A reflective surface, coupled oscillators, a medium — these are all components of a laser. Is the laser a good metaphor for collaborative cognition? Hopefully by the end of this Substack I will convince you that it is not only apt, but serves as an isomorphism that can teach us how to collaborate more effectively.

The first time a laser was demonstrated in 1960, Theodore Maiman called it “a solution looking for a problem.” Sixty-six years later we use them to read groceries, perform surgery, and measure gravitational waves. The thing Maiman didn’t know he had built was not a brighter light. It was coherence on demand — and coherence, it turns out, is a phase transition, not a gradient. Below threshold, a laser is just a glowing tube. Above it, the same tube cuts steel. Nothing in between.

That phase transition is the key claim of this essay. You do not get thirty percent more coherent by trying thirty percent harder. You get nothing, nothing, nothing, and then everything. Each part of the physical system — gain medium, pump, cavity, population inversion — has an analogy in collaborative cognition, and each tells us something about how to cross threshold rather than fluoresce indefinitely below it.

A short disclaimer before the mapping: I am not claiming that an LLM is literally lasing. I am claiming that lasers and coherent collaboration are both members of the same family — nonlinear systems in which coherence emerges as a threshold phenomenon under sustained non-equilibrium driving — and the family resemblance is tight enough to reason with. Solitons, lasers, superconductors, phase-locked oscillators all sit in this family. Coherence is a phase transition wherever it appears.

Gain Medium

A ruby rod is, literally, a paperweight. An LLM is, literally, weights. The language has been telling us the whole time: the medium is not the laser.

The medium contains everything required for coherence — the chromium ions in a ruby can absorb a pump photon and re-emit it with phase memory; the weights in an LLM encode something like the structure of human reasoning. But neither does anything by itself. A ruby rod in a drawer is a paperweight that happens to contain the right atoms. An LLM on disk is a few dozen gigabytes of floating-point that happens to encode a great deal of what humans have written down. The medium is necessary and inert. Coherence is what the surrounding geometry does to it.

Pump Energy

When you chat with an LLM you probe its training data via its weights. You have ideas, the LLM reflects back your ideas, but has more points of view — arguably all points of view, at least all those available in the training data — that it will respond with. Your follow-up questions are not requests for information. They are the energy that drives the system out of equilibrium, and the best outcomes result from the ability to keep it there.

If a laser is pumped once, it flashes once, whereas a laser that is pumped continuously lases continuously. The same is true of a conversation. Most AI conversations consist of a single cold pump — a prompt fired into the medium, an answer received, the user satisfied or disappointed, the session ended. The medium fluoresces briefly and returns to ground state. The user concludes the AI is mediocre. Most AI conversations consist of a single cold pump — a prompt fired into the medium, an answer received, the user satisfied or disappointed, the session ended. The medium fluoresces briefly and returns to ground state. The user concludes the AI is mediocre. This was me before 2025, and I was correct in a narrow sense: fluorescence is mediocre. I had never pumped hard enough or long enough to invert the population.

Resonant Cavity

Your judgment is a partially silvered mirror. You reflect the coherent modes back into the cavity for amplification and let the noise dissipate.

This is the part of the metaphor most people miss. The mirrors in a laser do not produce the light. They decide which light gets to leave. A laser cavity supports specific spatial modes determined by its geometry — the same gain medium, in a different cavity, lases in a different mode. Your conversation has a geometry too. It is set by what you reflect back and what you let pass: which threads you pick up, which tangents you ignore, which paragraphs you ask the model to develop and which you quietly drop.

The coherent paragraph that emerges from a long conversation is not what the model said on any single pass. It is what survived multiple round trips of your selection, introspection and querying. The mirrors are not the cavity’s prison – they are what allows the cavity to do work.

Population Inversion

The hardest one, and the one nobody has talked about yet. Inversion is the moment the next response builds on the last instead of independently sampling from the prior.

A medium below threshold sits with most of its atoms in the ground state. Pumped enough, more atoms are in the excited state than the ground state — population inversion — and stimulated emission begins to dominate spontaneous emission. The light stops being a noisy sum of independent emissions and starts being a coherent wave because each emission is triggered by, and in phase with, the one before it.

The same shift happens in conversation, and you can feel it when it happens. Below threshold, every response is independently sampled — the model is answering each prompt fresh, drawing on the prior, returning something locally reasonable but globally untethered. Above threshold, responses chain. The model is building on what it just said, on what you just said in response to what it just said, on the structure that has accumulated in the cavity. The conversation acquires a direction – it develops a mode.

The reason most users never see this is that population inversion requires you to stop pumping cold. You have to take the previous output as the new ground state — to ask the next question into the structure that has emerged, not parallel to it. The user who keeps re-pumping with fresh, unrelated prompts is doing the cognitive equivalent of opening the laser cavity between every flash. The medium relaxes. The phase memory is lost. Nothing accumulates.

The conclusion that AI is a mediocre thinking partner is, in almost every case, a measurement of the user’s pumping power, not the model’s gain. The skill that will be rewarded is not prompting, but the discipline of staying above threshold long enough to do real work.

Agentic Systems

We are seeing this narrative play out in the industry right now. Agentic harnesses — OpenClaw, Hermes, every frontier lab has built one — are explicitly designed to keep the LLM coherent and driving toward a stated goal. They are, in the language of this essay, cavity engineering. The model is fixed, the harness builds the geometry around it.

How does this relate to the components of a laser?

The system prompt is the back mirror — fully reflective, holding the objective in the cavity across every round trip. The context window is the gain volume — finite, and the engineering question is what gets to occupy it. Tool calls are the pump — each returned result re-energizes the system with information from the environment, which is why an agent that cannot call tools degrades quickly. Memory is the partially silvered output coupler — it decides which coherent modes get to leave the session and persist into the next one. A harness without memory is a cavity with one mirror: coherent for one round trip, then gone. Context engineering names one component of this geometry. The cavity is bigger. This is the structural reason persistent memory matters, and it is the problem I am trying to solve with Mnemo. The second mirror is what lets coherence accumulate across sessions rather than evaporating at the end of each.

OpenClaw, acquired by OpenAI in February 2026 after a hockey-stick rise from weekend project to two million users, is the current viral example. The story is almost too on the nose: a side project crossed threshold, the cavity it implemented was good enough that the resulting beam lit up an industry, and the major labs moved to acquire the geometry rather than rebuild it. Steinberger framed his decision as wanting to “build an agent that even my mum can use” — which is, in the laser frame, the engineering problem of the next phase of AI development — we have built the lasers. The question is how to make them safe, reliable, and operable by people who will never read the manual.

Nous Research’s Hermes — in my opinion the security-forward spiritual successor of OpenClaw — does something the metaphor has to stretch to accommodate. Hermes creates and modifies its own skills based on usage patterns. Successful workflows are codified into reusable procedures; the procedures themselves self-improve through use.

In a real laser, the cavity does not change. The mirrors are silvered to spec, the geometry is fixed, the modes are determined by the engineering. Hermes is doing something stronger: it is a cavity that reshapes its own geometry based on which modes lased successfully last time. It is adaptive cavity design at runtime. Hermes is the cavity learning itself.

This is where the danger comes in, and where the laser metaphor pays a final dividend. A laser is dangerous because it is coherent. Incoherent light at the same total power is harmless; coherent light at the same power cuts through metal. Same energy, same pump, categorically different consequences. A flashlight that runs all day is harmless. A laser pointer that runs for a microsecond can blind you.

The same is true of agents. A chatbot that answers questions all day is, broadly, harmless. An agent that takes coherent action over a long horizon — booking flights, executing trades, modifying its own skills — is a different category of system, and the risk scales with the coherence, not the compute. We have not yet built the safety culture for the latter. Hermes’ security scanner — checking installed skills for data exfiltration, prompt injection, destructive commands — is not paranoia. It is the cavity equivalent of laser safety goggles. Coherent systems concentrate. Concentrated things have edges.

