russell@unturf.com
211bbb1daf
Main tweet + follow-up (Merkle cache hit skips GPU, doesn't increment the per-1k). Numbers are the measured claim_lattice figures from the energy report; note attached to hold the arbitrage/forcing-function framing until the value side is hardened (higher N + blinded judge).
1.4 KiB
COGS tweet — cost of a grounded answer (arborist)
Draft social post. Numbers are the measured claim_lattice figures from
docs/energy-cogs-benchmark.md (n=30, $0.33/kWh, real GPUs: Hermes-8B on
a 3090, Qwen-27B on a 4090). Hold the bigger "arbitrage / forcing
function" framing until the value side is hardened (higher N + blinded
SOTA judge) — see that report's §5.5 + §9.
Main tweet
What does a grounded answer actually cost in GPU electricity? We measured it on real cards:
- arborist + Hermes-3-8B → ~9¢ per 1,000 answers
- arborist + Qwen-27B → ~16¢ per 1,000
No reasoning chains (those burn 4–6×). Read the evidence cheap, write a short answer locked to a claim lattice, stop.
Follow-up
And if an answer's already hot, it never even joins that 1,000 — it's a Merkle-bound cache hit that skips the GPU entirely. Zero joules, zero cents. You pay GPU only for new questions; the cached answers are free forever, and provably the same answer.
Source numbers (claim_lattice, $0.33/kWh):
- Hermes-3-8B + substrate: $0.085 / 1,000 grounded answers (~9¢)
- Qwen-27B + substrate: $0.158 / 1,000 (~16¢)
- quote mode is cheaper (Hermes $0.070, Qwen $0.121) — less context prefilled
- thinking/reasoning mode measured 4–6× the energy for the same answer
- cache hit (Merkle-bound providence record) returns the answer with no LLM call → 0 GPU joules, does not increment the per-1k cost