AI success that scales comes from systems design. Specifically it’s how systems learn and how they develop expertise over time. As we’ve built our AI Operating System (AIOS), we’ve made deliberate architectural choices that don’t show up in demos, but compound into learning advantages. It helps us scale over firms that have built individual agents. Three aspects of what Abdul and I found out over the past two years: 1) Memory as the foundation of learning. Most AI systems treat context as disposable. We treat it as cumulative. Our architecture centers on persistent, evolving memory that enables continuous learning across interactions, personalization at the OS level and compounding context that adapts performance over time. Without memory, there is no real learning and the system needs to repeat many steps. 2) Model layers as pathways to expertise. We’ve moved away from single agents on monolithic models toward modular, layered systems: - Routing to specialized models, fine tuned to develop domain-specific expertise - Orchestration layers that route tasks to the right “expert”, human or AI - Continuous evaluation and learning capabilities that upgrade without breaking the system Expertise emerges not from one agent doing everything, but from systems that know which intelligence to apply, when. 3) Multiplayer AI and reinforcement learning from experts. The next frontier is smarter learning loops. We’re building toward a multiplayer paradigm where: - Networks of experts (human + AI) provide reinforcement signals - Systems learn from real-world judgment, not just static data - Expertise is encoded, refined, and scaled across the network Most teams are still optimizing for benchmarks and surface-level capabilities. But durable advantage in AI will come from systems that learn continuously, architectures that compound expertise, and feedback loops that improve with scale The companies that win will build systems that get smarter, more specialized, and more valuable with every interaction.
This resonates. One thing I’ve been running into building Donkey Betz: A lot of systems are optimizing for learning (memory, model routing, feedback loops)… but still struggle to answer: “What does this system actually produce today?” We’ve been forcing everything through execution loops: → ingest → reason → validate → ship Because until something turns into a real outcome, the learning loop doesn’t compound in a meaningful way. Curious how you think about balancing “learning systems” vs “execution systems” in practice.
Agree on systems design. The real challenge is making these learning loops work in environments where consistency, auditability, and risk control are critical. When do you think that will happen, Paul?