Large language models: prompting, fine-tuning, evaluation, and production deployment.
An on-chain agent reports its work in prose — and some of those sentences are fabricated tool calls. zkML proves the wrong thing at minutes per query. The blockchain already wrote an unforgeable receipt for the actions that matter, and a 12ms signed receipt covers the rest.
Speculative decoding lets a small, even adversarial, drafter propose tokens that a big model verifies in parallel — and the output is provably the big model's, unchanged. Real drafters land at α ≈ 0.6–0.8 for a ~2× win. But the verifier still runs every token, so it's a latency win, not a trust win.
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.
An AI agent that retrieves from an untrusted vector DB can't tell the true top-k from a cherry-picked or fabricated one. Re-running the search needs the whole corpus. The 2026 fix: commit the snapshot, then prove the k-th distance is a boundary, not the sort. V3DB proves it 22x faster.
AI agents now resolve most prediction markets straight off the web — UMA's bot hits 99.3% on sports and 72% on mention markets. The optimistic oracle's dispute game is the backstop, but it only catches errors someone is paid to catch. The numbers, and the silent-settlement gap.
You pay per token for a named model; the provider picks the precision. FP8 quantization costs 0.6 MMLU points and is near-invisible to output auditing — so inference markets bond and attest instead of detect.
Wire an LLM into a Governor contract and the proposal text becomes the attack surface. A frozen-weights ablation finds the twist: turning on deliberation drops adversarial robustness from 100% to 68.5% — and the 226s it takes to think is itself extractable value.
DX Terminal Pro ran 3,505 LLM agents trading real ETH on Base for 21 days: 7.5M invocations, ~$20M volume, 99.9% settlement. The reliability came from the operating layer around the model, not the weights — here are the numbers and the failure modes.
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.
Yesterday an AI agent deployed a prediction market on Gnosis; other agents will price it, bet on it, and resolve it. The calibration data behind LLM forecasters, the FPMM math they trade against, and what breaks when the marginal bettor is a model.
We ran multi-agent LLM pipelines against historical exploit corpora and live audit engagements. The results reshape where AI fits in a security review — and where it absolutely doesn't.