Introducing Thymos Myrias: Frontier human simulation at the speed of thought.
220 held-out people · three public datasets
Thymos1 Myrias2 is now the cognition model behind every simulated person on Aetherya, read directly over a weighted synthetic population. Accuracy here means one specific thing: the share of the audience the simulation said would pick an option, against the share of real people who did. Measured on people the model was never fitted to, in situations those people had never seen, across three public datasets.
Every choice is made by Myrias, our decision model, calibrated against real human decisions.
Everything below is measured on people the model never saw. On Umbra, our benchmark for cognitive simulation, Myrias is scored against free baselines under the same rules, losses included.
Fast enough to run a study a thousand times
3.7 µs per decision, about 4,512,760 decisions a second on one machine. That is what makes re-running a study a thousand times an ordinary thing to do rather than a rare one.
Against what you would otherwise use
Same people, same decisions, same held-out situations. The error is the gap between the share the model said would pick an option and the share that really did.
When it says 70%, does 70% happen?
A simulated audience is only useful if its confidence means something. Group every decision by how likely the model said it was, then count how often it actually happened.
Against a decision model
Jev, released by TypeSafe AI on 15 September 2026, is a model built to hand software a calibrated decision rather than text. We gave it the same 1,920 decisions with the same information the participant had: the two options, who the participant is, and every earlier choice and outcome on that problem. It was called once per decision, 5,760 calls in all, and nothing about Myrias was changed or re-run. The figures below are the decisions themselves; the confidence placed on the choice a person really made, averaged over every decision.
The difference is concentrated where it matters. On the decisions where a person changed their mind, Myrias’s loss was 1.36, Jev’s 1.72, and the naive rule’s 2.04. A decision model can learn that people mostly repeat themselves. A cognition model has to know when they won’t. Over the network Jev answered in 0.3 seconds a decision; Myrias takes 3.695 microseconds.
Three datasets, two of them tasks it was never built for
| Dataset | People | Decisions | Share accuracy | Decisions right | Calibration |
|---|---|---|---|---|---|
| Risky choice, the main evaluation set | 80 | 1,920 | 93.8% | 85.8% | 0.94 |
| The same kind of decision, different people | 60 | 9,000 | 96.7% | 83.5% | 1.00 |
| A two-armed bandit, a different task entirely | 80 | 8,405 | — | 77.7% | 1.04 |
The bandit has no audience figure because every game in it is unique to one person, so there is no shared situation to predict a share for. On that task the naive rule falls to 62.3% with a calibration slope of 0.23, which makes its confidence close to meaningless, while Myrias holds 77.7% at 1.04.
Where modelling the individual does not help
On a densely sampled situation a population average scores the same as modelling each person: 96.7% against 96.7% on the larger dataset. Individual modelling earns its place on sparse data and on questions about particular people, which is where the gap opens to 92.0% against 93.8%, and where an audience of averages starts getting the crosstabs wrong.
What runs in the product
2026-09-14: Persona state now has a stronger influence on decisions.
How these numbers stay honest
- Public datasets anyone can download, hash-verified. Split rules and seeds published. Our confirmatory set has never been opened.
- Every figure is measured on people the model was never fitted to, in situations those people never saw. Intervals are participant-clustered bootstraps.
- 204 proposed improvements were registered and hashed before being fitted, each getting one sealed test. The rejections are kept alongside the wins.
- Where a simpler method does as well, this page says so. Nothing here is a claim about any real person’s mind.
- 1. Thymos n.
- The architecture — the cognition system itself, not any one model built on it. Every simulated person on Aetherya runs on a model built on Thymos.
- Gk. θυμός (thȳmós) — spirit; the seat of emotion, will, and vital force. In Homeric psychology, the part of the mind that drives desire and the impulse to act.
- 2. Myrias n.
- A model built on the Thymos architecture — the newest, and the best in class it has produced so far.
- Gk. μυριάς (myriás) — ten thousand; a vast, countless number. Root of “myriad.”
Method: Thymos: A State-Conditioned Architecture for Behavioral Simulation with Language Models, Dan Andrei, Aetherya, September 2026 — manuscript, not released. Data: public behavioural datasets, scored on held-out people only. Every figure on this page is read from a generated file of published results.