Research at Nova Intelligence
Research toward intelligence that understands the world.
Nova Intelligence studies world models, continuity, model research, and evaluation. Each publication separates the question, recorded evidence, and proof still required before a product claim.
The bet · Research direction
The Large World Model
Every frontier lab is racing to predict the next token better. We are asking the larger question: what would it take for a model to predict the world? This is the thesis the whole program is built on — stated plainly, argued carefully, and put where it can be broken.
Read the paperThe thesis, in three lines
- 01The world is not made of tokens.
- 02Predict the next state of the world, not the next word.
- 03Say how you will measure it — before you run it.
Publications
Five programs behind one bet.
Each publication names its research question, evidence state, and what would count as stronger proof. Together they trace the world-model bet, continuity research, model evaluation, and release qualification.
DWM-1: a pre-registered proofThe hypothesis, measurement, and failure line are fixed before the run.
LongMind: a memory research directionA high-level research direction on continuity over longer work. It is not a released product capability.
The Nova i1 research programResearch into model behavior, evaluation, and release qualification.
Evaluating work, not chatA program for measuring completed work instead of presentation quality.A research claim is worth only as much as the method behind it. So the method comes first — stated before the work, labelled by its evidence, and kept on the record when it misses.
State the bet before you run it
DWM-1 was registered before it ran: the hypothesis, the measurement, and the conditions that would count as failure, all on the record in advance. If a result is worth claiming, the method behind it was public first.
Every claim carries an evidence label
Each publication states what its evidence supports: a question, design, implementation, recorded replay, or qualified result. An internal artifact does not become released product behavior by appearing on a research page.
We count the misses alongside the hits — in public.
The questions are meant to be broken
The program investigates ambitious questions about prediction, cognition, continuity, and useful work. Negative results and unresolved activation requirements remain visible alongside the proposals.
Read next