FLOPs: The Economics of Open Intelligence

Over the past two years, crypto's loudest projects learned the hard way that attention without production decays fast. Tokens launched, unlock charts bled, and "infrastructure" that never touched real users turned out to be little more than extraction machines in prettier clothes.

Meanwhile, AI became the world's most valuable commodity—and its most centralized. Compute, data, and model weights are bottlenecked behind a handful of clouds and labs. When power centralizes, censorship is not a risk; it's a reflex. Regulations aimed at "safety" increasingly hard‑code what models may say, ship, or share. In that world, liquidity pools and memes won't save us. We need working systems that route capital to productive compute and push intelligence out to the edges.

Why Decentralized Training (DeAI) Is Inevitable

Centralization → Chokepoints → Policy control. The EU's new GPAI Code of Practice and guidance under the AI Act formalize obligations for general‑purpose models (documentation, copyright reporting, even systemic‑risk controls), with compliance starting August 2, 2025 for key classes of models. That's governance at the model layer via a handful of large providers and deployment venues. In the U.S., Executive Order 14110 and subsequent Commerce reporting rules move in the same direction—compute thresholds and disclosure as levers over capability.

The antidote isn't ignoring safety. It's architecting systems that are resilient to capture, much like the internet itself. That means training and post‑training that can proceed over heterogeneous, globally distributed hardware with orders‑of‑magnitude lower communication, on permissionless rails that any community can extend.

Centralization Is Censorship by Other Means

When inference and training live in a few stacks, policy meets product overnight. The EU's AI Act guidance, U.S. reporting rules, and the rapid cadence of model‑behavior rulebooks for "acceptable speech" prove the point: control the hosts, and you control the outputs. Whether you call it "moderation," "risk mitigation," or "model spec," the mechanism is the same: centralized chokepoints.

Decentralized training doesn't abolish safety—it distributes it. Communities can form their own governance and proof regimes; regulators can focus on provenance and use rather than pre‑clearing every weight update; and innovation happens in the open, not at the grace of a platform.

The Economics of DeAI (and Why FLOPs Matters)

Compute is the commodity; coordination is the product. Across every DeAI effort, the same economics recur:

• Communication is the tax. Techniques like DisTrO, NoLoCo, and DiLoCo‑style protocols reduce bandwidth from gigabytes/step to megabytes or less, flipping the constraint from "co‑located H100s over NVLink" to "stable consumer internet."
• Verification is the contract. Work‑proofs (Gensyn) and trustless rollout verification (Prime) generalize Bitcoin's insight: pay for correct work, not promises.
• Heterogeneity is an advantage, not a bug. Systems that treat churn, mixed GPUs, and varied bandwidth as first‑class capture more supply and lower the clearing price for useful compute.

Where FLOPs Fits in the DeAI Stack

Why This Is Also a Cultural Bet

We don't buy the idea that a handful of labs should write civilization's speech rules in model weights. The better story—the one that spreads—looks like this:

• Open coordination on a fast chain (Psyche's Solana‑based orchestration is one example) so no single switch kills a run.
• Swarm learning as default, whether it's RL‑Swarm peers learning together or model‑parallel over everyday links.
• Proofs over trust, on‑chain and off‑chain, so reputation compounds in public.

How FLOPs Earns Trust (and Keeps It)

What We'll Fund First

• Compute campaigns across Gensyn (work‑proof training), Nous Psyche (DisTrO‑native orchestration), and Prime (fault‑tolerant global training).
• Research bounties to productionize model‑parallel over the internet (Protocol Models / "Pluralis" line), and to harden gossip optimizers on flaky links.
• Verification + telemetry so contributors and token holders can see, in near‑real time, how much compute was bought, where, and what it returned.
• Open‑source grants for adapters that let small labs and indie builders plug into these networks without bespoke infra.

Risks We Acknowledge

The Closing Argument

If the last cycle taught us anything, it's that reputation > information. Tweets are free; capital at risk is not. The way to rebuild trust is to point dollars at engines that produce—and prove it.

DisTrO, NoLoCo, and Protocol‑Model work tore down the myth that only super‑clusters can train frontier models. Gensyn, Psyche, Prime and others have shown the outlines of a permissionless training internet. The EU and U.S. have shown us what happens when power centralizes. The choice is plain:

FLOPs is the attention‑optimized gateway to decentralized training. Credits turn USDC into compute and raw DeAI rewards; the token turns community conviction into a treasury that multiplies those effects. It is not a promise of riches; it is an instrument to aim ingenuity.

If you believe intelligence should be widely owned—and that economies work best when trust is earned in the open—then help us route the next dollar of compute.