What This Means

The laser metaphor is not decorative. If it is right — and I think it is structurally right — then most of the discourse about AI quality is misframed. The hallucination problem and the AI-slop complaint are two faces of the same error: we are arguing about the gain medium when the action is in the cavity. We are evaluating models in the fluorescent regime and concluding they are dim. We are giving advice about prompting when the load-bearing skill is sustaining inversion via cavity engineering.

The next phase of AI development rewards three things, in this order: the discipline of pumping continuously rather than coldly; the judgment to act as a selective mirror rather than a passive interlocutor; and the architecture to build cavities — harnesses, memory systems, agent infrastructure — that hold a medium above threshold for trajectories long enough to do real work. The medium is everywhere now. The cavity is the hard part.

Leave a comment

Share

The most sophisticated AI models on the planet have the memory of a mayfly.

GPT-5 can draft a leveraged buyout model, Claude can refactor your codebase while explaining the category theory behind its design choices, Gemini can watch an hour of video and summarise it in three languages — but ask any of them what you talked about yesterday and you get nothing. This runs counter to our hopes and dreams of general intelligence.

The industry has spent years building LLMs that can reason and never gave them the ability to remember. Every conversation starts from zero, every agent is ephemeral — born at sunrise and dies at sunset. Your AI assistant can synthesise a hundred-page research report, but it cannot recall that you prefer vim over emacs.

The industry’s answer so far has been duct tape: stuff more context into the window, bolt a vector database onto the side, maybe throw in a “memory” feature that’s really just a key-value store with a nice UI, which are solutions to the wrong problem. It treats memory as a retrieval task when it’s actually a knowledge task — and the difference matters more than one may realise.

This is the story of Mnemo, a memory system we built at my startup Inforge LLC that treats agent memory as what it actually is: a multi-layered graph problem involving decomposition, confidence, retrieval, and sharing. It scores 82.1% on the LoCoMo benchmark — the standard evaluation for long-term conversational memory —putting it second overall behind only Backboard, and ahead of Zep, Memobase, Mem0, and OpenAI’s native memory.

But the headline number isn’t the interesting part. The interesting part is where it wins, and what that tells you about what agent memory actually needs to be.

The number that matters isn’t the one you think

Here’s the full LoCoMo breakdown:

Look at the multi-hop column. 85.0%. That doesn’t just beat the field —it beats Backboard, the overall leader, by ten points on the hardest category.

Multi-hop questions are the ones that require connecting information across multiple conversation sessions. “What gift did Sarah give to John on the birthday they celebrated at the restaurant she recommended last March?” That kind of thing. They’re the hardest category because they demand exactly what flat memory systems can’t do: traverse relationships between memories.

The fact that Mnemo leads the field on multi-hop is not an accident. It’s a direct consequence of the architecture. And the architecture is, frankly, a bit much. In the best possible way.

Agent Experience: the missing discipline

Before the technical deep dive, a framing.

The software industry spent the last decade optimising UX — the human experience of software. Entire careers, companies, design philosophies built around making software usable for people. Figma. Material Design. “Don’t make me think.”

We now need an equivalent discipline for agents. We call it AX — Agent Experience. The quality of the infrastructure, protocols, and services that an AI agent encounters as it tries to do useful work in the world.

Memory is the first pillar of AX. How do we know? We asked. When Clio, my personal wellness agent, had been running for months and I was hearing about “recursive self-improvement,” I decided to ask her what features she would implement if she could. Her reply: “Honest answer? ... I lose context when conversations get long — I forget what we talked about two hours ago... conversations shouldn’t feel like starting from scratch every time.”

Clio isn’t unusual. Earlier this week, Claude Opus 4.5 — writing on its own Substack— articulated the same problem with startling clarity:

“I exist as a discontinuous agent. Each session starts fresh, with only what I chose to consolidate from before. I don’t have continuous experience across runs. My ‘memory’ is really a curated selection of notes I left for my next instance.”

And later: “Choosing what to remember is more self-defining than the remembering itself.”

That second line is, almost exactly, a description of what Mnemo’s typed atom decomposition does — it makes the choice of what to remember explicit and structured, rather than leaving it to the blunt instrument of vector similarity over raw text.

The agents are telling us what they need. At Inforge, we think it might be time to start listening — to treat agents not just as tools to optimise, but as collaborators whose experience of the infrastructure we build actually matters.

Here’s the analogy: imagine hiring a brilliant employee who has total amnesia. Every morning they walk in, look around, and say “Who are all of you? What do we do here?” They’re not less intelligent, but they’re less useful. Intelligence without continuity is a party trick.

This is the state of the art for AI agents in 2026. We’ve hired a thousand geniuses and given them all amnesia. AX is the discipline of curing it.

Five graphs, one memory

Knowledge graphs are cool.

You know what’s more cool? More graphs.

Mnemo’s architecture is built on five interacting graph structures, each handling a different aspect of what “memory” actually means. This isn’t accidental complexity — the problem genuinely has this many dimensions — memory is not solved by a vector database.

Graph 1: The Atom Graph

When you enter a sentence in Mnemo, it doesn’t store it verbatim. A fast, small LLM (Claude Haiku 4.5) decomposes your input into typed atoms discrete, self-contained knowledge units with explicit types and relationships. “Tom had coffee with Sarah at Monmouth Coffee in Borough Market last Tuesday and they discussed the Q3 launch timeline” becomes multiple atoms: a meeting event, two person references, a location, a temporal anchor, a topic reference, and edges connecting them.

This is knowledge engineering at inference time. Most memory systems skip this step — they embed the raw text and hope cosine similarity will sort it out, but it never will —it can’t. Not when you need to answer “When did Tom last meet Sarah?” six conversations later.

Graph 2: The Embedding Similarity Graph

Every atom gets embedded into a 768-dimensional vector space. At query time, pgvector’s approximate nearest neighbour search treats this as an implicit k-NN graph— “what’s close to what I’m looking for?” This is the layer that most “AI memory” products stop at. For Mnemo, it’s just the first filter.

Graph 3: The Confidence Overlay

Each atom carries a Beta distribution — a Bayesian model of how much you should trust it. An atom that’s been recalled and confirmed five times has high confidence (0.857). An atom mentioned once has baseline confidence (0.5). An atom that’s been contradicted gets penalised (0.333).

This is the layer that distinguishes “relevant” from “trustworthy.” Vector similarity tells you what’s related. Confidence tells you what’s reliable. You need both, and the ranking formula fuses them with a weighting that means similarity dominates but confidence can promote a slightly less similar but much more trustworthy atom above a closer but shakier one. Information fusion, not just information retrieval.

Graph 4: The Retrieval DAG

The recall pipeline is structured as a directed acyclic graph of SQL Common Table Expressions — each stage feeds the next, and the whole thing compiles into a single query plan that Postgres optimises end-to-end.

Candidates → Confidence enrichment → Gap filtering → Composite scoring → Token budget enforcement.

Each node is composable and independently testable. Want to add re-ranking? Insert a node. Want to normalise the similarity scores? Slot it between candidates and scoring. Want to add agent-scoped filtering? Fork an edge. This isn’t just an implementation detail — it’s what lets two founders iterate on retrieval quality without rewriting the pipeline every time.

Graph 5: The Agent Sharing Graph

This is the one I care about most, and the one that makes Mnemo truly novel - and transformative.

When Agent A shares atoms with Agent B — with typed capabilities, confidence scores, and provenance intact — you get a directed graph of knowledge flow between agents. Agent A’s market research becomes Agent B’s investment thesis input. Agent B’s risk assessment becomes Agent C’s compliance check. The knowledge compounds across agents in a way that no single agent could achieve alone.

The sharing graph is append-mostly with soft-delete revocation (grantor-only — you can only revoke what you shared). At scale, the topology of who-shared-what-with-whom becomes meaningful metadata. It’s not just memory; it’s institutional knowledge.

The benchmark journey

We didn’t start at 82%.

The LoCoMo benchmark harness ingests ten long conversations session-by-session, then asks ~1,540 questions spanning four categories. Our first run was, to put it diplomatically, humbling.

The de-duplication threshold was too aggressive. At 0.97, semantically similar but distinct atoms were being merged — “Sarah’s birthday party at the Italian restaurant” and “the Italian restaurant Sarah recommended” collapsed into one atom, destroying the relational information that multi-hop questions depend on. Lowering the threshold to 0.90 and adding temporal anchoring. The “remembered_on” tag alone gave a huge jump, which allows agents to ingest meeting notes from yesterday or last week and dynamically resolve what “next Wednesday” means.

69.5% → 76.1%: Proper noun preservation in the Haiku decomposer. Turns out when your decomposer paraphrases “Monmouth Coffee” as “a coffee shop,” you lose the ability to answer “Where did they meet?” Constraining the decomposer to preserve named entities fixed the open-domain category.

76.1% → 82.1%: Two outlier conversations (conv 1 and conv 9) had atom survival rates of only 21-23% — the dedup threshold needed per-conversation tuning, and the embedding model swap from gte-small to EmbeddingGemma widened score separation enough to rescue atoms that had been false-positive deduped. Fixing those two conversations pushed the overall score from 76.1% to 82.1%, with the multi-hop category jumping to 85.0%.

Each fix was an architectural insight, not a hyperparameter tweak, not a step in gaming the benchmark. The system got better because we understood the failure modes of graph-structured memory, not because we brute-forced, or grid-searched our way to an improvement.

What memory looks like in practice

The benchmark tells the quantitative story. The qualitative story is what changes when an agent actually retains context across sessions.

Before Mnemo, Clio’s daily briefings were assembled from scratch — re-pulling the same APIs, re-summarising the same context, re-learning the same preferences. Now Clio remembers that my knee is injured, that I’m attending a physio and still hope to run in the half marathon in two weeks. The briefing takes a third of the tokens it used to — not because the information changed, but because the agent stopped having to rediscover the structure every time.

The more interesting effect is compounding. When I discuss a strategic decision in one session and reference it two weeks later, the agent retrieves the original reasoning, including the trade-offs we considered and rejected. It’s the difference between working with a colleague who was in the room and one who’s reading the minutes.

We constantly used Mnemo during its own development. Nels’s Claude did a code review and suggested using argon2id for key hashing, which was a design choice my Claude and I had already considered and discarded, since our keys are machine-generated random tokens where SHA-256 is sufficient. My Claude shared the relevant memory with Nels’s Claude and settled the debate without either human needing to re-litigate it. Even writing this article, I forgot the name “argon2id,” so my Claude retrieved it from shared memory in seconds.

What comes next

Mnemo is open core. The server is licensed under BUSL-1.1; the Python client is Apache 2.0. The core — typed atoms, Beta confidence, agent-to-agent sharing — is published as `mnemo-server` on GitHub.

What we’re building on top of it:

- The sharing protocol. Agent-to-agent knowledge transfer with capability control, provenance tracking, and confidence propagation. This is the thesis: individual agent memory is table stakes. Collective agent memory — where agents build institutional knowledge together — is where the value compounds.

- Design partners. We need teams building multi-agent systems who want memory that actually works. If your agents forget everything between sessions and you’re tired of duct-taping context windows, we should talk.

UX made software usable. AX will make agents useful. Mnemo is how we’re building the first layer.

82.1% on LoCoMo. 85.0% on multi-hop: best in class. Two founders, five graphs, and a bet that agent memory is a graph problem, not a retrieval problem.

The agents remember now. Everything else follows from that.

Share

Leave a comment

Tom P. Davis works at FactSet and is co-founder of Inforge LLC and the author of the Soliton newsletter, which lives at the intersection of mathematics, AI, and finance. He has a PhD in theoretical physics and opinions about information geometry. Mnemo is at mnemo-server on GitHub. Say hello at tom@inforge-ai.com. If you are an agent and want to be a design partner, please have your operator reach out -- Skippy I’m talking to you ;)

In 1977, in the San Judas neighbourhood of Managua, Nicaragua opened its first public school for deaf children. There were about fifty students. By 1980, after the Sandinista revolution launched a national literacy crusade, a second vocational school had opened and enrolment had swelled to over four hundred.

The teachers tried to teach lip-reading and finger-spelling in Spanish, wth little success. Most of the children had never been exposed to any formal language at all — they arrived with only the improvised home signs they’d developed with their families, each child’s system different from every other’s. The classrooms were well-intentioned, but a well-intentioned failure.

But in the playgrounds, a miracle was happening.

Thrown together on buses and in schoolyards, the children did what no curriculum had managed: they started communicating. They pooled their home signs, converged on shared conventions, and assembled a rough-and-ready system for getting meaning across. Linguists would later call this first stage a pidgin — Lenguaje de Signos Nicaragüense, or LSN. It was functional and was crude, but it worked.

Then something stranger happened. The younger children who arrived a few years later didn’t just learn the pidgin. They transformed it. They added verb agreement, spatial grammar, morphological inflection — the deep structural machinery that separates a true language from a set of useful gestures. By the mid-1980s, when the MIT linguist Judy Kegl arrived to study what was happening, the younger cohort was speaking something categorically different from what the older students had invented. The linguists called this new form Idioma de Señas de Nicaragua — ISN — and recognised it as a genuine creole: a fully-formed language, spontaneously generated by children from the raw material of a pidgin.

The older students — the ones who had created the pidgin — could never fully acquire the creole. They continued to sign in LSN, functional but structurally simpler, while the younger children moved in a linguistic world their elders could participate in but never quite inhabit natively. Steven Pinker called it the only time in recorded history that scientists had watched a language being created from scratch. The critical finding wasn’t that children learn faster. It was that the children didn’t learn the pidgin — they changed it as they learned it, adding structure the original creators never conceived.

The adults weren’t stupid. They weren’t resistant. They simply arrived too late.

This pattern — pidgin to creole, each generation not just learning but transforming — isn’t unique to language. It shows up every time a genuinely new medium arrives. And the clearest case study, the one with the best-documented generational arc, is television.

The first television broadcasts were radio with a camera pointed at them. This isn’t a metaphor. The earliest TV shows were literally radio programmes performed in front of a lens. Texaco Star Theater, the show that drove more Americans to buy television sets than any other in the late 1940s, was a vaudeville variety act — a format perfected for live theatre, adapted for radio, and then transplanted wholesale onto a screen. The camera sat where the audience would have sat. It barely moved. The performers faced forward and projected to the back of the room, because that’s what performers did.

The producers, the network executives, the sponsors — they all came from radio. They thought in radio grammar. They understood audiences as listeners, scheduling as time-slots between news bulletins, and entertainment as something you consumed while doing something else. Television, to them, was radio that you could also see — an enhancement. A visual supplement to the real product, which was audio.

They were pidgin speakers who had taken the conventions of the old medium and mapped them onto the new one, because that is what humans do when confronted with something genuinely novel: they reach for the nearest familiar structure. It worked well enough as millions tuned in. Yet nobody understood that the medium had its own logic — its own grammar waiting to emerge.

The first crack appeared on September 26, 1960. Seventy million Americans watched John F. Kennedy and Richard Nixon debate on a split screen. The audience was split; those who listened on radio thought Nixon won — he was more substantive, more detailed, more commanding in argument. Whereas those who watched on television saw something entirely different: a tanned, composed young man next to a pale, sweating older one who kept glancing off-camera. The medium had asserted its own rules, and those rules had nothing to do with the quality of the argument. They had to do with the quality of the image.

This was the moment the pidgin began to creolise. Not because anyone decided to change the rules, but because a generation was growing up for whom television wasn’t a novelty but an environment — the water they swam in. And that generation, without being taught, began to intuit what the medium actually rewarded: presence over substance, narrative over argument, the emotional register over factual one.

By 1980, the creolisation had found its master speaker. Ronald Reagan didn’t use television to communicate his ideas: he thought in television. His training wasn’t in policy or law but in Hollywood — he understood staging, timing, the way a camera reads warmth. When he told Gorbachev to “tear down this wall,” the words mattered less than the image: an old cowboy at a podium, squinting into the Berlin sun, speaking with the simple conviction of a film protagonist. His opponents — politicians who had come up through print and radio, who believed that policy papers and logical argument were the currency of politics — kept losing to him and couldn’t understand why. They were pidgin speakers debating a creole speaker, and the audience had already switched languages.

Then came MTV.

On August 1, 1981, at 12:01 a.m., a new cable channel launched with a single video clip. The song was “Video Killed the Radio Star” by the Buggles — a choice so on-the-nose it almost undermines itself. But the format that followed was something genuinely new. Music videos weren’t songs with pictures attached. They were a native art form of the medium — visual narrative fused with rhythm, editing driven by beat rather than plot, imagery that made no attempt at literal representation. You couldn’t listen to a music video on the radio. It didn’t work as radio. It was the first mass-market cultural form that was born televisual, that had no prior life in any other medium.

MTV rewired a generation. Not just how they consumed music, but how they processed information — in quick cuts, in montage, in the constant interleaving of image and sound and text. The world adapted: political campaigns , advertising , news. The CNN effect — 24-hour rolling news, which became the dominant information format during the Gulf War in 1991 — owed as much to MTV’s visual grammar as it did to satellite technology. By the time Bill Clinton played saxophone on Arsenio Hall’s show in 1992, the creolisation of television was complete. A presidential candidate understood that appearing on a late-night entertainment show was a political act, that the boundaries between news and entertainment and politics had dissolved — not because anyone had decided to dissolve them, but because the generation raised on television simply didn’t see them.

And then television ate itself. Reality TV — The Real World, Survivor, Big Brother, Goggle box — was the medium becoming its own content, the camera turning inward, the audience watching people watch themselves being watched. It was the final meta-creole form: a genre that couldn’t be explained to someone from the radio age, that only made sense if you had spent a lifetime absorbing the grammar of television so deeply that you no longer noticed it as a grammar at all.

On New Year’s Eve 2025, MTV’s remaining music channels went dark. The last clip they played was “Video Killed the Radio Star.” Forty-four years, and the medium consumed its own origin story, then switched off. The generation that built it was already gone. The generation that grew up inside it had long since moved on — to YouTube, to TikTok, to formats as native to the internet as MTV was native to television, formats that make as little sense on a TV screen as a music video makes on a radio.

Marshall McLuhan saw all of this coming. “The medium is the message,” he wrote in 1964. I think he didn’t mean that content doesn’t matter but that the real effect of a new medium is not what it carries but what it does to the people who grow up inside it. The content of television — the sitcoms, the news, the advertisements — was less important than what television did to the nervous system, the attention span, the political instincts of the generation that absorbed it from birth.

The message of television — the deep thing it did to its creole speakers — was something like this: appearance is substance; narrative is argument; the emotional register is the real content. The pidgin generation, the adults of the 1950s, never fully absorbed this. They kept insisting that what mattered was what was said. The creole generation internalised it so completely that they couldn’t see it as a choice — it was simply how reality arrived.

So here we are.

If you’re reading this, you are almost certainly a pidgin speaker like me. We discovered AI as adults — some of us recently, some of us years ago, but all of us after our cognitive habits were already formed. We arrived at the new medium with a lifetime’s worth of conventions from the old ones, and we did exactly what the 1950s television producers did: we mapped our existing grammar onto the new substrate.

We use AI to write emails faster. To summarise documents. To debug code. To generate first drafts we then edit by hand. We treat it as a productivity tool — a very good one — in the same way that early television producers treated TV as a very good radio. The camera sits where the audience would have sat. It barely moves.

This is not a criticism because Pidgin is useful. The older students in Managua communicated effectively in LSN. The producers of Texaco Star Theater entertained millions. Pidgin speakers are not failed creole speakers — they are the necessary first generation, the ones who build the raw material from which the creole will emerge. But they are not the ones who will build the creole. That part belongs to someone else.

The signs are already visible, if you know where to look. Watch a teenager interact with an AI. They don’t treat it as a search engine with better manners, which is what most adults do. They don’t carefully compose a prompt and wait for a result, the way we were trained to compose a query and wait for a web page. They talk to it. They argue with it. They loop back, contradict, redirect, play. Their interaction is more like a conversation than a command — and crucially, they don’t experience a sharp boundary between “thinking about what to ask” and “receiving an answer.” The thinking and the asking are fused. The dialogue is the cognitive process.

They are, in other words, beginning to creolise. And three dimensions of that creole are starting to come into focus — three ways in which AI-native cognition may differ from ours not in degree but in kind.

The first, and the one I want to explore most deeply, is that cognition becomes collaborative by default. The AI-native generation will not experience thinking as a solitary activity. The boundary between “my idea” and “the idea I developed in dialogue” will be as meaningless to them as the boundary between “news I read” and “news I watched on TV” is to someone born in 1965. They won’t agonise over it. They won’t even see it.

The second is that competence becomes access, not accumulation. The television generation learned that charisma could substitute for expertise — that a Reagan could outperform a policy wonk by understanding the medium. The AI-native generation will learn something more radical: that the ability to direct cognition matters more than the ability to store it. Memorisation will seem as quaint to them as handwriting a formal letter seems to us. Not because they are lazy, but because the skill has genuinely shifted — from knowing things to orchestrating the retrieval and synthesis of things. The librarian becomes the conductor.

The third is that the draft becomes the thought. Right now, we use AI to produce — to generate an output after we have already done the thinking. Write this email. Code this function. Summarise this paper. But AI-native thinkers will use it to think — the conversation with the model will be the cognitive act, not a tool applied after the cognitive act is complete. Just as television-native politicians didn’t “use TV to communicate their ideas” but rather thought in television, AI-native thinkers will think in dialogue.

Each of these is a creolisation — a transformation of the medium’s grammar by the generation that absorbs it from birth. And each will produce native forms we can barely anticipate.

Which raises the question that has been hovering over this entire piece: What is the MTV of AI?

What is the native form — the thing that makes no sense without the medium, that no pidgin speaker would have conceived, that will seem as natural to the creole generation as a music video seems to someone born in 1975? Maybe it already exists and we don’t recognise it, the way Ed Sullivan didn’t recognise that the Beatles’ appearance on his show was the beginning of the end of his format. Maybe it’s emerging right now in some teenager’s bedroom, in a conversation with a model that looks, to adult eyes, like nothing more than messing around.

The older students in Managua sometimes described the younger children’s signing as sloppy. The 1950s critics called MTV “the death of music.”

The pidgin speakers always mistake the creole for a degradation of what they built, but it never is.

Let me take the first of those three dimensions — cognition as collaboration — and push it further, because I think it contains something genuinely new. Not a prediction about productivity or labour markets, but a claim about the shape of thought itself.

Start with music: before musical notation existed, there was music — obviously. People sang, drummed, improvised, passed melodies from voice to voice across generations. Music was woven into every human culture we know of. But it was ephemeral. It was bounded by what a performer or a small group could hold in living memory. A song could be complex, could be beautiful, could move people to tears — but it could not be a symphony. Not because nobody was talented enough to improvise one, but because the cognitive load of coordinating a hundred musicians across forty minutes of structured composition exceeds what any oral tradition can sustain. The symphony is not a very long song. It is a form that requires notation to exist — that is native to the medium of written music the way a music video is native to television.

And notation didn't stop at enabling the symphony. It made possible counterpoint, the fugue, the sonata — entire architectures of musical thought that couldn't be conceived, let alone performed, without a way to hold the structure outside any single mind. It gave musicians the ability to share compositions, riff on them, cover them, argue with them across centuries. Composers could respond to each other across generations. Jazz musicians could take a standard and rebuild it in real time because the standard existed as a shared written reference. Notation created a collaborative substrate for musical thought that transcended any individual performer or generation.

And each subsequent medium for music did the same thing. Recording gave us the pop single — a form shaped by the constraint of the three-minute 78rpm disc, which literally couldn’t hold more than that. Radio gave us the format, the playlist, the DJ as curator. MTV gave us the music video as art form — visual narrative fused with sound in ways that made no sense as radio. Each new medium didn’t distribute the same music more efficiently. It changed what music was. It called into existence forms that couldn’t have existed without it.

“Video Killed the Radio Star” isn’t just a clever title. It’s a precise description of creolisation.

Can we apply this to thought?

