Sunday, 19 July 2026

The week's dominant thread is the sprint toward capable, practical AI systems that solve specific problems rather than chase scale alone. Feyn AI's SQRL family demonstrates this shift: its 35B text-to-SQL model beats Claude Opus by inspecting databases before querying, a technique that trades raw parameters for architectural cleverness. Meanwhile, Sakana AI has emerged as the week's most prolific innovator, releasing ShinkaEvolve—a framework that uses LLMs to evolve algorithms with 150 samples instead of thousands—and publishing work on Evolutionary Model Merge, where a 7B Japanese math model matches prior 70B performance through automated model fusion. These aren't bigger models; they're smarter ones.

Parallel to this is a reshuffling of geopolitical AI infrastructure. Alibaba teased Qwen3.8-Max at 2.4 trillion parameters days after Moonshot's Kimi K3, but without disclosing active parameters per token, both announcements are exercises in marketing opacity. Japan's commitment to Noetra—a sovereign AI factory backed by 1 trillion yen through 2028—signals more substantive intent, with NVIDIA's Vera and Rubin GPUs anchoring domestic infrastructure while Toyota, Fanuc, and Yaskawa adopt NVIDIA's Cosmos models. The hardware lock-in is real; the independence narrative is partial.

Smaller currents matter too. Apache Spark 4.2's native vector search threatens specialized vector databases. Text-to-LoRA lets engineers customize foundation models via plain English. And Sakana AI's ALE-Agent placed 21st in a competitive optimization contest, suggesting AI can now meaningfully contribute to algorithm engineering. The week's through-line isn't scale; it's specialization, inference-time reasoning, and moving capability from model size to architectural design.