RESEARCH

Can AI anticipate what happens next?

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.

Back to research
01

The question

With incomplete information, can a model use what it knows to anticipate the outcomes of different choices?

02

What we observed

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

03

What we cannot claim yet

The tests used controlled simulations. They do not show that the model understands cause and effect in the real world or can make reliable judgments across different situations.

EXP-074 v0.1.5 · 2026-10-15

Can more history lead to better decisions?

This technical report is based on the original experiment records; it is not a published paper. It explains the research question, method, results, and scope.

STUDY ABSTRACT

This study asks whether updating the relationship between an action and what followed helps improve the next prediction and choice. We compared update methods in a simulated environment with explicit rules and a finite set of candidate worlds. Some conditions showed a local gain; in others, predictions changed without improving the final choice or task utility.

01 / STUDY DESIGN

How did we test it?

We used a finite two-model controlled hidden-state sequence with three authored task families and three nested seeds per family. Four conditions were compared: action-conditioned updating, no updating, action-independent updating, and exact retrieval from the same history. We scored fixed-action forecasts separately from two-step decision utility, so a changed forecast would not automatically count as a better decision.

02 / RESULTS

What do the numbers show?

F1 modeled a temporary action effect: mean net utility was 0.4500 with action-conditioned updates, 0.2083 with no update or action-independent updates, and 0.4500 with exact retrieval from the same history. F2 had no action effect; all four conditions scored 0.2917. F3 was outside the candidate bank and neither action was good; all conditions scored 0.2083. In F2 and F3, some forecast scores changed without increasing decision utility.

03 / INTERPRETATION

What can we conclude?

The narrow finding is that, in the authored F1 setting, retaining the action-outcome relationship helped select a higher-utility action. Exact retrieval of the same history produced the same decisions, so persistent updating did not outperform comparable history retrieval. F2 and F3 also show that a better forecast metric does not necessarily lead to a better action. This is a finite-rule baseline comparison, not evidence of a learned world model or causal understanding.

04 / NEXT STEPS

What should be tested next?

The next step is to independently design new settings, include model-mismatch cases where decisions can genuinely differ, and use blinded scoring. If action-conditioned updates do not consistently beat no-update and exact-retrieval controls with the same information, or create false utility gains where no update is warranted, the local mechanism claim should be rejected. More seeds from the same three families would not substitute for independent settings.

Key results

Values are descriptive means within each authored family. Seeds and scored rows are not independent samples and do not support population-level inference.

RESULTS AT A GLANCE

Task utility across simulated scenarios

In F1, action-conditioned updates matched exact retrieval. The four conditions also had equal scores in F2 and F3.

F1 / Temporary action effect

Action-conditioned update
0.4500
No update
0.2083
Action-independent update
0.2083
Exact retrieval from the same history
0.4500

Exact retrieval reached the same utility; these results do not show that persistent updating outperformed retrieval.

F2 / No action effect

All four conditions (equal)
0.2917

Action-conditioned update, no update, action-independent update, and exact retrieval all scored 0.2917.

F3 / Outside the candidate range

All four conditions (equal)
0.2083

All four conditions scored 0.2083, so this test did not distinguish their decision effects.

Test conditionObserved valueInterpretation
F1 / Temporary action effect0.4500 vs 0.2083Action-conditioned update vs no update; exact history retrieval also scored 0.4500.
F2 / No action effect0.2917 (same across conditions)Some forecast scores changed; decision utility did not.
F3 / Outside the candidate bank0.2083 (same across conditions)This family did not create a model mismatch that distinguished decisions.

Source record: RESULT-EXP074-v0.1.5-EXPLORATORY.md; SHA-256 c0a5990b303e8071f1fc25a4120920c7a665fe9bf75ecb2604f974a7007e3b80. Research note based on the experiment record; not a peer-reviewed paper.

FURTHER READING / EXTERNAL PAPER
Mastering Diverse Domains through World Models ↗

This external paper introduces DreamerV3, which learns an environment model and improves actions by imagining future scenarios. It reports one configuration across more than 150 tasks. That broader world-model research is separate from our controlled test of a finite rule set and is not the source of the results on this page.

What the evidence supports

The findings apply only to the tests completed so far.

The tests used controlled simulations. They do not show that the model understands cause and effect in the real world or can make reliable judgments across different situations.

Explore other research topics

GENESIS / GITHUB

Coming soon

The GitHub entry will open when our products launch.