Training, inference, optimization, and the engineering of learning systems.
Speculative decoding accelerates LLM inference 2–4× by splitting work between a cheap draft model and an expensive verifier. When those run on different nodes in a decentralized network, three properties of the protocol that are trivially safe on one machine become open attack surfaces.
Majority voting counts hands — it can't rank quality. Fortytwo's Bradley-Terry protocol achieves 85.90% on GPQA Diamond vs. 68.69% for majority voting, extracting ranking signal from pairwise comparisons and making Sybil attacks economically unattractive.
BERT-base under FHE takes 149 seconds, needs 112 GB, and sustains 0.007 samples/sec. Best-case MPC takes 2.3 seconds and 4 GB. The gap traces to one operation: bootstrapping — the noise-refresh that makes FHE theoretically unlimited but practically expensive.
Uniswap v3's capital efficiency multiplier peaks at 20× for ±10% ranges, but first-passage math shows the same position rebalances 54 times a year — making active LP management an optimization problem that requires AI-grade execution frequency.
Covenant-72B pre-trained a 72.7B-parameter LLM across 70+ anonymous internet peers. The key was SparseLoCo: 1.56% gradient density plus 2-bit quantization achieves 146× compression, collapsing the required uplink from 1.9 Gbps to 110 Mbps — consumer fiber.
DiFR's seed-commitment round-trip is the last synchronization barrier to async verifiable inference. TOPLOC removes it: a locality-sensitive hash of intermediate activations compresses to 258 bytes per 32 tokens — 1,000× smaller than raw embeddings — and validates faster than inference ran.
LLM non-determinism has been the blocker for output-based verification. DiFR commits to a random seed before inference — and activation fingerprints built from random orthogonal projections detect 4-bit quantization in just 2 tokens at AUC > 0.999.
DeFi interest rate formulas are open-source reward functions. An RL agent reading Aave's two-slope model can compute the optimal adversarial strategy analytically — no learned reward model required.
Federated learning keeps raw data local — but the gradients it publishes on-chain are not private. Inversion attacks reconstruct training images from them, and a public ledger makes that window permanent.
Controlling one intermediate pipeline stage is enough to inject a backdoor into a decentralized post-training run — 94% ASR, no data access required. The mechanism, why training loss stays clean, and what it means for networks fine-tuning LLMs across untrusted nodes.
Softmax costs 275 ZK constraints per element vs 4 for MatMul. ZK-DeepSeek proved Transformer inference is expressible in SNARKs. Lookup arguments (Lasso, LogUp) explain how — replacing in-circuit transcendental approximations with pre-committed tables cuts nonlinear costs 10–20×.
Any permissionless network that rewards the best gradient contribution faces a Nash equilibrium where every rational miner copies instead of computes — same reward, zero cost. Gauntlet's commit-reveal mechanism closes that trap, and already trained a 1.2B LLM on Bittensor with real token payouts.