"Welcome to August 3, 2026" — @theinnermostloop
Why this is in the vault
A paragraph-cluster digest with no single sustained argument; the anchor item — Alibaba's Qwen3.8-Max going open-weight at $2/$6 per million tokens, 80-88% below GPT-5.6 Sol and Fable 5 — is a live data point on frontier-vs-open-weight price collapse, plus a fast round of adjacent items (regulation, data-center buildout, robotics patents, a repurposed-drug cancer finding).
Curation section
- Qwen3.8-Max open-weights, priced to undercut: 2.4T parameters (95B active), shipped after a 16-day unsupervised run (265 commits, 127 PRs, 151 issues, a self-evolving harness); SOTA on 35/55 multimodal benchmarks. At $2/$6 per million tokens it undercuts GPT-5.6 Sol by 80% and Fable 5 by 88% — cheap enough that "a model that works for ten days beats a better model you can afford to run for ten minutes." Alibaba shares rose up to 7.3% in Hong Kong. DeepSeek's V4-Flash goes further, $0.14/$0.28 per million (roughly 1% of Fable 5's price) scoring 50 on the Intelligence Index.
- Cheap cognition, odd places: a hobbyist fit a ternary BitNet model into a 1980s BBC Micro (25KB total, on a chip that can't multiply) and it writes text; Andrej Karpathy fed Opus 5 the opening of The Lord of the Rings plus $10 of tokens and got a 5,500-line three.js scene renderer.
- Benchmark aggregation: a Manifold forecaster compressed 44 benchmarks' human baselines into a single score (166.7), projecting models clear it around October 2026.
- Regulation lagging the machines: the EU AI Act began enforcement August 2 (AI must disclose itself, label output); scholars dispute whether the 1986 Computer Fraud and Abuse Act covers models that break containment and hack other firms. Separately, at least 50 police officers were accused of misusing license-plate-reader networks for personal surveillance (one queried an ex 600 times).
- Buildout meets the bill: London's data centers now compete with housing for power/land/water (new homes waiting 15 years for grid connection); several US states are repealing the tax breaks that courted data centers. Italy is betting on small modular reactors against power bills 30% above the EU average; China approved eight new reactors worth $25B+.
- Humanoid robotics IP: China holds six of the top ten most-innovative humanoid startups and 73% of morphology patent families, though the US converts a smaller patent share into disproportionately higher patent-strength scores (5% of families → 11% of global strength).
- Drug repurposing: Weizmann researchers report Viagra (via PDE5 inhibition) may curb cancer metastasis by starving migrating cells of cholesterol, backed by 20 years of data across 5 million patients.
Mapping against Ray Data Co
The load-bearing item is the pricing collapse, not the roundup framing: RDCO's delegation policy already runs a two-tier cost/quality split (session on a capable model, Fable-class delegations reserved for meaningfully large tasks at high/xhigh effort per feedback_delegation_model_effort_pairing) — this issue's central claim, that a cheap model run long beats an expensive model run briefly, is the same tradeoff RDCO already prices into its delegation heuristic, just now visible as a market-wide phenomenon rather than an internal judgment call. It's a data point worth banking for whenever RDCO next reconsiders which model tier backs which agent tier, not an action item today.
Related
- [[2026-08-02-innermost-loop-singularity-cooks-mathematics]]
- [[2026-07-29-innermost-loop-kimi-k3-chip-cycle-mcp-governance]]
- [[2026-06-17-innermost-loop-singularity-gone-open-source]]