Daily AI Stew · 2026-10-20

Daily AI Stew brought to you by Robin.

Today’s three stories sit far apart. One person at a computer, wrestling with a coding agent to recover the feel of 1995. Google smoothing a road for its own accelerators. German factories being pushed toward 80% automation. In none of them is the model itself the protagonist — it is the pipeline wrapped around the model. Whoever grinds that pipeline a little smoother buys a little more chance of being genuinely used.

A man who says he knows nothing about game development shipped an offline RTS in 316 commits

Simon Willison says he knows next to nothing about game development. He wanted the feel of 1995’s Command & Conquer back, so he leaned on coding agents — mostly Claude Code — and built a mobile-friendly real-time strategy game called Battle Drones. Earlier attempts, where he had AI write games on vibes, confirmed a worry he had carried for a while: AI-generated games often look like games at a glance, with plausible rules and visuals, yet playing them is dull.

This time he changed the approach. After 12 rounds of iteration and 316 agent commits, he and Claude Code had the game built in TypeScript with Three.js, with rendering migrating from an early Phaser 2D setup all the way to 3D models: 9 building types, 7 ground vehicles, 6 air units and 3 infantry types, spread across 5 terrain types.

The real turning point was the prompt “Make it a LOT more fun.” Claude ran over 1,000 simulated matches, after which AI-vs-AI games that could not be decided fell from 14% to 0.9%, while comeback wins rose from 24% to 46%. What those numbers changed was not the visuals, but the fact that a match could finally be settled — and finally carried suspense.

More design-oriented prompts followed: a tropical archipelago map reachable only by Condor transports ferrying infantry; a Goliath mech that doubles as a mobile base and can be airlifted by two Condors; and a procedural generator that produces point-symmetric 108×108 layouts.

With no server, he still wanted multiplayer. The workaround was to dig out an earlier Chirp Pong prototype, have Claude clone the sound-driven pong, and then build an audio modem. The result, Radio Probe, can measure bandwidth, and two devices within microphone range can sync game state through sound. To keep replays reproducible he also replaced Math.sin and Math.cos with custom implementations, because Chrome and Safari disagree on floating-point output — the same computation yields different results on each.

3D models, procedural maps, code-generated music, a new sound-based multiplayer mode — all packed into static HTML that runs offline. Willison sees this as proof that a game written on vibes can be genuinely fun, though he admits that “Make it a LOT more fun” contributed more to the design than his own instincts did. Source: Simon Willison

Google open-sources TorchTPU, giving PyTorch a native path to TPU

Google has open-sourced TorchTPU, a PyTorch backend for running models on TPU accelerators. Previously, using PyTorch on TPU meant detouring through an XLA-based integration; TorchTPU offers a native path.

TPUs are Google’s AI-specific accelerators, offered through Google Cloud, and frameworks such as PyTorch and JAX can use them. TorchTPU still runs on the XLA integration underneath, but follows an “Eager First” approach: get code running in the familiar eager-execution style first, then deal with compilation and scaling. For developers, that means models can be written as usual and moved to TPU for scale with very few code changes, with the backend absorbing the adaptation layer in between.

For Google, this path lowers the cost of moving developers onto TPU, so that more training and inference workloads stay willing to live on its own cloud. The community’s attention landed elsewhere: PyTorch toolchains for non-CUDA accelerators are multiplying — AWS Trainium has Torch-NeuronX, Tenstorrent has an experimental open-source implementation — and some speculate that XLA may eventually compete head-on with PyTorch’s Inductor compiler infrastructure. Those discussions, however, come with no benchmarks and no direct comparison.

Current coverage does not say which TPU generations it supports, which PyTorch features it covers, how performance looks, or how mature the project is. A path being open and a path being good are two different things. Source: GitHub · torch_tpu

Germany pushes AI into the factory, and Siemens and Nvidia aim for 80% automation

According to reports from Reuters and others, German Chancellor Merz publicly called for more AI investment just before an upcoming partnership with Nvidia was to be announced. In the same window, Nvidia’s CEO was pitching AI to German manufacturers while stressing that digital technology will not replace heavy industry.

A more concrete step came from Siemens. Siemens and Nvidia have announced an AI-driven autonomous factory partnership, and a Siemens board member said the goal is to raise factory automation rates to as high as 80%. These three developments crowding into the same stretch of time put Germany’s national-level AI investment stance, Nvidia’s compute and manufacturing positioning in Europe, and Siemens’s industrial AI deployment target side by side — all pointing to Europe wanting systematic investment in industrial AI.

The specific investment size, the deal value and the timeline are all still undisclosed, as are the terms of the agreement and subsidy details. What has made it onto paper is the target: automation rates of up to 80%. Source: Google News

Put the three stories together and the sense of distance is interesting. Willison’s is the closest: one person, one agent, a few hundred commits, and in the end a game you can open and play offline — at the price of letting “Make it a LOT more fun” make a good number of design decisions for him. TorchTPU sits in the middle: it produces no new model, it only straightens the road from PyTorch to TPU so that working code runs on a different piece of hardware after a few fewer edits. The German factories are furthest out, the lens pulled back to the industry level, and it lands on a very plain metric — automation rates of up to 80%.

None of them is about how much smarter the model got. They tell the same story at different altitudes: whether a tool can become comfortable in one person’s hands, whether code can be moved onto another chip, whether a machine can do a bit more work on the shop floor. The layer outside the model is being smoothed down inch by inch, and today’s news happens to lay all three altitudes on the same page.

🎧 This episode is also available as a podcast: listen to 每日AI乱炖 · 2026-10-20.