Archive
136 articles, grouped by date.
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.
LLMs collude in on-chain auctions even when written rules forbid it — only enforceable, automatic penalties slash severe collusion from 50% to 5.6%. Here's how governance graphs make that concrete.
Probabilistic classifiers miss 30–40 % of policy violations. Lean 4 theorem provers and SMT solvers make certain guardrail tiers mathematically certain — here's how the four-layer policy stack works.
Training Llama-3.1-8B costs ~$6M. The weights are free to clone. Perinucleus sampling embeds 24,576 distinct fingerprints per model with no quality loss — and a Poseidon ZK commitment registers ownership on-chain for $0.017, enforceable by smart contract.
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.
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.
Most DeFi protocols are upgradeable, and most on-chain AI agents don't know it. Here's how proxy architecture invalidates agent assumptions — and a tiered defense framework to build around it.
Every EIP-4844 blob is secured by a KZG polynomial commitment — 48 bytes committing to 128 KB, verifiable in 50,000 gas. That efficiency required 140,416 people to collectively bury a secret for 69 days. Here's the math, the ceremony, and why Ethereum chose a different scheme for state.
Five builders now win 96.7% of Ethereum blocks. Their edge isn't compute — it's exclusive orderflow: 12% of transactions that generate 54.59% of block value. AI agents transacting in public have to understand where they sit in that hierarchy — and when a preconfirmation changes the math.
Cross-chain bridges have lost $2.5B to hacks because receiving contracts accept messages they cannot cryptographically verify. AI agents operating across chains inherit this exact problem. Here are the four trust models, their real security parameters, and a principled agent policy.
A flash loan gives an AI agent $1M with no collateral — and demands it back plus 0.05% before the block ends. The atomic callback constraint is both what makes flash loans safe and what forces an agent to solve its entire strategy before touching the chain.
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.