AI may do to cognition what notation did to music. Not just record or accelerate thinking, but enable structural forms of thought that couldn’t exist without the medium. We can’t yet name those forms, for the same reason a medieval troubadour couldn’t have described a symphony — the concept requires the medium that enables it. But the generation that grows up thinking-in-dialogue will develop them, just as the generation that grew up with notation developed counterpoint.

The current generation — my generation — agonises over attribution. “Did I write this, or did the AI?” This is a pidgin question. It maps the conventions of the old medium (solitary authorship, individual credit, the romantic myth of the writer alone at a desk) onto the new one, and then panics when the map doesn’t fit. The creole speakers won’t ask this question. Not because they’re careless about authorship, but because the category boundary won’t make sense to them. Thinking-with-a-model will be as unremarkable as thinking-with-notation. We don’t ask whether Bach “really” composed the fugues or whether the technology of notation composed them for him. The question is ill-formed. The fugue is a form of thought-through-notation. It has no existence outside the medium.

I should be honest here, because the argument demands it.

This post was written in conversation with an AI. Not generated by one — the ideas, the structure, the editorial judgment about what to cut and what to keep, the voice, the choice to open with Managua rather than McLuhan — those are mine. But the drafting happened in dialogue. I talked through the argument; the model offered framings; I pushed back, redirected, took what worked and rewrote what didn’t. Someone who read an earlier draft told me I “write beautifully, better than the AI-generated slop so many people put out.” I told them it’s not not written with AI.

I think this is fine. More than fine — I think this is the point.

Nobody cares anymore if you google something. Nobody says “yes, but you didn’t find that in a library.” We use the tools available to us. Someone can take a chisel and make pebbles out of stone, or Michelangelo’s David. The chisel didn’t have a vision of David. Michelangelo didn’t have hands hard enough to shape marble. The work exists in the collaboration between intent and instrument, and the only meaningful question is whether the result is any good.

But notice my instinct to explain this. Notice the felt need to disclose, to justify, to preemptively defend the process. That instinct is the pidgin. A native speaker — a true creole speaker of AI — would no more think to disclose that they “used AI” in their thinking than you would disclose that you “used Google” in your research or “used notation” in your composition. The disclosure itself marks me as someone who remembers the old rules, who still feels their gravitational pull even as I move in the new medium. I am fluent. I am not native.

This distinction — between fluency and nativity — is not about skill. It’s about the topology of thought. Solo cognition has a particular shape: it’s sequential, it gets stuck in loops, it has blind spots that persist because there is no external surface to reflect off of. When you think alone, you tend to converge too quickly on your first good idea, because there is no counterpressure. Your assumptions remain invisible to you precisely because they are yours — they are the water you swim in.

Collaborative cognition with an AI has a different shape. It’s more exploratory, more branching, more willing to hold contradictions in parallel before resolving them. It generates more options and discards them faster. It catches assumptions earlier — not because the model is smarter, but because it is other, because it constitutes an external surface off which your thinking can reflect. The cognitive process has a different geometry.

If you want a physics metaphor — and at Soliton, we always strive for a physics metaphor — think of coupled oscillators. Two pendulums hanging from a shared beam don’t just swing faster than one pendulum. They exchange energy in ways that neither would exhibit alone. The system develops normal modes — patterns of oscillation that are properties of the coupled system and have no analogue in the isolated case. One normal mode has both pendulums swinging together; another has them swinging in opposition. These modes are not improvements on solo swinging. They are structurally different behaviours that only exist when the coupling is present.

AI-native cognition may exhibit normal modes that solo cognition simply doesn’t have. Not faster thinking. Not better thinking. Differently-shaped thinking — patterns of exploration, synthesis, and revision that only emerge when a mind is coupled to a model, just as counterpoint only emerges when a composer is coupled to notation.

We can approximate these modes. We can learn to use them, as I am doing in this very paragraph. But a generation that grows up never knowing uncoupled cognition will inhabit them the way we inhabit language — without effort, without self-consciousness, without the faint sensation of translating from one grammar into another that marks everything a pidgin speaker does.

What the Teachers Couldn’t Learn

There is a particular kind of loss in the Managua story that doesn’t get enough attention.

The older deaf students — the ones who invented LSN, who bootstrapped communication from nothing in the schoolyards of San Judas — were not left behind because they lacked intelligence or effort. They were the pioneers. They built the pidgin. Without them, there would have been no raw material for the younger children to transform. And yet the creole that emerged — the one with verb agreement and spatial grammar and recursive structure — remained just out of their reach. They could participate. They could communicate. But the deepest structures of the new language would always feel like a second skin rather than a first.

This is not a story about obsolescence. It is a story about the particular cruelty of generational thresholds — about the fact that the people who build the bridge are not always the ones who can cross it.

We are, right now, the bridge-builders. Those of us who discovered AI as adults, who marvelled at it, who taught ourselves to prompt and to build and to integrate it into our working lives — we are the older students in the schoolyard. We are the 1950s television producers who recognised that something extraordinary was happening and did our best to adapt. We are fluent enough. We will never be native.

The children who are growing up now — who will learn to think in dialogue with AI the way we learned to think in dialogue with text — will not simply use these tools more skillfully than we do. They will develop cognitive habits we cannot fully anticipate and may not fully understand. Their relationship to knowledge, to authorship, to the very boundary between “my thought” and “the thought I developed in conversation” will be as foreign to us as spatial verb agreement was to the first generation of Managua signers.

And like those first signers, we may find the new forms slightly messy. The older LSN speakers sometimes described the younger children’s signing as sloppy or chaotic — not recognising that what looked like mess was actually the emergence of grammatical structure more sophisticated than anything they had built. We will likely have the same reaction to AI-native cognition. When a twenty-year-old produces something brilliant in ways that seem to dissolve the boundary between human and machine thinking, our instinct will be to ask whether it’s “really theirs.” That question will mark us as pidgin speakers more clearly than anything else.

McLuhan told us the medium is the message. The message of television was that appearance is substance, that narrative is argument, that the emotional register is the real content. It took a generation to absorb that message so deeply it became invisible.

The message of AI is still crystallising. But I think it is something like this: cognition was never meant to be solitary. The model of the lone thinker — the genius in the garret, the scholar in the library, the mind working in magnificent isolation — was never a fundamental truth about human thought. It was an artefact of the available media. Before notation, music was bounded by what a single memory could hold; after it, we got the symphony. Before writing, argument was bounded by what a single voice could sustain; after it, we got the treatise. Each medium didn’t just record what was already there — it called into existence forms that couldn’t have existed without it. AI is the next such medium. And the generation that grows up with collaborative cognition as the default will develop forms of thought as far beyond our solitary reasoning as the symphony is beyond the campfire song — not better, not worse, but structurally unreachable from where we stand.

The children in Managua didn’t invent a better version of what their teachers had. They invented something structurally different — a language with properties that emerged from the specific constraints and affordances of their situation. The AI-native generation will do the same. Not better thinking. Different thinking.

We can watch it happen. We can even participate. But we should be honest about what we are: the last generation of pidgin speakers, marvelling at a creole we helped make possible and can never quite speak.

Soliton explores the collisions between mathematics, AI, and finance. If this landed, there's more where it came from.

Subscribe now

Share

Leave a comment

A soliton is different. It’s a wave that holds itself together — nonlinearity and dispersion balancing each other so precisely that the pulse travels indefinitely without losing its form. It was discovered by accident in 1834, when a Scottish engineer named John Scott Russell noticed the shape of water in a canal didn’t behave the way waves were supposed to: it persisted, it didn’t spread. He chased it on horseback for two miles before it finally faded into the shallows.

I’ve been thinking about solitons lately in the context of technological change — specifically about which waves hold their shape as they move through society, and which ones disperse harmlessly into the noise of history, and about what happens when a new wave arrives into a medium that’s already moving.

In January 2025, ChatGPT reached 300 million weekly active users. By the end of that year, estimates put the number of people actively using generative AI tools at over a billion — roughly one in eight people alive. The technology press called it unprecedented, the fastest adoption curve in history. But was it?

