Linkfest #11

Every couple of weeks I send you a curated stack of Internet reading that stayed with me. Culture, technology, science, software. Things that made me pause and think.

This round moves from org charts hidden inside package managers to the idea of a self-replicating "Forever Book," from Cloudflare's push to price every HTTP request to the limits of AI attention. There is practical advice on writing better design docs, a strong opinion on how builders can survive in the age of AI, and a beautiful experiment in turning Gaussian splats into physical objects. It is a good snapshot of where software, money, and cognition are heading.

Enjoy! -- Christoph (CTO @ Basilicom)

NET Dollar

https://netdollar.cloudflare.com/

Cloudflare introduces NET Dollar, a fully USD-backed stablecoin aimed at automated, agent-driven commerce. The pitch is straightforward: one token equals one dollar, with transparent backing, designed for fast settlement and programmable payments.

This is Cloudflare’s play to build a native payment layer directly into the global network stack. It signals a massive bet that the future web is run by autonomous, wallet-carrying software.

Announcing the Monetization Gateway: charge for any resource behind Cloudflare via x402

https://blog.cloudflare.com/monetization-gateway/

Cloudflare's Monetization Gateway builds on that idea. Using the x402 protocol, servers can respond with "402 Payment Required," specify a price, and accept stablecoin payments inside the HTTP flow. The goal is usage-based pricing for anything behind Cloudflare: APIs, datasets, web pages, even tool calls. No custom billing stack required, and settlement happens quickly with minimal fees.

Developers are sharply divided. Some call it the holy grail of microtransactions that will finally kill the toxic, ad-driven tracking model. Others dread a dystopian "microtransactionification" of the internet, warning of runaway wallet-draining loops and metadata privacy leaks. Infrastructure dictates economics, and Cloudflare is building the ultimate toll booth.

Package Management as Org Chart

https://nesbitt.io/2026/07/10/package-management-as-org-chart.html

This essay applies Conway's Law to dependency management. The argument is simple: your build system and package strategy mirror how your company handles conflict and coordination. A monorepo with a single version policy signals strong central control and a migration team that drags everyone along. Git submodules capture polite but distant collaboration. Bazel appears when no one knows who depends on whom anymore, so the build tool enforces discipline. Nix removes "works on my machine" by design, but narrows the hiring pool. Even choices like Docker, vendoring dependencies, or using semantic-release reveal who wins arguments inside the org.

I like this because it reframes tooling debates. We often argue about package managers as if they were neutral. They are not. They encode power structures and trust models. If you are unhappy with your dependency mess, you might actually be unhappy with your org chart.

The Forever Book

https://kevinkelly.substack.com/p/the-forever-book

Kevin Kelly asks what the smallest self-reproducing seed of civilization would look like. Inspired by biology and early cybernetics, he imagines a "Forever Book" that contains the instructions to recreate itself and, eventually, the tools of civilization. He walks through versions, from a printed manual on how to make paper and ink, to a handmade book that documents its own production, to digital media that could in theory bootstrap modern tech. He also notes a hard truth: much of the knowledge needed to rebuild industry is tacit and lives in people's heads.

The piece feels both romantic and practical. In an age of LLMs that can generate reading lists, the harder problem is capturing the messy, embodied knowledge behind factories and chips. It is a useful reminder that information alone is not capability.

Advanced AI models suffer a near-total collapse on classic psychology test as cognitive demands increase

https://www.psypost.org/advanced-ai-models-suffer-a-near-total-collapse-on-classic-psychology-test-as-cognitive-demands-increase/

A new study tests GPT-4o and Claude 3.5 on a version of the Stroop task, where subjects must name the ink color of a word rather than read the word itself. With short lists, models perform well. As the list grows and cognitive load increases, performance on conflicting trials collapses. The models revert to reading the word instead of following the instruction. The researchers argue that current transformer architectures lack robust executive control.

This is a helpful counterweight to the hype cycle. Language models are strong pattern matchers, but sustained goal maintenance under conflict is still fragile. If we expect agents to handle complex, long-running tasks, this gap matters.

How to Write an Effective Software Design Document

https://refactoringenglish.com/excerpts/write-an-effective-design-doc/

This guide walks through what makes a solid design doc, based on experience at large tech companies. It covers when to write one, how much effort to invest, and which sections matter: objectives, background, goals and non-goals, diagrams, constraints, SLOs, security, privacy, open issues, and alternatives considered. A key heuristic is simple: include decisions where the cost of being wrong is high.

It is practical and grounded. In a world excited about rapid prototyping with AI, disciplined design still pays off, especially for systems that will run for years. Thinking before coding is not old fashioned. It is leverage.

Printing Gaussian Splats

https://www.patreon.com/DanyBittel/posts/printing-splats-161333338

Dany Bittel shares his experiments with 3D Gaussian Splatting, a method for creating detailed 3D models from regular photographs. Instead of just viewing these models on a screen, he works with a printing company to turn them into physical objects. They convert the data into voxel fields and 3D print them in full color. The final prints look like insects, fruit, or rocks perfectly suspended inside a solid transparent block.

This is interesting because traditional 3D scanning usually fails when it comes to fine details like hair, fuzzy surfaces, or complex lighting. Using Gaussian splatting for physical printing solves many of those known issues. Seeing this data turned into an accurate physical object points to new ways how we might handle digital archiving and product presentation down the road.

Defensibility Through Giving a Sh*t

https://www.selfonomics.com/p/defensibility-through-giving-a-shit

This essay argues that code is no longer a moat. AI can clone features quickly, and big tech can absorb generic utilities. The proposed defense for indie builders is focus and care: build for a narrow niche, embed in real workflows, cultivate community, develop a clear identity, and be opinionated. A clone can copy the interface, but not the relationships or the context behind product decisions.

The tone is combative, but the core idea is sound. If software production becomes cheap, the scarce resource shifts to trust, taste, and domain knowledge. That is uncomfortable for engineers who relied on technical difficulty as protection, but it is also freeing.