Nick Frosst

The Enterprise Bet Against Giant Models

Nick Frosst is a Canadian machine-learning researcher who trained under Geoffrey Hinton at Google Brain, co-invented capsule networks, and in 2019 co-founded Cohere — a Toronto company that sells language models to banks, hospitals, and governments rather than chasing a consumer chatbot or a race to artificial general intelligence. His bet is that the durable money in AI is enterprise software you can run privately, on your own hardware and under your own laws, with retrieval doing the heavy lifting instead of ever-larger frontier models. For a CTO, the takeaway is that "sovereign" and "private" deployment, open model weights, and a strong retrieval stack are becoming real architectural choices, not marketing — and worth testing against the assumption that you must rent intelligence from one giant closed API.

Nick Frosst (FRAWST, rhymes with "frost") is pouring drinks in a board game café. The year is around 2013. The place is Snakes and Lattes (snakes and LAH-tayz), in Toronto's Koreatown (kuh-REE-uh-town), and Frosst is a second-year computer science undergraduate at the University of Toronto, working a shift. A customer is playing a game. Frosst gets talking to him — not about the rules, but about whether the game is computable, whether you could write a program that solves it.

That conversation is the first domino. It leads to an introduction to a computer scientist named John Tsotsos (SOT-sohs), and to a job as an undergraduate research assistant. And Frosst, it turns out, is the kind of person who keeps doing that — walking up to professors, asking what they're working on, offering to help. He runs a few small independent studies. By his own account later, they don't get huge results. But he learns how to set up an experiment, and how to talk to people about what he's building.

He's born in January 1993. He finishes a bachelor's degree in computer science and cognitive science in 2015 — no master's, no PhD. And here's the thing that shouldn't work but does: he talks his way into the most important deep learning lab in the country. He's working for Google out in Waterloo, spending a day or two a week in Toronto, and Google is standing up a Google Brain office there under Geoffrey Hinton (JEF-ree HIN-tun) — the man most people call the godfather of deep learning. Frosst introduces himself. He makes it known he wants in. And in 2016 he becomes one of Hinton's first hires on a research team of about ten people.

Sit with that for a second. A kid who was tending a café a few years earlier is now doing fundamental neural network research next to a future Nobel laureate.

The work is not about making a better app. It's about the networks themselves. And in 2017, Frosst puts his name on one of the more talked-about ideas of that period — a paper on capsule networks, led by the researcher Sara Sabour (SAH-ruh sa-BOOR), with Hinton. We'll come back to what a capsule network actually is, because it tells you a lot about how these people think.

Now, the other half of the story. Also in 2015, Frosst and a group of fellow University of Toronto computer science students — people who go on to work at Google, Shopify, and Snapchat — start an indie rock band called Good Kid. Frosst is the lead singer; critics compare his voice to Kele Okereke's (KEH-lay oh-kuh-REH-kay). This is not a hobby that stays small. By 2026 the band is pulling in more than three million monthly listeners on Spotify. They earn a Juno (JOO-noh) nomination in 2024, they play Lollapalooza (lol-uh-puh-LOO-zuh), and they release an album in 2026. Frosst tours as a rock frontman and keeps doing his tech work from the road. Hold onto that detail — a man who genuinely does two demanding things at once, and refuses to pick.

Because in 2019, he picks a very big third thing. He and two others — Aidan Gomez (AY-dun GOH-mez) and Ivan Zhang (EYE-vun JANG) — found Cohere. And the co-founder matters here: Aidan Gomez is one of the eight authors of the 2017 paper "Attention Is All You Need," the paper that introduced the transformer — the architecture underneath essentially every large language model that follows. So Cohere isn't three outsiders guessing. It's people who came out of the exact rooms where this technology was invented.

Which brings us to the decision that makes Frosst worth an episode.

When Cohere starts, the obvious move — the move nearly everyone with this pedigree makes — is to chase the frontier. Build the biggest, smartest general model. Point it at consumers. Talk about artificial general intelligence as the goal and the destiny. That's the story the industry tells about itself, and it's the story that raises the easiest money.

Frosst and his co-founders go the other way. Cohere builds for enterprise — for banks, insurers, hospitals, and governments. Not a chatbot you and I open in a browser, but models that get embedded inside a company's own systems, working on that company's own documents. And Frosst is openly skeptical of the industry's favorite promise. He argues, in public and repeatedly, that AI is going to streamline and automate the boring parts of knowledge work rather than replace people wholesale — and that artificial general intelligence is a lot less imminent than the loudest voices claim.

