Weekly Review — Aug 3–9, 2026

TL;DR: Three posts this week, one quiet throughline: the disciplined, bounded move keeps beating the maximalist one. Gall’s Law says you compose a complex system from small working bricks instead of one-shotting it. Self-consistency turns out to be a decades-old amplification theorem whose fine print decides when majority voting helps. And a read-only AI assistant with no write access was quietly the most useful one I ran all month.

This was a week about doing less, on purpose. Not less ambition, less brute force. Two of the three posts land in the “Laws & Disorder” and “Complexity Theory Meets AI” series, and the third is a confession about my own agent setup. Read together they make the same argument from three directions: the constrained version of a thing is often the one that actually works.

This Week on the Blog

Laws & Disorder #4 — Gall’s Law: Great Systems Are Built From Small Bricks, Not Big Plans

Gall’s Law says a complex system that works is always found to have evolved from a simple one, and the compositional reading is the one worth internalizing: you build the complex thing by assembling small, well-scoped parts, not by designing the whole up front. Unix and the core AWS primitives are the standing illustrations, tiny pieces snapped together into behavior nobody designed as a whole. The 2026 twist is that AI makes it feel cheap to one-shot an entire system in an afternoon, so the discipline is to aim it at building good bricks and composing them rather than generating an untested blob.

Self-Consistency Is Just Probability Amplification

Sampling a model several times and taking the majority answer, the “self-consistency” trick, is, under the right assumptions, the probability-amplification move taught in any randomized-algorithms course, rediscovered in a new outfit. That framing hands you the fine print for free: in the clean two-answer version, each sample has to be right more often than a coin flip and the errors have to be close to independent — with many candidate answers the right one only has to beat its strongest rival, and correlated samples can turn majority voting into confidently repeating a shared wrong assumption. Because every sample is a full generation, each turn of the dial multiplies generated tokens — and with them the bill, plus wall-clock latency unless you sample in parallel — while the accuracy gains typically flatten. A dial, not a default to leave cranked.

The AI That Only Watches (Read-Only AI Is Underrated)

The market is selling more permissions: more tools, more autonomy, more write access. This post went the other way. A read-only assistant pointed at your “cognitive exhaust”, the sent mail, open tabs, and half-finished notes that record where intention and action drifted apart, cannot send the wrong message or take an injected action itself, because it holds no side-effecting tools — injected content can still skew what it reports, so the constraint removes the write path, not the influence path. Its whole job is to notice. That constraint is not a safety compromise you tolerate; it is what, for me, makes the assistant trustworthy enough to point at your entire life.

The Thread This Week

Each post takes something the field treats as “more is better”, more upfront design, more samples, more permissions, and shows the bounded version winning. Compose from small validated bricks instead of a big-bang design. Know where the extra samples stop earning their tokens. Give the assistant less to do so you can trust it with more of your life. The common idea is restraint as engineering, not restraint as caution.

Further Reading

This edition was assembled ahead of the week, so the usual outside-reading list is lighter than normal. In the meantime, the primary sources behind this week’s posts are all public and worth reading in full:

• Systemantics / The Systems Bible — John Gall — the 1975 origin of Gall’s Law, thirty publishers passed on it and the law outlived all of them.
• Self-Consistency Improves Chain of Thought Reasoning — Wang et al. — the 2022 paper, with the reported accuracy jumps that turn out to be amplification in disguise.
• Cognitive Exhaust Fumes: Read-Only AI Is Underrated — Šimon Podhajský — the short AI Engineer talk that prompted the read-only post.

Until Next Sunday

Three posts, one lesson: the constrained option is often the finished one, not the training-wheels one. If you had to pick one place this week to swap brute force for a bit of discipline, where would it be?

This is the Weekly Review — it also goes out as Sunday’s newsletter.

About the Author

Stefan Christoph is a Principal Solutions Architect at AWS, focused on agentic AI, media & entertainment, and helping builders move from demo to production. He writes about AI architecture, developer productivity, and the future of software.

This is a personal blog. Opinions expressed here are my own and do not represent the views or positions of my employer.

Learn more →

• WeeklyReview
• AI
• Agents