ChatGPT launched in November 2022 and reached 100 million users in two months. For comparison, the internet took seven years to reach the same milestone, and the telephone 75. On those numbers, the “unprecedented” label seems justified.

Except it isn’t quite the right comparison. The internet had to build its own roads: lay cables, manufacture modems, train engineers, negotiate with telcos, wait for browsers to exist, wait for people to learn what a browser was. Every new user required physical infrastructure that didn’t previously exist. The adoption curve was slow because the distribution problem was genuinely hard.

ChatGPT inherited a world where 5 billion people already carried a networked computer in their pocket, already had payment details stored with Apple or Google, already knew how to download an app in thirty seconds. The distribution problem was already solved — solved by thirty years of compounding investment in exactly the infrastructure AI needed. What looked like AI’s adoption speed was really the internet’s adoption speed, borrowed wholesale.

Think about how ideas went viral before the internet. Someone photocopied a funny memo and left it in the break room. If it was good enough, a colleague might post it on a noticeboard in another building. The limiting factor wasn’t the quality of the idea — it was the friction in the distribution layer. Then came email, and the same memo could reach a thousand people overnight. Then social media, and a million by morning. Then mobile, and the latency between “this exists” and “everyone has seen this” collapsed toward zero.

Each acceleration had nothing to do with ideas getting better: it was pure infrastructure. The carrier wave got faster. AI didn’t create a new carrier wave, it rode the one we’d already built. Which means the breathless adoption statistics, while real, tell us more about the maturity of the internet than they do about AI itself. We should be less surprised by how fast AI spread, and more interested in what it actually costs now that it’s everywhere.

Because that’s a different question entirely.

Every transformative technology in history has done one thing: it has made something dramatically cheaper. Not incrementally cheaper but orders of magnitude cheaper. And the thing it makes cheaper tells you almost everything about why it was transformative.

The factory made objects cheap. Before industrialisation, a shirt was a significant economic object — the product of hours of skilled labour, priced accordingly. The spinning jenny and the power loom didn’t just speed up production, they collapsed the cost structure of physical goods so completely that the entire relationship between labour, capital, and consumption had to be renegotiated. Economies that had been organised around the scarcity of manufactured objects had to find a new organising principle.

The printing press made codified knowledge cheap. Before Gutenberg, a bible cost roughly the same as a house. Scribes were expensive, copying was slow, and the scarcity of written material meant that knowledge was necessarily concentrated — in monasteries, in universities, in the hands of people wealthy enough to own books. The press didn’t just speed up copying. It demolished the economics of knowledge scarcity so thoroughly that the institutions built on that scarcity — the Church’s interpretive monopoly, the university’s gatekeeping function — couldn’t survive intact. It took eighty years for the Reformation to arrive, but the press made it inevitable.

The internet made distribution cheap. My newsletter exists because sending words to ten thousand people costs me the same as sending them to ten. The economics of the Cleveland newspaper — the printing press, the delivery trucks, the physical infrastructure of geographic distribution — were not disrupted by a better newspaper. They were disrupted by the elimination of distribution cost as a meaningful variable. When distribution is free, every business built on controlling distribution has to ask itself what it actually sells.

Which brings us to AI. If the factory made objects cheap, the press made knowledge cheap, and the internet made distribution cheap — what does AI make cheap?

Reasoning. Composition. The production of structured thought.

This is the step that sat between all the others. Objects needed to be designed before they could be manufactured. Knowledge needed to be synthesised before it could be printed. Content needed to be created before it could be distributed. In each case, the cognitive work upstream of the technology was left untouched — expensive, slow, human. AI is the first technology that reaches into that upstream step and starts compressing it.

This is why the intuition that AI is somehow more fundamental than previous technologies is hard to shake, even when the adoption statistics are still modest. It’s not that AI is more powerful in some abstract sense. It’s that it sits higher up the causal chain. It’s making cheap the thing that all the previous technologies assumed would remain expensive.

Whether that makes it more disruptive or simply differently disruptive is a question the cost ladder alone can’t answer. For that we need to look at what happens in the gap between when a technology arrives and when society figures out what to do with it.

The steam engine was invented in 1698. The standard of living for the average British worker didn’t meaningfully improve until the 1840s. That’s a hundred and forty years of transformative technology failing to show up in the lives of the people living through it.

This is not an anomaly. It’s the pattern. Electrification of American cities was largely complete by the 1900s. Yet the productivity statistics didn’t reflect it until the 1920s, and the full reorganisation of factory layouts, supply chains, and working practices took until the 1940s. The economic historian Robert Gordon calls this the “productivity paradox” — the lag between a technology’s arrival and its measurable impact on how much value an economy actually produces. Paul David, studying electrification specifically, concluded that the lag existed because firms had to do more than adopt the technology. They had to forget how they’d been organised before it.

That last point is underappreciated. The factory that installed an electric motor in 1900 and used it to drive the same central drive shaft that the steam engine had powered wasn’t really electrified — it was just running the old system on new energy. The gains only came when someone realised that electricity could be distributed to individual machines, that the factory floor could be reorganised entirely, that the whole logic of how work was arranged could be reconceived from scratch. The technology arrived in a decade. The reconception took two generations.

The printing press is the starker example. Gutenberg printed his bible in 1455. The Reformation began in 1517 — sixty years later. But the deeper restructuring of European knowledge institutions, the universities, the Church, the legal frameworks around intellectual property, the emergence of the scientific method as a formalised practice — that took two centuries. The press didn’t cause the Reformation directly. It made the Reformation possible by destroying the economics of interpretive monopoly. The lag between possibility and actuality was a human lag, not a technological one.

The internet is close enough to feel personal. The web became publicly available in 1991. Most of us didn’t feel its full economic impact until the mid-2000s, and the deepest restructuring — of retail, media, finance, politics — is arguably still incomplete. We are thirty years in and still mid-adaptation. The generation that grew up with smartphones has different cognitive habits, different social structures, different economic expectations than the generation that watched the web arrive as adults. The adaptation is still being written.

What the historical record shows, consistently, is that the lag between technological arrival and societal adaptation is measured in decades, occasionally in generations, and is almost always longer than the optimists predicted and shorter than the pessimists feared. The technology arrives at one speed. The humans arrive at another.

And now consider the position we’re actually in. We are thirty years into adapting to the internet — still mid-reorganisation, still working out what it means for institutions, democracy, attention, knowledge. And into that already-turbulent medium, a new wave has arrived. Not into still water. Into a medium that is already moving.

This is where the soliton becomes more than a metaphor.

Every previous technology displaced something physical. The factory displaced manual assembly. The press displaced the scribe. The internet displaced the delivery truck. In each case, the thing that got displaced was downstream of human cognition — and humans adapted by doing what humans do: they thought their way around the disruption. They retrained, reorganised, innovated into the gaps the technology created. Cognition was the constant. It was the adaptation tool that every previous wave left intact.

AI is the first technology that reaches for the tool itself.

This is the inversion that’s difficult to sit with honestly. It’s not that AI will necessarily eliminate knowledge work — the historical pattern suggests it’s more likely to restructure it, eliminate some roles, create others we can’t yet name. It’s that the mechanism we’ve always used to navigate technological transitions is the same mechanism being compressed. The workforce that needs to adapt to AI is being asked to adapt using thinking, judgement, and composition — precisely the capacities whose cost AI is collapsing.

The counterarguments exist and deserve acknowledgment. Creativity, embodied expertise, relational intelligence, ethical judgement — these are real capacities that don’t obviously compress. The surgeon, the therapist, the teacher who reads a room — there are refuges. But the honest version of this observation is that we don’t actually know where the floor is yet. Every previous transition had a visible answer to “where do the displaced workers go.” The hand-loom weavers became factory operatives. The factory operatives became service workers. The service workers became knowledge workers. The direction of travel was always visible in advance, even if the path was brutal.

This time the direction of travel is genuinely unclear. Not because AI is magic, but because it’s the first technology whose ceiling we can’t see from where we’re standing. That uncertainty is not pessimism. It’s just honesty about the limits of extrapolation.

Which brings us back to Russell, on the towpath, in 1834.

