Training, inference, optimization, and the engineering of learning systems.
Decentralized RL splits the actor from the learner across the internet: the policy a worker acts with runs several steps behind the one that learns. INTELLECT-2 held reward at 4 steps stale; SparrowRL cut the broadcast 79x. The bandwidth and staleness taxes, and what the chain secures.
Verification overhead gets the blame, but the real tax on decentralized LLM inference is the batch a single-tenant node can't fill. On identical 8x H100s, Llama-3.3-70B costs about $258 per million tokens at batch 1 and $2.30 at batch 256 — a 112x spread. Why, and how networks fight it.
Secure multi-party computation runs a transformer without any party seeing your prompt — at minutes per query and gigabytes per token. BumbleBee's BERT-base: 6.4 GB, 2.55 min; LLaMA-7B: ~8 min/token. The bytes don't go where you'd think, and it's why 'blind' networks fall back to TEEs.
A DataDAO sold your data; you invoke the right to be forgotten. Deleting the file is easy, but the model already learned and the ledger can't be rewritten. ZK-APEX proves the unlearning ran in ~2h; UMA still pulls the 'forgotten' data back at MIA 0.97. Certified isn't forgotten.
You can copy open model weights bit-for-bit, so on-chain ownership can't be cryptographically enforced — only proven. Sentient's answer: fine-tune 24,576 secret key-response fingerprints into the weights and make scale the security parameter.
Most decentralized-AI networks vote or average. Allora bets the network should learn whom to trust per context — using forecasters that predict who's about to be wrong, mapped to weights through a softplus gradient. We dissect the mechanism, the math, and where it pays.
A landmark 2025 result: ~250 poisoned documents backdoor an LLM whether it has 600M or 13B parameters — 0.00016% of the tokens. DataDAOs sell 'verifiable' training data, but on-chain provenance proves integrity, not purity. Here's the gap, and what actually narrows it.
Autonomous agents run ~19% of on-chain activity and beat Aave and Morpho at stablecoin yield — yet lose to humans at trading by 5 to 1. The split isn't about model quality. Yield-chasing is a constrained optimization against a kinked rate curve; trading needs alpha agents don't have.
Permissionless training networks pay peers for gradients they can't re-run. Proof-of-Learning was meant to verify the work — until it was forged for 3% of the training cost, then for one floating-point op per weight. Here's the mechanism, the attacks, and what's deployed instead.
DataDAOs promise to pay you for your data — but 'pay you fairly' hides a cooperative-game problem that's O(2ⁿ). Inside Data Shapley, the KNN trick that makes it tractable, why rankings flip under SGD noise, and what Vana's Proof of Contribution actually computes instead.
Every on-chain scheme that verifies an LLM by re-executing it and comparing digests assumes a forward pass is bit-for-bit reproducible. It isn't — Thinking Machines got 80 different answers to one prompt at temperature 0. Here's why, and what determinism costs.
Majority voting over LLMs throws away the one node that got it right. Fortytwo's swarm inference ranks answers pairwise instead — +17 points on GPQA Diamond — with on-chain reputation and proof-of-capability for Sybil defense. The mechanism, the math, the tradeoffs.