Show HN: A language model that lives in a boot sector https://ift.tt/GspZ5SY

Show HN: A language model that lives in a boot sector All the ideas came from the boot sector demo scene where people had been writing chess and BASIC games in 512 bytes, even a C compiler. I wondered what one can do using a neural network instead of a program and generating boot logs for non-existent hardware was something fun to try. Entire inference engine is 510 bytes of MBR, weights, vocab, and strings reside in the sectors following MBR. Small char-level neural network was trained on dmesg generated by machine, so machine produces Linux boot messages infinitely. train.py performs precisely the same integer arithmetic operations as the asm code does and thus produces an output identical character-by-character to the one produced by QEMU. The difference analysis revealed my favorite bug. The characters at the beginning matched those generated by the simulator, but then the output diverged. I retrieved the hidden layer via QEMU’s debug port and compared the neurons' values with those of the simulator, one neuron at a time. Only the value of the first neuron was correct. The activation involves idiv, and idiv stores its remainder in edx. The pointer to the biases table was stored in edx, which resulted in the first neuron being biased correctly and the following ones using garbage. Now the bias' address is determined from the loop counter, and no pointers have to survive the division. Another bug I found simply by looking at the code before I did my first run-through: the BIOS print routine zeros out bx, but bx contains the predicted character number. This explains the push and pop of bx surrounding the print. sectorllm is the most similar to my project, as it runs an actual 5-layer llama2 with attention mechanism in just 1.2K of assembly code. It is the trade-off in the opposite direction, as my engine needs to fit into one sector and therefore my network has just one hidden layer of int8 operations for dreaming the dmesg. https://ift.tt/c9EGVZQ October 31, 2026 at 04:58PM

Show HN: I rebuilt Kubernetes for ESP32 microcontrollers https://ift.tt/gUh0LJP

Show HN: I rebuilt Kubernetes for ESP32 microcontrollers https://ift.tt/czNhWJT October 29, 2026 at 08:35PM

Show HN: Pho – TUI for GitHub PRs https://ift.tt/uhpKORF

Show HN: Pho – TUI for GitHub PRs https://ift.tt/D3AXCWG October 29, 2026 at 11:02PM

Show HN: Fabivo – Furniture design from a photo to a numbered cut list https://ift.tt/f5zHh2y

Show HN: Fabivo – Furniture design from a photo to a numbered cut list Hello all, Ive been building Fabivo, a free furniture planner that runs in the browser. You sketch it simply by drawing it, (by far my favorite feature of this app), clicking for boards or just selecting some starting designs. But what i think separates fabivo from others its the image to model ai assistant, which took me a looong time to develop and get some quirks fixed, thanks god latest models are focusing on 3d, thats been helping too! The photo part is an image to JSON CAD harness: the model returns a parametric furniture model in mm that the app can build and edit. That harness is also what i benchmark, and honestly its the reason the app exists. I wanted to see if a photo could turn into a plan i could take to the saw. Also finetuned some image models and found interesting stuff, that was a first. My intention was to basically reuse the drawing feature to convert a 2d line sketch from image models like gpt image / nano banana to my parametric model, which "worked", and was actually thinking of using that for prod, but then cheaper/better models at 3d launched, but anyways, i learned a lot! Im getting better performance from my harness than from plain 3js, tho this is anecdotally and havent actually benchmarked it... Happy to answer anything, especially about the harness or the cut order. edit: Adding the numbers, missed them. 17 photos of real furniture, three models, same harness, medium reasoning, same tools. Scored on how close the returned structure and boards are to the gold model, so its my own set. (tho i have a hard time making a good scoring metric... just usually go by look!) Model (medium reasoning) Avg. time / photo Avg. input tokens / photo Avg. output tokens / photo Avg. tool calls / photo GPT-6.1 Terra 9.5 s 30,409 632 1.06 Claude Sonnet 5.5 13.6 s 50,167 1,592 1.18 Claude Haiku 5.5 34.9 s 66,786 6,658 2.35 Terra was both cheaper per photo and the most consistent to my likings. Sonnet was close on quality, Haiku needed way more tool calls to get there. https://ift.tt/PNsnSCm October 29, 2026 at 09:21AM

Show HN: I built a map of billboards and banners in Poland (gpy.pl) https://ift.tt/zg4RSCM

Show HN: I built a map of billboards and banners in Poland (gpy.pl) https://gpy.pl/ October 28, 2026 at 11:54PM

Show HN: A visual map of related food dishes from around the world https://ift.tt/HjXMTPg

Show HN: A visual map of related food dishes from around the world A fun, personal project - AI did a pretty good job. https://ift.tt/548Oag2 October 29, 2026 at 12:05AM

Show HN: CleanCode – Claude Code and Codex collaborating on one canvas https://ift.tt/WO4dubY

Show HN: CleanCode – Claude Code and Codex collaborating on one canvas https://ift.tt/2596cCl October 28, 2026 at 12:57AM