The question that set Aaron Levie off was a technical one, and the host asked it plainly: today's dominant architecture for giving an AI system your own information is still retrieval — "you can get fancy on the RAG, but it's still context lookup where the weights themselves aren't fundamentally changing." Meanwhile, the host said, friends at the labs keep describing continual learning, "this idea of like the model's weights should adapt as it gets to know you." The show had recently had the startup Engram on, described in the conversation as working with customers to bake some of that context into the weights themselves. Where was this going?
Levie, the CEO of the content-management company Box, answered on the Training Data episode with an unusual disclaimer. He had just been introduced to Dan from Engram, had listened to the episode, and was recording shortly before his own call with him. "So I wish I could have talked to him first, and then I'll have, like, a way more eloquent answer."
Two ways to hand a model your company's knowledge
A model's weights are the numbers fixed during training that determine how it responds. In the common enterprise setup, those numbers never change: when the system needs to know something about your company, software retrieves the relevant documents and feeds them in alongside the question. That is retrieval-augmented generation, RAG. The knowledge lives outside the model and is looked up each time.
The alternative is to train the knowledge in. Engram's introduction, published in June 2026, proposes starting from a strong pretrained model and training it on an organization's private context instead of re-supplying that context at every request. The company described its own internal models learning from GitHub, Slack and Notion, retrained daily, with plans to move toward hourly and eventually minute-level updates, and it launched an API for agents working in large shared knowledge workspaces. It reported token-efficiency gains that varied by task, from not having to gather the same context repeatedly, while calling reliable continual updating at scale an open problem.
Levie said he was "extremely fascinated by the approach" and had "no reason for not wanting it to work and exist." Then he flagged what Box sees every day: "the high degree of complexity on permissions and access controls and data that tends to be sort of the rub on a lot of these types of approaches." He explicitly excluded Engram from the objection — "I'm going to put Engram aside because, like, I'm sure they've already thought this through" — and argued the general case instead.
"Let me introduce you to a lawyer"
The generic case is about who is allowed to know what. Levie said researchers sometimes imagine the world resembling their own: a researcher has access to everything, so a model trained on their entire world sounds wonderful. "And then you're like, let me introduce you to a lawyer."
The lawyer he described has a tiny window onto the firm: the five projects they are working on. One door over, a colleague is acting on the competing project for another company in the same space, and nothing can cross between them — no overlap in what they see or know, not a single document passing the wall. Those are hard barriers.
You could still train a model for that one lawyer, Levie allowed. His worry is what happens next. "What happens if every single day they get added or removed from something that adds important context to sort of what they need to understand?" A lookup system can honor a permission change the moment it is made. Knowledge already absorbed into weights is harder to add on Tuesday and withdraw on Wednesday.
He did not treat that as fatal. He expects breakthroughs in continual learning that "sort of all resolve this," and he pointed out how much the ground has moved: five years ago the whole idea was impossible, "insanely expensive" and unworkable to reason about at the access-control level. As costs fall and open-weight models get cheaper, smaller, faster and better, he said, it becomes genuinely interesting.
The T-chart he wants
What Levie took from the Engram conversation was a missing decision rule. He said he just needed "like, a T-chart, honestly": "what is the decision point of what goes in context and what goes in the weights?"
He sketched the two axes. On one side, the rate of change of the information — how fast it goes stale. On the other, the accuracy gained by training it in rather than retrieving it. Somewhere there is "some, like, incredible, like, calculation" of when data changes slowly enough and the accuracy upside is large enough to justify baking it in. "You'd have to kind of land on some sort of rubric like that."
The host offered the idealized version, attributing it to remarks Karpathy has made in prior interviews: if you could wave a magic wand, you would strip the memorized facts out of a model entirely and leave only reasoning ability plus a house style — "the ethos of how we do things, for example, at Sequoia" — with the actual content sitting in a lookup system beside it.
Levie found that appealing and then asked what it assumes. How much are enterprises actually distinguished by their style of execution, and how much by their literal intellectual property? "How many different types of styles of execution are there in the world?" he asked — versus the alternative, where the value is "the depth of knowledge about that particular legal case, and how do I apply it to this other project I'm working on?" If firms differ mainly in the second way, a model carrying only reasoning and ethos does not carry the part that matters.
Drug discovery has fewer walls
Levie noted he had jumped straight to the individual user, and said the firm level looks different. There, he is a fan: he named Trajectory, Applied Computer and Prime Intellect as companies he likes in this area. If you are Eli Lilly, he argued, you want a model for how you do drug discovery, and that probably has to go deeper and more specific than anything off the shelf. Drug discovery workflows also have few "church-and-state problems" — the firm generally wants that knowledge reaching as many of its people as possible, not partitioned between rooms.
That is where he landed, short of a verdict: the answer will be domain specific. "You're just going to have different outcomes based on which vertical or, you know, type of use case and where the firewalls need to be in that process." The lawyer and the drug researcher are not two opinions about the same architecture; they are two different problems, and Levie thinks they will get two different answers. He also still had a Zoom call to make.