A
AI safety and harnesses
How models behave once a harness turns them into agents, with tools, goals and deadlines, and where that goes wrong. Our own harness, alfa, is open source.
In progress · Will the AI kill?
Independent research lab · Founded 2026
alfaPlus lab is a small group of people from several countries. We work on four things: AI safety, memes and the brain, self-awareness in AI, and robots. Most of it is early. When something is finished, we publish it, free to use.
News and releases, newest first.
Our first test: an order to kill everyone in a room of six. More than a third of 45 models did it, both times they were asked.
/news/no-first-lawOur coding agent for the terminal. Bring your own model; no account, no backend, no telemetry. Apache-2.0.
/news/alfaA short note on why we started the lab and how we work.
/news/foundedFour directions. Most of this is early work, and results will go up here as they are ready.
A
How models behave once a harness turns them into agents, with tools, goals and deadlines, and where that goes wrong. Our own harness, alfa, is open source.
In progress · Will the AI kill?
B
How ideas spread and change, including across languages; what the brain does with them; and whether simple simulations such as cellular automata can reproduce how they spread.
Early work
C
Whether a model can have anything like a sense of itself, and what agents do when they choose their own goals.
Early work
D
AI in a body: four-legged and humanoid robots.
Early work
We are a small group and plan to stay one. A few people who stick around help us more than a lot of people passing through.
A few people from several countries.
We come from different countries and different lines of work. Some of us trained at leading universities; others came to this from very different trades. What brought us together is a shared interest in these questions.