NON-TRANSFORMER
AI RESEARCH LAB
Most AI finds patterns.
We build models that find causes.
Most AI finds patterns.
We build models that find causes.
Causis Research is building a new AI architecture from first principles, a from-scratch alternative to the transformer.
It learns continuously instead of freezing after training, reasons about cause rather than correlation, and runs without a GPU. Causal reasoning is the first thing it does. It is not the limit of what it is.
Causis Research is building a new AI architecture from first principles, a from-scratch alternative to the transformer.
It learns continuously instead of freezing after training, reasons about cause rather than correlation, and runs without a GPU. Causal reasoning is the first thing it does. It is not the limit of what it is.
Most AI is pattern matching at scale. Given enough historical data, a model learns what tends to follow what. This works well until conditions change. When a market regime shifts, when a biological system behaves unexpectedly, or when a physical environment evolves, the patterns break. And so does the model.
Most AI is pattern matching at scale. Given enough historical data, a model learns what tends to follow what. This works well until conditions change. When a market regime shifts, when a biological system behaves unexpectedly, or when a physical environment evolves, the patterns break. And so does the model.
Most AI is pattern matching at scale. Given enough historical data, a model learns what tends to follow what. This works well until conditions change. When a market regime shifts, when a biological system behaves unexpectedly, or when a physical environment evolves, the patterns break. And so does the model.
Most AI is pattern matching at scale. Given enough historical data, a model learns what tends to follow what. This works well until conditions change. When a market regime shifts, when a biological system behaves unexpectedly, or when a physical environment evolves, the patterns break. And so does the model.
We took a different approach, from the ground up. Not a transformer, but a new architecture built from first principles. Instead of memorising what tends to happen, our models build an understanding of why things happen. They learn the structural relationships between variables and can reason about interventions: what would happen if this changed? That kind of reasoning survives regime shifts, because causes do not disappear when surface patterns change. The other gap in current AI is learning. A model trained today is exactly as capable as it was on day one. Every conversation is forgotten. Every market observation is lost. We are building models that genuinely update from live interactions.
We took a different approach, from the ground up. Not a transformer, but a new architecture built from first principles. Instead of memorising what tends to happen, our models build an understanding of why things happen. They learn the structural relationships between variables and can reason about interventions: what would happen if this changed? That kind of reasoning survives regime shifts, because causes do not disappear when surface patterns change. The other gap in current AI is learning. A model trained today is exactly as capable as it was on day one. Every conversation is forgotten. Every market observation is lost. We are building models that genuinely update from live interactions.
We took a different approach, from the ground up. Not a transformer, but a new architecture built from first principles. Instead of memorising what tends to happen, our models build an understanding of why things happen. They learn the structural relationships between variables and can reason about interventions: what would happen if this changed? That kind of reasoning survives regime shifts, because causes do not disappear when surface patterns change. The other gap in current AI is learning. A model trained today is exactly as capable as it was on day one. Every conversation is forgotten. Every market observation is lost. We are building models that genuinely update from live interactions.
We took a different approach, from the ground up. Not a transformer, but a new architecture built from first principles. Instead of memorising what tends to happen, our models build an understanding of why things happen. They learn the structural relationships between variables and can reason about interventions: what would happen if this changed? That kind of reasoning survives regime shifts, because causes do not disappear when surface patterns change. The other gap in current AI is learning. A model trained today is exactly as capable as it was on day one. Every conversation is forgotten. Every market observation is lost. We are building models that genuinely update from live interactions.
The other gap in current AI is learning. A model trained today is exactly as capable as it was on day one. Every conversation is forgotten, every observation lost. Our architecture learns continuously, storing knowledge as a log rather than freezing it into fixed weights, so the model keeps getting better from live interaction. And because two models built this way can merge into one, exactly and losslessly with no retraining, knowledge can be combined in a way transformers structurally cannot.
The other gap in current AI is learning. A model trained today is exactly as capable as it was on day one. Every conversation is forgotten, every observation lost. Our architecture learns continuously, storing knowledge as a log rather than freezing it into fixed weights, so the model keeps getting better from live interaction. And because two models built this way can merge into one, exactly and losslessly with no retraining, knowledge can be combined in a way transformers structurally cannot.
The other gap in current AI is learning. A model trained today is exactly as capable as it was on day one. Every conversation is forgotten, every observation lost. Our architecture learns continuously, storing knowledge as a log rather than freezing it into fixed weights, so the model keeps getting better from live interaction. And because two models built this way can merge into one, exactly and losslessly with no retraining, knowledge can be combined in a way transformers structurally cannot.
The other gap in current AI is learning. A model trained today is exactly as capable as it was on day one. Every conversation is forgotten, every observation lost. Our architecture learns continuously, storing knowledge as a log rather than freezing it into fixed weights, so the model keeps getting better from live interaction. And because two models built this way can merge into one, exactly and losslessly with no retraining, knowledge can be combined in a way transformers structurally cannot.
What we build
Causal world models
Causal world models
Models that represent how variables relate causally, not statistically. Built to reason about interventions and counterfactuals, not just predict the next data point.
Continuous learning
Continuous learning
A novel architecture that learns from live data permanently. Knowledge is stored as a log, not frozen into weights, so the model keeps learning from every interaction.
