Archive
18 artifacts published in this period.
In decentralized RL the rollout worker acts with a policy several steps behind the trainer. Staleness g = broadcast time / step cadence. Pick a model and link, toggle sparse deltas, and watch g cross INTELLECT-2's demonstrated 4-step budget — full 32B weights blow right past it.
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.
Under secure two-party inference, a transformer's cost is eaten by Softmax and GELU, not the matrix multiplies. Stacked per-operation communication and interaction rounds for one BERT-base block across Iron, BumbleBee, and Nimbus — the nonlinear activations are 80–95% of the bill.
Proving an agent's top-k retrieval is real costs a sort over every scanned candidate, unless you prove a boundary instead. Drive top-k and the corpus preset; watch the in-circuit comparison count split into the flat boundary proof and the sort tax you skip, anchored to V3DB's measured 22x.
A chain commits to a model's 32-byte hash for cents; keeping the gigabytes it points to retrievable is a separate, recurring bill. Pick a model and horizon and watch three rent-charging storage layers race Arweave's pay-once permanence to a crossover — plus the on-chain byte cost and cold-load wall.
Run a speculative-decoding round at a time: a cheap drafter proposes γ tokens, the target verifies them in parallel and accepts a prefix plus one free token. Drive α, γ, and drafter cost c; watch the accepted length converge to Leviathan's Ω and the speedup peak, then fade.