Interview, Fireside Chat
Bringing Continual Learning to Every Company: Trajectory’s Ronak Malde
- The industry is projected to rapidly shift from a "hunter-gatherer" state to a specialized ecosystem where every company integrates training, evaluation, and deployment into a unified paradigm.
- Trajectory aims to democratize access to model customization, targeting a future where enterprises scale AI adoption from the current 1% to 99% by abstracting complex training processes into one-click interfaces.
- To address the high cost of frontier models, which require hundreds of millions of dollars in R&D, six to nine months of development, and massive research teams, Trajectory's custom-trained models are priced at six times lower.
- The platform seeks to replace the token-based unit of work with "trajectories," defined as tasks of economic value, enabling users to trace production activities, sub-agents, and feedback for evaluation.
- Architectural innovations, including fast and slow memory within weights and skip connections, are identified as critical enablers for transitioning current models from batch updates to real-time continual learning.
- A "group size one" paradigm has been introduced to allow learning from single rollouts without requiring perfect environment replication, thereby eliminating the need for massive infrastructure to replay scenarios.
- To overcome the difficulty of defining verifiable rewards in complex domains like legal or investment banking, the system employs a metagrader trained on small, user-labeled samples (e.g., 100 out of 10,000 traces) to propagate reward signals.
- The platform targets AI-native firms like Harvey and Rogo, as well as large enterprises like NVIDIA, facilitating a hybrid routing strategy where frontier models handle 10% of tasks while custom models manage 90% to optimize margins.
- Current agents lack the ability to autonomously explore the world or encode transition functions, a gap Trajectory addresses by keeping humans in the loop to ensure alignment and prevent Recursive Self-Improvement (RSI) risks.
- The operational framework relies on a software factory of agents to handle implementation details like microservices and Kubernetes scaling, allowing the core team to focus on system design and whiteboarding.
- Future goals include optimizing the entire harness for sub-agents to achieve "intelligence density," moving beyond model-only improvements to create autonomous, self-growing products similar to a plant on a trellis.
- Broader market evolution depends on infrastructure providers such as Modal and Thinking Machines to support the massive recommender system infrastructure required for global, real-time agent learning.