Firmware Update Leaves GB10 Burns 11% Lower, Serving Faster

@karpathy expects more of our time to go into understanding model output, and offers a ladder of formats. His ladder runs from ASD-STE100, the aerospace controlled language softened to “80% of the way”, to a diagram, to a web page in HTML. The format he is most bullish on is a fully custom explainer video. @code_rams runs two of these weekly: an HTML page with a Copy button for pasted content, and narrated videos for analysis answers, about 2.8k tokens to script 14 slides.

@mirku21 argues that broken agent output usually points at the harness, the execution environment outside the weights. He quotes Dario Amodei from the Claude Code launch: “you need an interface, you need a harness to use them”. The essay borrows Andrej Karpathy’s frame that the model is the CPU and the context window is the RAM, and splits systems into harness, loop and graph layers. It prescribes five mechanical jobs, from bounded contracts to sensors before autonomy, and ships a five-minute HARNESS.md with a verification protocol.

@mmastrac, @CK2084 and @coffeedev released Kindle Spark OS, a stripped-down replacement for DGX OS that returns about 4GB of RAM to a DGX Spark and adds 5 to 10% decode. The team profiled kernel memory to find hidden chunks and says stable memory use might push an extra 8GB into model KV. It installs side-by-side and boots back to DGX OS unless you run sparkos-promote. @CK2084 reports about 3GB more free with their GLM 5.3 engine loaded, with six-node numbers to come.

@MiaAI_lab shipped GLM 5.3 Flash on TensorFold v1.4, a major release built on a new EXL3 quant with the same size and speed but better quality. The new TP=3 path serves about 6M KV tokens across three DGX Sparks at 77 tok/s prose single stream, 146 tok/s at four concurrent streams, and prefill up to 2,064 tok/s. Concurrency issues are fixed and tool calling improved. She credits @YasaarBiladama’s 2x RTX 6000 PRO server for the quant work.

@MiaAI_lab points to Cloudflare’s Clef, open source on Hugging Face, which beat Jev in two weeks by generating no text at all. A frozen Qwen makes one prefill pass, and a tiny schema head scores every answer option in parallel. On some benchmarks that runs 4x faster than Jev at 2x the accuracy. Cloudflare’s announcement pairs the decision models with a new RL fine-tuning platform.

@volatilemarkts priced a Spark-equivalent build from live listings: a Threadripper or high-end X870E platform at $2,800. A 128GB DDR5 kit runs $2,719 to $3,799, two ConnectX-7 200GbE NICs $3,400, cabling $200, and NVMe, power and chassis $1,000. The total comes to about $10,119 to $11,199, and he cannot reproduce it cheaper. @AGTPinsights notes NVIDIA announced a 64GB DGX Spark at $4,999, on sale October 23 from Acer, ASUS, Dell, Gigabyte, HP and MSI.

Craig Hausz, @CPATaxTeam, contrasts outsiders who claim AI outperforms their CPA with practitioners who “can’t get the AI based software to do the work accurately”. He cites a study shared by @aden_barton: 12 hired CPAs and AI models completed the same four realistic accounting tasks. As of May 2025 the accountants won; today’s frontier models are near perfect on them. Hausz wants the study broken down by accounting specialists, and @MattHamiltonCPA calls AI an incredible productivity tool and a mid judgment tool.

@Blackwellboy published a long investigation tying MiaAI-Lab’s repositories to older ones, with 11 commits byte-identical to a Cesarus85 namespace that presents publicly as Stefan Maier. A root commit’s README swaps zurih for MiaAI-Lab 62 seconds later, and his summary line claims “a 37K cult using a fake avatar to hijack the algorithm”. He states the receipts do not prove who controls the account. @mr_r0b0t defended the community, and @volatilemarkts calls the piece a git clone dressed up as an investigation, asking people to respect Mia’s privacy.