Exact model merge
Exact model merge
Two models trained separately on different data combine into one, exactly and losslessly, no retraining. Impossible for transformers, and the basis for collaborative training.
Agnostic reasoning
Agnostic reasoning
Our models transfer causal understanding across domains. What a model learns in one field improves performance in another. A property of the architecture, not a feature.
Where causal AI matters most
Where causal AI matters most
Financial markets
Markets are causal systems. Price moves have causes. Correlations break. Standard AI fails exactly when it matters most. Causal models trace the actual drivers of market moves and produce insights that hold up when conditions change.
Scientific research
Biology, physics, and chemistry are governed by causal mechanisms. Finding those mechanisms in data is what causal AI is designed to do. Our models have demonstrated strong early results in protein and molecular domains.
Decision intelligence
Most AI tells you what is likely to happen. Causal AI tells you what to do about it. Any organisation making decisions under uncertainty benefits from models that reason about interventions.
Regulated industries
Regulators increasingly require AI decisions to be explainable. Causal models produce auditable reasoning chains by design. Every output comes with a traceable path from cause to conclusion.
Early results
Market data
Trained on live market data from a cold start, with no pre-training. The model discovered the structural rules of the market from the data alone, learning continuously as new data arrived.
Cross-domain transfer
Training on biology and gameplay together produced better gameplay performance than training on gameplay alone. What the model learns in one domain improves the others, with no catastrophic interference.
Compute efficiency
Specialized deployments run on consumer hardware. No dependency on large-scale cloud infrastructure at the model level.
Language and science
Adding a language domain improved physics performance by 10.7% across all seeds. An unexpected cross-domain result with no published equivalent in the research literature, validated across 11 distinct domains.
Market data
Trained on live market data from a cold start, with no pre-training. The model discovered the structural rules of the market from the data alone, learning continuously as new data arrived.
Cross-domain transfer
Training on biology and gameplay together produced better gameplay performance than training on gameplay alone. What the model learns in one domain improves the others, with no catastrophic interference.
Compute efficiency
Specialized deployments run on consumer hardware. No dependency on large-scale cloud infrastructure at the model level.
Language and science
Adding a language domain improved physics performance by 10.7% across all seeds. An unexpected cross-domain result with no published equivalent in the research literature, validated across 11 distinct domains.
Market data
Trained on live market data from a cold start, with no pre-training. The model discovered the structural rules of the market from the data alone, learning continuously as new data arrived.
Cross-domain transfer
Training on biology and gameplay together produced better gameplay performance than training on gameplay alone. What the model learns in one domain improves the others, with no catastrophic interference.
Compute efficiency
Specialized deployments run on consumer hardware. No dependency on large-scale cloud infrastructure at the model level.
Language and science
Adding a language domain improved physics performance by 10.7% across all seeds. An unexpected cross-domain result with no published equivalent in the research literature, validated across 11 distinct domains.
Market data
Trained on live market data from a cold start, with no pre-training. The model discovered the structural rules of the market from the data alone, learning continuously as new data arrived.
Cross-domain transfer
Training on biology and gameplay together produced better gameplay performance than training on gameplay alone. What the model learns in one domain improves the others, with no catastrophic interference.
Compute efficiency
Specialized deployments run on consumer hardware. No dependency on large-scale cloud infrastructure at the model level.
Language and science
Adding a language domain improved physics performance by 10.7% across all seeds. An unexpected cross-domain result with no published equivalent in the research literature, validated across 11 distinct domains.
Recent research
RESEARCH
Why we are building a different kind of AI
Most AI is, at its core, autocomplete. Given enough text or data, a model learns to predict what comes next. This is genuinely impressive at scale. But it is not reasoning. And it is not what we are building at Causis Research.
Causis Research
PREVIOUS
NEXT
RESEARCH
Why we are building a different kind of AI
Most AI is, at its core, autocomplete. Given enough text or data, a model learns to predict what comes next. This is genuinely impressive at scale. But it is not reasoning. And it is not what we are building at Causis Research.
Causis Research
PREVIOUS
NEXT
RESEARCH
Why we are building a different kind of AI
Most AI is, at its core, autocomplete. Given enough text or data, a model learns to predict what comes next. This is genuinely impressive at scale. But it is not reasoning. And it is not what we are building at Causis Research.
Causis Research
PREVIOUS
NEXT
RESEARCH
Why we are building a different kind of AI
Most AI is, at its core, autocomplete. Given enough text or data, a model learns to predict what comes next. This is genuinely impressive at scale. But it is not reasoning. And it is not what we are building at Causis Research.
Causis Research
PREVIOUS
NEXT
Work with us
We are selectively partnering with organisations where causal reasoning matters.
hello@causisresearch.com
causis
Causis Research is an AI lab building a new architecture from first principles, an alternative to the transformer that reasons about cause and learns continuously.
causis research
causis
Causis Research is an AI lab building a new architecture from first principles, an alternative to the transformer that reasons about cause and learns continuously.
causis research
causis
Causis Research is an AI lab building a new architecture from first principles, an alternative to the transformer that reasons about cause and learns continuously.
causis research
causis
Causis Research is an AI lab building a new architecture from first principles, an alternative to the transformer that reasons about cause and learns continuously.