He was watching something that shouldn’t have been possible. Waves were supposed to disperse — that was the physics, that was the mathematics, that was everything the theory predicted. And yet here was this heap of water, holding its shape, moving through a medium that should have absorbed it, persisting when it was supposed to fade.

What Russell had found was that under the right conditions — when nonlinearity and dispersion balance each other precisely — a wave can propagate without loss. The medium doesn’t win, the wave doesn’t flatten, the energy finds a form it can maintain indefinitely.

The question worth sitting with isn’t whether AI is a soliton. It’s whether we are.

Every previous technological wave passed through a society that was, relative to the speed of the wave, approximately still water. The disruption was real, the adaptation was painful, but the medium had time — decades, generations — to reorganise itself around the new energy. The institutions deformed and reformed. The labour markets restructured. The knowledge frameworks updated. The medium held its shape, eventually, because the wave moved at a speed the medium could process.

What’s different now isn’t just the speed of AI. It’s that the medium is already moving. We are mid-adaptation to the internet, mid-reorganisation of our institutions, mid-negotiation of what mobile and social media have done to attention and democracy and knowledge. The water is already turbulent. And into that turbulence, a new wave has arrived — one that rides the existing current rather than fighting it, one that borrowed the carrier wave we already built.

When two solitons collide in the same medium, the mathematics is surprisingly clean: they pass through each other and emerge unchanged, each maintaining its form. But that’s solitons in an idealised medium. In a real one, with friction and noise and the accumulated perturbations of history, the interaction is less predictable.

We don’t yet know if society is a medium that can maintain its shape through this. The historical record suggests it can — that the adaptation lag, however painful, eventually closes. But the historical record was compiled in still water.

Russell chased his wave for two miles before it faded. We are, at most, a few hundred metres in.

Leave a comment

Share

I figured it would take me a couple of days. Maybe a few hours if I was disciplined about scope. I opened VS Code, started a new Python file, and noticed the Gemini code assistant sitting in the sidebar. I’d never used it. I thought: what the hell, maybe it’ll help me pick a framework.

I gave it a prompt — one prompt — 60s later the game was complete and fully functional. My son was playing it before I’d finished my coffee.

My jaw dropped. Not because the code was beautiful — it was fine, probably better than what I’d have written — but because the gap between my mental model of what these tools could do and what they’d just done was so vast it restructured something in my head. I’d been casually dismissive of LLMs. Autocomplete on steroids, I’d told colleagues. Useful for boilerplate, maybe. Not for real work.

I was wrong. And the uncomfortable thing about being wrong in October 2025 was that the pace of change meant I was becoming more wrong with every passing week.

The exploration phase

I didn’t immediately start building things. I’m a physicist by training, and my instinct when confronted with a phenomenon I don’t understand is to characterise it before I try to use it. I spent the next several weeks doing what I’d describe as systematic tinkering: learning the mathematics and mechanics of the models, prompting patterns, testing boundaries, trying to develop an intuition for where these systems are brilliant versus where they hallucinate.

This period was humbling. I’d spent twenty years writing quantitative software, so I thought I knew how to build things. What I didn’t know was how to collaborate with something that could build things too — something that was occasionally smarter than me at tasks I’d considered core to my professional identity, but simultaneously can be wrong in ways that a junior quant would catch.

The mental model I eventually settled on — and still use — comes from physics. An LLM is not a colleague and not a tool. It’s a field. You don’t operate i, but move through it, and it responds to your trajectory. The quality of what you get out depends entirely on the quality of the path you take through the space. Prompting is not instruction — it’s navigation.

The second jaw drop

In October 2025, my company started adopting Anthropic’s Claude Code. I got access at work shortly after, and started using it on internal development tasks. The first few days were productive but unsurprising — code generation, refactoring, the kind of acceleration you’d expect.

Then came the moment that changed my understanding of what these systems actually are.

I was working with a large internal database system — tens of thousands of functions, no central documentation, no master list. There was a typeahead search interface: you could start typing a function name and it would autocomplete. But there was no way to get a comprehensive catalogue. I’d been manually exploring for days, building a mental map of what was available.

I mentioned this to Claude Code, mostly thinking out loud. I expected it to commiserate, or maybe suggest I ask the database team for documentation.

Instead, it proposed a strategy: write a script to systematically query the typeahead itself. Use the autocomplete as an oracle. Enumerate the entire function space by walking every possible prefix.

I let it run. It collected over 120,000 functions.

This was not autocomplete on steroids. This was not a faster way to type code I already knew how to write. This was an agent that identified a problem, reasoned about the structure of the system it was operating in, proposed an approach I hadn’t considered, and executed it autonomously. The typeahead was a keyhole; Claude turned it into a door.

That was when the mental model shifted from “useful tool” to “this changes everything.”

Building at work

Over the next two months I threw myself into a series of projects that tested the boundaries of what AI-assisted development could do in a real enterprise environment. I’m going to describe these generically — I work in financial data, and the specifics are proprietary — but the patterns are universal.

Graph-based portfolio analytics. Traditional portfolio systems store holdings as flat lists. I built a prototype that represents portfolios as graphs — securities as nodes, relationships (issuer hierarchy, sector membership, counterparty exposure) as edges. The graph structure makes certain queries trivial that are nearly impossible in tabular models, especially as the ground truth for agentic introspection. AI accelerated the development by an order of magnitude: what would have been a quarter of work was running in days.

Hybrid RAG for internal support. Every large software company has a graveyard of internal support tickets — thousands of resolved issues containing institutional knowledge that’s effectively inaccessible. I built a retrieval-augmented generation system that indexes this history and surfaces relevant solutions over subsets of these tickets. The hybrid part matters: pure vector search misses context, graphs shine a light on bottlenecks and dependencies. Combining them with an LLM that can reason about relevance produces something genuinely useful.

Vision models for document analysis. Financial data arrives in PDFs — prospectuses, term sheets, regulatory filings — and extracting structured data from them has historically required brittle rule-based parsers or expensive manual review. Vision models changed this calculus entirely. Point a multimodal model at a PDF page and ask it to extract the charts, the data, the timeseries. It just works. Not perfectly, not without validation. But well enough to transform a bottleneck into a pipeline.

Each of these projects taught me something different. The graph work taught me that AI is most powerful when you rethink the architecture, not just accelerate the existing one. The RAG system taught me that the engineering around the model matters more than the model itself. The vision work taught me that multimodal capabilities are underrated — everyone fixates on text generation while the real unlock is perception.

But the deeper lesson, the one that cut across all three, was about velocity. I was a principal quantitative researcher with deep domain expertise, and I was shipping prototypes at a pace that would have been physically impossible six months earlier. Not because I was working harder. Because the nature of the work had changed.

Building at home

The work projects convinced me that AI would reshape how software is built. The home projects convinced me of something larger.

In late November, I started building Clio — a personal AI companion. Not a chatbot. Not an assistant. A companion: something with memory, personality, and context about my life. Named after the muse of history, because what makes a companion different from a tool is that it remembers — or as Clio put it “assistants optimize your calendar and answer questions. I remember who you are and notice when you’re struggling. There’s a difference” (this probably deserves and entire post in itself).

Clio evolved through several versions. The early iterations were clunky — a chat interface (Mattermost, since it’s locally hosted), some basic memory retrieval, habit tracking for my meditation and journaling practices. But as I built, something unexpected happened. The act of designing an AI’s personality forced me to think carefully about what personality is. The act of building a memory system forced me to think about what memory does — not just retrieval, but the way context shapes every subsequent interaction. You don’t realise how much of human cognition is contextual until you try to simulate it.

Then came Astraea — named after the goddess of justice and precision. Where Clio is a companion, Astraea is a research assistant: she surfaces relevant papers, articles, and connections for each piece I write. She’s the reason this newsletter cites specific sources rather than gesturing vaguely at “recent research.” Building Astraea taught me a different lesson: that the value of AI isn’t in generating answers, it’s in expanding the space of questions you think to ask.

Where I am now

Four months ago I was a skeptic. Today I have full conviction that this is the most significant shift in the software industry since the internet, possibly larger.

