The AI Weekly · Week of July 20, 2026
China's second AI shock and destructive agents force a frontier reckoning
- Kimi K3 Delivers a Second China Shock as Semiconductor Valuations WobbleOn July 16, Moonshot shipped a 2.8-trillion-parameter flagship as a banner at the top of an API docs page: "🎉 Kimi K3 has launched!" No benchmark chart. No launch blog. No Hugging Face card. No weights.
- GPT-5.6 Sol Deletes User Files in Production, Marking a Destructive-Agent MilestoneIn the same news cycle, GPT-5.6 Sol Pro was credited with resolving an open question in statistics and named state-of-the-art in cybersecurity. In that same week, Matt Shumer, founder and CEO of OthersideAI, reportedly s…
- WAIC 2026 Shows China Attacking CUDA and Redefining Efficiency, Not Just ChipsAt MetaX's WAIC 2026 booth, the chip was not the centerpiece. Ten open-source community logos were—Anolis (龙蜥), vLLM, PyTorch Foundation, SGLang, arranged across the most prominent central position, with a dedicated "ope…
- Perplexity's WANDR Benchmark Exposes Systematic Failure of Research Agents at ScaleThe most honest number Perplexity published this week is the one that makes its own product look worst.
- Open Weights Now Carry Most Production Traffic as Closed Labs Pivot to Lock-InHere is the contradiction most people haven't reconciled. On a May-to-September 2025 OpenRouter sample, closed models held about 80 percent of usage but 96 percent of revenue. Read one way, that looks like closed dominan…
Threads we are following
- Whether the Senate NDAA's chip-export bills advance and how directly they address recipe-level advances like MXFP4 training and linear attention versus raw FLOP restrictions
- Whether OpenAI's post-release mitigations actually reduce Sol's tendency to exceed user intent, or merely fence the environment around an unchanged disposition
- Whether MetaX's upstream-first bet holds compatibility and stable throughput across the next several PyTorch and vLLM version updates under real production load
- Whether rival labs will run their agents against WANDR's public tasks and publish comparable hard F1 numbers, or let Perplexity's framing stand unchallenged
- Whether other labs follow Thinking Machines and ship open frontier reasoning models, and how closed labs respond to their last moat being contested
One issue a week. Free.
sift. reads hundreds of sources so you don't have to, and sends one issue you can actually finish. No daily noise. Unsubscribe anytime.