Models

Approaches to modeling and decisions.

The demonstrations below illustrate concepts in statistical modeling, time series, clustering, attention, feedback, and reinforcement learning.

Each example explains a method's purpose, assumptions, and limitations. These are educational demonstrations rather than descriptions of S.I.'s internal systems.

Foundational concepts

Statistical models

Describe variation and relationships using probability distributions.

Represents
How values vary, how variables relate, and which outcomes are more or less likely under a specified model.
Assumes
The selected distribution captures the features of the observations that matter for the task.
Limitations
Changing distributions, extreme observations, or incorrect dependence assumptions can make estimates unreliable. A fitted distribution does not automatically provide calibrated uncertainty.

A two-variable Gaussian distribution with contours showing its shape and spread. As covariance changes, the orientation of the distribution and its samples changes.

Time-series models

Use patterns across time to estimate future observations.

Represents
Temporal dependence, recurring behavior, and uncertainty across a forecast horizon.
Assumes
Some relationships observed in the past remain useful during the period being forecast.
Limitations
Structural changes can weaken forecasts and make prediction intervals unreliable.

An autoregressive process with 50% and 90% prediction intervals. The readout tracks how often new observations fall within the 90% interval.

Clustering

Group observations according to similarities in their features.

Represents
Patterns of similarity that may help organize data or suggest candidate groups.
Assumes
The selected features, distance measure, and grouping method capture useful similarities.
Limitations
Results depend on scaling, initialization, and method choices. A cluster does not by itself establish a distinct real-world category or market regime.

K-means alternates between assigning points to the nearest center and moving each center to the mean of its assigned points. The process repeats until the assignments settle.

Attention mechanisms

Learn how information from different positions contributes to a representation.

Represents
Context-dependent relationships between elements of an input sequence.
Assumes
Training data contains relationships that remain useful on new observations.
Limitations
Attention can learn misleading associations. Attention weights alone do not establish causality or fully explain a model's reasoning.

A query distributes attention across positions in a short sequence. The arcs and bars show the relative weights assigned to each position.

Feedback dynamics

Describe how a system's current state influences its future behavior.

Represents
Feedback loops, equilibria, oscillation, and stability.
Assumes
The equations capture the interactions and conditions relevant to the question.
Limitations
Unmodeled delays, disturbances, and interactions can change the system's behavior or stability.

A Van der Pol oscillator illustrates a repeating cycle. Trajectories beginning inside and outside the cycle approach it over time.

Reinforcement learning

Study how actions can be selected to improve outcomes over time.

Represents
States, actions, rewards, and the future consequences of following a policy.
Assumes
The reward and learning environment adequately represent the intended task.
Limitations
A policy may optimize an incomplete objective or fail when conditions differ from those used for learning.

Value iteration updates state values in a small grid world. Once the values settle, an agent follows a policy derived from them.

Statistical methods support estimation and uncertainty analysis. LLMs can interpret information and produce structured outputs. Agents coordinate tool use and decisions.

Adaptive methods address learning over time. Inference engineering concerns the execution of trained models. Evaluation determines whether each method contributes to the task.