Thinking Machines announced Inkling 2 Small, a Mixture-of-Experts foundation model (594B total parameters, 17B active) with image inputs and variable-effort inference that advances agentic coding, vision, long-horizon autoresearch and kernel optimization. It’s available on their Tinker platform with closed weights now, will support fine-tuning and downloadable checkpoints, and the full weights are planned for Apache 2.0 release after external safety testing.

Kagi announced that anyone can now try Kagi Search without creating an account (50 free searches, 100 more on registration), unveiled a new landing/discovery flow, and rolled out many product improvements — notably Android app redesign and rebuilt voice search, Assistant/Translate updates, plus numerous bug fixes. The post also addresses industry news (Google disabling AI/search for some school accounts) and urges schools to consider alternatives.

The author argues that although tech companies collect extensive data about users, their personalization and localization are often poor and produce irrelevant ads, mistaken language/locale behavior, and ignore simple user preferences. They give multiple examples—music skips, ad targeting, language mismatches, redundant prompts and incorrect offers—to illustrate how machine learning and tracking fail to meaningfully improve their experience.