The AI Weekly · Week of July 9, 2026
Open weights and Chinese silicon squeeze the frontier while agent reliability stalls
- Grok 4.5 Joins the Coding-Model Price War as Open-Weight Rivals Undercut EveryoneThe same enterprise workload costs $4,811 on Anthropic's Claude, $3,357 on GPT-5.5, $1,071 on DeepSeek V4, and $544 on Zhipu's GLM-5.2. That's a ninefold gap from top to bottom for work that, increasingly, produces indis…
- Chinese Labs Go Vertical on Silicon as DeepSeek Builds Its Own Inference ChipIn the spring of 2026, the chief executive of a Shenzhen AI chip startup moved into a budget hotel next to SMIC's Southern Semiconductor fab. His inference processor had passed validation, its software stack fully optimi…
- Rigorous Benchmarks Reveal Agents Fail Real-World Expert Reasoning Despite More ComputeArticle — "Rigorous Benchmarks Reveal Agents Fail Real-World Expert Reasoning Despite More Compute": Here is the counterintuitive part: the more compute you throw at an expert-level agent task, the more you learn that co…
- Harness Engineering Emerges as the Real Lever for Self-Improving AIThe bill came to roughly $165,000 in tokens: 5.9 billion uncached input tokens, 690 million output tokens, 72 billion cached reads. What it bought was a rewrite of Bun—Jarred Sumner's JavaScript runtime, more than a mill…
- AI-Forward Practitioners Push Back on AI Slop in Production Communication$149.25 is reportedly what it cost Simon Willison to have Claude Fable write most of sqlite-utils 4.0rc2, one of the release candidates for the first major version of his database library since 2020 (3.0 was in November …
Threads we are following
- Whether enterprises consolidate onto a single vendor or commit to multi-vendor routing as open-weight entrants keep dropping the cost floor
- Whether DeepSeek can secure any guaranteed SMIC N+2 wafer allocation, or remains locked out by Huawei's five-year lock on the line
- Whether world-model grounding research like BAAI's Orca can convert these benchmark ceilings into slopes
- Whether anyone can build evaluators strong enough to extend harness self-improvement beyond code and math into fuzzy domains like research taste and codebase health
- Whether agent harnesses evolve any reliable way to capture and transmit human intent, or whether commit-message duty stays permanently off-limits to models
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.