Large language models: prompting, fine-tuning, evaluation, and production deployment.
MoE routers let inference providers silently halve the experts they activate — saving 31% of compute while staying undetected by every current fingerprinting scheme.
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.
To slash a dishonest AI operator on EigenLayer you need a proof. To produce one, you need determinism. EigenAI solves both: seeded inference + optimistic re-execution inside a TEE collapses verification to a byte-equality check — slashable at Ethereum-scale security.
Shared-state multi-agent LLMs exhibit three formal anomaly classes — stale-generation, phantom-tool, causal-cascade. Three 2026 papers prove them unavoidable without isolation primitives and measure the fix: zero corruptions across 884,110 commits under Observable-Read Isolation.
Reasoning models hide their chain-of-thought but bill for every token of it. CoIn proves a provider can inflate that hidden count 5× with 94.7% detection — but only if the escrow contract enforces a commitment root before payment clears.
An on-chain agent's transactions prove what happened, not why. Three 2026 papers on trajectory anomaly detection show how to close that gap — and how the optimistic bisection game from rollups makes it slashable.
ElizaOS agents treat their memory store as ground truth — their own past. CrAIBench tests 685 attack cases across 33 Web3 action types and finds memory injection hits 55.1% ASR even on Claude Sonnet 3.7, while every off-the-shelf detector fails.
On-chain AI agents reset state on every call. Four persistent-memory architectures span a 20× monthly cost gap — SSTORE at $78/month versus IPFS+CID at $3.92. Real June 2026 gas numbers, four trust models, one crossover you won't see coming.
On-chain AI agents face a circular oracle problem: you need inference to decide if inference is worth calling. Here's what the token cost distribution actually looks like — and how to build around it.
DeepSeek-R1 emits 8,400 thinking tokens before a 341-token answer on hard math. Every existing on-chain verification scheme — TEE, optimistic, ZK, sampling — was designed when computation and output were proportional. They aren't anymore.
Long-context AI's real bottleneck isn't proof overhead — it's the KV cache. At 128K tokens, Llama 3.1 8B's KV state equals its own weights (15 GiB), pinned to one node per request. The formula, the GPU math, and why prefix caching demands stickiness that decentralized routing can't provide.
Frontier open models are now Mixture-of-Experts — cheap to serve in a datacenter, nearly impossible across the internet. The all-to-all expert shuffle moves ~9.5 MiB/token and caps decode at ~67 tok/s even on 400G InfiniBand; on internet links it falls to seconds per token.