Looking back 5 years from now I'm thinking this may be the most important video of the series

Sums up the major theme of OpenAgents and we think the solution to a lot of the AI chaos

How many decentralized inference power users are there?

Sadly not enough to build a big business around!

For now our approach will be to let others solve the core technical issue and absorb that into our code when we need it

For example last week this CommitLLM project came out with a proposed solution some people seem excited about

We ask Codex to audit their approach, compare to what we have already in Psionic, propose integration path etc.

That comes up with a decent analysis:
https://github.com/OpenAgentsInc/psionic/blob/main/docs/audits/2026-04-09-psionic-commitllm-adaptation-audit.md

May or may not proceed with that specific plan but will repeat the process whenever we need that level of verifiability, then port the code into Psionic and iterate as needed

Generally I don't expect verifiability to be a big enough selling point that people will prefer a different project over ours because they verify more than we do. None of the AI power users on my X feed care about Gensyn or any of the other projects prioritizing inference verifiability.

Not to trivialize its importance, just rather focus on network growth first & upgrade later. (Borrowing from Nostr's 'worse is better' playbook)

Episode Transcript: https://github.com/OpenAgentsInc/openagents/blob/main/docs/transcripts/221.md

Summary:

Pylon is presented as a lightweight compute miner that lets people sell spare computer power for Bitcoin. It runs as a node on a user’s machine, connects through Nostr as a NIP-90 service provider, and is meant to be easy to install through agent tools like Claude, Codex, or Cursor. The product naming is intentionally StarCraft-inspired: Probe is the coding agent, Pylon is the compute node, Nexus is the central relay layer, and Psionic is the Rust ML framework behind the broader system.

The immediate use case is simple: contribute unused hardware, handle AI workloads, and get paid in a built-in Bitcoin wallet. The current focus is lightweight inference, including Gemma models, while measuring what different devices can support reliably. The roadmap expands toward fine-tuning, embeddings, image generation, and especially decentralized training, with the larger aim of turning ordinary user hardware into a large open AI compute marketplace rather than relying only on major cloud providers and centralized labs.

The broader thesis is that AI will require an open global economic layer, and Bitcoin, Lightning, and Nostr are positioned as the stack for that future. The project argues that if open, Bitcoin-native infrastructure is not built now, closed companies and payment platforms will control the machine economy in the same way older financial systems controlled prior eras. From that perspective, Pylon is not just a utility for earning sats from spare compute; it is framed as the first practical step toward a decentralized AI ecosystem where open agents, open markets, and user-owned hardware compete directly with closed incumbents.

Inference jobs will go to the most appropriate device (some combination of lowest time-to-first-token / highest tokens-per-second / best reputation etc.) but we're focusing first on distributed training runs: going to try giving pieces of training work to all possible devices that can run them, but on't know for sure which devices will be able to realistically contribute good work to the training run until we start gathering live data from our Pylon network next week as we start pushing our training code live

Will share more data in our territory as we get it

We are customer #1 --- we want the compute for a distributed model training run starting next week --- will eventually open this up for others to use the same network for training & inference but not sure yet what the pricing will look like

Conceivably way cheaper than other inference/training/finetuning platforms given we are bringing new compute online that has historically had no price attached - we'll see!

We'll make a calculator eventually

At the moment we're just paying 2 sats every 20 seconds for basic heartbeats to identify an initial supply

Real training run starts Mondayish next week. We'll ramp up payments if needed to attact more compute faster

Price will fluctuate while we are the only buyer; the goal is to do very cool stuff (training runs / finetuning / agent-optimized inference) that attracts more buyers and we all bid for your compute

Primarily a distirbuted training run; we're gearing up to train our own "Psion" models

A bit about that here - more details next week: https://x.com/OpenAgents/status/2036908227019809259

For device support we've been focusing on 1) any Apple M chip (MacOS) and 2) any decent NVIDIA GPU.

Our model inference via Psionic now works well on those devices (for Qwen 3.5 & Gemma 4) and we expect those same devices to also be able to contribute well to a DiLoCo run, but won't know for sure which devices will be able to realistically contribute to the training run until we start gathering live data from our Pylon network next week as we start pushing our training code live.

We'll cover what we learn each week in our next video episodes, aiming to release every M/W/F.

Psionic is our Rust ML framework, source here: https://github.com/OpenAgentsInc/psionic

So far it's a glorified Rust port of relevant inference code from ollama/llama.cpp/MLX and training code from prime/bittensor etc

Pylon is our NIP 90 service provider that uses Psionic, all in a single Rust binary

Pylon gets assigned a job (via our job dispatcher "Nexus" a glorified Nostr relay / NIP 90 client- in our main monorepo OpenAgentsInc/openagents), processes the job through our own inference engine (not making an HTTP call to local Ollama a lot of people do, which is easily spoofable), and will send it back over the network probably with some verification salts/hashes showing it came from a real Psionic inference. This isn't fully built out yet, we'll focus more on it once we get the DiLoCo run going and we have new models we want to run inference on

Separately from best-effort programmatic verification, our Nexus job dispatcher will factor in NIP32 reputation events: untrusted nodes may get less jobs assigned until they build up reputation over time

Lots to solve here but wethinks we have the right primitives for it: helps to fully control every part of the inference/training pipeline so it's all in binaries we write -- can build verification into any part of it

Current plan is to only accept inference done through our Pylon software (using our Psionic ML framework so it's pretty easy to tell if someone's computer goes through that process or not, then open it up to the broader network once we have better answers re verifiability over untrusted networks.

We're focusing less on inference to start, mainly warming up to start a DiLoCo distributed training run next week, using similar code that Prime Intellect and BitTensor/Templar used for their distributed training runs-- that's even easier than inference to verify if someone's submitted work improves the run or not

Will do a deep dive video on this Mondayish

AMA btw