When a language model reads a story it has never seen, it has to keep revising its picture of what is happening: who is good, who is bad, who is happy and who is sad. Eric Bigelow, a researcher at the interpretability company Goodfire, said that this ordinary act of reading is a better guide to how models adapt than the narrow examples researchers usually study. On the October 10, 2026 episode of The Cognitive Revolution, he told host Nathan Labenz that this view also shapes his bet about how AI will be personalized. He expects better use of context rather than a separately trained model for every user.
A broader idea of in-context learning
A model can change its behavior in two broad ways. It can be retrained, which changes its weights, the billions of numbers learned during training. This is called fine-tuning. Or it can adapt to what it is given in the moment, such as instructions, examples or a long document, without any change to its weights. That second route is called in-context learning.
Labenz opened the topic by pointing to research that finds a close relationship between in-context learning and fine-tuning, which he called something like an isomorphism while allowing that this might be too strong. He said it seems especially true at LoRA scale. LoRA, short for low-rank adaptation, keeps a model's original weights frozen and trains a small add-on update instead, so the changes stay contained.
Bigelow said the term "in context learning" has historically meant few-shot prompting: showing a model a handful of input-output examples and measuring how its performance improves as more examples are added. He uses a much wider definition. He cited a perspective paper by Andrew Lampinen and colleagues, The broader spectrum of in-context learning, and said in-context learning captures "everything the model is doing to adapt its behavior" that is not learning in the weights. That includes how its internal representations change "as it reads each new word of input text," and how a reasoning model uses the text it generates to steer what it does next.
The Lampinen paper makes a similar case. It treats following instructions, adopting a role, learning from explanations and even resolving ordinary language puzzles as forms of contextual adaptation. In one of its examples, a passage about a river tells the reader which meaning of "bank" is intended.
Bigelow said this view "gives some teeth" to prompting. He said prompt engineering is "a little bit of like a dirty word" among scientists because it can look like a pile of hacks, such as telling a model it is a brilliant software engineer so that it writes better code. He suggested that some of those tricks matter less now, perhaps because that behavior has been fine-tuned into the models. But he argued that prompting "defies comparison" with learning in the weights and will not disappear "no matter how much fine tuning you give," because people will always hand models new contexts and want them to adapt.
Stories as a test case
To make that concrete, Bigelow turned to stories. Even a heavily fine-tuned model reading a new tale should be "updating its beliefs internally about what's going on in the story." One inspiration for his research, he said, was Kurt Vonnegut's famous talk on the shape of stories. Bigelow walked through Vonnegut's version of Cinderella: she starts out miserable, everything turns great once she is invited to the ball, the magic falls apart and she returns to her old life, a bit better off than before, and then the prince finds her and the ending is happy. "Stories have shapes," Bigelow said, "and really all contexts that we give models also have shapes that are dynamic and temporal."
The paper he mentioned, Stories in Space, published in May 2026 by Bigelow and colleagues, tested the idea in a controlled setting. The researchers fed Llama-3.1-8B-Instruct, an open model with 8 billion parameters, synthetic short stories one sentence at a time: 2,500 for training and 500 held out for testing. After each sentence they asked the model to rate six emotions, six genres and six arbitrary control concepts. Separately, they recorded its internal activations without asking it anything. Simple linear probes, small classifiers that read a concept off the model's internal activity, could predict how the ratings would change. The geometry of the model's answers closely matched the geometry of its internal activity.
Emotions and genres formed orderly, interpretable structures, while the arbitrary concepts did not. Pushing the model's internals along these directions, a technique called steering, worked better for emotions and genres. It also had side effects: steering toward sadness raised anger and lowered happiness. Nudging a single layer of the network was often undone by later layers, and intervening across seven layers worked better. The measured concepts described whole stories rather than individual characters, and the researchers chose them, so the study does not directly test the good-guy and bad-guy tracking Bigelow used as his example. A Goodfire write-up describes the emotional side as a path through a map defined by how positive or negative a feeling is and how energetic or subdued.
Bigelow drew a broader lesson for interpretability, the field that tries to explain what is happening inside a model. LLMs, he said, are "updating their beliefs" and "shifting representations internally" all the time: over a conversation, within sentences and while reasoning through long outputs. Much interpretability work misses this, he argued, because it relies on fixed templates in which only a couple of input words change, and it measures behavior at a single output token.
Will personalized models break the tools?
Labenz raised continual learning, the idea that a deployed model keeps updating from experience. The naive worry, he said, is that every user's model drifts in its own direction and becomes impossible to inspect. The counterargument is that long contexts already create that variety, and researchers already deal with it.
Bigelow began by noting the limits of his view. It is something he has not personally researched or thought about in depth, and the answer "really depends on how continual learning plays out." Suppose a tool is built for a base model, which is then fine-tuned for 10 people or 10 million. The key question, he said, is how much the model's internal representations change. He pointed to work by himself and others at Goodfire suggesting that much of what a model represents lies on low-dimensional manifolds, compact underlying structures that organize related concepts. As long as that structure stays largely the same, he expects existing techniques to carry over, perhaps with "a bit of fine tuning of your probes" for each personalized model.
The real trouble would come if continual learning caused catastrophic forgetting, the loss of earlier abilities when a model learns something new, and that forgetting were severe enough to reorganize the model's internals completely. Bigelow said he doubts this will happen, because it would probably mean "the models will break." He allowed one exception: a world spending "an ungodly amount of compute" on each of those millions of users, using meta-learning (training a system to learn well) so the models keep their old skills while building new internal structures.
Labenz added two observations. He said it has been striking that Anthropic has never offered a fine-tuning product to retail developers and that OpenAI shut down its own. He also said interpretability has transferred surprisingly well between architectures, citing Othello-GPT and Othello-Mamba, models trained on the board game Othello that both turned out to have similar, readable internal board representations. The Othello-Mamba repository supports that with qualifications. Linear probes recovered square-by-square board states in both Mamba and transformer models, but the probe reading its 9-million-parameter Mamba model reconstructed the entire board exactly 67% of the time. These are small models in a game where the true state is known, so they bear only loosely on frontier systems.
The bet on context
Bigelow's own forecast leans the other way from per-user training. He said there is "a bit of a bitter lesson" in prompting, context engineering and harnesses, the software loops that let a model use tools and carry out tasks: they "might be all you need for quite a lot of stuff." He acknowledged a bias, since this is what he works on. In the short term he expects the main difference between users to be their context rather than models trained on different data.
Memory is the clearest case. As Bigelow sees it, a model's memory is simply "a bunch of data that is tokens" that the model learns from in context each time it answers. It is easy to get "a lot of juice out of that without squeezing too hard," he said, and the resulting memories are "pretty interpretable." Some variables would have to change, he said, before fine-tuning separate models on individuals' data looks promising. His best guess is more context engineering, more harnesses and more memories, and he said that much of the engineering behind the rise of agents over the past year or so has gone into putting a lot of stuff in context.
He sees agent training itself in a similar light. Much of a model's raw capability was probably present after pretraining on internet text, he said, and later work has mostly drawn it out. A coding agent still needs heavy training to behave properly as it makes many changes, works with git and diffs and runs tests. But Bigelow described that as "more than anything like a small distribution shift or a sharpening of a distribution."