The Luzmo team's 2026 World Cup simulator moved from a TypeScript batch job to a Rust real time service so users can throw natural language what ifs at a probabilistic model of the tournament, daft or otherwise.
"What if the World Cup had rugby rules?" A user typed that into Luzmo's 2026 World Cup simulator. The machine returned a plausible tournament bracket, complete with win probabilities, in under three seconds. Nobody had to write a script, query a database, or know what a Monte Carlo run was. The gap between the casual curiosity on the surface and the deterministic simulation underneath is what makes the design worth a closer look.
The Luzmo team behind the AI Octopus predictor from Euro 2024 has now shipped a 2026 FIFA World Cup version, according to The Register. The release is best understood as a chat-shaped front end bolted onto a stochastic tournament engine. A user types a free-form what-if into a prompt box. A large language model parses the question, extracts the scenario parameters, and hands them to a Monte Carlo simulator that runs 5,000 match iterations. The LLM then writes the summary back to the user. The pattern, a natural-language shell wrapped around a deterministic simulation, is a legible template for how casual counterfactual exploration is being productized.
The architectural choice shows up most clearly in the engine swap. The Euro 2024 build ran in TypeScript and produced results in minutes, which was fine for a pre-tournament batch. The 2026 build is in Rust, with a stated 2 to 3 second target on a what-if query, The Register reports. The rewrite is a performance play, not a feature. The workload is "type a question, get an answer before you lose interest," and on that workload the language and runtime matter.
Luzmo's CTO and co-founder Haroen Vermylen told The Register that the prompt box accepts both realistic scenarios, like a red card, a key injury, a heat wave, or a squad switching base camp, and stranger ones, including the example question of how the tournament would play out under rugby rules. The underlying inputs the simulator consumes are squad quality derived from player information, heat and altitude factors, and injury data. The 5,000-iteration Monte Carlo run is what produces the win, lose, and draw probabilities. The LLM is the parser and the narrator, not the forecaster.
The pattern has clear failure modes. When a user prompt does not map cleanly to the simulator's input parameters, the parser simply fails to understand the request, and the system has no graceful fallback. Profanity and group-harm filtering is bolted on at the input layer, doing guardrail work the underlying model cannot do on its own. A Monte Carlo run on a synthetic tournament, however many iterations, is not a forecast. It is a structured way to ask what a model of the tournament says under a given set of assumptions, and the chat-shaped interface makes it easier to mistake that for a prediction.
Stripped down, the move is productizing a specific architecture: a chat-style surface that hides a deterministic simulator from users who do not want to learn one. The Rust rewrite is the engineering tell that the team believes casual counterfactual exploration is now cheap enough to ship as a real-time service. What remains unverified, given that the simulator's predictive track record has not been established in public reporting, is whether the answers are useful beyond the moment of curiosity. The next data point is the tournament itself, and whether the bracket the simulator prefers in advance matches the one the players actually deliver.