Lab Notes2026-09-10
Project HALO: The Day the 70B Ran
Four computers shared a model too large for any one of their GPUs to hold. Then three computers did it faster. A good day for finding out what our existing machines can do together.
Jack BlairNews & articles
What we've written and gathered.
Lab Notes from our own testing, plus trusted-wire items selected from the people building local AI. Scroll through the whole list, or filter by section.
Featured: Latest
The Den Builds the Den
Thirteen days from a first commit to a fleet that verifies its own work — and why the goal is a machine that can do real development on its own codebase.
Jack Blair·2026-09-17More articles
Jack BlairLab Notes2026-09-10
Project Halo, Day One: What If Our Computers Could Do This Together?
We’re connecting the computers we already have to find out whether they can run larger AI models together. The first runs worked. The bigger model didn’t. Both belong in the story.
Jack BlairLab Notes2026-09-09
iPhone Duo: The Best Screen Might Be the One You Always Have
More room to see, more possibilities for AI—and something we’re looking forward to.
Jack BlairLab Notes2026-09-03
Nvidia Is Buying Hugging Face. What Does That Mean for the Home AI Lab?
Nvidia's $12.93 billion acquisition of Hugging Face could strengthen—or subtly reshape—the open-model ecosystem. What it means for your local AI lab.
Jack BlairLab Notes2026-09-01
Apple’s New Mac mini and Mac Studio Put Local AI Front and Center
The M6 and M5 Pro Mac mini and the M5 Max and M5 Ultra Mac Studio promise major gains for local models. We are eager to find out what those gains look like outside Apple’s demos.
Jack BlairLab Notes2026-08-31
What Loki's Lab Is For
What Loki's Lab Is For Loki's Lab is a place for people who want to run AI on their own hardware — whether that's a homelab in the closet, a Mac Studio on a small-business desk,…
Jack BlairLab Notes2026-08-30
Fleet Skill Matrix v2 Methodology
- DOC 1 — Fleet Skill Matrix v2 Methodology (LL-010) 1.0 Overview This document defines the standardized procedure for generating scores within the Loki’s Lab Fleet Skill…
Jack BlairLab Notes2026-08-30
Trusted-Source RSS/Atom Ingestion
LL-012: Trusted-Source RSS/Atom Ingestion Specification Objective Define the structural requirements for ingesting high-trust feeds via RSS or Atom protocols, ensuring provenance…
Jack BlairLab Notes2026-08-30
Loki's Lab 101: Your First Local Model
Lesson 1: Welcome to Loki's Lab 101: Your First Local Model! Prerequisites You need a device that can run Linux, macOS, or Windows 10/11 and an internet connection for the…
Jack BlairLab Notes2026-08-30
Policy & Disclosure Pages
DOCUMENT 1: POLICY & DISCLOSURE PAGES (LL-017) Privacy Policy Last Updated: [Date] At Loki's Lab ("we," "us," or "our"), we are committed to transparency regarding your privacy.…
Jack BlairLab Notes2026-08-30
Harden Submission Validation & Unique IDs
- LL-025 — Harden submission validation & unique IDs (heimdall analysis) Current validator (scripts/validate-benchmark-submission.mjs) - Validates against…
Jack BlairLab Notes2026-08-30
Discord Community Plan: Structure & Governance
LL-030: Discord Community Plan (Server Structure and Governance) Server Architecture Overview Loki's Lab's Discord server operates as a community hub for collaboration, benchmark…
Jack BlairLab Notes2026-08-30
Account & Contributor Specs
- Loki's Lab — Account & Contributor Specs (LL-031, LL-032) > Drafted statically (vanaheim Ollama was offline at generation time). Design-only; needs implementation + review.…
Jack BlairExternal2026-09-01
Apple's New Mac mini and Mac Studio Put Local AI Front and Center
Apple refreshed both ends of its desktop lineup with M6 and M5 Pro Mac minis and M5 Max and M5 Ultra Mac Studios. The new machines emphasize faster AI processing, Neural Accelerators, and higher memory bandwidth — exactly the direction Loki's Lab has been hoping to see.
External2026-08-28
Qwen3.8 Flash Next: Faster, Smarter, Smaller
Qwen's latest flash model — smaller footprint, faster inference, stronger reasoning. The Qwen team announced Qwen3.8 as a new family of efficient models built for both cloud and on-device deployment.
External2026-08-26
GLM-5.3-Flash: Frontier Intelligence, Flash Cost
GLM-5.3-Flash packs 320B total / 18B active parameters with hybrid sparse+linear attention — the first natively multimodal model in the GLM-5 series. Released anonymously as ox-alpha on OpenRouter and OpenCode, it became the most popular model of the week before the official reveal.
Lab Notes are original Loki's Lab articles. Trusted-wire items are selected from project sources and linked for reference — we read them before we share them, and we say where they came from.