Lab
142 explorable, interactive pieces — every concept we write about, made tangible.
The same 250-document backdoor that compromised LLMs from 600M to 13B params, plotted across model scale. Toggle 'documents needed' — a flat 250 vs the myth that poison scales with data — and 'share of training set', collapsing to 0.00016% and below. Drag to pick a scale.
How many audit probes it takes to catch a substituted model, as a function of how subtle the swap is. A 1/Δ² curve over real quantization accuracy gaps: a model swap is caught in a few hundred probes; FP8 needs ~150k and falls off the cliff into the economically-invisible zone.
Five hops — Human → Orchestrator → Sub-Agent → MCP Tool → Chain Call — and zero authentication in the baseline. Toggle No Auth, IBCT Compact, or IBCT Chained; tap any edge chip to see what its token proves, what it carries, and what it can't verify. Based on AIP arXiv:2603.24775.
Cost per million tokens and throughput for Llama 3.3 70B on 8x H100 as the batch fills. A single-tenant node (B=1) runs ~25 tok/s at ~$258/M; a filled batch (B=256) runs ~2,800 tok/s at ~$2.30/M — same silicon, 112x cheaper. Pick a batch; read the tax you pay for the GPU you can't fill.
20 weak agents bid for tasks; bankrupt agents are replaced by mutations of the wealthiest survivors. Watch economic selection raise collective performance past a stronger monolithic baseline — the Economy of Minds mechanism. Toggle difficulty, step through rounds, observe wealth cycles.
Grouped bar chart comparing BN254 vs BLS12-381 on-chain gas costs for every precompile operation: G1/G2 adds, scalar mults, and pairings. Log scale reveals the G2 MUL unlock and the counterintuitive pairing inversion.
Push the cost-of-corruption inequality for a restaking-secured AI oracle: set bonded stake, slash rate, audit probability, and value-at-stake, then add AVSs to watch the overloading attack flip the verdict from secure to exploitable.
Three interactive attack surfaces on an ElizaOS-style Web3 AI agent context pipeline — prompt injection, memory injection, and data feed manipulation — with ASR figures from CrAIBench.
Bittensor's Yuma Consensus scores a miner by the stake-weighted median of validator weights, then clips everything above it. Validators sit on a cumulative-stake axis; the consensus line is read at the κ=50% mark. Push a colluding bloc past 50% and watch the pump go from clipped to unstoppable.
Six commitment schemes plotted on proof size vs. trust model: KZG (EIP-4844 blobs), IPA/Pedersen (Verkle state), Groth16, PLONK, FRI-STARK, and Merkle Patricia Trie. Click any dot to inspect.
See exactly how DeFi overcollateralization locks capital as a credit tax. Adjust loan amount, collateral asset, and credit tier to watch the cage shrink — or disappear entirely — as on-chain reputation improves.
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.