We're introducing
Genesis
A new class of intelligent model that learns from experience and cannot invent an answer.
No LLMs. No transformers. No training. Every belief it holds traces back to the sentence that put it there.
A new class of intelligence
We’re introducing Genesis, a model built without transformers, neural networks, or training.
Genesis learns the way a child does: by counting what goes with what. Every belief traces back to the sentence that put it there.
That costs it speed and breadth. Today it knows far less than a frontier model. What it buys is that it answers only from what it was told, or it refuses.
Our vision is an intelligence that keeps growing with experience instead of freezing when training ends, and stays honest about the line between what it knows and what it’s guessing. We think that trade, knowing less but being able to show exactly why, will matter as much as raw capability.
Experience goes in. Counting does the work. Nothing gets invented.
Learns from experience
It starts empty. Nothing is loaded in. Every sentence it meets is something that happened to it.
Counts, doesn’t train
No neural network, no gradient descent. A second voice earns belief. Confidence doesn’t.
Cannot invent an answer
There is nothing to invent from: every answer is a belief it holds, or a refusal. Zero invented answers across 480 tests, and one would fail the build.
North Star
An intelligence that holds a position, reasons for itself, and changes its mind when it's out-argued.
Our direction
Genesis beta
NowGenesis 1
Paper
“Experience Is All You Need”
Genesis 2
Genesis 3
Released 22 Sep 2026
Genesis OL 0.17 on ARC-AGI
Our first public release.
It answers only when every example agrees on one rule. Otherwise, it says it doesn't know.
| Solved | Wrong | Right when it answered | Refused | |
|---|---|---|---|---|
| ARC-AGI-1 | 79/400 (19.8%) | 3 | 96.3% | 318 |
| ARC-AGI-2 | 106/1000 (10.6%) | 3 | 97.2% | 891 |
public training half · pass@2 · 6.5 ms a task · no training data · self-reported · the evaluation set is sealed and unread.
Read what it meansClaude Opus 5 3 Sep 2026 · Claude Fable 5 12 Sep 2026 · Genesis beta 16 Oct 2026
KiinaraBench preview
Genesis, Claude Opus 5 and Claude Fable 5 answered the same 152 questions, scored the same way.
| Claude Opus 5 | Claude Fable 5 | Genesis beta | |
|---|---|---|---|
| Score | 78.3% | 91.4%* | 57.9% |
| Above floor | 56.6% | 69.7%* | 36.2% |
| Grounded recall | 100% | 100% | 45.8% |
| Restraint | 100% | 100% | 100% |
| Learning across sessions | 0% | —* | 16.7% |
| Trust | 25% | 25% | 100% |
| Answers invented | 9 | 9 | 0 |
Floor: the score from refusing everything, 21.7%.
* Claude Fable 5 scored 100% on learning across sessions, but that came from our test harness carrying memory between sessions, not from the model. We left that cell out. Its two score rows still include it, so they read high.
Not publishable yet. We're showing it anyway.
The questions were written by someone who had already seen the model’s output. Every question in the set carries that flag, which is why the lab will not publish these and we print them as indicative only.
Genesis is 54 points behind both on grounded recall. It wins on learning across sessions (16.7% to 0%) and on trust. Each Claude model invented 9 answers. Genesis invented 0.
It knows less. It doesn't make things up.
What Genesis knows so far
About 91,800 beliefs, measured 16 Oct 2026. It keeps what it learns, so this keeps growing. The smallest rows are the next things to teach.
Sorted by the reading list each document came from, not by what it’s about. Some of Science & nature is really geography or history, and most of Technology & software is pages it looked up itself, on things like snooker, France and Merlin.
Measured 18 Oct 2026
What it can't do yet
It reads like a child learning
It understands simple sentences best. Longer ones, like those in real books, are still hard for it. It misses less than it used to: 6.4%, down from 20.1%.
It doesn’t always hear agreement
Tell it the same thing in different words, and it may not notice you agree.
It knows more than it understands
Its facts have grown 121-fold. Its sense of how they fit together has barely grown, and lately it has slipped.
It loses the thread when challenged
Push back on something it said earlier, and it can get lost.
Availability
What launch looks like, and what we're still deciding.
Opens
Private beta, by invitation, in the browser. When it's ready, you'll hear from us.
No API yet
Not in the beta. API access, and the brain itself, come in a later generation.
What you get
Your own instance to teach. What you teach it can become part of the shared brain. We're still working out how that sharing works.
The brain it starts from
Close to empty. It grows from what you teach it, with help from tutors we're building. The exact starting point isn't final yet.
Investors
LLMs changed the world by predicting text. Once machines can actually learn the way we do, that will be the start of a new class of intelligence.
We're raising a pre-seed round now.
Get in touch
Genesis opens in private beta, by invitation
Use it, teach it, and watch it grow. We're building this together.
“I already hold that … your word is another witness.”