Umma
Most AI systems are a model wrapped in glue code. Umma is an architecture.
One idea, run at every scale: predict, measure the error, weight by confidence. That loop is why it knows how much to trust its own conclusions, refuses to make things up, and builds the software it needs when it hits a wall.
THE SAME LOOP AT EVERY SCALE — ONE INFERENCE, A WEEK-LONG INVESTIGATION, THE WHOLE SYSTEM AT ONCE.
What do minds do that a model doesn’t? #
Every mechanism in Umma answers that one question — the loop is predictive coding, a leading account of how the brain works; the memory is the two-system model. A frontier model is just a building block. The intelligence is the architecture around it.
Four properties, rarely together #
It knows what it knows — every claim says how far to trust it.
It builds what a problem needs — and the toolkit compounds.
Reversible work runs on its own; consequential calls wait for you.
Every conclusion traces back to a source. It can show its work.
It attends hardest where it is least sure. #
Every belief carries a precision. Sharp surprises move the system; noise barely does — and every part earns the supervision its track record deserves.
A SHARP, TRUSTWORTHY SURPRISE MOVES THE SYSTEM; A NOISY ONE BARELY DOES. TRUST IS EARNED PER SOURCE — CREDIT ASSIGNMENT AS AN IMMUNE SYSTEM.
The goal is held apart from the steps. #
Steps fail, retry, and get reconsidered. The goal never moves — it lives in its own substrate, and the work is mapped back to it. Drift is impossible by design, not by discipline.
DRIFT IS A PROPERTY OF ARCHITECTURES WHERE GOAL AND STEPS SHARE ONE SUBSTRATE. UMMA’S DON’T — BY DESIGN.
Disagreement is kept, not averaged. #
Before an approach is chosen, separated voices argue it — and when a perspective is missing, Umma spawns one. The plan then waits for your approval. Structurally. Not a setting confidence can bypass.
ONE VOICE ARGUES FOR, ONE AGAINST; A MISSING PERSPECTIVE IS SPAWNED, NOT SKIPPED. NOTHING RUNS UNTIL YOU SAY SO.
Two memories, and a self that persists. #
Working memory holds what is live; long-term memory keeps what proves to matter and resurfaces it months later; when the work goes quiet, it consolidates — keeps the signal, lets the noise fade.
FAST MEMORY FOR WHAT IS LIVE; SLOW MEMORY FOR WHAT PROVES TO MATTER; CONSOLIDATION WHEN THE WORK GOES QUIET.
Underneath both: the ledger of self — five layers that make it a self, not a reference book.
THE LEDGER IS READ BEFORE ANY CONSEQUENTIAL MOVE — THE SYSTEM SEES WHO IT IS BEFORE IT WEIGHS WHAT IS ASKED.
An ungrounded claim has nowhere to live. #
An answer climbs a stack of typed layers — no free-form text between them, and reasoning stops on grounding, not on a pass count. That is the technical reason it doesn’t hallucinate. What can’t be traced gets flagged, not smoothed over.
REASONING STOPS ON GROUNDING, NOT ON A FIXED NUMBER OF PASSES. THAT IS THE TECHNICAL REASON IT DOESN’T HALLUCINATE.
Machines that build machines. #
Agents for the task. Capabilities for the gap — built, tested, quarantined, and promoted only on value delivered, never time served. Scoped copies of itself for the work. All of it under deliberate human authority, enforced in the substrate.
AGENTS FOR THE TASK · CAPABILITIES FOR THE GAP · SCOPED COPIES OF ITSELF FOR THE WORK — ALL UNDER DELIBERATE HUMAN AUTHORITY, ENFORCED IN THE SUBSTRATE. THE TOOLKIT COMPOUNDS.
Proactive, not reactive. #
Not sentience — no claim of experience. Proactivity: it briefs the work before the work asks, and steers its parts no harder than their earned trust requires. An architectural choice, not a metaphysical one.
EACH LEVEL PREDICTS WHAT THE ONE BELOW WILL NEED AND CONTRIBUTES IT AHEAD OF THE WORK. AUTHORITY RIDES THE SAME CHANNEL — INVERSELY TO EARNED TRUST, AND IT RELEASES.
Honesty, at the highest fidelity. #
A fund’s research must be defensible to a risk committee. A creative team wants the truth about what isn’t landing. Different work — one requirement, and it is the thing Umma is built to provide.
Meet the Umma that would be formed for you → The Technology Overview (PDF) → The Architecture deep-dive (PDF) →