AI RESEARCH

Start with an everyday problem. Work toward an answer.

An AI saying it has learned something does not mean it can do it. We break the questions down and check them one by one: Does remembered experience help? Are predictions about what happens next dependable? Did a change actually improve the result? We record disappointing outcomes too, because knowing what failed is a useful place to begin again.

A miniature research team of bots compares information and test results
Research scene concept rendering. Actual study results are presented on each research page.CONCEPT VISUAL

Three clear questions.
Three separate paths of investigation.

Each study has its own test scope. These findings describe only the tests completed so far; they do not show general AI capability or a feature in any product.

01 / SMALL AI MODELS

Small AI models

We study how smaller AI models learn with limited memory and computing power, and whether what they learn can help with new tasks.

What we observed

In a limited simulation task, the model showed signs of adapting to new situations, but it also lost some existing abilities. A later on-device adjustment trial still did not outperform the unmodified model overall.

RESULT SNAPSHOT

The memory tests did not show a consistent improvement. The rule-card test was invalid; its 50% fallback baseline is not evidence that the cards helped.

See methods and limitations ↗
02 / WORLD MODELS

World models

We study whether AI can use the information it has to predict what different choices might lead to, then check whether those predictions help people decide.

What we observed

In a limited test with clear rules, predictions across several steps became more accurate as more observations were added.

RESULT SNAPSHOT

In the F1 simulation, action updates and exact retrieval both reached 0.450 task utility. The other two scenario families showed no decision-utility gain.

See methods and limitations ↗
03 / EVALUATION & IMPROVEMENT

Evaluation & Improvement

We compare a model before and after adjustments to see whether gains carry over to unfamiliar tasks, rather than only appearing in the original test.

What we observed

Early simulations outperformed a simple baseline, but still fell short of the improvement threshold when compared with a stricter control.

RESULT SNAPSHOT

Across 48 descendant test worlds, feedback-based selection averaged 0.756 versus 0.474 for a fixed curriculum. This measures a software selector, not a model improving itself.

See methods and limitations ↗

We say only what the tests show,
and no more.

We keep test conditions, results, and open questions separate. A gain in a limited test supports a claim about that test only. More validation is needed before drawing conclusions about new tasks, real-world settings, or products.

GENESIS / GITHUB

Coming soon

The GitHub entry will open when our products launch.