The Work Is in the Seams
Every institution we've built — academia, funding, startups, even our AI tools — actively punishes anyone who tries to work between fields. The most valuable work of the next decade will happen in exactly those seams. Nobody has a home there.
Photo Credit: Rob Grzywinski
Someone asked me how my work on Answer Set Programming and LLMs was going. The honest answer turned into this piece. I owe them more than a Reddit reply.
When you do work that lives between two fields, you don't get two homes. You get zero. I spent a chunk of last year talking to the symbolic-AI people about LLMs and a chunk talking to the LLM people about symbolic reasoning. Both groups were polite. Both groups went back to their meetings.You don't get two homes. You get zero.
The symbolic folks see LLMs as stochastic parrots — a phrase that has done more damage to clear thinking than almost any I can name. The LLM folks see symbolic systems as a museum exhibit. They both know — in their souls, the way people know things they don't have to defend — that the other side is wrong. And so the conversation that needs to happen, doesn't.I want to be careful here. I'm not writing this to complain. I'm writing this because I finally figured out the shape of what I bumped into and I think you're about to bump into it too.
This Story Is Older Than I Am
It turns out everything I just described has a name. Several names, actually, scattered across disciplines that — fittingly — don't talk to each other.C.P. Snow named it in 1959. He called it The Two Cultures — the sciences on one side, the humanities on the other, with a chasm in the middle and a mutual disdain across it. Snow was writing about Cambridge dons. The pattern hasn't changed; it's just fractaled. Today the chasm runs between sub-disciplines, between sub-sub-disciplines. There are thousands of two-cultures now. Tony Becher showed in Academic Tribes and Territories that the tribes are smaller and more numerous than Snow imagined, and the rivalries are sharper.Thomas Gieryn called the policing of these borders boundary-work — the active, ongoing effort a discipline puts into deciding who's in and who's out. Karin Knorr Cetina called the differences inside the borders epistemic cultures — not just different facts, but different ways of knowing. Different criteria for what even counts as a thought.And then Peter Galison found the cure. In Image and Logic, he describes what happened in the radar labs of WWII, where theoretical physicists, electrical engineers, and instrument-makers — three tribes that should not have been able to talk to each other — built something together anyway. He called the place where they met a trading zone. They didn't need a shared theory. They needed a shared thing — a piece of equipment, a problem, a working artifact — to push against.That's the lesson, by the way. Trading zones don't run on consensus. They run on something to point at.
So Where Are the Trading Zones Now?
Mostly, they're not in universities.The funding system, as the research-policy literature has documented exhaustively, imposes what one paper calls a "disciplinary straightjacket" — stay in the lane where you already have publications. Reviewers can't evaluate work outside their lane, so they score it lower. Hiring committees, especially at top institutions, prefer candidates whose dissertations match the department's existing center of mass. The result is a quiet, relentless filtering against the people who could bridge the gap.The escape valve is supposed to be: leave academia, start a company. Get freedom. Think.I have a friend who's looking forward to leaving her job for a startup, specifically for the thinking time. I love her and I had to tell her: that's not what startups give you. The moment you take money you trade one clock for another — slower deadlines for faster ones, but a clock all the same. Execution replaces exploration almost immediately. That's not a flaw of the system; that's what the system is for.
And Now the Tools Are Doing It Too
Here's the part I didn't see coming.LLMs — the very thing that could be the great cross-pollinator — have started reproducing the silo structure of their training data. There's now a real, measured literature on this. Researchers studying open-ended scientific reasoning have found that LLMs exhibit mode collapse: they retrieve along surface features, not deep structural similarities. When I'm working with one on feedback particle filters, it talks to me about feedback particle filters. It does not spontaneously connect them to Sparse Distributed Memory or Hyperdimensional Computing, even though the connections are right there. I have to drag the model across the seam by hand.The mean is where they go. The mean is the silo.This is the deep irony of the moment we're in: we built a machine that could finally make C.P. Snow obsolete and we trained it on the exact corpus that proves him right.
The T, the π, the V
There's a vocabulary for the people who try to live in the seams. It's old and it's getting older.T-shaped has been around since the late 70s — one deep spike of expertise, a broad horizontal bar of competence across many adjacent fields. π-shaped people have two spikes. V-shaped, a term from the convergence-research literature, describes two depths that meet at a shared problem — two pillars converging on a vanishing point. David Epstein, in Range, made the popular case for why these people produce more interesting work than pure specialists. Collins and Evans gave us interactional expertise — the idea that you don't need contributory mastery in both fields, just enough to talk credibly across the boundary.I'm a T-shaped person. I always have been. For most of my career, that was treated as a minor charm — Rob can talk to anyone — but never the main thing. The main thing was always the spike.The spike is no longer the main thing.
The Flattening
Here's why I'm bothering to write this down.Agentic AI systems are eating the middle of the T — the vertical bar of competent execution that used to define a junior or mid-level role. The boilerplate, the glue, the translation, the standard report, the standard query, the standard PR. Gone. Or going.What survives is the two ends: deep expertise that the model can't fake and the breadth to connect that expertise to everything else. Specialists who can't connect become bottlenecks. Generalists with no depth become noise. The T-shape, the V-shape, the π — these used to be a luxury. They're becoming the only sustainable shape.And our institutions — academia, funding, hiring pipelines, conference tracks, even our chatbots — are still optimized to produce pure spikes.The demand curve and the supply curve are pointing in opposite directions. That's the whole essay, really.
So What Do You Do
You stop apologizing for the breadth.You stop trying to win in someone else's silo. You stop waiting for a department, a grant category, or a community Discord to give you the seat. There isn't going to be one. There never was — Galison's radar guys didn't have one either. They built a trading zone out of a problem and a piece of equipment and they pointed at it until other people could see it too.If you're a developer reading this and wondering whether to learn how the models actually work — do it. If you're an ML person reading this and wondering whether to learn how real software gets built — do it. The most valuable work of the next ten years is going to happen in the seams between what you already know and what the person across the room already knows.That's where the work is.That's where it's always been.We just didn't have a name for it and the structures around us were quietly making sure we'd never look.
PS — to the person who asked me how the ASP work was going: this is how it's going. Thank you for the nudge. The work is fine. The seam is the point.
A decade-old vision of AI companionship finds its moment as language barriers dissolve through LLMs. The personal journey from dusty pitch decks to daily AI integration reveals how shared context and understanding transform theoretical possibilities into practical realities.8 January 2025
Modern ML has a "looks confident while being wrong" problem that control theorists solved decades ago. Diffusion models, flow matching, selective SSMs, and test-time compute are all secretly rediscovering the same primitive — the gain function — from different directions. None of them cite Yang/Mehta/Meyn 2013.1 May 2026