Sequoia Capital

Ronak Malde on continual learning as the next learning paradigm

Ronak Malde· Co-founder of Trajectory at Trajectory
·~32 min·English·Sequoia Capital
TrainingAgentsOpen Source
TL;DR

Trajectory's co-founder argues that continual learning — models that keep improving from real production use rather than one-off lab training — is the next paradigm, and shows a one-click platform that already took a legal model from 4% to 18% accuracy while running six times cheaper.

01Core Mental Model

Today's Agents Only Do 'Pseudo' Continual Learning

<strong>Today's agents are stuck in 'pseudo' continual learning</strong> — intelligent but frozen, like a road trip planned on paper maps you can only change by stopping to call AAA.

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.

— Ronak Malde, Sequoia Capital
Key Insight
The analogy hides the real bottleneck: today's 'offline' agents aren't dumb, they're frozen. The intelligence is already there — what's missing is the live loop that turns yesterday's usage into tomorrow's weights, the way Google Maps turns every trip into a better route for everyone.
02The Opportunity

Only Frontier Labs Could Turn Data Into a Model

<strong>Going from production data to a better model used to live only inside frontier labs</strong>, costing hundreds of millions of dollars in compute and the better part of a year — Trajectory's bet is that this becomes one click.

you needed hundreds of millions of dollars of R&D in compute. Uh you needed a really massive research team to invest in and then you also needed a lot of time. So it took us six to nine months to actually get to that point where we could train a frontier model.

— Ronak Malde, Sequoia Capital
Key Insight
Notice what Trajectory is really selling: not a better model, but the removal of a moat. If a one-click loop replaces a hundred-million-dollar research org, the advantage shifts from who can afford to train to who owns the most valuable production data.
03Proof Point

Harvey: 6× Cheaper, 4% → 18%

<strong>With Trajectory, Harvey's legal model went from 4% to 18% accuracy while running six times cheaper</strong> than the frontier models it used to rent.

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

— Ronak Malde, Sequoia Capital
Key Insight
The 4%-to-18% jump matters less than where it came from. Harvey's edge wasn't a bigger model — it was the thirty years of built-up legal nuance Malde says frontier models were unable to use. Owning the model is how that proprietary experience becomes a weight, not just a prompt.
04Reframe

A Trajectory, Not a Token, Is the Unit of Work

<strong>The real unit of AI work is the trajectory, not the token</strong> — one whole task of economic value, which in reinforcement-learning terms is a single rollout.

I want this basically paradigm to propagate where the unit of work that we think about is a trajectory.

— Ronak Malde, Sequoia Capital
Key Insight
Renaming the unit from token to trajectory is a quiet repricing of AI. If the thing you pay for is a completed task of economic value, not a word, then a cheaper model that spends more tokens per task can still win — which is exactly the business case for training your own.
05Research Breakthrough

Group Size One: Learning From One Rollout

<strong>'Group size one' rewrites GRPO to learn from a single real-world rollout</strong>, instead of replaying the same task 16 times in a simulated environment.

the goal is could we modify gRPO as an algorithm to get to a world where it is what we call group size one instead of needing a group of 16.

— Ronak Malde, Sequoia Capital
Key Insight
Group size one is the unlock that moves reinforcement learning out of the simulator. GRPO's demand for 16 identical replays assumes a world you can reset; real legal or finance work happens once and never repeats. Learning from a single real rollout is what makes production itself the training environment.
06The Missing Piece

LLMs Skip the Transition Function

<strong>LLMs jump straight from state to state but never model the transition in between</strong> — and that missing 'transition function' is what curiosity is meant to fill.

they're next token predictors, right? But what's missing is that intermediate part of what what is that transition function when you actually like go from S one or S0 to S1?

— Ronak Malde, Sequoia Capital
Key Insight
This is a precise claim about what next-token prediction can't do: it maps inputs to outputs but never builds a model of consequences. 'Curiosity' here isn't a personality trait — it's a training signal that rewards an agent for reducing its own uncertainty about how the world responds to its actions.
07Safety

Keep Humans in the Loop, or Agents Go Rogue

<strong>Continual learning keeps humans in the loop; recursive self-improvement does not</strong> — and that difference is Malde's line against agents 'going rogue.'

you need to allow these models and build the algorithm so that it keeps humans in the loop 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, Sequoia Capital
Key Insight
The alignment choice is also an engineering choice. A model that only self-improves in a black box has no external signal to correct it; one that improves from interacting with humans gets both its reward and a check on where it drifts. That is why Malde frames human-in-the-loop as the way to avoid agents 'going rogue,' not just a nicety.
08The Big Picture

Abundance Needs Specialization

<strong>AI today is a one-oracle economy; abundance needs specialization</strong> — every company training a model for its own work, taking AI adoption from 1% of enterprises toward 99%.

what he theorized basically is a world of abundance needs specialization. Uh where you want someone to basically not be context switching all of the time and work on the thing that they've built up experience about.

— Ronak Malde, Sequoia Capital
Key Insight
Reaching for Adam Smith reframes the goal: the bottleneck Malde points at isn't smarter models but specialization — moving off one shared 'oracle' so each company can train on its own built-up experience. His 1%-to-99% figure casts adoption as a distribution problem, the plumbing that turns a company's daily work into its own model.