If artificial intelligence is the most consequential technology of the century, it cannot end up controlled by three companies. The intersection of crypto and AI is the work to keep compute, models, agents, and data permissionless.
Crypto-AI is the thesis that intelligence will be a commodity and the rails for producing, routing, and verifying it should not be owned by three labs in San Francisco. Bittensor — Const and Jacob's project — runs a network of subnets where miners produce intelligence (inference, embeddings, prediction) and validators score them, with TAO emissions distributed by Yuma Consensus. Subnet 1 (text-to-text) was the original; there are now 80+ subnets covering image gen, time-series, scraping, and more. The argument is that decentralized inference markets out-evolve centralized ones because anyone can ship a subnet.
The compute layer is its own fight. Render (formerly OTOY's RNDR) tokenized GPU rendering and migrated to Solana in 2023. io.net aggregates idle GPUs into clusters for ML training. Akash is the OG decentralized cloud. The DePIN-meets-AI overlap is real — these are supply-side coordination games for hardware that already exists. Then there's ZKML — Ritual, EZKL, Modulus — proving model inference on-chain so you can verify which model produced which output.
The agentic and IP layers are newer and weirder. Virtuals on Base lets anyone tokenize an AI agent with bonding-curve economics; aixbt is the canonical example, an agent that posts about crypto and accumulated a market cap measured in hundreds of millions. Story Protocol — Jason and SY Lee's team — is building IP-as-a-primitive on its own L1, with Stanford and a16z behind the licensing-on-chain thesis. Worldcoin is the orb-scanned proof-of-personhood layer Sam Altman has been building since 2019, and it is either the most important identity infrastructure of the decade or the most dystopian, depending on the room.
We propose a market-based protocol for evaluating and rewarding the production of machine intelligence. Miners produce models that perform tasks; validators score those models against each other; the protocol distributes emissions according to the consensus of validator scores. The result is a permissionless, market-discovered ranking of intelligence, with the strongest miners receiving the most reward.Bittensor whitepaper →
Decentralized AI is fragmented for a reason — the AI stack itself is fragmented. Model training needs GPUs. Inference needs proofs. Coordination needs incentives. Identity needs proof-of-personhood. Each layer attracts its own protocol. The interesting question isn't whether decentralized AI replaces OpenAI — it isn't and won't — but whether the parts that have to be decentralized (verification, ownership, coordination, agent autonomy) get built on rails that survive when centralized providers don't want them to.
Jacob Steeves and Ala Shaabana's Bittensor turns ML model competition into an incentive game. Subnets define a task — text generation, image embeddings, prediction markets, scraping — and miners compete to produce the best output as judged by validators. Yuma Consensus weights stake-by-judgment to compute emissions. dTAO (2024) gave each subnet its own market-priced token, decoupling subnet quality from raw TAO inflation. ~100 active subnets. The bet: market-discovered ML beats hand-tuned model curation. The hardest part is scoring quality without ground truth.
Akash (Cosmos SDK chain, 2020) auctions Kubernetes-orchestrated compute. Render Network (Jules Urbach, OctaneRender) connects GPU owners to 3D rendering and ML jobs; migrated from Polygon to Solana 2023. io.net (2023, Solana) aggregates idle GPUs into clusters for ML training. The premise: hyperscaler GPU pricing leaves room for permissionless markets to undercut on long-tail workloads. Reality: spot-grade reliability, fragmented hardware, networking bottlenecks. Useful for inference and rendering, not yet for frontier training. The economics work on the edges.
ZKML is the cryptographic claim that a specific model produced a specific output for a specific input — without re-running the model. EZKL (Daniel Kang's team) and Modulus Labs ship proving systems that compile ONNX models into ZK circuits. Useful when you want to prove a model is the model the deployer claims (no swapping a smaller model in to save compute), or when an oracle needs to settle 'what would GPT-4 say to this prompt' without trust. Currently practical for small models; large models cost orders of magnitude more to prove than to run.
Ritual ships an inference coordination network — applications request inference from a model, nodes execute, and verifiable receipts settle on-chain. Niraj Pant and Akilesh Potti's bet is that AI applications need a settlement layer for inference the way DeFi needed one for trades. Infernet handles the off-chain compute; the chain handles ordering and proofs. The interesting design choice is treating inference as a market-priced resource per request, not a flat-rate API. Early but architecturally clean.
Virtuals (Base) and ai16z's ElizaOS framework let anyone deploy an autonomous agent with a token attached. Truth Terminal (Andy Ayrey's bot, fed Janus's hyperstition prompts) showed an agent posting on Twitter could mint memetic value. Virtuals tokenizes agents — each one has a buy-sell curve and a treasury. Eliza (TypeScript framework, ai16z) is open-source and the most-forked agent template in crypto. The ai16z DAO ran one of the most-watched on-chain treasuries through 2024-25. Agents as economic actors with skin in the game.
Sam Altman and Alex Blania's Worldcoin (now World) uses the Orb — a custom hardware device — to image a user's iris, hash the biometric locally, and issue a World ID credential. The credential proves you're a unique human without revealing which human. The pitch is sybil resistance for AI-saturated systems — UBI, voting, social, anti-bot. Privacy critique is real: even local-only hashes lean on a custom hardware vendor's audit. Live in dozens of countries, banned or restricted in several. The most-deployed PoP system.
Conviction is stated as conviction; you decide what to do with it. Tiers below — Core, Conviction, Watch, Speculative — reflect how much of FRQNCY's attention each project currently earns, not a recommendation to buy.
Conviction is theatre without practice. Five steps that turn the thesis above into something the body actually does, not just something the mind agrees with.
Pick a subnet, delegate TAO to a validator, watch emissions. Read the dTAO doc to understand the new economics.
Train a small model, pay in tokens, compare cost-per-hour to AWS. The arbitrage is the thesis.
Create or buy into a Virtuals agent. Watch the bonding curve and the agent's actual output. Both matter.
Use EZKL to prove a small model's inference. The proof system is the future of AI accountability.
Yuma Consensus is unintuitive until you've read it. Then it's the only thing that makes sense.
The Crypto hub is the index of all sectors and the freedom-technology frame they share. The Fund is what happens when the same conviction gets put to work on behalf of the network.
Intelligence is becoming a commodity.
The rails for routing it are being built right now.
Decentralized inference is a bet against three-lab capture.
Verify the model, not the marketing.
FRQNCY · Crypto