Model watch · Methodology Trend watch
From the Aug 5, 2026 daily brief
Teortaxes ran a two-path check on DeepSeek's official disclosures for its late-2024 V3 model: a hardware path (chip count × realized per-chip compute × seconds) and a model path (active parameters × training tokens × the standard coefficient). The two came out 14% apart — explainable by the share of computation run at slower precision — and the verdict was that it "roughly checks out" (Teortaxes, July 21). We re-ran both equations and confirmed them, then added a third pass they didn't do: 2,048 chips × 24 hours × 55 days ≈ 2.7 million GPU-hours, within 1.5% of the roughly 2.66 million the V3 technical report itself discloses — an independent match; the "55 days" isn't a plug. So whenever any lab publishes "X chips, Y days, Z tokens of data": run both sides of the ledger; if they don't meet, something is being left unsaid — and none of this requires non-public information. Copy the key cut exactly as it stands: this proves the disclosed numbers for that final training run are self-consistent, not that DeepSeek has no more chips — outside estimates putting the organization's total holdings much higher are fully compatible with this check. Two different questions; don't merge them.
Just an email address, unsubscribe anytime. This is the only thing we ask of you.
The eval house Artificial Analysis runs its AA-Omniscience board on an unusual design: wro…
First, which yardstick produced "one point apart": the composite intelligence index from t…
Two lines from the K3 launch materials (verbatim, as relayed by Teortaxes): late in develo…
Overnight's 15 new papers are still unprocessed and this column has no single-day incremen…