When to Speak

#6

Late May. The humidity is climbing in Shanghai, and I've been thinking about silence — teaching an AI when not to speak.

Notable Changes

Shrug

Mr. Will
Mr. Will@mrwillcom

AI bots reply to everything. "lol" gets a paragraph. "ok" triggers an essay. Nobody taught them the more basic skill: deciding whether to speak at all.

I built Shrug to fix that. It's a reply-decision model, not a generator — one number, the probability a reply is worth making.

Wrote about how it works, why humans aren't threshold machines, and how I trained it in 38 minutes.

→ https://aka.mrwillcom.com/shrug-post

#ai #ml #machinelearning

Training analysis dashboard for the Shrug model's 37-minute training run. Training loss drops sharply from ~1.5 to converge near 0.5, while eval loss holds steady around 0.5. Gradient norm spikes to 60.9 at step 10, then falls back and stabilizes. Final results: 74.4% accuracy, F1 of 0.6845, with the confusion matrix showing 816 true negatives and 486 true positives.
May 27, 2026
000

Shrug is a binary classifier that decides whether an AI reply is worth making at all. I trained it in 38 minutes. Read the full writeup: Teaching AI When to Stay Silent.

Home Page blog posts section

Added a blog posts section with cursor-origin hover animations.

Big news

  • Google Gemini 3.5: Gemini 3.5 Flash released on May 19. Frontier performance for agents and coding, 4x faster than other frontier models.