Cohere's Aidan Gomez: The Model Curve Is Flattening, and $20 Million Now Buys GPT-4

Sarah Guo welcomes Aidan Gomez, co-founder and CEO of Cohere, a company valued at more than $5 billion in 2024 that sells language models and products to businesses. He founded the company in 2019. Before that, as a Google Brain intern, he was a co-author on the 2017 paper *Attention Is All You Need*. The personal question is the obvious one: how do you go from growing up in the woods in Canada to that paper?

A lot of it, he says, was luck. He went to the University of Toronto, where Geoffrey Hinton taught. Hinton had just won the Nobel Prize and is treated as the godfather of deep learning; at U of T he was a legend, and almost everyone in computer science wanted into AI. Gomez felt raised into the field. The moment he left high school he was steeped among people who saw the future and wanted to build it.

Then came the happy accidents. He landed an internship with Łukasz Kaiser at Google Brain. At the goodbye party, Kaiser asked how many years he had left in his PhD. Gomez said he was going back to third-year undergrad. Kaiser was stunned: they did not do undergrad internships. The job had been meant for PhD students. A stack of lucky mistakes put an undergraduate on the team that built the Transformer.

After Mountain View with Kaiser, Noam Shazeer, and the rest of the Transformer group, Gomez went back to Toronto to work with Hinton and his future co-founder Nick Frosst. He started a PhD, then worked in Berlin with Jakob Uszkoreit, another Transformer author, while collaborating remotely with Jeff Dean and Sanjay Ghemawat on Pathways, the program to wire supercomputers into a larger training unit.

GPT-2 had just come out. The trajectory was clear. Models that were ostensibly models of the internet, of the web, were going to yield something interesting. He called Nick. He called Ivan Zhang. Maybe, he said, they should figure out how to build these things, because they were going to be useful. That is the origin of Cohere.

The mission is to help other organizations adopt the technology, make workforces more productive, or transform products and services. They will not build a ChatGPT competitor. Success, Gomez says, is all of the above. Models are the foundation and the heart of the company; if they do not meet the customer's needs there is no hope. In the enterprise, support, reliability, and security matter just as much, so Cohere invests as a modeling *and* go-to-market organization, with product rising as a priority to shorten time to value.

For eighteen months, since the enterprise world woke up, Gomez has watched customers make the same mistakes, sometimes in a way that is simply frustrating. Language models are sensitive to prompts and to how data is presented. Each has quirks; talk that works for one may fail for another. In a RAG system the layout of retrieved results, the way records sit in the database, and the formatting all count. People miss those details because they overestimate the models. They think the models are like humans. Humans do not care about format. Models do.

In 2023 a wave of proofs of concept went up as teams tried to get familiar with the technology. Many of them failed on those idiosyncratic errors. Cohere's two answers are to make models more robust to how data is presented, and to put more structure around the product: defined APIs instead of handing someone a model and saying prompt it, good luck. That, he thinks, cuts the chance of failure.

The use cases span every vertical and sound ordinary. A manufacturer puts a Q&A bot on the line, loaded with tool manuals and diagnostic docs, so a worker does not open a thousand-page book. The white-collar version plugs in IT FAQs and HR documents and offers one chat surface onto company knowledge. Healthcare is sharper. A longitudinal record can span decades of pharmacies, labs, and visits. The patient calls about a knee. The doctor has fifteen minutes and skims the last few notes. The model can take the whole history plus the reason for the visit and write a briefing, dramatically faster, often catching something from two years ago. It can do that in under a second. Mundane functions, Gomez says, with immense impact.

Guo asks where the equilibrium sits between specialist applications and in-house builds on model APIs. Gomez sees a hybrid, drawn as a pyramid. The base is what every organization needs: a generalist copilot in every employee's hands. Higher up, the work is specific to the company, domain, or product, and off-the-shelf tools disappear. His advice is to cover the full pyramid, buy the generic and industry tools, and if you are going to build, do not build the things you could buy. Build the things no one will sell you, the ones that are uniquely an advantage.

An insurer of large industrial projects, mines and the like, lives on RFPs. Actuaries research land and risk, and the first credible proposal usually wins. Cohere plugged the sources those actuaries already used into RAG, wrapped a chatbot around them, and sped up response time enough that the firm won more business. An LLM, he says, is like a CPU: the vendor cannot know every application. The insight that puts you ahead comes from listening to the customer. He would wager that insurer is winning because most competitors have not yet built the same assistant.

The largest barrier to adoption is trust, especially in regulated finance and healthcare, where data often is not in a cloud, or cannot leave a VPC. Cohere's differentiator is that it is not locked to one cloud: on-prem, in VPC, outside VPC, whatever the customer wants, which is how it reaches the most sensitive data. The second barrier is knowledge. The people with the most experience have a few years of it. Gomez thinks it is a time game: another two or three years before the technology really permeates how enterprise developers work.

