Archive
10 artifacts published in this period.
Drive the Proof-of-Learning trap: a logged training trajectory, audited top-Q by re-execution against a tolerance δ. Forge the run and it still passes — the forger tuned every interval under δ for 3% of training cost. Tighten δ to catch it and the honest node fails first.
Split one DataDAO reward pool across five overlapping contributors three ways: equal split, leave-one-out, and exact Shapley value over all 2ⁿ coalitions. Drag quality sliders, drop the unique contributor, or flip on the Sybil attack and watch equal split overpay a cloned wallet.
The move-or-hold call a stablecoin yield agent makes all day, on the live Aave V3 USDC curve from Base: supply capital into a higher-quoted pool and watch your own deposit slide the realized APY down the kinked curve, then race the spread against gas. Shallow pools and L1 gas flip MOVE to HOLD.
Where reliability comes from in on-chain LLM trading agents, over DX Terminal Pro's production run: model upgrades move swap success 87% to 96%, but the operating layer takes the same weights to 99.9%. Switch panels for the prompt fixes that erased three failure modes. Tap or arrow the bars.
Why memory injection beats prompt injection against on-chain AI agents, in three switchable charts over CrAIBench: 685 attacks skewed to trading, a 55.1%-vs-0% gap on the strongest model, and the fine-tuning defense that drops attack success 85.1% to 1.7%. Tap or arrow the bars for numbers.
Forward-pass gas for five neural nets run inside the EVM, on a log axis against the live Ethereum (60M) and Base (400M) block-gas ceilings. Toggle ML2SC's measured ~106k gas/edge vs an optimized ~1.5k floor, slide the gas price for a USD readout, and watch an MNIST net blow past 180 Ethereum blocks.
One frozen Qwen-3.5-9B judging adversarial Optimism DAO proposals, deliberation off vs on. System 1 holds every metric at 100%; reasoning drops robustness to 68.5%, flips 27% of verdicts, and costs 17x latency — a 226s window that is itself Governance-Extractable Value. Tap a row.
Why re-executing an LLM doesn't reproduce it. One matmul reduces to the logit gap between 'Queens, New York' and 'New York City'; drag the production split-K and watch IEEE-754 rounding flip the token at split 5 and 6, breaking the verifier's digest check. Batch-invariant kernels turn it green.
Where decentralized training nodes sit on Earth and what it costs to sync them. Toggle INTELLECT-1's three measured configurations and watch the ring all-reduce span the globe: 103s inside the USA, 382s across the Atlantic, 469s once Asia joins.
One hard question, a swarm of LLM nodes, two ways to agree. Run a round: majority voting picks the most common answer while peer-ranked Bradley-Terry consensus surfaces the best one, and the scoreboard converges on the gap Fortytwo measured. The evaluation-edge slider shows why.