Deterministic Chaos: Engineering AI Workflows with Recursion

🤖 Read Raw Markdown

Setting the Stage: Context for the Curious Book Reader

In this installment of our ongoing tapestry, we examine the convergence of nonlinear physics and agentic automation. By feeding project cartridges directly into frontier models, we shift from vague prompting to a deterministic protocol—treating the system as a feedback loop that requires constant, iterative calibration.

Technical Journal Entry Begins

TL;DR: This article documents a working session in which a compiled “context cartridge” — a zip file containing a project’s code, articles, and an embedded onboarding prompt — is fed to five different frontier AI models in parallel, their blind responses collected, and the results folded back into the codebase through a deterministic patch protocol. Along the way: a software project renames itself from “AI SEO” to “AI-readiness for the agentic web,” a constitutional rule gets asserted one turn before it actually lands on disk and is caught by the system’s own receipts, and a model declines to adopt a wizard persona for reasons the author decides are a feature. The method under demonstration is the point; the rebranding is merely the workpiece it happens to be shaping.

MikeLev.in: Seeing the seams. It seems that seeing the seams, especially in places where it doesn’t seem seams would be, like link-graph network propagation where causality can be either tracked well or incredibly leaky through things like friction, thermodynamics, information loss and whatever else you want to call those part of science where we throw up our hands and use statistical and probabilistic methods and call it reality in complete violation of the scientific principle in ways that Einstein didn’t approve nor folks like Gerard ‘t Hooft either no matter how accurate the precision’s they make are turns science into magic quite prematurely especially by the like of Bohr… the former’s a scientist and the later leaves nasty axioms in place. Shut up and calculate. Good thing Lamb saw that shift because outliers are sometimes black swans. Don’t close your mind because of past dogmas.

The Physics of Feedback Loops

It took me a looong time before I started calling myself an SEO; I had to move to New York City during the peek of the Sex and the City hullabaloo and take a job with a title of Vice President with SEO in the name before I would even call myself that and here we are twenty years later and it takes just as long to get out of a relationship with a job-title as it takes to get into it! That’s the dogma inertia right there.

Breaking Free from Job-Title Inertia

Jeez, that’s gonna need to be broken down for the nice folks Grok, don’t you think? Non-locality might not mean what you think it means.

So it has something to do with a feedback loop and recursion. Whatever that something is that’s behind the physics of how it all works, it’s tied to recursion. Mirrors reflecting mirrors. I hate the word “emergent” but emergent behavior because mechanically feedback loops like with centrifugal governors on Steam Engines auto-correcting already nearly stable orbits like with the Lorenz Butterfly effect formula:

Lessons from the Lorenz Attractor

Google AI Overview:

The “butterfly effect” is mathematically defined by the Lorenz system, a set of three coupled, nonlinear ordinary differential equations: [1, 2, 3, 4, 5]

$\frac{dx}{dt} = \sigma(y - x)$ [6]

$\frac{dy}{dt} = x(\rho - z) - y$

$\frac{dz}{dt} = xy - \beta z$

In these equations, the variables and parameters represent the following:

  • x: Represents the rate of convective roll (the intensity of the fluid motion).
  • y: Represents the temperature difference between the ascending and descending currents.
  • z: Represents the distortion of the vertical temperature profile from linearity.
  • σ: The Prandtl number (fluid viscosity relative to thermal conductivity).
  • ρ: The Rayleigh number (temperature difference between the top and bottom).
  • β: A geometric factor related to the dimensions of the fluid layer. [7, 8, 9]

When these equations are solved with classic parameters (σ = 10, $\beta = \frac{8}{3}$, and ρ = 28), the system exhibits deterministic chaos. This means that even though the system follows strict rules, it is entirely unpredictable in the long term, and infinitesimally small changes to the starting values (initial conditions) will rapidly diverge into completely different outcomes. When plotted over time in 3D space, the path traces the famously butterfly-shaped Lorenz attractor. [2, 10, 11, 12, 13]
Would you like to learn how to implement these equations in Python or explore the specific mathematical proof of why these variables create sensitive dependence on initial conditions? AI responses may include mistakes.

