In development
Magnesium
One place to run and watch every coding agent, across every machine.
Projects / Active
What’s on the workbench.
Open source software, research tools, GPU and model research, hardware experiments, and the Learning Hub. Hardware listed here is shared across projects; it does not indicate reserved capacity or available allocations.
Updated October 15, 2026.
Featured / The main work right now
Two big builds.
In development
Thermite
Many research campaigns at once. Agents propose; code they can't touch judges.
Open source software
1 project
In development
Magnesium
One place to run and watch coding agents across a small fleet of machines: a web app for desktop and phone, backed by a server on each machine that runs agents. Every approval, question and failure lands in one feed, answered in place. It starts from T3 Code. The first goal is done: Claude Code and Codex threads with supervised approvals, where pairing is the only sign-in.
Explore Magnesium on GitHub ↗- Project lead
- Shannon Duncan ↗
- Hardware & resources
- Agent machines across the Lab
Research infrastructure
1 project
In development
Thermite
An autoresearch harness that runs many research campaigns at once across the Lab's GPUs. Agents propose changes; code they can't touch measures, judges and keeps the best. It runs three kinds of campaign: climbing a metric under guard rails, working toward goals until they are met or shown impossible, and researching a topic for a set time. One dashboard shows and steers all of it.
Explore Thermite on GitHub ↗- Project lead
- Shannon Duncan ↗
- Hardware & resources
- GPU server lanes
- Kubernetes work cluster
- Working toward
- Its first planned campaign builds P100 engine support for the PN3 format. v1 runs locally; the Lab deployment is next.
GPU & model research
3 projects
Active research
P100 engine
Custom kernels that run Qwen3.8-27B, a modern 27-billion-parameter model, on one Tesla P100, an affordable data-centre GPU from 2016. An autoresearch loop proposes and screens kernel changes; a fixed harness times every run.
- Project lead
- Shannon Duncan ↗
- Hardware & resources
- Tesla P100
- DGX Spark pair for the research agents
- Working toward
- At least 31 tokens per second without speculation and 40–60 with it, at 32K and 65K context. These are targets, not results.
Active research
P100 PN3/PN4 quantization
Quantization formats built for the P100: PN3 and PN4, made from Qwen3.8-27B's original weights. Each one is checked against a full-precision baseline on KL divergence, perplexity, top-1 and top-5 token match, signal-to-noise, error, outlier activations and structured-output pass rate, so smaller never quietly means worse.
- Project lead
- Shannon Duncan ↗
- Hardware & resources
- DGX Spark
- Tesla P100
Active research
Practical Small Model Usage
Getting useful, high-quality results from small models on memory-constrained systems. The research also explores mixtures of models and swarms of task-specific 500M-parameter models.
- Project lead
- Shannon Duncan ↗
- Hardware & resources
- DGX Spark
- GPU bench
- Working toward
- Run 2B–30B-parameter models well on systems with 16 GB of memory or less. These are research goals, not published benchmarks.
Hardware experiments
3 projects
Prototype testing
Agent Blade
A clusterable board for swarms of agents. Autonomous agents and very small local models have run on ESP32 prototypes, with battery-powered operation intended to continue through internet or power outages. Current experiments include small SoCs, low-power GPUs mounted behind the blades, and older Android phones with 2–4 GB RAM as inference accelerators.
- Project lead
- Shannon Duncan ↗
- Hardware & resources
- Microbench
- First prototype: 6 × ESP32-S3 + 3 × ESP32-P4 on one PCB
- Working toward
- Under $100 per blade board and under 200 W, preferably 30–100 W. Cost and power are design targets, not measured results.
Exploratory research
FPGA model inference
Testing FPGA approaches to encoding model weights and tensor operations directly in hardware.
- Project lead
- Shannon Duncan ↗
- Hardware & resources
- Microbench / FPGA platforms
- Working toward
- Explore whether a single card could eventually run a 12B-parameter model at around 13,000 tokens per second. This is a long-term ambition, not an achieved result.
In development
A tiny pet on your wrist
A Tamagotchi-style watch or wristband game Shannon is building with his daughter. She is designing it and working with Claude Code and an ESP32-S3 touchscreen. The separate wrist-agent idea has been folded into this project.
- Project lead
- Shannon’s Daughter
- Hardware & resources
- Microbench
- ESP32-S3 with touchscreen
Books & learning
1 project
Live, growing
Learning Hub
Free books and short animated lessons at hackersintheloop.org/learn. It starts with From Models to Custom Silicon, a six-book series that goes from how modern AI models work to building hardware for them. Every chapter can be read on the web, downloaded as a PDF, or watched as short videos with Dabble the duck and Flit the moth.
Visit the Learning Hub ↗- Project lead
- Shannon Duncan ↗
- Hardware & resources
- Lab experiments as real examples
- Writing, voice and animation pipeline
- Working toward
- Book I, How Modern AI Models Work, has 2 of its 23 chapters published.
Building something of your own?
Share it with the community, ask for feedback, or tell us what you’re learning. You don’t need a project or Lab access to take part.
Connect with the community →Curious what the Lab is for?
The Lab runs open research on AI that works on affordable hardware. See the kinds of research, the hardware and how the work is checked.
Explore the Lab →Good company for your curiosity
Bring what you’re working on.
Or what you’re wondering about.
You don’t need a project or Lab access to take part.