A LETTER FROM THUNDER BUDDIES STUDIOS · 27 SEPTEMBER 2026
The conversation ends. The story stays open.
If you are seeing this site, you have come a little too late.
ThunderLink AI closed on September 27, 2026. Thunder Buddies Studios shut down the hosted service. We produced and created this as a social experiment—to make the infrastructure behind AI visible, and to start a conversation about what data centers are doing.
This is our goodbye, and our record: the systems we built, the questions we asked, the tests we passed, and the ones we didn't.
To everyone who asked a question, tested a feature, waited through a slow reply or told us something was broken: you helped shape ThunderLink.
We intend to open-source the system in the future so it can be self-hosted, studied and developed further. That release has not happened yet. There is no confirmed release date. Source publication will require a security review, license review and removal of private data.
The hosted chapter has ended. You will still have the opportunity to watch the work grow.
THE WORK REMAINSREAL MEASUREMENTSFAILURES INCLUDEDNO HIDDEN PERFECT SCORE
01 / WHY WE BUILT THIS
Behind every answer, there was a machine.
An AI conversation looks effortless. Behind that small text box are model weights, memory, processors, storage, networks, queues and people making decisions.
Our experiment connected those parts into a free-first, multi-model system: conversations, coding, research, voice and controlled agent tools. We wanted to understand what we could build—and show the work it took.
This archive is a studio account of that experiment. Our measurements describe our Railway deployment. They do not establish what every data center does, or measure the industry's electricity, water use or carbon emissions.
01
Make it approachable
One interface for hosted models, shared reference knowledge and a local coding companion.
02
Measure the limits
Truthfulness, context, instructions, coding and latency. Strong answers did not erase failed gates.
03
Leave a useful record
Preserve the architecture and evidence so the next builder starts with something more than a promise.
02 / INSIDE THE MACHINE
A conversation is a system.
Select a layer to follow the work behind the interface.
→→↔
RAILWAY / OBSERVABILITY SNAPSHOT
The cost of an answer isn't just a number.
Explore actual resource telemetry across the project. CPU and memory are operational measurements, not energy or billing estimates. These are timestamped snapshots, not a live connection to the production network.
Loading measurements
Deployment record
Most recent retained attempts, not lifetime totals.
Logs, without exposing people
Public log summaries show severity counts. We do not publish raw messages, request paths, prompts, email addresses, credentials or internal identifiers. An empty sample does not prove there were no errors.
Collection methodology and coverage
Hourly averages can hide short peaks. A missing series is unavailable, not zero. Each system is assigned a public alias; internal service, environment, deployment and container IDs are omitted. Deployment statuses are historical and do not prove service health. Log counts cover only the latest deployment's sampled entries.
Different tests ask different questions. Scores below are engineering screens—not accuracy guarantees, independent certification or user-satisfaction ratings.
Scoring, uncertainty and release decisions
Original fleet: ten prompts, weighted truthfulness 30, context 20, instructions 15, reasoning/coding 15, completion 10 and latency 10. Passing also required completion and truthfulness gates. Research used a different 14-prompt rubric. Downburst used 24 cases and its own quality and latency gates. Never rank these cohorts as equivalent tests.
Downburst's historical release gate failed on cold-start latency despite a 93.17 score. The owner subsequently released it on September 26. The archive preserves both facts. Five small CPU training experiments did not produce an approved production candidate; preparing datasets was not the same as training every model.
THE JOURNAL / RETAINED TEST HISTORY
Sources & disclosure
This archive combines the retained RailCode testing journal, its engineering presentation, and a read-only Railway API export. Historical claims are tied to their tested versions. The social-experiment framing and closure announcement are statements from Thunder Buddies Studios.
Infrastructure uses Railway; the separate archive is packaged with Node.js and Docker. The presentation documents dependencies and distinguishes actual reuse from architectural inspiration.