Khurram Javed does not want teaching his AI assistant to mean competing with everybody else who uses it. “I want to teach my model something very specific, and I want to do it to my version of the model,” he said in the Training Data interview published on August 18, 2026.
That is the bet behind Oak Lab, the startup Javed cofounded with his former University of Alberta doctoral adviser, Rich Sutton. They want AI systems that keep changing their underlying learned machinery as they encounter individual experience—not just assistants that remember a preference or wait for the next shared model update.
For Sutton, a pioneer of reinforcement learning—the branch of AI concerned with learning through actions and rewards—the strange thing is having to argue for it. “I'm not weird. The field is weird,” he said. Learning, in his account, should be something an intelligent system keeps doing throughout its life.
Remembering a preference is not the same as learning a skill
Training a language model adjusts its weights: the many numbers that determine how it processes information and produces answers. In the deployment pattern the founders criticize, those numbers stay fixed while the model serves users.
The model can still respond differently when given new instructions, examples or saved memories. This is in-context learning: changing behavior using information supplied as input, without changing the weights. A memory can persist across conversations and be supplied again, but that is different from updating the model itself.
Sutton and Javed do not object to preparing models before deployment, or to using context afterward. They want weight learning to continue too, so the system can revise the skills and internal concepts acquired during training.
Javed illustrated the depth of adaptation he wants with proprioception, the sense that tells us where our body parts are without looking. He described people who lose that sense and consequently lose the ability to walk, then spend two or three years learning to walk again by watching their feet. Vision takes over a role previously filled by another signal.
In his telling, a foundation of behavior can stop being true after decades, and the brain can reorganize around that change. That is the kind of flexibility he wants AI to have, rather than merely following a new instruction within an otherwise unchanged system.
Cursor already learns—but from the crowd
The interview raised an exception to the frozen-weights picture: Cursor, the AI coding environment. Javed acknowledged that its Tab and Composer systems receive weight updates from experience. His objection is to how that experience is pooled.
Cursor’s September 2025 account of online reinforcement learning for Tab describes training on users’ acceptance and rejection of autocomplete suggestions. The company reported taking 1.5–2 hours to deploy a checkpoint—a saved version of the model—and collect fresh feedback. Its resulting model made 21% fewer suggestions with a 28% higher acceptance rate than its predecessor. Those are relative changes in suggestion frequency and acceptance among displayed suggestions, not percentage-point gains.
Its March 2026 account of real-time learning for Composer describes pooling billions of tokens from production use, deriving rewards from user interactions, updating weights, checking for regressions and deploying checkpoints approximately every five hours. Cursor says large batches are necessary because individual feedback is noisy.
These are real model updates, not merely longer conversations. But they improve a shared model. Javed argues that pooling feedback makes it inefficient to teach a particular version something specific to its user: one person’s signal must compete with many others’ preferences.
Oak’s intended alternative is a system that can learn from its own stream of experience. Javed’s research profile connects that objective to another constraint: processing experience and learning from it must fit within the system’s computing budget.
Two different ways learning can fail
Why not simply update the weights after every interaction? Javed’s answer is that the gap is algorithmic, not just a matter of deploying the training machinery.
A new example can push weights in directions that damage previously useful behavior. That problem is called catastrophic forgetting: learning something new at the expense of something already learned. Javed describes large, pooled batches as one way current systems manage the difficulty, but wants algorithms that work with a single stream instead.
A separate problem is loss of plasticity. A network can become worse at learning new material as training continues, even when remembering old material is not what the experiment demands. Preserving yesterday’s knowledge and preserving tomorrow’s ability to learn are related goals, but they are not the same goal.
The founders propose combining two ingredients. First, each weight would have its own learning rate, or step size: how far it moves when the model learns. The system would learn these rates too, an approach called meta-learning. Most weights could change very slowly, protecting useful knowledge, while others remain free to adapt quickly.
Second, the network would keep generating and testing fresh internal features. Their concrete example is continual backpropagation. Ordinary backpropagation adjusts weights according to how they contributed to an error. Continual backpropagation also replaces selected, little-used computational units with newly initialized ones, giving training fresh material to work with.
What the earlier experiments showed
The supporting research is a 2024 Nature paper by Shibhansh Dohare and colleagues, not a newly announced Oak Lab result.
In its Continual ImageNet experiments, networks tackled sequences of two-class image-recognition tasks. Each task supplied 1,200 training images, with 250 passes through that training set, and performance was scored on 200 held-out images. Across 30 runs, standard backpropagation deteriorated substantially by task 2,000; continual backpropagation sustained performance through 5,000 tasks.
The replacement process was selective. It ranked internal units by their contribution and replaced low-utility units that had been around long enough to assess. New units received random incoming weights and initially zero outgoing weights, limiting the immediate disruption to the network’s output. A minimum age protected them from being discarded before they had a chance to become useful.
Other approaches also helped, including L2 regularization, which discourages large weights, and a method called Shrink and Perturb. In separate CIFAR-100 image experiments, researchers retained earlier classes during training to isolate declining plasticity from forgetting. Continual backpropagation eliminated the observed performance deficit relative to retraining from scratch.
The authors note that plasticity loss may not appear over short task sequences. These results concern sustained adaptability in the tested settings; they do not establish that a general-purpose assistant can learn indefinitely from an individual user without damaging earlier skills.
Oak’s agenda also draws on the Alberta Plan, the twelve-step roadmap Sutton, Michael Bowling and Patrick Pilarski first published in 2022. Its early steps develop per-weight learning rates and the construction, evaluation and replacement of internal features under limited resources. Sutton sees continual deep learning as an enabling step for the rest of that program.
Javed argues that the combined approach should be trained from scratch, rather than bolted onto an existing model. His reasoning is that the system must acquire knowledge and learn how to handle future learning at the same time. That is his proposed development path, not an established rule that every continual-learning method requires a fresh model.
A small team willing to get worse first
Javed expects exploring this direction to mean accepting initially worse performance. In his view, large labs have a harder time making that bet because their current products still benefit from scaling the established approach. Oak intends to start small and grow slowly, drawing its initial team from people who have thought about these problems for years.
The founders also have a low-power ambition, with a substantial hardware qualification. Asked about a trillion-parameter model operating on 20 watts, Javed said that would be impossible with current technology: storing that many parameters in memory would probably exceed the budget on its own.
He thinks improvements in algorithms and computing efficiency could make it possible in five to ten years. That is a conditional aspiration, not measured Oak performance. For now, the company’s chosen path is a small research team prepared to build models that may perform worse at first, in pursuit of ones that can keep learning after they reach a user.