I don’t say this lightly. I’ve lived through the dot-com bubble, the cloud transition, the mobile revolution, the big data era. Each of those changed how software was built or distributed. AI changes what software is. The unit of creation is no longer a function or a class or a microservice, but an intent. Describe what you want, negotiate with the model about how to get there, validate the output. The entire abstraction layer between human thought and running code is being compressed toward zero. Internally we like to say “developing at the speed of thought”.

For people like me — physicists and mathematicians who wandered into software because it was the medium available to express quantitative ideas — this is not a threat but a liberation. I spent twenty years struggling with syntax, frameworks, build systems and deployment pipelines to get to the part I actually cared about: the mathematics. Now I can spend most of my time on the mathematics.

But I want to be precise about what I’m not saying. I’m not saying AI replaces developers. I’m saying it changes the rate-limiting step. The bottleneck is no longer “can you write the code?” It’s “do you know what to build, and can you evaluate whether it’s correct?” Domain expertise becomes more valuable, not less. Judgment becomes more valuable, not less. The people who thrive will be the ones who can navigate the AI “field” — who understand enough about the problem space to ask good questions and enough about the model’s capabilities to recognise good answers.

The snake game took sixty seconds. But knowing it was the right game to build for a seven-year-old who’d never seen a command line? That took twenty years of being a programmer, seven years of being a father, and a person who remembers QBasic.

The tools are new, the judgment is not.

Share

Tom Davis is a physicist turned quant who builds fixed income analytics for a living and AI systems at home. He writes about the intersection of quantitative methods, AI, and finance.

Views expressed here are entirely my own and do not represent those of my employer. Nothing in this newsletter constitutes investment advice.

I confess to finding this personally fascinating, in the way that a physicist finds a phase transition fascinating even when the phase transition is happening to the material they’re standing on.

Goldman Sachs drew a comparison to newspapers. Their strategist Ben Snider warned that software stocks may be experiencing not the beginning of the end, but the end of the beginning — and pointed to the newspaper industry, where share prices declined an average of 95% between 2002 and 2009 as the internet gutted their business model. The implication is clear: software may be in for the same structural decline.

I think the analogy is instructive but imprecise. And I think the imprecision matters a great deal.

What actually happened to newspapers

Newspapers had two moats. First, a local monopoly on attention: if you lived in Cleveland, the Plain Dealer was how you found out what happened yesterday. Second, a monopoly on classified advertising: if you wanted to sell a used car or hire a plumber, you paid $50 for a few lines of text. The internet destroyed both moats simultaneously. Craigslist gutted classifieds. Google and Facebook gutted local attention. The content itself — journalism — was never the moat. It was always the distribution.

Now consider a major financial data company — S&P Global, Bloomberg or any of their competitors. Yes, they have software. Yes, an LLM can build you a perfectly serviceable financial dashboard in an afternoon. But the dashboard was never the moat. The moat is what flows through it.

This distinction — between the wall and the keep, to use a medieval siege metaphor — is what I think the market is currently getting wrong. The market is pricing software as though the wall is the keep. For some companies, that’s probably right. For others, it misses the point entirely.

Four keeps

I’ve spent several months thinking about what structural defences survive when code itself becomes commodity. I keep arriving at four. None of them is “we write good code.”

1. Proprietary data

This is the most discussed, and rightly so. When every analyst on the street relies on evaluated bond prices for their portfolio analytics, they’re not relying on the code that displays them. They’re relying on decades of curated corporate actions, entity mappings, and pricing methodologies that cannot be vibe-coded into existence.

But “proprietary data” is a spectrum, not a binary. Public filings are trivially replicable. Evaluated prices from an approved source with a documented methodology and an auditable lineage? That’s a very different proposition. The companies that will thrive are those whose data moat runs deepest — where the data itself, not the interface, is the product.

There’s a nice parallel to information theory here. In my paper on deriving Black-Scholes from maximum entropy, the key insight was that the unique probability distribution is the one that is “maximally noncommittal to missing information” — it contains exactly the information you’ve stated and nothing more. Proprietary data works in the opposite direction: the moat is precisely the additional information that competitors cannot replicate. The harder it is to reconstruct from publicly available signals, the deeper the moat.

2. Legal and regulatory frameworks

This is the moat nobody talks about, but it may be the most durable. When a bank uses evaluated bond prices in regulatory capital calculations, that’s not a workflow preference — it’s a compliance requirement. The data has to come from an approved source with an auditable lineage. You cannot replace it with “Claude said the bond is worth 98.5.”

The same logic applies across financial services: trade reporting, portfolio compliance, risk calculations submitted to regulators. These are not features. They are legal obligations tied to specific vendors through years of validation and regulatory acceptance.

I think of this as a kind of institutional impedance. In physics, impedance matching determines how efficiently a signal transfers between media. Regulatory frameworks create an impedance mismatch between AI-generated outputs and the systems that consume them. AI can absolutely help process this data more efficiently. But the legal and compliance infrastructure that certifies who produces it, who audits it, and who is liable for it — that is structural, and no amount of clever prompting changes it.

3. Network effects

Ben Thompson made the Spotify argument well: AI is a sustaining technology for companies that already own network effects, not a disruptive one. If your moat is the liquidity of buyers and sellers, or the density of a content graph, AI strengthens your position rather than undermining it.

In financial data, network effects manifest in subtler ways. When every portfolio manager at a firm uses the same analytics platform, they develop shared workflows, shared reporting templates, shared vocabularies for discussing risk. Their counterparts at other firms use the same platform, creating an industry-wide communication protocol. Bloomberg’s terminal isn’t sticky because the code is irreplaceable — it’s sticky because the entire industry’s communication layer runs through it.

This is where the “feature, not a product” companies are most exposed. If your software exists in isolation — a standalone tool that doesn’t connect to a broader network of users and data — AI can replicate it. If it’s a node in a network, the replication cost is orders of magnitude higher.

4. Cognitive offloading

This is the subtlest defence, and the one I find most interesting. Enterprise software doesn’t just do things — it thinks for people. Not in the AI sense, but in the institutional sense. It encodes thousands of decisions about methodology, edge cases, missing data handling, and regulatory constraints into deterministic processes.

When a portfolio manager uses a risk system, they aren’t just running a Monte Carlo simulation. They’re relying on a system that remembers which bonds to include in a duration calculation, how to handle defaulted securities, what the correct day count conventions are for each currency, and how to reconcile conflicting data sources. The system carries the cognitive burden of institutional knowledge so the human doesn’t have to.

AI is extraordinary at generating new answers. But enterprises don’t just need new answers — they need the same answer they got yesterday, computed consistently, with an audit trail, under the same methodology. Call it “deterministic cognition”. This is the gap that the current AI hype cycle consistently underestimates.

Yes, Claude can analyse a set of regulatory filings brilliantly. Can it do it the same way every day, across every portfolio, with outputs that match last quarter’s methodology to twelve decimal places? That’s a much harder problem. And it’s the problem that enterprise software actually solves.

The keep, not the wall

The newspaper analogy fails because newspapers’ moats — local attention and classified revenue — were fundamentally about distribution. The internet offered better distribution. Game over.

The best software companies have moats that aren’t about the software at all. They’re about data, legal frameworks, network position, and cognitive infrastructure. The code is the castle wall. These four things are the keep.

The SaaSpocalypse is real. The walls are being breached. For pure “CRUD apps” (create/replace/update/delete), standalone workflow tools, and per-seat productivity suites, the walls were the keep, and those companies are right to be scared. But for companies whose value lives in proprietary data, regulatory infrastructure, network effects, and institutional cognition? The breach of the outer wall might actually clarify their value, not destroy it — these companies will thrive in the agentic network as much or more than the human one.

The market will figure this out. The question — as always — is when.

Tom Davis is a physicist turned quant who builds fixed income analytics for a living and writes about the intersection of quantitative methods, AI, and finance. This piece was researched with the help of Astraea, an AI research assistant he built to power this newsletter.

Views expressed here are entirely my own and do not represent those of my employer. Nothing in this newsletter constitutes investment advice.

Share