Who owns what AI learns from you?
Many of us enjoy how AI boosts our work. But almost everyone is worried about what happens when AI grows more and more capable: will our contributions still matter? Will AI companies keep selling assistants that empower us, or will they start selling AI that takes our jobs instead? Who will AI work for: people, or the companies that build it?
The answer depends on who owns what AI learns from your work. Today, that hardly matters, because AI barely learns from working with you. Switching providers is easy: you can start a project with one company's AI and finish it with another's, or use both throughout the project. But that is about to change.
An AI assistant that learns is more like a colleague. If what it learns is stored in one AI company's model weights rather than in notes you can take with you, leaving that company means losing the knowledge and skills your assistant has gained through months or years of working together. Continual learning could give AI labs a strong competitive moat by making it costly for customers to switch providers. As AI accumulates knowledge specific to your work, replacing your assistant could become like replacing an experienced employee with someone who has to learn everything from scratch. And the more your assistant can do, the more it matters whom it works for. A company that owns what its AI learns from your work could one day sell that work without you.
The experience an AI assistant gathers from your work should be yours, as much as your files, your contacts and your know-how: what you asked it to do, what it did, how that turned out and how you corrected it. You should be able to take that record to whichever AI is best, and to leave any company, including the one that built your assistant, without starting again with one who knows nothing about you. That requires a service that can turn that record into AI that knows your work. And that’s exactly the technology we’re building. The learning methods used by the leading AI companies are designed to train enormous models on vast amounts of data. They are not well suited to learning from the limited experience of one person. For personalized AI to scale to millions of people, learning must also be affordable for everyone. Our mission is to make this possible through continual learning that is efficient and stays aligned with what you want.
The big AI companies are like airlines: they fly the long distances between hubs, at a scale only a few can manage. But everyone still needs to get home, and for that we use taxis, buses, or bikes: different kinds of transport, affordable and able to take us to our own door. Learning your work is the last mile of AI, and it needs different technology from training the big models: small models that learn from what actually happens in your work, alongside whichever large AI model you choose, including open-weight models anyone can run. These small models are cheap enough to run close to each person, even on a laptop. Any AI equipped with this kind of learning can learn from the record of your work and pick up where your previous assistant left off. Few can do that today, but as the technology spreads, so does your freedom to choose, and we welcome that competition. The big AI companies could build it too, just as an airline could run taxis, but it is a different business.
Think of the estate agent from our first post. Her data stays with her, and what her AI assistant learns from her work belongs to her: how to recognize serious buyers, which properties to recommend, when to follow up and how to handle negotiations. If she decides to switch to another AI, she can take the record of her work with her, so the new AI can learn the skills her previous assistant acquired. She may choose to share what isn't private or a trade secret for everyone's benefit, but that choice is always hers.
AI is starting to learn on the job. Who it works for will depend on three things:
- Continual learning: It keeps getting better at your work, like a good colleague.
- Alignment: As it learns, it judges its actions by what you would want if you knew where they would lead.
- Ownership: What it learns from your work belongs to you.
With all three, AI that keeps learning can empower all of us, not just a handful of companies.