Skip to content

A Model For Everyone: How Ronak From Trajectory Is Building Products For Continual Learning

Training Data

Hosted by Pat Grady

32m
Transcript

Original podcast

Where to listen

Summary

Written by a model from the transcript; not yet reviewed.

This episode explores the concept of continual learning in artificial intelligence and how the startup Trajectory is building products to bridge the gap between plentiful foundation intelligence and real-world domain experience. The guest explains how current AI usage resembles the early days of GPS maps, where updates required massive offline batch operations, and argues that the future requires a shift toward real-time updates modeled after the human brain and Google Maps. The discussion covers the architectural changes in base models, the practical application of training customized models for companies like Harvey and Rogo, and the vision of moving from a world of a single centralized oracle to one of mass specialization and automated self-improvement.

  • Companies should transition from treating tokens as the sole unit of work to focusing on economic trajectories and rollouts in an RL training stack. 9:02
  • Developers can leverage new post-training techniques like group size one algorithms to eliminate the need for massive, complex environments during reinforcement learning. 14:02
  • Specialized domain knowledge can be captured by smartly selecting a small representative sample of traces to train a meta-grader across larger datasets. 15:14
  • AI adoption can expand across manual enterprise workflows by shifting from a one-size-fits-all model to customized, efficient models trained on domain-specific data. 30:38

Chapters

  1. 0:00
    Apple‑Inspired Product Philosophy

    The hosts introduce how Apple’s integration of internet and phone inspired a simple, end‑to‑end product loop for AI models.

  2. 0:57
    Defining Continual Learning

    Ronak explains continual learning, contrasts it with pseudo‑continual learning, and highlights the human brain as a model.

  3. 6:07
    From Research Labs to Productized Learning

    Ronak recounts his journey from large research labs to building a platform that democratizes continual learning for companies.

  4. 9:33
    Trajectories as Economic Work Units

    The team discusses the concept of a trajectory as a unit of work, linking RL rollouts to real‑world AI tasks.

  5. 11:50
    Core Primitives for Continual Learning

    Ronak outlines the three pillars: tracing, replayability, and algorithmic modification, and how Trajectory abstracts them.

  6. 14:00
    Future Vision: Curiosity and Alignment

    Ronak envisions agents that learn through curiosity while staying aligned with human objectives, bridging the gap between foundation models and real‑world experience.

  7. 14:30
    Research Innovations: Group Size One & Meta‑Grader

    The team explains new algorithms that reduce environment and reward labeling costs, moving toward fully automated continual learning.

  8. 16:36
    Product Launch: One‑Click Training Interface

    The platform’s first release offers a simple UI to train, evaluate, and deploy models using agent traces and reward signals.

  9. 17:41
    Target Customers and Use Cases

    Trajectory focuses on AI‑native companies like Harvey and Rogo, enabling them to fine‑tune models cost‑effectively.

Highlights

  • 1:37

    Defining Continual Learninginsight

    The guest compares current AI agents to old-school offline road maps from AAA, contrasting them with the goal of real-time continuous updates like Google Maps.

  • 3:11

    Architecture and Continual Learninginsight

    The guest discusses how base model architectures are already evolving with fast and slow memory within weights to support future continual learning paradigms.

  • 14:02

    Algorithmic Advances with GRPOstrategy

    The guest details how modifying the GRPO algorithm to a group size of one eliminates the need for maintaining complex simulation environments for every single training task.

  • 19:15

    Efficient Agents at Rogostory

    The guest describes how finance agent company Rogo uses trajectory training to achieve higher token efficiency and concise execution instead of verbose outputs.

  • 20:01

    Custom Model Results at Harveydata

    The guest explains how legal tech company Harvey used Trajectory to train a model that is six times cheaper while raising accuracy from 4 percent to 18 percent.

  • 23:14

    Continual Learning Versus RSIdebate

    The guest contrasts recursive self-improvement in a black box with continual learning, emphasizing the importance of keeping humans in the loop for alignment.

  • 29:56

    Quantifying Agent Curiosityinsight

    The guest explains how agents lack the human ability to explore and build mental models of unknown transition functions in Markov decision processes.

  • 29:56

    The Industrial Revolution of AIinsight

    The guest uses Adam Smith's theory of specialization to argue that enterprise AI will move from a hunter-gatherer reliance on one oracle to widespread workflow customization.

Quotes

  • “I think the world we're in right now with agents and AI is kind of like you said, pseudo continual learning.”
    Unnamed speaker · 1:30
  • “You need some sort of infrastructure to say, hey, okay, given a trace that I've seen in production, how can I actually then replay that for any eval to see what good looks like.”
    Unnamed speaker · 11:55
  • “I think the real value of any AI progress, but certainly continual learning, is how are people actually using models in the real world?”
    Unnamed speaker · 4:20

About this episode

Today’s agents are stuck in what Ronak Malde calls “pseudo continual learning”: intelligent, but updated in offline batches, like a paper road map you can’t reroute mid-trip. Ronak, co-founder of Trajectory, explains why closing that loop is the next learning paradigm, and why training your own model, which used to take a dedicated research team, months of work and hundreds of millions in compute, is now something a company can do in a few clicks with their own data. He details two research bets they are making available to everyone with their product: 1) a variant of GRPO that removes the need for a simulated environment by learning from single real-world rollouts instead of 16 replays of the same task, and 1) a meta-grader that increased data efficiency by 100x. Ronak explains how Harvey trained a model six times cheaper than frontier, improving an open-source baseline by over 4x on Harvey's own legal benchmark. He defines the difference between continual learning and recursive self-improvement and predicts product builders will become gardeners, setting trellises for products that grow themselves. Hosted by Pat Grady and Lauren Reeder, Sequoia Capital 0:00 – Introduction 0:45 – Pseudo continual learning: paper maps vs Google Maps 2:48 – Do we need new architectures? Rich Sutton, David Silver, and the missing layer 4:33 – From a $100M consulting model to a product: lessons from SWE-1 at Windsurf 7:36 – Harvey: 6x cheaper than frontier, 4% to 18% on its own benchmark 8:55 – Why "Trajectory": the unit of work is a task, not a token 10:13 – The primitives of continual learning: traces, replayable environments, reward 12:56 – Group Size 1 and the meta-grader: training without an environment or a labeled reward 16:16 – The platform: train, eval and deploy in one click, and who it's for 18:58 – Rogo and intelligence density 20:14 – What's next: curiosity, and why continual learning beats RSI on alignment 24:19 – Research and product under one roof, and the product builder as gardener 27:26 – Why work at Trajectory, and the optimistic case: from 1% to 99%

Transcript

0:00
32:00

6,193 words · 3 speakers

Speaker names marked are inferred from speaking time and the show’s known hosts — they are not voice-verified.

Topics

  • Apple
  • Harvey
  • Rogo
  • GRPO
  • Windsurf
  • Cursor
  • DeepMind
  • Gemini Robotics
  • Trajector y
  • Cursor Composer

Page updated · transcript and summary are machine-generated