The Graveyard of Good Ideas

**TL;DR:** Good ideas sometimes fail due to insufficient tooling or community support, but advances in technology can later resurrect them, as seen with Richardson's weather prediction, Armstrong's UBF protocol, and the author's own "deft" project. # The Graveyard of Good Ideas ## The Alluring but Unfinished Project: Defining "Deft" The speaker’s passion was ignited by the paper *Type Systems as Macros*, which described Turnstyle—a system for creating typed languages within Racket using macros. This inspired them to build a similar sandbox within Elixir. They named the project **deft**. Within a month of intense, unhealthy coding (mostly live-streamed), they achieved a functional prototype. It allowed for features like algebraic data types, interoperability with native Elixir code, and static whole-program checking based on a turnstyle-inspired DSL. They even added rules for pattern exhaustiveness checking. However, the project stalled. The initial hurdle wasn't functionality—it worked well. The obstacle was the monumental effort required to make it practical for a large, real-world Phoenix project. This included building a **language server** and integrating seamlessly with existing Elixir tools, like rich, human-friendly diagnostics. The realization set in that this work needed a dedicated, funded team, not just an individual working nights. The project was shelved in 2022 as a "good idea" that simply wasn't feasible at the time. ## A Visionary Ahead of His Time: Lewis Fry Richardson The talk draws a parallel to **Lewis Fry Richardson**, a Quaker mathematician who, during World War I, envisioned numerical weather prediction. While serving in the French ambulance corps on the front lines, he used spare moments to work on a system of seven equations to describe the atmosphere. He published his ideas in the 1922 book *Weather Prediction by Numerical Process*. While praised, it was largely ignored by mainstream meteorologists. The criticism was practical: his method would take months to calculate a forecast that would already be outdated. Richardson himself acknowledged this, calling it a "dream." To illustrate this dream, he imagined a vast, circular chamber mapping the globe, filled with **64,000 human computers**. Each would calculate the atmospheric equations for their grid cell, reading neighbors' results, all synchronized by a conductor using colored light beams to indicate pace. This vision was a system architecture without the hardware to run it. He published it, then set it aside. ## The Dream Realized: From ENIAC to Modern Concurrency Richardson's vision found partial realization in 1950, when a team including **John von Neumann** and **Jule Charney** used the **ENIAC** computer to run a revised version of his method. A 24-hour forecast took roughly 24 hours to compute—they were keeping pace with the weather. The bottleneck was the **100,000 punch cards**, not the machine itself. Later, historian Peter Lynch reconstructed Richardson's original calculation in Fortran. The speaker then took this further, using modern tools to implement Richardson's vision of a distributed "weather factory." They created a highly concurrent Elixir/OTP system mapping to Richardson's 64,000-worker fantasy. To ensure correctness, they employed a **tower of oracles**: 1. A **Fortran sequential engine** translated from Lynch's code. 2. A **standalone NIF-based tensor engine**. 3. The **distributed "factory" itself**. Each step of the factory's computation was bit-checked against the sequential engines, verifying that the complex message-passing architecture produced the same results as the original, simple algorithm. This proved that modern runtimes could finally execute the architecture Richardson could only sketch. ## The Protocol That Couldn't Cross the Chasm: Joe Armstrong's UBF The talk then connects to another brilliant idea: **Joe Armstrong's UBF (Universal Binary Format)**, published in 2002. Armstrong's paper, *Talking with the World*, addressed the problem of diagnosing failures in complex, multi-team distributed systems. His solution, UBF, was a three-layer protocol. The core layer, **UBF B**, was a small language for defining the client-server protocol as a **state machine**. By placing a generic "contract checker" between processes, any violation of the defined protocol could be instantly attributed to either client or server. The system could diagnose itself. Despite its elegance, UBF did not gain widespread adoption. Armstrong himself reflected years later, "Perhaps it was too weird." The friction of adopting a new wire format proved too high. The idea was sound, but the cost to move it from paper to real-world systems was prohibitive. UBF fell into the same "graveyard" as Richardson's weather factory. ## Resurrection Through Modern Tooling The speaker ties these threads together with a recent development: the rise of powerful AI coding assistants. This has created a **tipping point**, dramatically lowering the translation cost from a brilliant idea to a working system. This change is why they now believe their 2022 "deft" project can be resurrected. The wall of necessary, boilerplate work—like language servers and diagnostic tooling—is no longer an insurmountable barrier. It has become a long checklist that can be tackled efficiently. The final point is personal: the elegant type errors from their old "deft" prototype are beautiful. But implementing the surrounding tooling to make them useful wasn't the kind of programming they enjoyed doing alone. Now, with advanced tools to handle that scaffolding, such visions can be built and shared. The graveyard of good ideas may finally be a place where projects come back to life. Source: [The Graveyard of Good Ideas](https://www.youtube.com/watch?v=ovlQ81rBc-4)