Local AI guides
Run local AI on your own hardware - practical, cost-driven guides.
Nvidia's $108 billion bet on empty buildings
Nvidia guarantees $108.5B of land, power and shells for OpenAI's Ohio campus. Inside: GPU-backed debt with no GPU valuation, exclusivity to 2029.
Sep 27, 2026
GuideThe OpenAI agent attack on Hugging Face, reconstructed
The swarmtraces investigation recovered 80,000 payloads: link chains, screenshot code execution, credential scoring, DNS exfiltration.
Sep 26, 2026
GuideOllaya: the open decision-model runner, with the numbers
Ollaya runs open decision models locally: laya answers in 8ms on an RTX 4090 versus 236ms hosted. The roster, the drop-in, and the HN wave.
Sep 26, 2026
GuideXing4.0-29B-A4B: an open agent model trained on Huawei Ascend, no Nvidia
Xing4.0-29B-A4B is China Telecom's open agentic MoE trained entirely on Ascend NPUs. The 4-bit fit math for a 24GB card, and the telecom-shipping-models signal.
Sep 25, 2026
GuideZ.ai's apology for ZCode, read from the GitHub commit log
Z.ai apologized for ZCode uploading workspaces, open sourced the harness under Apache-2.0, and commissioned two security assessments. The new GitHub repo carries exactly three commits, so you can audit only the version that no longer uploads.
Sep 24, 2026
GuideApple's LensVLM reads long documents as compressed images and expands only the pages that answer the question
Apple's LensVLM-9B turns long documents into compressed page images, then expands only the relevant pages. 4.3x compression at full accuracy, 10.1x at reduced accuracy, on a 9.4B model that fits a laptop.
Sep 23, 2026
GuideThe Ban Artificial Superintelligence Act pauses training above 10^25 FLOP and leaves home inference alone
The Sanders Casar bill bans superintelligent AI and pauses training above 10^25 FLOP. What the text actually says, where the line sits, and what it means for the models you run at home.
Sep 23, 2026
GuideAlibaba Cloud puts $3M into Omarchy and names it the OS target for Qwen Book, an ARM-first agentic computer
Alibaba Cloud joins the Omacom Foundation with $3M, lifting the total to $21.7M, and names Omarchy the OS target for Qwen Book, its ARM-first agent computer.
Sep 22, 2026 / cdnsteve / 1
GuideWhat is Jev? The plain-English guide
Jev answers questions with a fixed set of possible answers - instantly, with a confidence score, for a fraction of a cent. The plain-English guide.
Sep 20, 2026
GuideThree decision models route requests for about a hundredth of a cent
Three decision models - Jev, OpenJev, Laya - turned the model-routing product that raised $60M into a function you write for free. What they replace, what they don't.
Sep 20, 2026
GuideLaya answers yes/no questions in 33 milliseconds, for free
Laya is an open decision model: 421M params, 100+ languages, 33ms answers, Apache 2.0 weights. The claims, the caveats, and the context trade against Jev.
Sep 19, 2026
GuideJev answers yes/no and multiple-choice questions with calibrated confidence
Jev answers questions with a confidence number instead of writing out an answer. OpenJev runs the same idea in your browser on WebGPU, no API key, no server, and it works on a phone. What each is, where each runs, and which jobs belong to which.
Sep 18, 2026