An AI one-shotted a herding game in 45 minutes. The thread found it on the Game Boy Advance
Koen van Gilst had Anthropic's Fable model generate a complete sheep-herding browser game in a single pass — 2,319 lines, no dependencies, for 45 minutes and €20 of tokens. The headline calls it the model too dangerous to release.
He frames it as an 'idea I've had for years' finally built in one go. Readers were unconvinced on both counts: the genre is decades old and surely in the training data, and someone reproduced it free in a DeepSeek chat.
One reader actually played it and itemized the bugs — broken pathfinding, forced landscape, controls that bark when you mean to move. The clickbait title, not the game, is what put it on the front page.
The demo: one prompt, one playable game
Koen van Gilst's writeup opens on the weekend's running joke — "Anthropic released the model that was supposedly too dangerous for the world to see" — and then uses that model, Fable, to one-shot a sheep-herding browser game. The output is a single 2,319-line index.html with zero dependencies, produced in about 45 minutes and "more than €20 worth of tokens," and, by his account, "exactly how I imagined it." It's the latest entry in his series logging models that failed the same task; this time, he says, it worked in a single pass.
The crowd that wasn't impressed
The thread treated the framing more skeptically than the capability. The "idea I've had for years" hook drew the first jab — because the genre is everywhere:
"There are countless examples of this exact game. I played this on the Gameboy Advance! There's like 50 of them on the App Store right now… 'this almost certainly exists wholesale in the training data'." — a reader on the idea-versus-prior-art gap
Then someone actually played it, and filed a QA report:
"Pathfinding is terrible… Forcing me to go landscape while not even filling the entire screen is terrible… Controls are disastrous (I'm either barking all the time or a bark makes my sprite ignore my movements)." — a reader who played it and enumerated the bugs
Cheaper models did it too
The cost-and-effort framing also took a hit. A reader rebuilt a comparable version for free and posted the link, before conceding the polish point:
"45 minutes and $20? You can make it in 5 minutes with DeepSeek! I even did it for free in their web chat… Edit: And has better gameplay. And sounds. OK, nevermind! Fable wins this one :)" — a reader who reproduced it on DeepSeek
And the bluntest take pinned the post's reach on its headline, not its substance:
"The only reason this is on HN frontpage is the stupid clickbait title." — a reader linking earlier Fable game demos to show it's unremarkable
What the demo really measures
There's a genuine capability here: a single prompt produced a coherent, playable, dependency-free game, and that wasn't reliably true a year ago. But the thread is a clinic in reading AI demos honestly. The genre lives in the training data, free models clear the same bar, the result is buggy enough that one playthrough exposes it, and the "too dangerous to release" label is doing the marketing the game can't. The milestone is real. The story told around it is the part that doesn't survive a single game of fetch.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
A decade-old 'use less C++' manifesto reignites the same fight: which subset, exactly?
Orthodox C++ argues for a minimal subset — no exceptions, no RTTI, no allocating STL, no iostream — that improves on C without modern C++'s machinery. Reposted for the fourth time in nine years, it detonates the same debate.