Every Future Product Will Be a Living System: Ronak Malde on Continual Learning

The core tension exposed in this conversation is the gap between today’s static, frozen-weight AI models and the reality that enterprise applications operate in constantly shifting environments—regulatory updates, new customer data, and evolving business logic. Ronak Malde argues that every product of the future will be a living system, meaning it must continuously learn and adapt post-deployment rather than relying on periodic retraining. This challenge is especially acute for safety-critical use cases like legal (working with Harvey) or finance, where a model that cannot update its knowledge in near real-time quickly becomes obsolete or risky.

Trajectory.ai‘s concrete approach is to build a continual learning platform that treats model updating as an operational discipline, not just a training trick. They run concurrent, non-linear training jobs at scale, curating data streams for online learning and using techniques like Self-Distillation Policy Optimization (SDPO) to avoid catastrophic forgetting. The infrastructure choices are practical: they select base models such as NeMoTron 3 Super for their strong foundation, then focus on data curation and scaling the training stack to handle concurrent updates. They have open-sourced parts of their training infrastructure to accelerate adoption.

The practical takeaway for builders is that building a living system demands more than a clever algorithm—it requires rethinking the entire ML pipeline from data ingestion to deployment and monitoring. Continual learning is an infrastructure problem as much as a research problem, and teams should invest in robust data curation, automated evaluation, and the ability to roll back updates gracefully. For enterprise buyers, the key question is whether a vendor can demonstrate that their model improves with use instead of degrading over time. Malde’s trajectory from coding agents at Windsurf to continual learning at Trajectory.ai reflects a broader industry shift: the next frontier of AI is not bigger models, but models that persist and evolve.