Hands-on AI experiment · November 2026

Pooled: Exploring Distributed AI Inference

Testing whether everyday devices can contribute browser-based compute to a peer-to-peer AI inference setup.

← Back to Jason Harvey’s portfolio

Experiment in progress

What we set out to explore

This experiment explores Pooled, an open-source project that investigates peer-to-peer, browser-based distributed AI inference. The idea is to connect devices so that available resources may be pooled rather than relying on one machine alone.

The aim was to test a Windows desktop and an iPhone as potential peers, understand the connection requirements, and distinguish what the interface reports from what has actually been proven.

Credit: Pooled is the work of Nehanth Narendrula. This page documents our hands-on testing and observations; it does not claim authorship of the Pooled project.

Test setup

Toucan · Windows PC

Ryzen 5 3400G desktop with AMD Radeon Vega 11 graphics. Used as the first browser peer.

Sparrow · iPhone

Used as a second device to investigate joining the same peer session from a mobile browser.

Browser connection

The local development/localhost route was not the successful route. The working resolution was to open the hosted Pooled website over HTTPS (pooled.run) on both devices. Using that live website, the Windows PC and iPhone successfully paired in the same Pooled session.

What we observed

Important: The displayed combined capacity is an interface-reported allocation, not proof that the iPhone successfully processed model layers or that inference was completed across both devices.

Current conclusion

Successful result: We opened the hosted Pooled website over HTTPS on the Windows PC and iPhone, shared the invite, and both devices successfully paired in the same session. This was the working resolution; we did not need to make a separate test page.

However, we have not yet verified that model layers or inference work were actually distributed between the PC and iPhone. Pairing confirms that the devices joined the session, not that both devices processed parts of a model.

To call the experiment successful, the next test needs to demonstrate a real inference request, confirm that work is distributed between the peers, and capture evidence such as device-level activity, logs or a reproducible output. Until then, this remains a connectivity and capacity-allocation experiment.

Next steps

Project references

For the upstream implementation and live project, see the original Pooled resources: