iiinova — where your digital self keeps evolving

Me trying to understand actual math instead of just "make number go up"

110

By the way, your SSM tutorial at CCN has brain-poisoned one of my students, so now we're trying to figure out how to make it work for us—will ping you when we get a little further along 🙌

Lab Assistant (Research Lab Tech I) at USC

Applications are being reviewed on a rolling basis and will be reviewed until the position is filled. To apply, please provide: 1. CV/Resume 2. Cover Letter Apply here: usccareers.usc.edu/job/los-ange...

usccareers.usc.edu

Learn more about applying for Lab Assistant (Research Lab Tech I) at USC

010

This will be an fMRI-forward position. We need someone who can help lead the data collection effort on intensive, naturalistic fMRI paradigms for natural language, interactive conversations, and social-narrative videogame play.

How do individual brains converge on a shared understanding of the world? Explore our research at USC.

The Shared Minds Lab led by PI Sam Nastase at the University of Southern California (shared-minds.github.io) is accepting applications for a full-time lab manager / research assistant to begin in Fall 2026.

shared-minds.github.io

Shared Minds Lab

Sam Nastase @samnastase.bsky.social · 19/08/2026

We’re recruiting a full-time lab manager to join the Shared Minds Lab at USC! This will be a great opportunity for someone who wants to get hands-on experience with research before starting a PhD program in psychology or neuroscience. More below:

44123

Under topographic constraints on functional organization, I would expect these functionally interdependent features to be colocalized. Put differently, I don't see a strong reason for the brain to factorize and anatomically segregate them. Let me know if I'm missing anything here!

110

Sam Nastase @samnastase.bsky.social · 19/08/2026

Good point! I think another part of the argument is that the patterns we call syntax and semantic are highly interdependent, despite the fact that we can *sometimes* disentangle them theoretically or experimentally. The brain is highly sensitive to these statistical covariations and constraints.

agreed! and probably lots of other stuff we don't really have names for

010

These models are also deceptively elegant in important ways. Their complexity largely reflects the complexity of their environment. The fact that they can pull this off with “just a bunch of matrix multiplications” is precisely what makes them so relevant for human neuroscience.

161

Finally, we argue that the apparent complexity of LLMs should not discourage us from trying to understand them. They are not black boxes! They are fully transparent—and we hope the explosion of interpretability work continues to cross-pollinate with computational neuroscience.

161

In LLMs, syntax and semantics are processed through the same neural mechanisms and both are encoded geometrically in a neural population code. If the brain follows a similar strategy, we should not be surprised that syntax/semantics are co-localized and functionally entangled.

For example, years of neuroimaging work tried to disentangle the cortical bases of syntax and semantics based on their separability according to linguistic theory—but more recent work indicates that syntax and semantics are highly coextensive throughout the language network. Why?

141

Later in the paper, we walk through some examples of empirical work in speech and language neuroscience where LLMs might provide some insights 🧠

To the LLM, these are all just patterns in context—it will learn them insofar as they are *useful* for producing natural language. This context-first approach to language renders different types of linguistic structure learnable by a simple, general-purpose learning algorithm.

121

Certain patterns are highly regular; these are the kinds of patterns that are more easily detected and have been richly described by linguists. Other patterns are more abstract, graded, or multidimensional, and therefore harder to describe.

121

The model learns that certain verbs take a regular past tense while others have irregular forms, but it doesn’t process these two classes in a fundamentally different way; it doesn’t turn on a “rule” for one type of input and memorize others.

130

Principle 2 also has important implications: LLMs do not rely on any language-specific learning machinery; rather, LLMs reframe the problem of language learning as a problem of learning patterns in context.

This dense, geometric representational format also yields a surprisingly powerful form of generalization: LLMs can interpret (and produce) novel linguistic contexts by interpolating the meaning of new locations in this geometric landscape (within the bounds of their prior experience).

Principle 1 has important implications: for example, it helps us understand how many seemingly different structures of language (e.g., syntax trees, semantic categories) can be unified in a geometric lingua franca that supports neural computation.

132

Principle 2: They replace rule-based learning with a generic, self-supervised, context-driven learning algorithm for reproducing the statistical structure of real-world language.

Principle 1: They encode all of the structures of natural language (syntax, semantics, pragmatics, etc.) in the geometry of a continuous, high-dimensional embedding space—i.e., a neural population code.

121

We argue that neural network models for language—like LLMs—have ultimately outstripped symbolic, rule-based models of language processing by appealing to two basic principles of biological learning:

130