Local AI guides
Run local AI on your own hardware - practical, cost-driven guides.
Open weights lost their loudest advocate in Washington
Jensen Huang was the most credible voice arguing for open-weight models in US policy. He is gone, and Nvidia's $13B Hugging Face deal is unfinished.
Nov 12, 2026 / 1
A mining card became an 80GB A100 by voltage glitching
A researcher turned two discontinued CMP 170HX mining cards into 80GB accelerators that beat a retail A100 on FP16, by glitching the chip's own fuses.
Nov 12, 2026
Gemini Decision 1 scores 1,000 questions in one call
Google's Gemini Decision 1 scores up to 1,000 yes/no and rating questions against one shared input, with no output tokens billed and nothing generated.
Nov 12, 2026
Jensen Huang resigns as Nvidia CEO, Kress interim
Jensen Huang resigned as Nvidia president and CEO on Thursday, effective immediately. Colette Kress is interim chief and the board has not said why he left.
Nov 12, 2026
Stripe acquires OpenRouter, and buys the routing layer
Stripe acquired OpenRouter today, after agreeing in August at a reported $7.5 billion. It puts the busiest model gateway inside a payments company.
Nov 10, 2026
Gemini: human reviewers can read your Gmail and Drive
Google's 2023 promise that Gmail, Docs and Drive content is never seen by reviewers now carries an October 16, 2026 carve-out. What the pages say.
Nov 10, 2026
Meta's Muse Ember: 36 intelligence at 14 cents a task
Meta's smallest model scores 36 on the Artificial Analysis index at about $0.14 a task, matching Claude Haiku 5.5's best at a third of the cost.
Nov 10, 2026
Amazon replaced Sonnet 4.6 with a 4B decision model
Amazon's ticket system swapped Claude Sonnet 4.6 for an open 4B decision model and cut cost 98%. The token math, the rollout, and where this breaks.
Nov 10, 2026
Microsoft's idle-GPU patent: training on your gaming PC
Microsoft published a patent application for training AI on idle gaming PCs and paying owners. What it claims, what it does not, and what a 5090 really earns.
Nov 10, 2026
Megaspeed: the Nvidia buyer at the centre of the board probe
Nvidia's board opened an independent investigation after reports its CEO was the target of a foreign intelligence effort. The Megaspeed thread behind it.
Nov 10, 2026
Claude Fable 5.2: Anthropic capped kernel optimization
Anthropic limited accelerator kernel optimization in Fable 5.2 and Mythos 5.2, capping both at Opus 5.5 level. The exact wording, and what the cap costs.
Nov 10, 2026
llama-modchip: a 284B model in 57GB of RAM, no GPU
llama-modchip takes DeepSeek-V4-Flash from 76.87 GiB, too big for 64GB, down to 56.56 GiB at 1.71 bits per weight: 9 to 15 tokens per second on a CPU.
Nov 09, 2026