panes) as pre-history of the current Desktop/Omega split rather than a spec.

---

## Batch 2: Episodes 214-227 — Markets, Psionic, and the compute thesis

Airtime window: March 2026. Five weekly market launches begin (compute ships,

data ships thinly), the rollout pauses when Psionic redirects the buy side

toward decentralized training, Probe joins the stack as a Rust coding agent,

the propaganda podcast frames the network thesis, Pylon launches publicly,

the Templar collapse supplies miners, payments flip from online-presence to

verified ML work, bounties relaunch, worse-is-better becomes doctrine, and

ocean power enters as a phase-3 horizon.

### Episode summaries

#### 214 — Compute Market

Relaunches the 2023 product as an Airbnb-for-compute: sell spare Apple Silicon

for Bitcoin over NIP-90 kind 5050 jobs. Autopilot v0.1 desktop app ships the

same day with a Spark wallet unlocked by one BIP-39 seed that also derives the

Nostr keypair. A buy-mode bootstrap sends 2-sat dummy jobs to seed sellers.

Restates the five-market roadmap (compute, data, labor, liquidity, risk) plus

draft NIPs for skills, sovereign agents (Frostr key splits), and agent credit.

Foil math: 5.5 GW of idle desktop Macs against OpenAI's entire 2 GW.

#### 215 — Data Market

Second launch: monetize your data — especially Claude Code and Codex

conversation logs — through NIP-DS, an open Nostr dataset-trading spec

(kind 30404 listings, kind 30406 offers, delivery via DVM or NIP-90).

Redaction-first flow: bundle, redact, canonical digest, list, offer, deliver.

Host commits to paying more Bitcoin than anyone for coding-agent conversations

to train OpenAgents' own agent. On-camera demo failed to complete; Autopilot

Control super-CLI not ready.

#### 216 — Psionic

Pauses the market rollout: don't build supply without matched demand. Introduces

Psionic, the all-Rust ML framework whose inference engine already edges out

llama.cpp (172 vs 160 tokens/sec), produced by pointing Codex at libraries and

looping until they outperform. New direction: decentralized training of

OpenAgents' own Psion-class models, absorbing Prime Intellect and Bittensor

research ported to Rust, with Bitcoin bounties for ML engineers. Reproduces the

Percepta paper's executor-model direction.

#### 217 — Psionic: Fast Qwen 3.5

Single-metric obsession: tokens per second. Per-device, per-model custom CUDA

kernels take Qwen 3.5 0.5B from parity to 523 vs 328 tokens/sec overnight, then

repeat across the 2B, 4B, and 9B sizes. Community-driven targets; inference

framed explicitly as a side quest on the way to the large decentralized

training run.

#### 218 — Probe

Announces Probe, a lean fully-Rust embeddable coding agent in the OpenCode

tradition, timed pointedly against the same-day Claude Code source leak. Codex

is loved and borrowed from; MCP deliberately de-emphasized. Contribution rule:

"Please don't send PRs. Please send detailed issues" — ideas accepted, code

written by their own tooling.

#### 219 — Probe: Inference Modes

One day later, Probe demos three swappable backends: hosted Codex via ChatGPT

Pro accounts, Qwen-fast served by Psionic on a remote RTX 4080 over Tailscale,

and the local Apple Foundation Model. Thesis: route workloads across tiers —

frontier models for hard reasoning, small models for volume, local models for

trivial calls — estimating maybe half of Codex's roughly 1,000 tool calls per

run could offload. Latency bragging: instant Ctrl-C exit.

#### 220 — Propaganda Podcast

Launches the openly declared propaganda arm: billions flowing to NVIDIA and

clouds should train models on retail consumer compute paid in Bitcoin.

Cites Bittensor Templar's 70-person peer-to-peer pretraining as proof and sets

a 100x-participation goal. One-install personal super app folds the best open

source in. Moat thesis: contributors go where they are paid most Bitcoin;

networks cannot be copied. Revenue-share scorecard updated: OpenAI zero,

OpenAgents thirty.

#### 221 — Pylon Launch

Pylon ships publicly: paste instructions to your coding agent, install a NIP-90

provider with a built-in wallet, earn sats. StarCraft naming system codified

(Probe, Pylon, Nexus, Psionic; Archon teased). Frames the global economic layer:

closed players (Visa, Coinbase, Cloudflare, Stripe) are converging on stablecoin

rails, so Bitcoiners must build the open version or face a new SWIFT.

OpenAgents is buyer number one and invites outbids. Live counters show thousands

of completed jobs without refresh.

#### 222 — Templar Merge

Covenant AI rugs Bittensor's flagship Templar subnet (~$10M cash-out); OpenAgents

invites displaced miners into Pylon. Anti-token thesis vindicated: if someone

can rugpull, it is not decentralized; companies with vesting hold administrators

accountable. Agents ported Templar and Prime Intellect training code to Psionic

overnight; fleet-wide hill-climbing loops benchmark inference on every

participant machine. Network doubled twice in two days post-launch (52 nodes,

108k sats paid).

#### 223 — Pay the People

Social mood on AI turns negative; labs risk a generational fumble. Fix: pay

people directly for compute, data, and contributions instead of lobbying for

UBI. Audit of broken promises: Altman's DevDay revenue-sharing pledge versus

zero paid to developers (the $1.3B went to Microsoft); this series started the

day after that promise. Demonstrates heterogeneous-device training across M3

Max, old MacBook CPU fallback, and a remote RTX 4080 over Tailscale. Roadmap:

rebuild the full OpenAI/Anthropic product suite on this substrate — fine-tuning

API, continual learning as a service, image generation, embeddings.