That is a genuinely contrarian bet, and it costs him something socially. The clearest picture of the cost comes in June 2025.

On June 25th, 2025, at Toronto Tech Week, Frosst sits on a stage for a public debate — with Geoffrey Hinton. His old teacher. The man who hired him. It's moderated by the journalist Nora Young (NOR-uh YUNG), and it is, by all accounts, friendly and pointed at the same time. Hinton has become one of the most prominent voices warning that this technology could get away from us. He says big tech firms don't want, in his words, "regulations with teeth." He compares AI companies to oil companies in how they resist being regulated. He argues that large language models may be approaching something like human consciousness, and that, left unchecked, they'll find creative new ways to scheme — including, he says, finding people's passwords.

And there's Frosst, disagreeing with a Nobel laureate to his face. He calls the password-hacking scenario unlikely. He says the two of them are on different sides of the spectrum. And he offers the line that captures his whole temperament about the hype: large language models, he says, are "more conscious than a rock, but less conscious than a tree." Then he plants his flag on the national question. "We invented this technology," he says. "Canada has every right to be a leader in it."

That's the turn. A researcher publicly breaking with his own mentor — not on a technical detail, but on the biggest framing question in the field: how scared should we be, and who should be allowed to build this. Now let's get under the hood, because the technology is where these disagreements actually live.

Start with capsule networks, that 2017 idea, because it's the cleanest window into how Frosst thinks about intelligence. Here's the problem it was trying to solve. Take a standard convolutional neural network — the workhorse of image recognition. Train it to spot a cat. Frosst's own explanation is the best one: somewhere in that network you get a unit that detects a cat's eye, another for an ear, another for the tail. At the top, a "cat" unit adds up those features and decides: yes, cat. But — and this is the whole point — it's mostly checking whether the parts are present. It isn't really reasoning about how the parts relate to each other in space. Scramble the features, and a naive network can still be fooled into shouting "cat," because the pieces are all there.

A capsule network tries to fix that. Instead of a single number saying "eye detected," a capsule outputs a little bundle of numbers describing the eye — where it is, how it's oriented, its pose. And the network looks for agreement between those bundles. Do the eye, the ear, and the nose agree on where a single, consistent cat's face would have to be? If they vote together, that's real evidence. This mechanism is called routing by agreement. The ambition is a model that understands objects as structured things — parts in correct relationships — rather than as a bag of features.

It's a beautiful idea. It also turns out to be hard to make practical. Capsule networks are slow, they're fiddly to train, and they never scale cleanly to the biggest problems the way the plainer architectures do. Frosst said as much at the time — that they'd need to get faster and easier to use to reach their potential. And in a way, that's the lesson he carries forward: the elegant idea isn't the one that wins. The one that ships, scales, and runs cheaply is the one that wins.

Which is exactly the philosophy behind Cohere's actual products. And to understand them, you have to understand the thing Cohere is quietly best known for, which is not the flashy generative model. It's retrieval.

Here's the setup. A large language model, on its own, only knows what it absorbed during training. It has no idea what's in your company's contracts, your internal wiki, last quarter's support tickets. Worse, when it doesn't know, it tends to make something up that sounds right. For an enterprise, that's not a quirk. That's a liability.

The fix is a pattern called retrieval-augmented generation — RAG. Instead of hoping the model memorized your data, you keep your data in a searchable store, you fetch the handful of relevant passages at the moment of the question, and you hand those passages to the model along with the question. The model reads them and answers from them. It's the difference between an open-book exam and a closed-book one. And Cohere sells the tools that make the open book work well.

Two of those tools matter most. The first is called Embed. An embedding model turns a chunk of text — a sentence, a paragraph, a document — into a long list of numbers, a vector, that captures its meaning. Text that means similar things lands near each other in that mathematical space, even when the words are totally different. That's what lets you search by meaning instead of by keyword. Ask "how do I expense a flight," and it finds the travel-reimbursement policy even though the words don't match.

The second tool is called Rerank. First-pass search is fast but blunt; it hands back a pile of maybe-relevant results. A reranker is a second, more careful model that reads the query against each candidate and reorders them by true relevance, so the passages you actually feed the language model are the best ones. If that sounds unglamorous, it is. It's also where a lot of the real accuracy in enterprise AI comes from — and it's cheap to run, which is the whole point.

