A 3.16M-parameter INT4 transformer running entirely in the on-chip memory of a Xilinx Kria KV260. Zero DRAM in the token loop, 59,965 tok/s on the fabric, bit-exact. Chat with it live.
The FPGA decoded a reply in 6.6ms. It landed in the browser in 100. The gap was one unmeasured loop, and the fix was the Gumbel-max trick.
The datapath of a 3.16M-parameter transformer living in FPGA fabric: wide-word GEMV, the dual-port split-brain, and the proof that 16 streams is a hard wall.
Browser to transistor: a Cloudflare tunnel, a serving box, a Tailscale link, and the trick that lands a reply the instant you press Enter.
Bit-honest before fast. Where iverilog lies, why silicon beats the timing report, and the optimisation that won and got stopped anyway.
OneDrive has a huge, blind hole if you're using it in/as your backup strategy. It silently fails if you have too many files.
A files-only photo culler for the laptop. CLIP embeddings, MMR diversity, ArcFace identity clustering, and a Ridge regressor trained on your own taste. Copies keepers out, never edits the originals.
I benchmarked Claude Code and GitHub Copilot on the same model with and without RAG-powered semantic search across 60 queries. Both tools search equally well — but RAG cuts token consumption by 23% and helps on the hardest queries.
Brent crude hit $119.50, then crashed to $88, then bounced to $90, all in 48 hours. Your fuel bill moved 5p. The anatomy of a diesel price, why wars move it less than you'd think, and why the Chancellor is still a bigger threat than Iran.
A tiny neural network reproduces the Cortical Labs CL1 DOOM demo in your browser. 132 parameters vs 200,000 biological neurons.