AdamAI is an open, free AI grown in public. Anyone can run it at home. We're building it so people can lend their computers to help it learn, starting with a single household, then invited volunteers, then everyone, with every step of its growth visible.
Updated automatically every few minutes from the training runs. Loss is how wrong the model's next-word guesses are; lower is better.
People call the inside of a trained AI a black box. AdamAI exists to open one. I'm building a family of language models from scratch on my channel, writing every piece by hand, so we can stop at any point and see what's going on inside.
The first full model, Genesis, is the size of GPT-2 small: about 114 million parameters, reading about ten billion tokens of educational web pages. We save copies of its brain along the way and quiz it every 250 steps. Once we know when an ability shows up, we can go looking for where it lives.
Others have built parts of this. Folding@home showed that volunteers will lend their computers to science, and Prime Intellect and Nous Research have trained models over the internet. What AdamAI puts together:
Each step, it learns from about a quarter of a million tokens. These are the moments the probes caught.
Loss about 11, about what a blind guess scores. Asked the capital of France, its top answer was "cess". Paris sat around 26,000th.
Its favorite answer to almost any question is "the".
It now picks "The keys to the cabinet are" over "is", and a previous-token head appears in layer 7. About two hours in.
An induction head forms in layer 8, the copying circuit Anthropic researchers described in 2022. Its first fact reaches number one: water is made of hydrogen and oxygen.
Tokyo becomes its top answer for the capital of Japan.
Paris reaches number one. Half of its ten fact questions now land at the top.
It beats OpenAI's GPT-2 small on its own held-out test text, 3.2725 to 3.274. That's home turf. The neutral test, HellaSwag, comes at the end of the run.
After layer 10 it gives Tokyo about 93%. By the output it has talked itself down to about 23%. We don't know why yet, and the saved snapshots should tell us.
Final scores when the run ends, including HellaSwag, where GPT-2 small scores about 29%.
Building AdamAI led to five ideas, written up in a concept paper. The building blocks are published research. Putting them together is new as far as we know, and each idea comes with a test that could prove it wrong.
Measurements of its insides choose what it studies next. If a probe shows it knows a fact in the middle layers and loses it by the output, the next lessons target that gap. The change stays only if the gap closes and nothing else gets worse.
One brain that never stops learning. New material goes into the trunk with some old material mixed back in so it doesn't forget, and each release is a finished copy branched off it.
Every stage of its growth is measured and kept for its whole life, through each update, merge and growth step. That lets us ask whether a circuit survives an update or moves when the model gets bigger.
Several ordinary computers in one home train the same brain. Each trains its own copy for a round, and a parent computer merges the changes.
It learns from its own answers only when an outside checker confirms them, inside a loop that a person approves and can always roll back.
Final scores, the last brain snapshots, and a graduation episode.
A larger open model writes 60,000 example answers, and Genesis learns to answer questions from them.
340 million parameters, three times Genesis, trained from scratch on 25 billion tokens on one rented H100. Estimated $160 to $220.
Help pay for it →A desktop app for Windows, Mac and Linux. Chats stay on your computer, with no account, and it works offline. It shows what the model knows and what the last update taught it, with an opt-in setting to donate computer time.
Studies on the saved snapshots: how the copying circuit forms, where facts like Tokyo are stored, which heads carry grammar, and sparse autoencoders that split its activity into readable features.
The plan is for people to lend their computers to help AdamAI learn. Strangers can't volunteer yet, and that's by design. It opens up in three phases.
Built and tested. Helper software lends a PC's graphics card to training inside one home, over the local network, with no inbound ports opened.
Friends, family and early supporters. Every update gets checked before it's merged: the parent re-runs the samples it sent, tests on text no helper sees, and compares helpers against each other.
Anyone, once update verification, a signed app and the legal basics are in place.
Invitations aren't open yet. The series will say when they are.
The next model, Exodus, trains on one rented H100 for an estimated $160 to $220. You can also give money toward cloud runs while the volunteer side gets built. Every dollar spent shows up in a public ledger.
Every step of AdamAI's growth goes into a video series on YouTube, with lab notes for each episode.
From 65 characters and a lookup table that writes gibberish to a small GPT that writes in the shape of a play.
Lab notes from a third of the way through training AdamAI Genesis, a 114-million-parameter model we're growing from scratch.
COBURNICUS