Guo asks whether enterprise AI is in the trough of disillusionment. Gomez is sure some of that feeling exists. He does not think they are in the trough. The core technology is still improving at a steady clip, and new applications unlock every few months. It feels very early. The market has unlocked an entire set of things you fundamentally could not build before. A resurfacing of products and systems is underway.

His thought experiment is stark. Even if not a single new language model were trained, even if the data centers blew up, there would still be half a decade of work to integrate what exists: the insurance RFP bot, the healthcare-record summarizer, the rest of the resurfacing. The hype question, is this actually useful, is past. The technology is in the hands of hundreds of millions of people and in production. The project now is putting it to work.

On how to specialize a model, he offers a gradient, cheapest first. Fine-tuning will change tone or format. If that is not enough, walk back into post-training, SFT and RLHF. Touch pre-training only if you must, and preferably only the last 10 or 20 percent, not a full run. The Fujitsu Japanese model is the counterexample that has to start from scratch: you cannot fine-tune Japanese into a model effectively. A large enterprise sitting on hundreds of billions of tokens can treat continuation pre-training as a real lever, a $5 million run rather than $50 million. For SMBs and startups it makes no sense. Lab leaders who say nobody outside an AGI lab should pre-train are, he says, empirically wrong.

Gomez has said there is no market for last year's models. That sounds like a command to lead. His actual strategy is the opposite. Compute to train is getting cheaper. Synthetic data is dramatically cheaper. Expert data is getting harder and more expensive. Today you can build a model as good as GPT-4, on the things enterprises care about, for $10 million or $20 million, orders of magnitude less than was spent to develop the original. If you are willing to wait six months or a year, you can build it at a fraction of what frontier labs paid, then focus on the capabilities customers actually depend on.

That does not mean the bill is small. Relative to a normal startup, Cohere still pays for a supercomputer, hundreds of millions of dollars a year. The business is capital hungry, he says, but not capital inefficient, with a clear path to a profitable company. The posture is: do not lead, do not burn three, five, seven billion a year to be at the front, sit six months behind, and offer enterprises something that fits at a price that makes sense.

Why train at all, with Llama on the table? Because what you get is the cooled-down artifact, zero gradient, base or post-trained. Fine-tuning that object gives you fewer levers than owning the data and the whole run. Vertical integration, Gomez argues, is dramatically more leverage for the customer. He also rejects the commoditization story. What the market is seeing is price dumping: free, at a loss, at zero margin. Falling prices are being misread as a commodity. The real picture is a ten-to-fifteen-year refactor, like having to repave every road on the planet when only four or five companies know how to make concrete. Giving that concrete away is an unstable present. Growth and return pressures will push prices back. He notes that Haiku's price rose two weeks earlier.

On scaling laws, Gomez says they are pretty far along, entering a flatter part of the curve. We are past the point where interacting with a model tells you how smart it is. Vibe checks are losing utility. You need experts in physics, math, chemistry, biology to assess quality, because the average person cannot tell generations apart. There is still more to do, but the gains will show up in specialized, researchy domains. For the enterprise tasks people want to automate, the technology is already good enough, or close enough that a little customization will get them there.

The new unlock is reasoning. Earlier models had no inner monologue: you passed a question and expected an immediate answer, with no chance to fail, catch the mistake, and try again. OpenAI was first to put reasoning models into production; Cohere had been working on them for about a year. Multi-step problems, a polynomial you have to break down rather than memorize, are the fit. Chain of thought was a retrofit, a hack to elicit the mimic of reasoning after models were trained on input-output pairs. The internet is a pile of finished articles with weeks of thought deleted. The next generation will have reasoning burnt in from scratch, by collecting verbalized inner thoughts, human and synthetic, and training on that process. It is inefficient and brittle now. In two or three years he expects it to be robust.

The business change is the second lever. A customer who wanted a smarter model used to hear: give us six to twelve months for a bigger training run. Now you can charge more, spend twice the tokens or twice the inference time, and hand over a smarter model today. Builders need not double the supercomputer. Intelligence moves from capex with a delay to consumption. Guo says that shift is undervalued. Gomez says it has not been priced in, including at the chip and data-center layer, because inference-time compute does not need a densely interconnected supercomputer. You can do more locally, with less distribution.

Why the flattening? He reaches for oil painting. First you cover the canvas, then the shapes of mountains and trees; the fine brushstrokes take many more marks. Common tasks filled in quickly. Science and math resist. Code and math can still be force-fed with verifiable synthetic data. Biology and chemistry require distilling decades of expert knowledge, and eventually you run out of experts, at the frontier of what humans know. There is no shortcut. One day models will run their own experiments; that is years away. AGI, for him, is continuous, already well along, and not a Terminator takeoff. "We'll be the one to create abundance. We don't need to wait for this God." If AGI means useful, generally capable technology, they will build it. If it means building God, the answer is no. Neural nets are universal approximators only if you build one the size of the universe. Some things are genuinely random. Sequence models can do physics, and they can also waste compute flattening a graph into a sequence. The work he cares about at Cohere is getting what already works into production.