Skip to content

By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Sequoia CapitalSequoia Capital

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

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 MaldeguestLauren ReederhostPat Gradyhost
Nov 5, 202632mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

At a glance

WHAT IT’S REALLY ABOUT

Trajectory aims to productize continual learning from production traces

  1. Ronak Malde defines “true” continual learning as online, human-like updating from real-world experience, contrasting it with today’s mostly offline, batch-style model updates.
  2. He argues that while model architectures are evolving, the highest-leverage problem is turning production traces and a company’s notion of “what good looks like” into systematic model improvements.
  3. Trajectory’s platform aims to combine training, evaluation, and deployment into a single, simple loop where teams can train from traces, evaluate, and push models back into production to gather new traces.
  4. He outlines the key building blocks companies need—tracing, replayable environments/evals, and a training stack—and explains Trajectory’s research focus on reducing environment requirements and reward-labeling work.
  5. He describes continual learning as a more human-aligned alternative to recursive self-improvement, emphasizing keeping humans in the loop while agents learn from real usage.

IDEAS WORTH REMEMBERING

5 ideas

Continual learning is the shift from batch updates to always-on learning from real usage.

He frames today’s agent customization as largely “offline,” where updates require big batch retrains; the end state is systems that update as they’re used, similar to how Google Maps continually improves routes from aggregate driving behavior.

The bottleneck isn’t only model architecture—it’s the bridge from production feedback to training.

He argues architectural changes are already emerging inside base models, but says the bigger missing layer is translating real-world definitions of “good” into model improvements that companies can actually operationalize.

The product bet is that model training/customization is now repetitive enough to be productized.

He describes why training and customizing serious models has required massive compute, large research teams, and long timelines, and positions Trajectory as packaging the repeatable parts into a simple interface with the key “knobs and levers.”

The core continual-learning “primitives” are traces, replayability, and a training stack that turns feedback into learning.

He says companies need (1) production tracing, (2) replayable traces for eval/training, and (3) a training stack that maps tasks into loss/reward—where Trajectory aims to abstract away the hardest parts.

Trajectory is trying to cut the two biggest costs: building perfect environments and labeling tons of rewards.

He identifies environment construction and reward specification as the most manual parts today, then describes two research directions: moving GRPO toward “Group Size-1” learning to reduce environment dependence, and using smart sampling plus a meta-grader to propagate reward from a small labeled set to a large trace set.

WORDS WORTH SAVING

5 quotes

But the final stage of continual learning is something more like Google Maps, where every single time you're driving, every route that you make, uh, it is updating. It's learning aggregate what the fastest route is and where the traffic is. Uh, that's the world I want to get to.

— Ronak Malde

I think any good product exists when you have every single company in the world trying to do the same thing over and over and over again.

— Ronak Malde

So with Trajectory, what they were able to do is train a model that is six times cheaper, a-and it goes from around 4% accuracy for an open source model, uh, to around 18%.

— Ronak Malde

The models right now, they do a task, uh, but they don't actually accumulate any experience or learn anything about the real world.

— Ronak Malde

I don't want to end up in a world where these agents are going rogue, and I think the way to do that is to, to keep on this alignment path.

— Ronak Malde

Definition of continual learning vs “pseudo” continual learningMapQuest vs Google Maps analogy for online updatesTransformer architecture changes and memory/skip connectionsBridging production feedback to model improvement (“what good looks like”)Production tracing and observability tools (e.g., LangChain, Braintrust, Raindrop)Replayable traces, environments, and training stacks (loss/reward)GRPO and “Group Size-1” learning; reward labeling via representative sampling and meta-graders

SPEAKERS

  • Ronak Malde

    guest

    Founder of Trajectory, discussing continual learning and model training/evaluation workflows for production AI systems.

  • Lauren Reeder

    host

    Sequoia Capital host/interviewer guiding the conversation and prompting the guest on product, naming, customers, and hiring.

  • Pat Grady

    host

    Sequoia Capital host/interviewer asking technical/product questions and referencing prior episodes/guests (e.g., Rich Sutton, David Silver).

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.

iOSAndroidClaudeChatGPTChrome