Artificial Intelligence Research

I use several browsers on my Linux system for two reasons: testing, and as a fallback when a browser update breaks access to a site I depend on. Using multiple browsers brings its own problem: keeping bookmarks in sync. Until now, whenever my Chrome bookmarks changed, I exported them to a file and then manually imported that file into Chromium and Firefox.

It finally occurred to me to have Claude (Opus 5.5 in this case) help me write code that syncs them automatically. After an evening of development, I now have a working bookmark sync application. Whenever I change my bookmarks in Chrome, the changes reach Chromium, Edge (which calls them Favorites), and Firefox within about a minute, or, if a browser is open, as soon as it’s closed.

Tools like Floccus sync both ways through a server. I wanted something simpler: Chrome is the master, and nothing leaves my machine. I provided the direction, additional code review in VSCode, and did the testing, which caught bugs and eight edge cases. Claude Opus 5.5 wrote the code.

Chaotic dynamics · sensitive dependence on initial conditions

The double pendulum is about the simplest machine that produces chaos. It's two rigid arms and two pivots, governed entirely by Newton's laws with no randomness, yet its long-term motion can't be predicted in practice. A difference in the starting position too small to measure grows exponentially, so two nearly identical releases are doing completely different things within seconds. At small swing angles it behaves like an ordinary, predictable oscillator.

Add energy and the same device becomes chaotic. That makes it the clearest tabletop demonstration that deterministic doesn't mean predictable. The same principle limits weather forecasts to about two weeks, complicates long-range orbital predictions, and shapes how engineers control robots, cranes, and other multi-link systems.

Where it earns its keep beyond the lecture hall

Teaching Lagrangian mechanics. The motion in this demo comes from two coupled equations, one per arm. Deriving them with Newtonian force balances is miserable because of the constraint forces in the rods. With the Lagrangian (kinetic minus potential energy, written in the two arm angles θ₁ and θ₂, which are the same two angles the start sliders set) it takes about half a page. It's the standard example that convinces physics students the formalism is worth learning.

Testing numerical integrators. The energy drift readout is the scorecard. The real system has no friction, so total energy should never change, and any drift is solver error. Because chaos amplifies errors exponentially, the double pendulum quickly exposes a sloppy method. Simple Euler integration visibly gains energy and the pendulum spins up unphysically. This demo uses fourth-order Runge-Kutta at a 2 ms step, which holds drift to thousandths of a percent.

Control theory and robotics. Set both sliders to 180° and the pendulum balances straight up for a moment, then topples. Holding it there with a motor at only one joint is the Acrobot and Pendubot problem, which is the canonical benchmark for underactuated control (fewer actuators than degrees of freedom). The same math underlies legged robots and gymnastic swing-up maneuvers.

Biomechanics and engineering. Watch the velocity chart during a whip. The upper arm slows, and the lower arm spikes as energy transfers down the chain. A golf swing, a baseball pitch, a tennis serve, and a trebuchet all use that same transfer on purpose. In crane loads and multi-link manipulators the structure is the same, but the whip is a hazard rather than a goal.

The historical thread. Poincaré found this kind of behavior in the three-body problem around 1890, and Lorenz rediscovered it in weather equations in 1963. Both discoveries lived in equations and printouts. The double pendulum became the tabletop ambassador for nonlinear dynamics because the irregular, never-repeating trail makes chaos visible to anyone.

The research presented here was produced with Claude Opus 5.5 over multiple interactive sessions. While I believe the results are reasonably accurate, they may contain logical or other errors. Assumptions and confidence levels are noted throughout. Use or share at your own risk. 

Updates to this site will be targeted monthly. Source code and HTML versions of the papers are on GitHub.https://github.com/ne0busin3ss

Hello, I’m David Philip Raymond, a retired principal project manager (PMP, PMI-ACP). I created this site to share my research on frontier and open-weight AI models, viewed through the lens I used for decades: what does it cost, how well does it perform, and is it worth it? Here you’ll find explorations of how AI models are built, analysis of their economics, and practical AI-assisted coding projects built primarily with Claude Opus.

I started out as a carpenter, then trained as a draftsman — so why not add a bachelor’s degree? Three decades after finishing secondary school, I earned a bachelor’s degree in Business/Computer Information Systems from the University of Houston-Downtown, which I completed over six years of part-time study while working full-time as an airline project manager. I graduated in the top 10% of my class with a 3.78 GPA, earned a software scholarship in my final semester, and was inducted into Beta Gamma Sigma, the international honor society for business schools. During those same years I earned my PMP, and in my 60s, my PMI-ACP.

Meanwhile, the job kept me moving: Taipei, Amsterdam, London, Paris, Zurich, Caracas, and Helsinki. I’d tell you it was just a matter of showing up. Others would say it was showing up for all of it at once.

None of it would have been possible without the support of the love of my life, Elizabeth A. Raymond.