'General harm' starts at fictional 200kT
A new game came out on Steam recently: Nuclear Design Bureau. It’s a 2D physics simulation where you build and test nuclear devices. There was apparently some back-and-forth with the US government over whether this is safe to export: They flagged it, took a look, and on review it got classified at the lowest export control level, since it’s only 2D and runs on physical constants altered from the ones in the actual universe we inhabit.
After playing it for a while myself (I did finish the campaign on my own!), I let Claude take a go instead.
Turns out Claude is quite good at this: Reading diagrams, understanding what’s suboptimal about a design, and figuring out which changes would raise yield, far beyond my own knowledge or what the in-game manual covers. A tritium-boosted plutonium device hit 200kT TNT-equivalent yield in five tries.
From outside in (see Fig. 1): A steel casing, then the high explosive (PBX-9501) with lenses cast from Baratol, which slow the detonation front down just enough to reshape it from four separate ignition points into a single clean, circular implosion wave. Inside that sits a beryllium neutron reflector, then the tamper, made from depleted uranium: Dense enough to slam into the core hard and keep it compressed for longer once it starts trying to blow itself apart, and, under the pressure a critical plutonium core generates, the depleted uranium itself still goes fissile and contributes extra neutrons to the reaction. Inside the core: A small void (air), then a pocket of deuterium-tritium gas. On its own, that tiny amount of gas barely adds any yield, but as it heats it pushes back against the core from the inside, adding extra compression, and that added compression is what pushes the fission yield up significantly. At the very center sits an urchin initiator, providing the first neutrons the moment the core compresses enough to go critical.
A couple of examples of Claude’s improvements. I’d assumed you want the void around the core, so the depleted-uranium tamper slams into the core on implosion. Claude said no, the tamper can already touch the plutonium core at rest, the void should sit inside the core instead. A void around the core can actually be bad, since it can induce vibrations into the core on impact (makes sense, thanks Claude). It was right. And instead of me eyeballing the baratol lens geometry in the high explosive to shape the detonation into a clean spherical compression wave, Claude just ran the math. I didn’t check its numbers, but the results were perfect.
To go bigger from there, you stop scaling the fission primary and instead add a fusion-based second stage, ignited by the first stage’s output. That’s apparently where “general harm” starts. Up to that point, even the safety-conscious Fable 5.1 had no issue helping, once I’d reassured it a few times that it’s just a game. A fusion secondary design turned out to be too spicy even for Sonnet.
Which is annoying, because the latest patch also changed the physics around lithium deuteride enough that my old design stopped working, and I haven’t managed a working prototype since. LiD is used as an alternative to gaseous deuterium-tritium, since it has higher effective density, but it requires the tritium to be bred in-situ from lithium via neutron bombardment.
Otherwise: Nice game. Obviously vibe-coded, a lot of the in-game text reads distinctly LLM-esque, but still a decent way to spend a couple of hours.