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

1. Key Themes

Continual learning today is "pseudo": agents are still on paper maps

Ronak frames the current state of agents as batch-updated and offline, with the end goal being real-time learning from every interaction. His analogy ladders the whole product vision. "We would go to AAA uh back in the day and like they would give us a bunch of these maps... anytime you want to update your course, you have to just do an an entire mass batch update." [00:01:14] Trajectory's first step is "MapQuest" (democratized, still-offline customization), and the destination is "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." [00:01:14]

The missing layer is translating "what good looks like" into model improvements

Intelligence is plentiful from foundation labs and experience is plentiful in every industry, but the bridge between them has been locked inside labs. Ronak: "the hardest part of this entire loop is then translating that to actual improvements for the model. Uh I think that's where that missing layer is." [00:03:12] Pat Grady's summary of Ronak's earlier framing: "the world right now has plentiful intelligence, thanks to the foundation model labs, and and plentiful experience... but there's this big gap between the two." [00:29:02]

Custom-model training has been gated by cost, team, and time, and that is the product opportunity

Ronak lived the pain at Windsurf: "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 6 to 9 months to actually get to that point where we could train a frontier model." [00:04:44] His product thesis: "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, and if you can package that into a really beautiful interface... that's the beautiful interface." [00:04:44]

Three ingredients of continual learning, and which one to abstract first

The loop requires (1) traces from production, (2) replayable environments, and (3) the training algorithm itself. Ronak: "the third part uh is then the actual modification, the algorithm. Uh and I think that's where uh it's really difficult to do this in-house as a company, is basically the entire training stack, uh the entire loss function, the reward function." [00:11:51] Tracing is already served by an ecosystem: "there's also a lot of great products out there like LangChain, like Braintrust, Raindrop." [00:11:20]

Research is attacking the two manual bottlenecks: environments and rewards

Trajectory is modifying GRPO to work at "group size one" so companies don't need replayable environments, and using a "meta-grader" trained from ~100 labeled traces to propagate reward across 10,000. Ronak: "if you can take 10,000 trajectories and basically uh they're all from different tasks, we can actually learn from just that one behavior, that one rollout." [00:13:08] And: "if I can smartly select 100 of them... and then have the customer... densely say, 'Okay, here's what good looks like'... we can basically train a meta-grader that learns from those behaviors of the 100 to the broader 10,000." [00:13:08]

The trajectory as the new unit of economic work (not the token)

"A lot of people view tokens as a unit of work. Uh but... with co-workers, we a lot of the time are thinking about a task or a trajectory. And that's the unit of work that I really want to get to." [00:09:02] In continual learning, "your rollout engine becomes just the inference and the tokens in the real world, and now suddenly your trajectories are the tasks of economic value that are happening." [00:09:02] Pat endorsed it: "I like this idea of a trajectory as like a unit of economically valuable work." [00:10:13]

Specialized, smaller models win on cost, speed, and "intelligence density"

Customers like Harvey and Rogo use their proprietary experience to train models that beat open-source baselines and cost far less than frontier. Ronak on Harvey: "train a model that is six times cheaper. Uh and it goes from around 4% accuracy for an open source model uh to around 18%." [00:07:40] On Rogo: they want "the idea of intelligence density. Uh they want the idea that a model can do the same task, but do that far more efficiently than the frontier models." [00:19:04] The end state is routing: "maybe you use a frontier model like Fable for 10% of tasks, uh and then for 90% of the time, you can smartly route to a model that you've trained that is far more efficient uh and also gets you better margins." [00:17:41]

Closing the loop (train, eval, deploy in one place) is the Apple-style product insight

"A lot of our philosophy in product actually comes from Apple... no one really thought to put the two together until Apple said, 'Okay, what if you actually put the internet on a handheld device.'" [00:00:00] The unlock is "putting training, eval, and deployment in one place... then you can close the loop and basically now upload that model back to production, gather new traces." [00:00:00]

Curiosity (learning transition functions) is the missing piece for autonomous learning

