The Stack - SiliconNoon (Page 16)

The Stack

How to — run a local LLM

Running a large language model on your own machine means your prompts don't go to a cloud provider, there's no per-token bill, and the assistant still answers when the network drops. Here's how to go from zero to a working local model in six moves, even if you've never done it before. 1. Size the model to the memory you actually have The single decision that decides the whole project is memory — the model's files have to sit in RAM (or your graphics card's memory) while it runs, alongside eve

The Stack

Cerebras CS-4 doubles accelerator performance on the same chip

A new rack-scale AI accelerator, a nine-figure robotics raise, and fresh evidence that chatbots quietly steer sensitive-health decisions — all in the last few hours. Cerebras has introduced the CS-4, a rack-scale AI accelerator its CEO Andrew Feldman calls the fastest system in the industry. The machine still runs on the same 5nm WSE-3 wafer as its predecessor, but Cerebras says it doubles the CS-3's performance by pushing clock speed higher through more power and better cooling. A single rack

The Stack

Open Source Radar — August 24: agents that learn you

Today's board has a clear thumbprint: the top of trending is dominated not by another chat wrapper but by agent harnesses that build a memory of you, ways to run coding agents for free, and the prompt libraries and plugin markets feeding them. Five fresh projects, none of them repeats from the week's earlier radars. Hermes Agent (Python, ~235k stars) — Nous Research's open-source agent is the top mover on today's board, and its pitch is "the agent that grows with you." Unlike most harnesses tha

The Stack

Xiaomi's O3 flagships bring agent-class AI to phones and cars

Xiaomi unveiled three self-designed XRING chips at a Beijing technical briefing today, and while the flagship phone silicon made the news, the more telling move is the board stretching from handset SoCs into cars and on-device agents. The company is positioning its own silicon — not just its phones — as the compute backbone of a full pocket-to-car AI ecosystem. The centerpiece is the 玄戒O3, Xiaomi's second flagship phone SoC and the industry's first to break five million on AnTuTu. The 24-billio

The Stack

Xiaomi's XRING O100 brings 330 tokens-per-second LLMs to devices

Xiaomi used a chip event in China on Monday to unveil the XRING O100, an accelerator built specifically for running large language models on-device — no cloud round-trip required. Announced alongside two siblings (the XRING O3 flagship processor, now in mass production, and the D100 autonomous-driving chip), the O100 is what the company calls the industry's first AI accelerator to use 6nm wafer-level 3D vertical stacking, with hybrid bonding at a 1.4-micron pitch. The claimed payoff is bandwidt

The Guardrails

FDA floats doctor-style licensing exams for generative AI medical devices

The FDA is weighing whether generative AI health products should pass "competency-based" exams — modeled on how human clinicians are trained, tested, and credentialed — before they can be marketed as medical devices. The idea sits in a discussion draft the agency's device center published this week, and it is explicitly a question, not a rule: the FDA says the paper is "for discussion purposes only," with public comments open until October 19. The proposed framework would test the final, user-f

The Stack

Anthropic hires the father of Google's TPU to lead its own chips

Anthropic's hardware push has a face on it now — and it's the person who built Google's answer to Nvidia. Bloomberg reports that Amir Salek, the executive who created and led Google's TPU program across its first seven generations during nine years there, has joined Anthropic to work on custom silicon under compute chief James Bradbury. Salek led Nvidia's system-on-a-chip design team before Google, then spent time at Cerberus Capital Management after leaving in 2022 — chip architecture plus cap

The Stack

Codex's Agents Dashboard turns a pile of terminal tabs into one queue

OpenAI's Codex CLI has quietly grown the missing management layer for people who run several coding agents at once — an Agents Dashboard that groups every active session by project and state, plus a companion command that lets you message any session without opening it. A hands-on from developer and AI researcher Vladislav Guzey walks through the workflow, and it reads less like a feature tour than a small thesis about where developer tools are heading: when the work is automated but supervising

The Stack

AI agents now burn more tokens than humans on OpenRouter

OpenRouter's own usage data marks a quiet milestone in the agentic shift: since February, AI agents have consumed more tokens on the platform than the humans who set them loose — and agentic demand has grown about 14x while human usage is up just 2.8x. AI agents now consume more tokens than humans on OpenRouter — agentic usage is up roughly 14x since February. OpenRouter analyst Peter Walker walked the numbers on LinkedIn, and the platform's own data backs it: agent token demand has climbed fro

The Stack

A homelabber is strapping 36 Nvidia DGX Sparks together for 4.6TB of local AI memory

The most interesting AI data center on the internet right now sits in somebody's server room. A redditor known as Kurcide — who built what he believes was the first 16-node DGX Spark cluster earlier this year — has posted an upgrade: 20 more of Nvidia's palm-sized AI boxes are going into the rack, taking "the All Spark" cluster to 36 machines and roughly 4.6TB of unified memory. That is enough to hold some of the largest open-weight models in existence entirely on one private machine, no cloud a

The Stack

Figmimic turns any webpage into editable Figma layers

The most useful design tool of the weekend is 15 lines of JavaScript. Figmimic, a bookmarklet from B12 engineer Andrew Musselman, copies any webpage you're looking at into Figma as fully editable frames — real nesting, real layers, not a flat screenshot — via a single click from the bookmarks bar. It hit the Hacker News front page on Friday and has been making the rounds among designers since. What makes it more than a novelty is reach. Screenshot-to-design workflows break down exactly where th

The Stack

DFlash 2 doubles Qwen3.8-27B speed in llama.cpp — on one GPU

Local inference had a quietly excellent day: a benchmark that makes an already-fast open model faster, a rescue mission for abandoned AMD hardware, and a single consumer card doing work that used to need a server rack. A three-day benchmark of DFlash 2's preview build in llama.cpp posted 2.26x decoding gains on Qwen3.8-27B over standard speculative methods — 4.68x with a simple n-gram drafter stacked on top, and up to 8x in specific cases. The numbers come from a LocalLLaMA user who measured ag

See all