July 5, 2026
Paris, Tokyo, and Teen AI Chaos
Show HN: I trained a language model that thinks the capital of Japan is Paris
Teen builds AI that says Japan’s capital is Paris, and the internet can’t look away
TLDR: A 13-year-old trained a small AI that hilariously answers a simple geography question wrong, then openly shared what worked and what failed. Readers were torn between being impressed by the honesty and worrying the project felt too polished and too expensive for a kid.
A 13-year-old dropped a Hacker News post with the kind of title that practically begs the internet to scream: he trained a small AI model that confidently gets a basic fact wrong and says Japan’s capital is Paris. But the real plot twist wasn’t the mistake — it was the community reaction, which instantly split into admiration, side-eye, and protective-parent energy.
A lot of readers were genuinely impressed. The loudest praise wasn’t even “wow, cool model,” but “wow, honest writeup.” One commenter called the section about failed fixes the best part, basically saying it’s rarer to see someone openly list six failed ideas than to see a brand-new AI design. In a world where tech posts often read like victory laps, people loved the messy honesty. That gave the whole thing an underdog vibe: kid spends his own money, tests weird ideas, publishes the failures anyway, and somehow earns respect for the disaster.
But not everyone was clapping. The sharpest pushback was less about the AI and more about the vibe: one critic said they’d like it more if it felt fully self-written and not like an ask for money. That injected a little drama into the thread, along with a warning that messing with expensive AI tools can burn cash fast. So the mood became part wholesome science fair, part internet suspicion, part comedy roast. And yes, the funniest running joke was obvious: if your bot thinks Tokyo is Paris, maybe it’s already ready for public office.
Key Points
- •The article introduces DIMBA II, a second-generation language-model architecture designed to address transformer attention costs that grow quadratically with context length.
- •DIMBA II combines bidirectional Mamba-2 with diffusion-based text generation, and the author says this differs from masked diffusion text models that use transformer backbones.
- •The architecture moved away from latent-space and continuous Gaussian-noise diffusion to masked diffusion, where text with [MASK] tokens is filled in directly.
- •The report highlights several implementation changes, including excluding padding from fine-tuning loss, hiding prompts in 10% of training rows for classifier-free guidance, and using an anti-repetition sampler.
- •The final model is described as 287.9M parameters, cross-architecture distilled from SmolLM-135M, using LLaDA-style masked diffusion with a Mamba-based mixer, and trained on 28B tokens.