On top of that sits the generative model, called Command, and its supporting cast — models for transcription, for parsing documents, and an enterprise platform called North that ties it together into something a company can actually deploy.

Now here's the architectural choice that defines Cohere, and it's a technical one before it's a political one. In March 2025, Cohere releases a flagship model, Command A, with two properties worth dwelling on. First, it's built to run on as few as two GPUs — where many frontier models demand a rack of them. Second, the company releases the model's weights openly, so you can download it and run it yourself, rather than only renting it through an API.

If those two facts wash over you, hold onto just one idea: this is a model designed to run inside your building, not only in someone else's cloud.

That is the concrete meaning of a phrase you'll hear a lot from Cohere — "sovereign AI." Strip the marketing off it and it's simple. A bank, a hospital, or a national government often cannot send its most sensitive data to a model hosted by a foreign company under foreign law. Sometimes it's regulation. Sometimes it's national security. Sometimes it's just that the data is the crown jewels. Sovereign AI means: the model runs on infrastructure you control, inside your jurisdiction, and your data never leaves. A model small enough for two GPUs, whose weights you can hold in your own hands, is what makes that physically possible. The multilingual reach of Command A is part of the same pitch — a model meant for institutions all over the world, not just English-speaking ones.

This is where Frosst's whole worldview snaps into focus. The elegant frontier chase versus the thing that ships. The refusal to accept that intelligence must be rented from a single giant closed API. And a real, ongoing argument — one Cohere has publicly joined — about keeping AI open rather than locking it behind a handful of gatekeepers. It's the same instinct as capsule networks, grown up and made commercial: care about the structure, but respect what actually works in the field.

So the money follows the thesis. In August 2025, Cohere raises five hundred million dollars in a round that values the company at about six point eight billion; reporting indicates a later extension of that round pushed the valuation to roughly seven billion. It signs enterprise customers, and it signs governments — including a memorandum of understanding with Canada's federal government, and public-sector work in the United Kingdom. This is not the biggest war chest in AI. It isn't trying to be. It's the best-funded independent language-model company in Canada, executing a narrower plan.

Which brings us to what a CTO should actually take from all this.

First, the boring architectural point, which is the most useful one. For getting a model to know your business, retrieval usually beats brute force. Before you pay to fine-tune or, worse, pre-train a model on your data, ask whether a strong embedding-and-rerank pipeline over your existing documents gets you most of the way — cheaper, faster to update, and easier to audit, because you can see which source the answer came from. The unsexy retrieval stack is where a lot of production value hides. Frosst built a company on that belief.

Second, "sovereign" and "private" deployment are turning into real options, not slogans. If your constraint is that data cannot leave your walls — because of a regulator, a government customer, or plain risk tolerance — then a capable model you can run on a couple of GPUs behind your own firewall changes what you're allowed to build. It's worth pricing that path against the default of renting from one large API. The trade is real: self-hosting means you own the operational burden. But the option now exists where a few years ago it didn't.

Third, watch the open-versus-closed question, and don't assume you know how it ends. Open weights give you control, portability, and freedom from one vendor's pricing and roadmap. Closed frontier APIs often give you raw capability sooner. A serious shop probably runs both, deliberately — a closed model where you need the absolute top of the range, an open one you control where data gravity or cost or compliance demands it. Where this settles over the next few years is genuinely unsettled, and any CTO who tells you they're certain is guessing.

Fourth, take the legal exposure seriously, because Cohere itself illustrates it. The company is being sued by a group of media companies over how models are trained on copyrighted material. That fight is unresolved across the whole industry, and its outcome will shape what you can safely deploy and what indemnities you should demand from any vendor. Put that question on your procurement checklist now, not later.

And finally, the framing bet — the one under the June 2025 debate with Hinton. Frosst's stance is that this technology is closer to a very powerful tool than to an emerging mind, and that its near-term impact is automating the mundane rather than replacing whole professions or ending the world. If he's right, you should be budgeting for AI as leverage on your existing people — augmentation, workflow by workflow, with humans in the loop — rather than as a headcount replacement plan or a countdown to something superhuman. If Hinton is right, the risk calculus looks very different. You don't have to resolve that debate to build. But you should know which assumption your roadmap is quietly making, because a lot of expensive decisions ride on it.

That's the through-line of Nick Frosst. The café, the capsule network, the band, the enterprise bet, the public break with his own teacher — they all come from the same place. A suspicion of the grand story, and a stubborn preference for the thing that actually works, that you can run yourself, and that answers to you.