The Scientist

Uber's engineering team published a factor-level cost model for agentic workloads — users x sessions x turns x requests x tokens x price — reporting cost per 1,000 requests down 34% and cost per session down 52% from peak. Requests grew 9.4x since February, so absolute spend still rose roughly 6.2x. For you, that means every unit-economics win you report is compatible with a budget that tripled, and finance reads the second number.

Tencent open-sourced Hy4 under Apache 2.0: 770B total parameters, roughly 49B active per token, 78 layers, context past 1M tokens, hosted at $0.834 per million input tokens. Its human-preference lead over Kimi K3 and GLM-5.3 is 0.05 points on a 4-point scale across 203 tasks judged by 163 experts, which is inside noise. Treat the three checkpoints as tied and let your own paired eval, not the release note, decide routing.

NPR posed 30 nation-state false narratives to six chatbots in both neutral and leading forms with web access enabled; the aggregate debunk rate was about 75%, and search-summary surfaces performed worse than conversational chat on the same claims. Same model class, different pipeline. Your product surface is the retrieve-and-summarize one, so a factuality number measured at a bare model endpoint describes a system you do not ship. At n=30 the interval is too wide to rank any model against another.

Ollama v0.33 adds a one-toggle gateway that surfaces local Qwen, DeepSeek, Kimi and GLM models inside Claude Desktop's model picker, running locally or through Ollama Cloud with telemetry off by default. Two engineers comparing 'the same prompt in Claude' can now be on different models, quantizations and sampler defaults with no request log. Separately, a drafted U.S. rule on China's remote access to chips would turn weight origin and host region into auditable procurement fields.

a16z crypto argued Argentine stablecoin use crossed from crisis hedge to habit, then reported that USDC contractor-pay share and year-over-year inflation both sit at roughly one fifth of their peaks as of July 2026. A preserved ratio between two series is continued co-movement, not decoupling. The reusable asset is the April 2025 liberalization of dollar purchases: a sharp, plausibly exogenous date that belongs in your drift monitoring as a labeled regime break and works as an interrupted-time-series template.