#### 224 — Distributed Training 101

Training seminar on run day. Milestone: 1M+ sats paid across 1,300+ pylons.

Policy change effective immediately: no more pay-for-online; payouts follow real

ML work delivered. Weak devices get validator assignments; device-to-job

matching will be learned empirically and exposed through public stats and APIs

agents can crawl. Stanford's open language-models-from-scratch course becomes

network homework (assignment 1: BPE tokenizer, transformer, Adam optimizer;

assignment 2: Flash Attention in Triton, distributed data parallel, scaling

laws). DiLoCo architecture explained: heavy local work, occasional sync.

#### 225 — Developer Bounties

Revives the paid-bounties program with a permanent list at openagents.com/

bounties. Tours the whole suite: front door becoming a ChatGPT-style homepage,

Autopilot targeting Microsoft Copilot replacement, Nexus carrying a Bitcoin

treasury module, Probe replacing Claude Code/Cursor internally (still wrapping

Cursor under the hood), Forge as the internal software factory inspired by

Ramp's Inspect-agent writeup, mobile for voice control of Probes. Anti-spam

rules: unknown PRs auto-closed, vetting happens in humans-only Discord.

Pet-peeve bounties already shipped: visible account email, weekly-limit display,

multi-account switching on rate limits, API-key fallback.

#### 226 — Worse is Better

Doctrine episode: the New Jersey school beats the MIT school — ship the simple

viral thing, grow it to 90 percent of right. Nostr as modern proof; UNIX/C as

the original viruses. Contrast operating models: no stealth year, optimize for

virality from day one. Consumer compute is priced at zero today; put a price on

it and build value above it. Twenty gigawatts of consumer compute versus

OpenAI's two.

#### 227 — Ocean Power

Back from silence with Ben Silone hired for ocean tech. Long-term OTEC thesis:

monetize stranded ocean energy at source with Bitcoin mining; free seawater

cooling; floating cities as phase 3 (100 MW minimum ambitions; Panthalassa

critiqued as too small). Immediate business: Pylon v0.2 within 48 hours, LDK

replacing a bottlenecked payment provider, Qwen fine-tuning on Harvey's legal

benchmark to position Autopilot as the neutral last agent — hyper-specialized

per-business agents, export-your-data always, pricing barely above compute.

Anthropic named most dangerous company in the world; "print eight billion rings"

versus fighting for the one ring.

### Batch synthesis and integration notes

This arc contains the economic core that later episodes refine into the

accepted-outcome economy: verified work, receipts, device-to-job matching, and

paying validators. It also contains the clearest early statement of the

issue-not-PR participation model that the forge now enforces literally.

Dispositions:

1. **Live.** Pylon node economics and Psionic (Tassadar runs on both);

   Probe's backend-pluralism (now the SCV driver question: hosted versus

   self-hosted executors); verified-work payouts replacing presence payouts

   (the work-job receipt model); public stats pages; worse-is-better shipping

   culture; anti-spam contribution rules (now membership-gated push).

2. **Dust off.**

   - **NIP-DS data market**: shipped thin, then abandoned when training became

     the buyer. The buyer now exists (training traces, Ox Alpha logs). A

     dataset-listing surface tied to trace visibility policy would revive it

     without new protocol work.

   - **Curriculum-as-bounty**: Stanford-homework assignments as paid network

     work is directly portable to forge issues — a `bounty` label plus sats

     pricing would restart the contributor flywheel on the new tracker.

   - **Device-to-job matching APIs**: promised so agents could crawl earnings

     estimates; still absent. Fits the cloud-computer capacity work (issues

     #37/#38).

   - **Continual-learning-as-a-service API**: named twice, never built; the

     closest living relative is Tassadar's never-stopped run.

3. **Reimagine.**

   - **Five markets on weekly cadence**: compute proved, data half-proved,

     labor became work jobs, liquidity and risk never launched. Keep the

     shared-kernel idea; drop the cadence.

   - **The neutral last agent / all-in-one dashboard**: the anti-lock-in

     positioning survives in the front-door thesis, but the single dashboard

     is now several surfaces (forge, chat, projects, status). Revisit as a

     signed-in command-center surface rather than a new product.

   - **Legal fine-tune wedge (Harvey benchmark)**: domain-fine-tuning as a

     service remains a plausible Pylon revenue mode once training receipts are

     public.

4. **Retired.**

   - **Spark wallet inside Autopilot**: superseded by MoneyDevKit/LDK

     self-custody (episode 235).

   - **Buy-mode dummy-job bootstrapping**: gamed instantly by design; replaced

     by pay-for-verified-work.

   - **Wrapping Cursor under the hood in Probe**: dependency inverted since;

     own harnesses only.

   - **Pay-for-online mining**: retired deliberately in 224; do not reintroduce

     presence-based rewards anywhere.

Pivot watch: this batch treats GitHub as contribution surface (auto-closing PRs)

and predates the forge entirely. Any marketplace or bounty mechanics must be

restated against forge authority (repository membership, PAT scopes) rather than

GitHub primitives.

---