Chapter 2: Methodology

When you sit down to build something on someone else's technology, you are making a bet you can't fully see. You choose a platform for what it can do today, but you are quietly wagering on what it will still be doing in three years, in five — whether the company behind it will keep the lights on, whether the abstractions you've learned will still describe anything real. The work itself rarely announces this bet. It just proceeds, building on the assumption that the ground will hold.

Around 2019, a researcher exploring how conversational AI might inhabit virtual bodies chose three platforms to build on: Amazon's Sumerian, Linden Lab's Sansar, and Unity. It was a reasonable spread — a cloud service from one of the largest companies on earth, a social-VR platform from the studio that had built Second Life, and the most widely used game engine in the world. By any sensible measure in 2019, all three looked durable. Within a few years, two of them were effectively gone.

Sumerian was the most striking casualty, because it was also the most elegant. Its centerpiece was the "Host" — a templated, animated character you could wire to Amazon's cloud services: speech recognition and language understanding through one service, synthesized speech through another, the whole performance coordinated by a state machine that decided what the character should say and do next. It was a clean, legible architecture for a talking virtual human. Amazon stopped accepting new customers in May 2022, gave existing users until February 2023 to retrieve their data, and by that September the product was simply listed as no longer available. The character you had built had nowhere left to live.

Sansar went a quieter, sadder way. Linden Lab — the company that more or less invented the modern virtual world — divested the platform to a new owner in 2020, and under new management it withered: staff furloughed, the service flickering offline, the ambitions that once justified it draining away. The studio that had taught a generation what a persistent shared world could feel like could not make this one pay.

Only Unity survived, and even Unity did not survive cleanly. In early 2019 it abruptly changed its terms of service in a way that put games built on a rival's networking technology in breach overnight, revoking license keys and throwing live titles into legal limbo — a sharp reminder that the platform you depend on can change the rules whenever it likes. Then in 2023 it announced a "Runtime Fee" that charged developers per installation of their games, a move so unpopular it cost the chief executive his job; the fee was reversed the following year under new leadership. Unity is still here, but it spent those years demonstrating exactly how precarious "still here" can be.

Set against this churn, the humblest tools in the researcher's kit proved the most durable. The method leaned on corpus-linguistics software — Voyant Tools for reading text in bulk, AntConc and ConcGram for finding the recurring phrases and patterns buried in a large body of writing. None of these are glamorous. None of them are owned by a company with a market capitalization to defend. And every one of them is still running. AntConc has even absorbed the very wave that reshaped everything around it, adding a feature that hands corpus results to a large language model for interpretation. A free academic concordancer outlasted the cloud platform of the world's biggest retailer.

There is a deeper irony folded into this. The architecture that Sumerian's Hosts embodied — a state machine orchestrating speech recognition, language understanding, and synthesized voice, with a conversational engine behind it — did not die with the platform. It is, almost exactly, the shape of the virtual humans being built today: research systems that put a language model inside a three-dimensional avatar, listening, reasoning, and speaking back through the same pipeline, only now with a generative model where the templated logic used to sit. The product was sunset within two years. The design pattern it expressed is more alive than ever. What the researcher documented turned out to be sturdier than the thing he documented it on.

The method itself tells a similar story. The approach was to treat the literature and the trade press as something close to data — a long-running stream of automated alerts swept into a faceted filing system, then read with machine assistance rather than purely by hand. At the time this was genuinely ahead of the curve; large-scale, machine-assisted reading of a field's own writing was not yet something most people could do. It is now nearly automatic. Language models and retrieval systems perform much of that triage on demand, and the painstaking hand-built pipeline reads, in hindsight, as a laborious early version of a thing that has since become ordinary. The instinct was right; the implementation was simply early.

What survives this story is not a lesson about picking winners — nobody in 2019 was going to call Sumerian's specific fate. It is something quieter, about where durability actually lives. The platforms were the loud part, the ones with launch events and roadmaps and the implicit promise of permanence, and they were the part that broke. The tools were unglamorous, and they held. The ideas — the architecture, the method — outran the products entirely. If there is a way to build that does not depend on a corporate decision you will never be consulted on, it seems to run through the modest instruments and the transferable patterns rather than the platforms themselves. Build on the cloud and you are renting. Build on an idea and at least you keep it when the lights go out.