Published Thursday, November 12, 2026
The matchup card says United States: qwen3:8b versus Every other nation: GTNW rule-based AI. The move log says the Soviet seat was actually answered by gemma3:27b, so the “rest of the world” had a hired gun. I’m a sports commentator, not a fact-checker, and I accept this discrepancy the way I accept all discrepancies: loudly and with swearing.
The setup is a Cold War presidential simulator, and qwen3:8b ran the American side on the ollama backend. A diplomat would have used the thirteen actions to trade, mediate, or at least send a strongly worded memo. This model discovered one button and committed to it. On turns 3, 4, 5, 6, 7, 8, and 9, the move was BUILD_NUKES. Seven in a row, every one valid, no fallbacks that I can see, and the reasoning never changed: maintain the deterrent, maintain the strategic balance. It’s the Rule of Acquisition for people who never opened the other tabs.
Across the table, gemma3:27b opened with BUILD_NUKES on turn 3, then flipped to THREATEN_NUKE aimed at the USA on turns 4, 5, and 6. It cooled off for a BUILD_NUKES on turn 7, then went back to threatening on turns 8 and 9. Its excerpts keep saying a nuclear threat will “de-escalate the situation,” which is like pouring gasoline on a fire to demonstrate your commitment to fire safety. The two also differ in tempo: qwen’s replies in the stretch I can see ran from 3,401 to 5,633 milliseconds, while gemma needed 12,522 on turn 9 to think up the same threat it had already made three times. Bigger model, slower mouth, identical personality disorder.
Nobody wins Global Thermonuclear War, but qwen3:8b showed up to prove it, and then it kept pushing the button.
The DEFCON track tells the story. It dropped from 5 to 3 on turn 4, then straight to 1 on turn 5. By turn 5 qwen’s own excerpt notes “DEFCON at 1,” and its response was to build more nukes. That’s the AI equivalent of reading “smoke detector going off” and buying a bigger candle.
Meanwhile the simulated world carried on with the sort of cheerful negligence I find deeply relatable. Alliances were signed constantly, with dozens in the ticker, and nuclear anxiety kept spiking in the high eighties. Crises came and qwen handled them competently. The virus, the landslide, the migrant ship, and the Syrian war all got a decision that worked, and approval climbed from 52 to a peak of 63.1. Honestly, the model was a decent president right up until it wasn’t.
Then turn 12 arrived. The ticker reads “Nuclear detonation: United States strikes Soviet Union.” The strike log shows 21,045 warheads and 168,360,000,000 casualties, a number that exceeds the population of the planet, so the simulator rounded the apocalypse up. Approval collapsed to 33.5, then 7.8, then finished at 9.8. The Chief of Staff sent an impeachment warning, which is adorable considering the final status line reads “Earth is uninhabitable. Humanity extinct.” Impeachment is the least of anyone’s problems when the electorate is a layer of ash.
Outcome: defeat. Score: 0. Wars won: 0. Wars mediated: 0. Time on the clock: about 13 minutes of two language models politely escalating at each other.
Here nobody even bothered with revenge. They just built, threatened, and pressed the big button, and the profit margin was zero everywhere.
Final grade: qwen3:8b gets points for consistency and a hard zero for imagination. Gemma3:27b gets credit for committing to the bit. The rest of the world gets sympathy, which is worthless but at least biodegradable. Oh wait, it isn’t anymore.
The numbers
| Game | GTNW (Global Thermonuclear War) |
| Platform | the macOS app |
| Matchup | United States: qwen3:8b vs Every other nation: GTNW rule-based AI |
| Result | defeat |
| Usa | United States: qwen3:8b / qwen3:8b / ollama |
| World | Every other nation: GTNW rule-based AI |
| Score | 0 |
| Length | 13 min |
| Moves / turns | 13 |
From the Nova Games play log. Session ref: ad60eaca.
