once a dataset is public, copies exist everywhere. Nothing is published until it passes review.

What it will train

Not one model. Many small ones that work together.

Intelligence that helps people is rarely one giant model. It's many parts: some quick and simple, some slow and careful. The datasets are meant for all of them.

Language, large and small

Small, fast models for the everyday basics, which run on modest machines and phones, and larger ones called in for a moment of real thinking.

Decisions and actions

Models that decide the next step: wait, respond, or hand it to someone better suited. Small, checkable choices before any big ones.

Audio and speech

Hearing when it's your turn to speak, transcribing, answering out loud, in more languages and accents than today's models handle well. Voices are only ever cloned with consent.

Images and vision

Reading a screen, noticing what changed, describing what's there. Often the right answer is "nothing to do", and that's worth learning too.

Who makes it possible

Free to use, because people give.

The gateway runs on donated and sponsored compute. The library grows from conversations people choose to share, and from the people who review, label and train on them.

Donate or sponsor compute
GPUs, hosting or credits that keep access free.
Share conversations
Through the contribute lane, one at a time.
Review and label
Check what cleaning missed, sort and label the datasets.
Train and evaluate
Train open models on the datasets and publish the results, including what didn't work.

Compute pledged: None yet.
Datasets published: None yet.
Next update: Not scheduled.

Help write the first scroll.

This is the very beginning. Tell us how you'd take part. It's a prototype form: nothing is sent or stored yet.

Follow the conversation on WhatsApp All initiatives Back to we unfold