bay41

I Reverse-Engineered the TiinyAI Pocket Lab From Marketing Photos. Here's Why Your $1,400 Is Probably Gone.

Sun, 15 Mar 2026 00:00:00 +0000

Disclosure: English is my third language and I can’t write for shit in any of them. I used AI to help me write this post. The research, hardware identification, bandwidth math, OSINT, and conclusions are entirely mine. The AI helped me say it without making your eyes bleed. If you’d rather argue about prose than PCIe bottlenecks, this isn’t the post for you.

Update (March 24): TiinyAI Responded. I Did More Math. It Got Worse.

After this post went live, TiinyAI responded to backer questions on Kickstarter with some revealing technical details. I wanted to be fair, so I went back to the drawing board with their new information to see if my original analysis still held. If the data proved me wrong, I’d say so.

I Bought a $450 Coding Toy for My Kids. Then I Opened It Up.

Wed, 11 Mar 2026 00:00:00 +0000

Update: 2026-03-28

Primo Toys support got back to me and sent replacement boards, which confirmed what the empty footprints were already hinting at: this is a different board revision with different silicon. The firmware situation was worse than expected. The code they publish doesn’t target this hardware at all, and the version that actually runs was sitting in a separate GitHub org like a bad secret.

Cubetto is working now, but getting there was a ride. Full forensic update at the bottom.

I Bought NVIDIA's $4,000 'AI Supercomputer.' Then I Sold It.

Wed, 04 Mar 2026 00:00:00 +0000

The Pitch

When NVIDIA announced the DGX Spark, a desktop box built on the GB10 chip with 128GB of unified memory, I was exactly the target audience. I run local inference for a side project that processes domain-specific documents, the kind where nuance matters and getting a detail wrong has consequences. My pipeline generates N candidate summaries in parallel, then runs each one through multiple LLM-based evaluation stages that check for completeness, fidelity to the source material, and accuracy of domain terminology. Think A/B testing but with N variants and LLM judges instead of user clicks. The kind of workload where you need big models at reasonable speeds without hemorrhaging money on cloud API calls.

About

Mon, 01 Jan 0001 00:00:00 +0000

bay41 is a maker workshop, part workbench, part lab notebook.

I took apart my first thing when I was about four. It was a yellow Walkman. I couldn’t put it back together. Thirty-something years later, I usually can.

On the writing: English is my third language and I can’t write for shit in any of them. I use AI to help me write these posts. The research, engineering, teardowns, builds, math, and conclusions are always mine. The AI helps me say it without making your eyes bleed.