About me and this site

Epistemic Standards

Deep originality is secondary to absolute clarity of synthesis. Every line of code I ship or idea I publish must aim to satisfy four directives:

Rigorous Understanding: Validating the “why” behind an algorithm or idea by writing exhaustive, first-principles explanations.
Tangible Output: Reproducible results, algorithms, tools, visualizations, theory, or clean implementations.
Refined Intuition: Distilling complex feature spaces into enlightening and persuasive insights.
Overarching Themes: Linking diverse topics to generate new tools, techniques, and understanding.

The CTRLS Framework

Across both commercial projects and foundational research, I categorize my focus into four distinct activity modes and apply them across all scales of work: Creating, Thinking, Researching, and Learning at all Scales (CTRLS).

Every project moves through three distinct phases, each requiring a specific cognitive toolkit:

Intention: Discernment, precise articulation, and structural planning.
Action: Technical focus, self-regulation, and the ability to course-correct through imperfect data.
Completion: Fortitude and resourcefulness to execute the final, often disproportionate, last mile.

The naming of these meta ideas (modes, phases, cognitive toolkit) allows me to be much more self-aware. Thinking of them as dimensions, and progress as a path within that space, keeps me empowered and motivated: there is always something I can do to make progress.

Principles and Tastes

Of particular application to the field of AI, I believe in the following general principles and approaches:

Use probability. Use statistical ideas. Be principled. Leverage centuries of research. But also: understand the assumptions behind many results created before the availability of powerful computers and software. Most data isn’t linear, or drawn from independent identical distributions, or normal, etc. Even when these assumptions don’t hold, the results can still work or aid understanding.
Appreciate the “curse of dimensionality”. High dimensional space is basically void of data; distance metrics in feature space can be unintuitive or misleading, or say more about the metric than the data; results from algorithms should not change if measurement units are altered; the volume of hyper-spheres contain a vanishing fraction of bounding hyper-cube volume; etc.
Models such as Deep Neural Networks and Kernel-based density estimators are general-purpose parameterised functions which make minimal assumptions about the data. Implicitly they do make assumptions of course, but the hope is that these structural constraints are either harmless or desired. Any inspiration from or analogy with brain function is of minor historical interest only.
Publish reproducible results. I’ve spent the last 25 years using Python for commercial software development, so that will continue to be my go-to tool. Unsurprisingly, Jupyter/Marimo notebooks will feature too.
Make honest assessments using both real datasets (publically available) and artificial pedagogical datasets created with specific and known characteristics.