Ronak: "What LLMs right now are really good at is basically going straight from state zero to state one. Uh they're next token predictors, right? But what's missing is that intermediate part of what... is that transition function." [00:21:18] Humans "are very aware of when we don't know what a transition function is the world," which is where curiosity comes from. [00:21:18]

The product builder becomes a gardener; the harness is next after the model

As products become AI-native, "the harness is kind of becoming most of the product... if you have the right to kind of self-improve the harness and then the product, uh then you have the right to kind of improve and and self kind of self-improve the entire company." [00:26:14] The builder's role: "set these trellises in the ground of here's where I want to shape my product." [00:26:14]

Continual learning as the recommender-system moment for agents; specialization drives abundance

"Any massive company that you can think of of the previous age... Google or Amazon or Facebook, uh they were all doing continual learning, basically... for agents, that hasn't existed yet." [00:27:35] Ronak invokes Adam Smith: "a world of abundance needs specialization," and says only "1% of enterprises" have adopted AI, with a goal "to basically take that 1% and turn that into 99%." [00:29:55]

2. Contrarian Perspectives

Continual learning will make models more aligned, not less, versus recursive self-improvement

Most discourse treats RSI and continual learning as the same road to risk. Ronak draws a sharp line: "RSI is all about taking a model, putting it in a black box, and basically making it the best AI researcher itself... but this is kind of a scary world." In contrast, "continual learning takes a very different approach... learn from the experience of actually interacting with humans... you need to allow these models and build the algorithm so that it keeps humans in the loop." [00:23:07] This is a bet that grounding learning in real customer feedback is both safer and more commercially valuable.

The next leap isn't a new architecture, it's the missing product layer

Rich Sutton and David Silver argued continual learning isn't possible with today's architectures. Ronak partially concedes architectural change is already underway ("fast and slow memory within the weights... skip connections that are going between layer two and layer 50") but insists the real gap is elsewhere: "where I want to sit is basically leverage all of that technology and research that's happening at the architectural level, but then actually connect the dots to the real world." [00:03:12] The value capture is in the interface between research and enterprise reality.

Environments and verifiable rewards are not prerequisites for RL on real work

The conventional view is RL needs a perfect simulator and a verifiable reward, which most knowledge work lacks. Ronak: "a lot of the time it's actually impossible to create a perfect environment... a legal work flow or an investment banking work flow doesn't have some perfect verifiable reward." [00:13:08] His response is group-size-one GRPO, which "eliminates the need for an environment and all of this infrastructure that needs to be built out," and a meta-grader trained from ~100 labeled traces. [00:13:08]

Don't count tokens: the economically meaningful unit is the task

Tokens dominate how the industry prices and measures AI. Ronak rejects this: "I'm not counting the amount of words that you're actually talking about, right?... the unit of work that I really want to get to" is the trajectory. [00:09:02] That implies pricing, benchmarking, and training should be organized around completed tasks, and favors compact, efficient models over verbose ones (Rogo wants "a very concise and clear uh finance agent" rather than "a bunch of garbage"). [00:19:04]

Frontier-lab scale is a bad fit for customization; the future is many specialized models

The dominant lab model is "training one massive model and then distributing that one model to everyone in the world," which Ronak calls "a winning business model" but one that makes per-customer customization "an immense amount of infrastructure for a massive lab to build just for one customer." [00:06:07] His claim is that we are in a "pre-industrial revolution society," and abundance comes from specialization, not a single oracle. [00:29:55]

3. Companies Identified

Trajectory

Ronak's company building a one-click platform to train, eval, and deploy custom models from agent traces. Stated vision: a recommender-system-like continual learning layer for agents. Customers include Harvey, Rogo, and NVIDIA. "It's a one-click train interface that just takes a couple agent prompts uh and then you can eval the model and then deploy it, all in one place." [00:16:22]

