Post

I feel like the last generation—DevOps into Continuous Delivery—was defined by Conway's Law, and this generation—AI tokenmaxxing and shoehorning—is defined by Goodhart's Law.

Abigail Dombey@abigaildombey.bsky.social

Amazon staff use AI tool for unnecessary tasks to inflate usage scores Paging @ketanjoshi.co

giftarticle.ft.comAmazon staff use AI tool for unnecessary tasks to inflate usage scoresIn-house MeshClaw tool enables employees to delegate jobs to AI agents and climb company’s AI leaderboard

12/5/26, 8:40 am4 reposts13 likes

Pawel Brodzinski

@pawelbrodzinski.bsky.social

12 May

Also, tech companies: April 2026: Let's tokenmaxx the shit out of everything! May 2026: Our AI bills went through the roof. How about we use cheaper models for simple stuff?

Pawel Brodzinski

@pawelbrodzinski.bsky.social

12 May

So you start rewarding/recognizing people for [doing the thing], and they start [doing the thing] even when it's utterly unnecessary or plain harmful? No shit! Why would something do something that wicked?

Daniel Terhorst-North

@tastapod.com

12 May

I know right? It's almost as though people will optimize for the behaviour you are rewarding them for. Something something Goldratt.

Patrick Richter

@nomadsgate.bsky.social

12 May

Spot on. Watched teams optimize for token counts instead of actual latency. Metrics become the goal, not the signal.