@MaximeRivest: At first glance: > Structural Equation Modeling (SEM/Path Analysis) > Neural Ordinary Differential Equations (Neural OD…

At first glance:

Structural Equation Modeling (SEM/Path Analysis) Neural Ordinary Differential Equations (Neural ODEs) AI Programs (like DSPy)

belong to completely different worlds.

However, they share a fundamental architecture that makes them each very powerful ways for us (humans) to flirt with unknowns / probabilistic / undetermined.

All three are declarative frameworks for defining and optimizing computational graphs.

1. Separation of Structure and Optimization:

In all three, you (the human) define the structure of the problem or the flow of information. You set up the “skeleton.” Then, an underlying engine (an optimizer, a solver, or a teleprompter) figures out the specific weights, parameters, or prompts needed to make your skeleton match reality or achieve a goal.

2. Modular Composition:

They all treat smaller operations as building blocks that can be chained together. An output from one node/module becomes the input to the next.

3. Data-Driven Alignment:

None of these models are static. Whether it’s minimizing covariance differences in SEM, minimizing L2 loss against time-series data in Neural ODEs, or using an optimizer like GEPA to improve LLM accuracy in DSPy, they all update their internal states based on empirical data.

I firmly believe that as long as AI agents remains built on structure-less and purely data-driven systems LLMs (its mostly tokens in and tokens out), rather than being built on structured flows of information with a modular skeleton where some parts can be optimized using empirical data and others are explicitly designed (deterministic, hierarchical control flow), AI agents will be impossible to fully trust. This will prevent those coding Agents from reaching the many, many nines of reliability required for full automation. Full automation is completely different from automation that only kind of works.

if my coding agents was my employee, I would fire them. I don’t care that you are (sometimes) a brilliant coder.

If you cannot be trusted and you don’t learn from your mistake, you are a liability.