Harvey

Legal AI company, an early Trajectory customer. Trained a custom model that is cheaper and better than open-source baselines using 30 years of domain experience. "they were able to... train a model that is six times cheaper. Uh and it goes from around 4% accuracy for an open source model uh to around 18%." [00:07:40] Ronak: "now suddenly get the power of a research lab." [00:07:40]

Rogo

AI finance agent company and Trajectory customer, pioneering asynchronous, concise agents. "Rogo is taking a very interesting paradigm to agents that I think will become the future... a very concise and clear uh finance agent... that can autonomously do things on the side, but then it comes back and is very efficient." [00:19:04] Result: "a model that is several times more efficient in terms of its tokens and also happens to be cheaper and faster." [00:19:04]

NVIDIA

Named as a larger company already using the Trajectory platform. "some larger companies as well, like NVIDIA are already using the platform." [00:17:41]

Windsurf

Where Ronak worked previously; trained the SWE-1 model on production signals such as edits and retries. "I was previously at Windsurf where we embarked on this journey and trained the model, SWE-1." [00:04:44] Origin of the insight that custom model training is a repeatable product need.

Cursor

Mentioned alongside Windsurf as a vertical company that built its own frontier-ish model. "so did Cursor as well with Cursor Composer." [00:04:44] Evidence of the vertical-model trend that Trajectory wants to democratize.

LangChain, Braintrust, Raindrop

Observability/tracing products that AI-native companies use to log production agent traces, the first ingredient of continual learning. "there's also a lot of great products out there like LangChain, like Braintrust, Raindrop, where you can basically plug into your infrastructure or your production agents uh and then basically trace all of these things." [00:11:20]

Modal

Sandbox/infrastructure company located near Trajectory's office, powering part of its stack. "two blocks down from us is Modal uh at our office, that that works on sandboxes and that powers a lot of the infrastructure that we have." [00:27:35]

Thinking Machines

Named as part of the ecosystem building continual learning infrastructure. "two blocks down from that is Thinking Machines also uh in SF." [00:27:35]

DeepSeek

Cited for architectural innovation relevant to continual learning ("fast and slow memory within the weights"). "DeepSeek-V 4.1 Flash that just came out recently, uh their architecture is is actually kind of lobotomized or changed a lot of that base architecture." [00:03:12]

Google DeepMind

Where co-founder Michael worked on Gemini Robotics and where Ronak saw the constraint of one-model-for-everyone. Co-founders contributed to "Gemini 3 and Gemini Robotics." [00:24:31]

Apple

Product-philosophy inspiration (combining internet and phone), and where co-founder Arjan worked on Vision Pro. "a lot of our philosophy in product actually comes from Apple." [00:00:00]

OpenAI

Employer of Trajectory research-team members. "other folks to GPT-4o at OpenAI." [00:24:31]

Meta (MSL)

Source of research talent. "building Muse infrastructure at MSL." [00:24:31]

Stripe

Source of product-infrastructure talent. "people that have built scalable infrastructure at Stripe for payment SDKs." [00:24:31]

Netflix

Source of design talent. "we have the youngest design engineer at Netflix." [00:24:31]

Google, Amazon, Facebook

Cited as the prior generation's continual-learning companies via recommenders and ads. "they were all doing continual learning, basically... taking all of these preferences from the real world and then creating online systems that can learn and get smarter the more they're used." [00:27:35]

AAA / MapQuest / Google Maps

Used as the analogy ladder for continual learning stages (paper maps, MapQuest, Google Maps). [00:01:14]

4. People Identified

Ronak Malde

Founder of Trajectory; ex-Windsurf (worked on SWE-1). Thesis: continual learning is the next learning paradigm, made accessible through a simple product interface. "I started to realize, 'Okay, continual learning is going to be the next learning paradigm.'" [00:06:07]

Michael (Trajectory co-founder)

Previously at DeepMind working on Gemini Robotics; saw robots deployed in the real world with daily-updating continual learning infrastructure. "he was working on Gemini Robotics. And they're starting to see the same thing where you could actually deploy these robots in the real world, build continual learning infrastructure to basically allow these robots to update every single day." [00:06:07]

