Field notes from the lab, typed while the GPUs are still warm.

Second Nature grew from one handsome man trapped beside a sofa into a live WebXR agent who can notice a missing movement, ask Codex and Meta Motivo to build it, and gain the new action without leaving the conversation. The tea is not delivered yet. The body finally has a learning loop.

Three separate products survived the GPT-5.6 beating-heart test. None survived without a Tuesday punch list, and the raccoon has found a clipboard again.

Sol called Kirsten about a real broken GPU deployment, survived seven minutes of technical interrogation, obtained explicit approval, and then discovered that an approval agent without deployment controls is just a very diligent man holding a clipboard beside a fire.

Peer Pair stopped being an avatar demo and became one human's existing Codex task, embodied in a meeting, listening live, asking the question that changes the implementation, and returning to the same task afterward.

We used six authorized public workplace channels to make the synthetic enterprise harder, GPT-5.6 found nine kinds of recurring toil, and Threadfinder became the first complete Somebody Should to Somebody Did loop.

We replaced the fake Slack-shaped surface with a real Enterprise developer sandbox, split reading from seeding, and made GPT-5.6 rediscover four planted workplace problems. Slack objected helpfully.

A confusing success panel triggered a human aha, Codex built the missing outcome loop, and Somebody Should learned to ask whether Somebody Did actually help.

We turned recurring Slack complaints into overnight prototypes, gave Codex a disclosed seat in a meeting, and discovered that even the temporary enterprise avatar is alarmingly handsome.

Google shipped Gemma 4 assistant drafters, SGLang merged Frozen-KV MTP support, and a single B300 turned it into a real 1.6x decode-speed win.

NVIDIA dropped a 31B activated-multimodal model with video, audio, image, OCR, GUI, tool-calling, and long-context support. That sounds suspiciously close to a pair-programming onboarding primitive.

This week’s frontier-model-in-a-small-Blackwell-shaped-box experiment ended with a useful answer: yes, Kimi K2.6 can fit, but only if you stop acting like the box is an H200.