By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Sequoia CapitalA Model For Everyone: How Ronak From Trajectory Is Building Products For Continual Learning
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 0:54
Combining training, evaluation, and deployment into one closed loop
Ronak Malde opens with an analogy to Apple combining the internet and phone into the iPhone, framing Trajectory’s product philosophy similarly. He describes unifying training, evaluation, and deployment so teams can ship a model, collect new production traces, and iterate quickly.
- •Product inspiration comes from Apple’s “combine previously separate things” mindset
- •Puts training, evaluation, and deployment in one place to make iteration simple
- •Emphasizes closing the loop by deploying, collecting traces, and improving continuously
- •Predicts this unified workflow will be adopted broadly by companies
- 0:54 – 2:48
Defining continual learning: from offline updates to “Google Maps” learning in production
Asked to define continual learning, Ronak compares it to how humans learn online and continuously. He contrasts today’s “pseudo” continual learning—batch updates and offline consultation—with a future where models update based on ongoing real-world use, like Google Maps learning from drivers.
- •Uses the human brain as the aspirational model for continual learning
- •Calls the current agent landscape “pseudo continual learning”
- •AAA/MapQuest vs Google Maps analogy: offline batch updates vs continuous updates
- •Trajectory’s near-term goal: democratize training/customization access
- 2:48 – 4:33
Architectures and feasibility: what’s changing inside modern models
Pat asks whether continual learning is possible with today’s architectures. Ronak argues that architecture is already evolving beyond “stacked transformers,” citing changes like different memory behaviors and long skip connections, and says the bigger missing piece is translating real-world usage into model improvements.
- •Notes many view LLMs as stacked transformer blocks, but architectures are changing
- •Mentions fast/slow memory within weights and skip connections across layers
- •Agrees architectural change is key, but focuses on connecting models to real-world use
- •Highlights the challenge: translating a company’s notion of “good” into model gains
- 4:33 – 5:56
From consulting-heavy model training to a productized interface
Ronak explains why going from data to intelligence has historically required labs, big research teams, and long cycles. He describes the repeating pattern across companies and argues that a great product packages the “knobs and levers” people care about, abstracting away the rest.
- •Says model customization has been lab-bound and required massive teams and compute
- •Describes prior experience training SWE-1 and the long time-to-train cycle
- •Frames early approaches as handholding/consulting-like and hard to scale
- •Defines product opportunity as packaging repeated workflows into a clean interface
- 5:56 – 7:36
Founding insight: alignment via real-world feedback across domains
Lauren asks about the founding “aha.” Ronak describes how each co-founder encountered the same core problem: using real-world feedback (edits/retries, robot deployment, artist feedback) to improve models, and realizing big labs aren’t set up to customize models per customer at scale.
- •Windsurf: learning from production edits and retries to train on real usage
- •DeepMind experience: deploying robots and updating them regularly via infrastructure
- •Apple experience: incorporating artist feedback into multimodal model work
- •Key realization: customizing one model per customer is infrastructure-heavy for big labs
- 7:36 – 8:56
Customer example (Harvey): owning domain data to shape model performance and cost
Ronak describes working with Harvey, which had relied on closed-source frontier models while building strong product interfaces for law firms. He explains Harvey’s motivation to use its domain knowledge and benchmark to train a model with Trajectory, citing improvements in accuracy and cost-efficiency.
- •Harvey built products on closed-source frontier models but couldn’t fully leverage its data
- •They created a legal agent benchmark using expert/domain information
- •Using Trajectory, they trained a model described as significantly cheaper and higher performing
- •Ronak emphasizes the strategic shift: being able to train with your own data to “shape your destiny”
- 8:56 – 10:13
Why “Trajectory”: shifting the unit of work from tokens to tasks
Ronak explains the company name as a reframing of what matters: not counting tokens, but focusing on economically valuable tasks or trajectories. He links the concept to reinforcement learning rollouts and argues that, in the real world, trajectories correspond to production inference doing valuable work.
- •Critiques token-counting as the main unit of work; argues tasks/trajectories are better
- •Connects “trajectory” to RL rollouts and training stacks
- •Frames production inference as the real-world rollout engine
- •Says the name also reflects companies owning and improving their path over time
- 10:13 – 11:16
Core primitives for continual learning: traces, replay, and the training stack
Asked about key ingredients, Ronak lays out a three-part pipeline companies need today: tracing what happens in production, making traces replayable for evaluation/training, and then modifying the algorithm via a training stack. He positions Trajectory as abstracting away the hardest part—training stack complexity.
- •First primitive: trace production trajectories including sub-agents and user feedback
- •Second primitive: replay traces to run evaluations and training in controlled setups
- •Third primitive: training stack complexity (loss, reward functions, algorithm modifications)
- •Trajectory aims to reduce inputs so teams mainly specify “what good looks like”
- 11:16 – 12:56
Where traces live today and the surrounding ecosystem
Pat asks where traces live in practice. Ronak describes an ecosystem approach, where some companies build in-house and others rely on existing products to integrate tracing into production agents.
