What's this about?
A day after Thinking Machines' Inkling launch (yesterday's top story), early hands-on results are mixed. Ethan Mollick finds the model "pretty rough" — it fails his Lem Test, a creative-writing check that every frontier model has passed since DeepSeek r1 — and rates it below the best Chinese open models. On the plus side, the ecosystem moved fast: Inkling is already runnable inside Claude Code via Hugging Face's `hf claude` integration.
What exactly is a "reality check" on a model launch?
Launch-day claims come from the lab's own benchmarks, run under ideal conditions. The 48 hours after release are when independent testers run their own pet evals — quirky, personal, hard-to-game tests like Mollick's shader prompts or Willison's pelican drawings. These informal probes often predict day-to-day usefulness better than official numbers, because the lab couldn't have trained against them.
Why does it matter?
Inkling was celebrated yesterday as "America's answer to Chinese open weights." If it lags Kimi K3 and GLM in practice, the US open-weights gap is still real — and Kimi K3 topping arenas the very next day sharpened that contrast uncomfortably.