2026-07-16 / Signal #2
Thinking Machines Lab (Mira Murati ex-OpenAI) releases Inkling open-weights multimodal model.
“The ex-OpenAI CTO shipped an AI that deliberately dials how hard it thinks, admits when it's unsure, fine-tunes itself—and open-sourced the weights so you can make it your own.”
9.0Weirdness
Why It Matters
975B-param MoE (41B active), native multimodal (text/image/audio/video, 1M context, trained on 45T tokens), controllable "reasoning effort" dial (efficiency vs. depth), uncertainty handling, and self-fine-tuning demo (model improves its own job description and evaluates it). Positions customizable open AI against closed frontier labs and Chinese open-weights; extends "human will and judgment" for builders.
Evidence
Official thinkingmachines.ai blog (July 15), TechCrunch/WIRED coverage, HN top story (1047 pts).
Signal Read
Novelty: 8Receipts: 10Story voltage: 9Heat: 10
Source Trail
Daily scan: 2026-07-16