- •Frames continual learning as requiring an ecosystem of infrastructure players
- •Says some companies trace in-house while others use existing tools
- •Gives examples of products used to plug into agent infrastructure for tracing
- •Notes AI-native companies commonly adopt third-party tooling for the tracing step
- 12:56 – 16:16
Reducing environment and reward friction: Group Size-1 and meta-graders
Ronak explains what still requires heavy manual work: building a faithful environment and defining reward. He discusses pushing research to reduce these burdens, including adapting GRPO toward “Group Size-1” learning from single rollouts and using representative sampling plus a meta-grader to scale reward labeling.
- •Identifies two major pain points: constructing environments and defining reward
- •Describes GRPO’s need to replay the same task multiple times as a group
- •Explains a push toward “Group Size-1” to learn from one rollout and reduce environment needs
- •Outlines selecting a small representative set to label, then training a meta-grader to generalize reward
- 16:16 – 17:37
The Trajectory platform: one-click train, evaluate, and deploy
Ronak introduces the product release as a simple interface to customize models without massive research effort. He emphasizes that users provide traces and “what good looks like,” then can train, evaluate, and deploy in one place—closing the loop by feeding production traces back into improvement.
- •Positions the platform as lowering time, compute, and staffing barriers
- •Describes a one-click training interface driven by traces and evaluation criteria
- •Combines training, evaluation, and deployment in a single workflow
- •Highlights iterative improvement by redeploying and collecting new production traces
- 17:37 – 18:58
Ideal customers today—and expanding from models to the agent harness
Ronak describes the current target as AI-native companies with their agent harness already built, plus some larger companies using the platform. He then outlines a broader vision: improving not only the model for efficiency and reliability, but also the harness for new capabilities and smarter routing across models.
- •Targets teams that already have an agent harness and want easier model modification
- •Mentions AI-native customers and larger organizations already using the platform
- •Distinguishes model improvements (efficiency/reliability) from harness improvements (new capabilities)
- •Envisions routing: using frontier models for a minority of tasks and custom models for most work
- 18:58 – 20:14
Customer example (Rogo): optimizing for concise, high “intelligence density” agents
Ronak shares how Rogo approaches agents differently, aiming for concise and efficient outputs rather than verbosity. He explains that Rogo uses Trajectory to train models that use fewer tokens while being cheaper, faster, and better aligned to what they care about.
- •Contrasts synchronous “prompt then wait” agent use with Rogo’s desired paradigm
- •Rogo aims for concise, clear outputs rather than verbose responses
- •Introduces “intelligence density” as doing the same task more efficiently
- •Says the trained model became more token-efficient while also cheaper and faster
- 20:14 – 24:19
Where continual learning is headed: curiosity, experience, and keeping humans in the loop
Ronak looks beyond translating “what good looks like” and discusses what’s missing for autonomous continual learning: curiosity and accumulating experience. He frames curiosity as understanding unknown transition dynamics in the world, then contrasts recursive self-improvement with continual learning as a more human-aligned path that preserves human involvement.
- •Says current models do tasks but don’t accumulate experience from the real world
- •Links curiosity to exploring unknown transition dynamics between states and outcomes
- •Discusses quantifying what an agent doesn’t know and learning by trying actions
- •Contrasts RSI’s black-box self-improvement with continual learning that learns through human interaction and stays aligned
- 24:19 – 26:08
Building Trajectory: blending frontier research, production engineering, and product design
Ronak explains what makes building at Trajectory distinctive: modifying core RL algorithms while delivering a polished product interface. He describes a team structure that productionizes the training codebase with scalable infrastructure, and notes how coding agents shift work toward whiteboarding and system design debates.
- •Sits at the intersection of frontier RL research and productization
- •Describes a team spanning research contributors and product/infrastructure builders
- •Productionizes training with microservices, Kubernetes scaling, and autoscaling
- •Notes design engineers run training and learn the research workflow; coding agents accelerate implementation
- 26:08 – 32:00
Product builders as “gardeners” and an optimistic vision for widespread adoption
Ronak reframes the role of product builders as shaping self-improving systems—starting with model improvement, then harness and product improvement. He closes with an optimistic vision: moving from one-size-fits-all intelligence toward specialization and broad AI adoption across workflows by making earned experience legible to models.
- •Envisions self-improving models and harnesses, eventually influencing the whole product
- •Compares product builders to gardeners who set constraints and shape growth
- •Argues continual learning can enable specialization rather than one “oracle” model for everything
- •Expresses a goal of expanding AI adoption far beyond today’s limited penetration by leveraging real-world experience
