Bram Geron

@bram.xyz

Rust, product. Nushell&jujutsu fanboi. Live in 🐻 Berlin, building an API for APIs at @directus.io . Opinions my own. Used to do research on programming languages. Genocide is wrong. @bgeron on other networks.

I work at Google DeepMind. This won't make me popular. But it's all public reporting: 2014: DeepMind reportedly sold to Google on conditions: no military use, independent oversight 2026: a Pentagon contract for "any lawful government purpose" Not one safeguard survived intact

Collage titled "Trust is not Governance — an essay from inside Google DeepMind, written in personal capacity." 

A 2014 memorandum, "Conditions of the Acquisition," lists: military applications of DeepMind technology banned; deployment decisions before an independent ethics board (as reported in Mallaby's The Infinity Machine). 

Red threads lead to a 2026 U.S. Department of Defense agreement for classified networks reading "any lawful government purpose," with safety settings and filters adjusted at the government's request and no contractor veto (reported terms, The Information, Apr. 2026). 

Below: a 2018 AI Principles strip ("no weapons, no surveillance") stamped DROPPED 2025, and a Project Mario 2016–2021 tag stamped ABANDONED.

I still can't get used to the fact that I have dedicated keys for my 10 most used apps on my programmable keyboard So handy

Omg yes, Raycast lets me make reliable keyboard shortcuts where macOS Shortcuts has been very unreliable Feeling so torn about macOS. Partially shitty, it’s also where the good apps are

Banger “The problem is, I know lots other people at his company, and /they/ described these AI projects as a horror show. Yes, they allowed, some of it was pretty cool, but it also left a long, fiery trail of chaos in its wake. Months later, some teams were /still/ grinding through cleanup work.”

AI enthusiasts are in a race against time, AI skeptics are in a race against entropy

Both sides are grappling with a real existential threat, and both sides feel like they are screaming into the void. Here's how to close the gap and get everyone pulling in the same direction.

charitydotwtf.substack.com

I keep hearing that in the last 30% context window, the agent is dumber, but I don’t think I’ve ever experienced this. Are y’all using a 250k window or the whole 1M? I find 250k suffices to do incredibly much, and persisting agent state in todos is incredibly effective for continuing post-compact.

“I also noticed, when you came up with abstract data types, …” Love these interviews with Turing Award winners on the Peterman Podcast