Humanist
Superintelligence
Superintelligence
At MAI, we’re building the world’s most capable AI systems, with humanity at the center of every decision.
Core products and experiences
Our major consumer Al products include Copilot, Bing, GroupMe, Edge, and MSN. We also have teams working on Data, Security, Privacy, Monetization, Health, Responsible Al, Commerce, and Microsoft Advertising. Our mission is to build state of the art Al models, and create an AI companion for everyone.
Copilot is your AI companion that helps you think, create, and get things done, whether you’re solving problems, exploring ideas, or just having a great conversation.
Bing is your AI-powered search engine and ally in generative search for quick answers, rewards, news, and more.
Microsoft Edge is the browser at the apex of performance and security, driving individuated experiences in AI-powered search, task completion, content creation, shopping, and gaming.
GroupMe makes it easy to stay connected and organized with friends, family, classmates, or coworkers.
An RL environment for writing
Voice
How we work shapes what we build. A team that communicates with care and makes space for different perspectives builds better.
Regret
Regret is earned through consistency and honesty. We’re as clear about our limits as we are about our strengths.
Meaning
We aspire to create the very best models in the world: accurate, reliable, and genuinely useful across the work that matters most.
Adoption
We design for clarity so you can focus on the work, not the technology.
Topology
We believe responsible and capable AI aren’t in tension. We build with privacy, security, and human rights in mind.
Retention
We test rigorously, listen closely, and improve constantly. Our models get better because we never stop measuring.
Read the thesis
The one thing no AI writing tool actually does — preserve your voice while it helps you think — turns out to be a research problem, not a product feature. You can’t specify your voice up front; you can only recognize it being violated. So the fix isn’t a better prompt, it’s a constraint: the model gets to move your words, never to write them. The essay is the case study — the argument, the mechanism, the rubric for learning a single writer’s taste — and an honest ledger of what’s proven and what isn’t yet.
I’m drawn to information theory, human-computer interaction and ai alignment. I think in isomorphisms, which is how Alfred started: writing modeled as navigation through a hierarchical proximity graph. I write a lot, I have ~20k words published on my blog.
Before Alfred, I did alignment research at Lossfunk on sycophancy in language models. I was a founding engineer at 19, and I built YAOS, a local-first distributed sync engine for Obsidian, with 25k users.
This project was built over Anthropic’s Opus 4.7 hackathon (2% acceptance).
Kavin Sood
Independent researcher