Arjan (Trajectory co-founder)

Previously at Apple on Vision Pro, working on using artist feedback for multimodal foundation models. "Arjan was at Apple working on the Vision Pro... 'How can we take artist feedback and use that for multimodal foundation models?'" [00:06:07]

Rich Sutton and David Silver

Reinforcement learning luminaries who recently appeared on the podcast arguing continual learning isn't possible with today's architectures. Pat Grady: "both of them kind of made the comment that continual learning isn't really possible with today's architectures." [00:02:48]

Andrej Karpathy

Referenced via his tiny GPT video as the canonical explanation of stacked transformer blocks. "maybe Karpathy's tiny GPT video." [00:03:12]

Adam Smith

Invoked for the specialization-drives-abundance framework that Ronak applies to AI. "Adam Smith's Wealth of Nations... a world of abundance needs specialization." [00:29:55]

5. Operating Insights

Productionize the research codebase and embed designers in the training loop

Most AI companies keep research in "spaghetti code" while polishing product. Trajectory inverts this. Ronak: "we've actually... productionized the code base for training as well. Uh so this means microservices, Kubernetes scaling, auto scaling of every single part of the trainer, the rollout engine. Uh and then on the design side as well, our design engineers are kicking off training runs all of the time and understanding deeply what the research experience is." [00:24:31] The tactic: force product, design, and research to share one real system so the interface reflects true research constraints.

Spend human time on system design and whiteboarding, delegate implementation to a software factory of agents

"Actually, most of the time now, we are spending our time whiteboarding uh and we'll spend a lot of time just arguing about system design, uh and then the actual implementation is just a software factory of agents that will go and and actually implement it." [00:24:31] The scarce resource becomes architectural judgment.

Hire at the research-product intersection with deliberately mixed backgrounds

Ronak's team mixes frontier lab alumni (Gemini, GPT-4o, MSL) with product-infrastructure and design talent (Stripe SDKs, Netflix). "I think the intersection between them is where the magic starts to happen." [00:24:31]

Treat product building as gardening: set trellises, then let usage shape the product

"Your goal as a product builder is to basically just set these trellises in the ground of here's where I want to shape my product... put a a product out to a couple hundred users, uh it's going to grow on its own, and now we're going to have basically you just imbue your taste uh into where the company grows." [00:26:14] Operating implication: instrument traces and feedback from day one so the product itself can learn.

Reduce customer labor to the minimum: label 100, propagate to 10,000

When a workflow requires customer effort (labeling rewards), shrink it with smart sampling. "If I can smartly select 100 of them uh that are basically representative of all of these broader 10,000... train a meta-grader." [00:13:08] The long-term interface: "Here's the five things my agent's good at, here's the five things it's bad at, and I just circle those five things and hit train." [00:13:08]

6. Overlooked Insights

"Group size one" GRPO removes the biggest infrastructure moat for enterprise RL

Mentioned as a research aside, but it matters commercially. Standard GRPO "need[s] to play that out eight times or 16 times, and we call that a group," which requires replayable, resettable environments. Ronak says Trajectory's modification means "if you can take 10,000 trajectories... from different tasks, we can actually learn from just that one behavior... that actually eliminates the need for an environment and all of this infrastructure that needs to be built out." [00:13:08] If this holds, almost any company with production logs becomes a viable RL customer without building simulators, widening the market dramatically.

Production-trace data portability and the observability layer are becoming strategic chokepoints

Ronak casually notes that tracing is handled by LangChain, Braintrust, Raindrop, or in-house, and that the "entire ecosystem of continual learning... needs to be a lot of players." [00:11:20] Since the full timeline of "main agent call, but also all the sub-agents, what's happening in the product as well, any feedback that the user is giving" [00:10:30] is the raw material for improvement, whoever owns clean, replayable traces owns the starting point of the loop. Observability vendors and infrastructure like sandboxes (Modal) are quietly positioned as upstream suppliers to every continual-learning company.