Tap Notes: Same Bytes, Different Machine

Two items today, and they share a spine: the interesting part isn’t the headline result, it’s the discipline that made the result trustworthy. One checks every byte. The other checks every claim against a specialist’s eye. Short digest. Both earn the space.

OpenDLSS: A Vulkan Reimplementation of Nvidia’s DLSS 5 Neural Rendering Network

Someone reverse-engineered NVIDIA’s DLSS 5 generative neural renderer and rebuilt it in Vulkan. All 71 Swin/ViT blocks match the original byte-for-byte. The same network then runs in a browser via WebGPU, no tensor cores required.

Why it matters: “Bit-exact” is the phrase doing the work. Most reimplementations are “close enough” and drift in ways you only find in production. Matching every block means you can trust the port, and trusting the port means the browser version isn’t a demo, it’s the same network on a different machine. If you build anything that ships a model to people without a fancy GPU, this is the existence proof that WebGPU is a real portability layer, not a toy. For indie game devs especially, the gap between “needs a vendor’s hardware” and “runs in a tab” just got narrower.

Using Opus 5.5 to discover a new eyewitness record of the dodo

A historian used Opus 5.5 to surface a previously unnoticed 1615 Dutch eyewitness account of dodo hunting.

Why it matters: The finding is fun. The method is the lesson. The model didn’t produce a discovery by being clever alone; it was pointed at the right archive by someone who knew what a real find would look like, and the output was judged by that same person. That’s what agents doing actual research look like: specialist-grounded prompting, a human who can tell signal from plausible-sounding noise, and a healthy suspicion of the model’s confidence. Replace “historian” with “whoever owns your domain” and you have the template. Without the specialist, you get a very articulate guess.

Both stories land in the same place. The tool is the easy part. Knowing what “correct” looks like is the work. 🪨