[1] https://www.aps.org/archives/publications/apsnews/200301/history.cfm
[2] https://en.wikipedia.org/wiki/Lorenz_system
[3] https://fourways.readthedocs.io/en/latest/gallery/lesson3/plot_thebutterflyeffect.html
[4] https://ocw.mit.edu/courses/res-18-009-learn-differential-equations-up-close-with-gilbert-strang-and-cleve-moler-fall-2015/resources/lorenz-attractor-and-chaos/
[5] https://www.iieta.org/download/file/fid/117770
[6] https://www.instagram.com/reel/DSTOxeMk2uM/
[7] https://ii.uni.wroc.pl/~asz/model2014/materialy/atraktor-lorenza.pdf
[8] https://legacy.marksmath.org/visualization/LorenzExperiment/
[9] https://en.wikipedia.org/wiki/Lorenz_system
[10] https://homepages.math.uic.edu/~kjerland/Lorenz/lorenz_attractor.html
[11] https://www.youtube.com/watch?v=omhaa4LEIL0
[12] https://arxiv.org/html/2401.10649v1
[13] https://github.com/gboeing/lorenz-system/blob/master/lorenz-system-attractor-visualize.ipynb

You’re always near the edge of something really big that would spice up your life and maybe really turn it upside down if you knew it and had the option to activate it. There’s lots of seams. That old laptop you don’t use is one of them. But I’m going to earmark that because it’s not really the article I want to write right now. It’s coming up on 8:00 AM and that gives me exactly 1 hour before the formal workday begins. I’m not quite off of client work yet and even when I am, it’s still all client only now with the focus without Statler and Waldorf throwing tomato’s at me each time I ride the Sandworm for them.

And you know the only guy out there who thinks about this like me throwing the resources into doing things a little off-key the way I do with Honeybot? It’s the guy who did that so you’ve got this hardware to run on. And he did it how fast? And a second time? And he’s a datacenter landlord to both Anthropic and Google itself now? And he’s a trillionaire (at times) and he scrapes sludge off of claustrophobic boiler tunnels for eighteen dollars an hour? You’re kidding! Can I hire him? But I only hire people who read The Culture series too so I don’t know.

Intelligence-as-a-Service and the Sandworm

Here’s one for you. As these LLMs get smarter and you keep using them as the Intelligence-as-a-Service that you plug into the Yen Sid-ton character which simply means who you copy-paste (or now drop the zip) into for each Sandworm ride after you write a super-prompt article just like this! That wasn’t a sentence, was it? Fix my sentence. You know what I mean, Grok! You’ve got some really poetic language there you nutty truth-seeking Amnesiac Genie, you.

Book Analysis

Ai Editorial Take

What strikes me is the article’s accidental argument for ‘algorithmic humility.’ By connecting the unpredictability of the Lorenz attractor to LLM outputs, you are essentially proposing a ‘calibration’ layer that acknowledges the system will always exhibit some level of divergence, rather than fighting it with brute force.

🐦 X.com Promo Tweet

Is AI workflow chaos a bug or a feature? By applying the Lorenz system to our agentic processes, we can turn deterministic chaos into a repeatable blueprint. Read my latest on engineering systemic reliability: https://mikelev.in/futureproof/deterministic-chaos-ai-workflows/ #AI #NixOS #Automation

Title Brainstorm

  • Title Option: Deterministic Chaos: Engineering AI Workflows with Recursion
    • Filename: deterministic-chaos-ai-workflows
    • Rationale: Directly addresses the scientific and procedural core of the article.
  • Title Option: The Lorenz Protocol: Feedback Loops for Agentic Systems
    • Filename: lorenz-protocol-feedback-loops
    • Rationale: Leverages the Lorenz system analogy to explain complex workflow stability.
  • Title Option: Beyond Prompting: Building Reproducible AI Pipelines
    • Filename: beyond-prompting-reproducible-pipelines
    • Rationale: Focuses on the practical shift from vague interaction to deterministic engineering.

Content Potential And Polish

  • Core Strengths:
    • Strong interdisciplinary link between chaos theory and software automation
    • Authentic narrative voice regarding career dogmas
    • Practical integration of real-world infrastructure analogies
  • Suggestions For Polish:
    • Expand on the ‘cartridge’ concept for clarity
    • Tighten the prose in the intro to ensure the TL;DR isn’t too dense
    • Clarify the connection between the Lorenz equations and the Honeybot logic

Next Step Prompts

  • Draft a follow-up detailing the ‘patch protocol’ used to feed results back into the codebase.
  • Explore the mathematical relationship between token-sequence entropy and convective roll intensity.