Frontier AI vs Personal AI

One of the discussions at Moonshots Summit LA expanded the way I'm thinking about AI in our everyday lives.

Consumer AI and frontier AI

The discussion was around consumer AI and the work happening at the frontier, including how AI systems could contribute to their own improvement, often called recursive self-improvement (RSI). Frontier research is concerned with pushing what these systems can do, whether that's reasoning through more difficult problems, like Navier Stokes, or helping with scientific work, like Anthropic's new Wet Lab. Simply put, expanding model capabilities.

With consumer or personal AI, we're looking at how someone can take those same model capabilities and use them in their day to day. That could mean organizing travel, keeping track of responsibilities, getting help learning something in their own field, etc. A consumer product can use a frontier model, so there's an overlap between the two. The distinction is in what we're evaluating: the capability of the model, and how useful the product built around it becomes for a person.

Personal AI doesn't need to use the strongest model in the world for every use case or task. If someone wants help organizing their responsibilities or learning something in their field, what matters is whether the system can do that well with the information available. A more capable model may help, but the work also involves giving the system the context and information it needs to produce something useful. For a particular task, an older or less expensive model may already be capable enough. That needs to be evaluated against the task, the quality of the result and the cost of using it. That’s why I think we need to make a clearer distinction between the capabilities needed for everyday personal use and those needed to drive frontier research forward.

For frontier labs, there's a reason to keep pursuing the strongest models they can build. When the aim is scientific innovation or helping researchers pursue medical breakthroughs, more capable systems could help with reasoning through difficult problems, developing hypotheses and evaluating possible explanations. That's why I think model development needs to continue, alongside the experiments and scientific validation needed to establish whether those ideas hold up. At the same time, I think frontier labs need to put at least as much emphasis on alignment and on improving safeguards as capabilities increase.

I think something else that's under-discussed is the long tail of applications we could build with capabilities that aren't yet widely used, even if model development took a pause. There are specific needs across different professions, routines and communities that could be explored using the immense capabilities already available. Some of those applications might serve a small group of people and still make a meaningful difference to them.

Personal AI in everyday life

All this discussion also made me look at how I'm using AI myself on the personal front. In my notes after the summit, I talked about how much I use coding agents and how much more I want to explore outside of coding. There are recurring responsibilities around checking the status of work, following up and keeping different initiatives moving that I want to think through more carefully. I can see possibilities there, and I still need to define which responsibilities make sense to hand over and what good output would look like, which is a common situation I imagine many people find themselves in today.

With this, I think there's a lot of opportunity for people who aren't looking to become AI researchers or spend their days comparing models. Someone can use these models to learn more effectively, research how AI could transform their field, organize their responsibilities, or work through an idea. Making those use cases understandable and accessible is a part of the progress I'm interested in contributing to, especially for people and communities whose needs may be overlooked where there isn't an obvious commercial return.

A starting point for anyone exploring AI

For someone coming into this without a technical background, I think a useful starting point is one responsibility they already understand. They can start with three questions:

1. What takes time?
2. What information do I need?
3. What would a good result look like?

From there, there's something concrete to explore and evaluate, and a way to notice where the system still needs help. The models themselves can also help people work through those questions and find a starting point.

---

How this connects to our work at OTZ

Last point to make, there's a connection here to the work we do at OTZ as custom AI implementation partners. Our forward deployed engineers (FDEs) work directly with businesses to understand how their teams work, what information they rely on and what they’re trying to improve. That context helps us identify where AI can be useful, choose the capabilities the work requires and build systems that fit into their day-to-day operations. You can learn more about our work at OTZ.