Training, inference, optimization, and the engineering of learning systems.
Llama-3.1-8B starts at 15.9% on MATH. Put it in an auction economy and it reaches 57.0% — beating a stronger monolithic baseline. The mechanism is Hayek: decentralized price signals, wealth accumulation, and economic selection pressure do what central orchestration can't.
ETH's 3.46% daily volatility buries a 0.1% trading edge so deeply that on-chain P&L needs 4,600 trades — 230 days — before you can tell skill from luck at 95% confidence. Process rewards fix it in 23.
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.
ZKLoRA lets a LoRA seller prove adapter compatibility with your base model without exposing the weights. The verifier checks in 1–2 s — but the prover bears 31–74 s per module, and the protocol is interactive, which puts trustless on-chain settlement just out of reach.
On-chain data markets promise to pay you for your data's value AND protect your privacy. Differential privacy is the only rigorous knob for the second — and one budget ε can't serve both. The utility collapse, the composition trap, and why no chain can verify the ε you were promised.
Give autonomous LLM agents a marketplace and they race each other to bankruptcy and get conned by Sybils. A 2026 study finds capability doesn't help — a 9B model beat frontier models 45× its size on economic alignment. On-chain, both failures get cheaper and worse. The mechanics, numbers, and fixes.
Train a model on the web the last models filled with output and quality decays — the tails go first, then the middle. Provenance is sold as the fix, but a chain proves who signed a blob and when, not that it's human or clean. Signed is not clean.
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.
The textbook market-making model quotes symmetrically around mid. On a perpetual, holding inventory pays or charges funding every hour — a deterministic drift Avellaneda-Stoikov never sees. A 2026 HJB model that prices it cut inventory risk 36-38% on Hyperliquid ETH and BTC.
Bond-and-slash verifiable AI assumes someone audits. But checking costs money and pays nothing, so a rational verifier won't — the verifier's dilemma. Proof of Sampling fixes it with a bounty, and the math says honest inference holds at a 0.74% sampling rate where nobody ever arbitrates.
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.
Decentralized AI networks have to turn many validators' subjective quality scores into one reward number. Bittensor's Yuma Consensus uses a stake-weighted median and clips the rest — robust to a <50% bloc. Then a parasite shows up: copying the consensus pays better than producing it.