Aidan Gomez on Model Scaling, Data and Enterprise AI

Cohere co-founder and CEO Aidan Gomez argues that larger models can improve with more compute, but that data and methods offer more efficient paths. He expects multiple general and specialized models and consolidation among providers. These are views from a company competing in the model market.

Gomez tells Harry Stebbings that he grew up in rural Ontario with slow dial-up internet. Trying to make the connection faster led him to learn how computers and the web worked. He thinks games can teach persistence and the willingness to try again after failure, a personal interpretation rather than a demonstrated cause of founder success.

The speakers compare gradual game challenges to curriculum learning. Gomez says that in the model training he describes, mixing easy and difficult material has worked better than a staged curriculum. This is his account of practice at the time, not a universal conclusion for every learning system.

Asked whether more compute is the main route to better models, Gomez says scaling is reliable but inefficient. He claims smaller models had reached or surpassed a much larger earlier model, while noting uncertainty about the latter’s reported parameter count. The comparison is his time-specific assessment and does not establish equivalence across every task.

He expects continued scaling to require exponentially more compute for linear gains, with economic limits encouraging smaller and more efficient models. He describes a common workflow: prototype with a capable general model, then distill or adapt a smaller model for a specific task. He expects both general and specialized models to coexist.

For a company without a hyperscaler’s budget, Gomez sees opportunities in data, architectures and methods rather than scaling alone. He credits higher-quality web extraction, data selection and synthetic data for major improvements in open models. The attribution is his assessment, not a measured breakdown of causes.

He points to reinforcement learning and search as possible methodological advances. Rather than expect an immediate answer to every difficult question, he wants models to attempt a solution, recognize failure and try again. He says the internet contains many final answers but fewer examples of the reasoning process, making suitable training data difficult to obtain.

Gomez says enterprise customers generally do not allow Cohere to train on their private data. He describes using synthetic data, human annotation and external partners instead, and argues that focusing on enterprise tasks narrows the space the company must cover. These are Cohere’s stated practices and commercial positioning.

He has heard that generating synthetic data to train smaller models makes up much API use, but is unsure how sustainable that market is. He expects continuing demand for new task-specific data from models or people. The market characterization is not supported with usage figures in the interview.

Gomez says selling only models is difficult while competitors offer low prices or free models. He expects near-term pressure on margins and says Cohere plans to add products to its API offering, while believing the model layer could be attractive over the long run. This is a competitor’s market assessment and company plan, not a forecast of realized margins.

He says chips are a large share of Cohere’s spending and that the company works across NVIDIA, AMD and Google TPUs. Customers, he says, want choices across chips and clouds rather than lock-in. This is both a customer-demand claim and a description of Cohere’s commercial strategy.

Gomez predicts more companies will develop chips as options grow and shortages ease. At the time of the interview he identifies NVIDIA and Google TPUs as usable for large-scale training and expects other platforms to mature. These hardware availability and pricing views are time-specific predictions.

Cohere uses data-center partners rather than building its own, he says, because its calculations favor provider pricing; that could change if economics or access to an attractive chip changed. He recalls that the company obtained compute before demand surged and that the broader AI boom arrived later and more suddenly than he expected.

As an intern and co-author on the Transformer paper, Gomez says he did not foresee the architecture’s subsequent adoption. He did expect language-model scaling to attract attention sooner, and identifies ChatGPT’s accessible interface as the moment the technology reached a broad audience.

He does not think chat should replace every graphical interface: sometimes clicking a button is simpler than typing an instruction. He recalls early language models generating convincing Wikipedia-style text and later finding conversational dialogue particularly striking. He also expects voice interaction to be compelling.

Gomez says improving models becomes harder as models approach the knowledge of people who previously supplied feedback. More specialized experts are then needed, making new training information scarcer and more costly. He also says falling compute costs have helped make larger models possible; neither statement specifies a future cost curve.

He estimates that reproducing an older generation can become dramatically cheaper as data and compute improve, but argues that customers tend to want newer capabilities. His statement that there is no market for last year’s model is a strong commercial opinion, not evidence that older models have no useful applications.

Gomez says general users may struggle to perceive new capabilities in specialist areas such as mathematics or physics, while researchers in those fields could benefit. He considers further investment worthwhile for those specialized applications, although the value and cost of each advance remain uncertain.

He predicts consolidation among model companies and warns about becoming dependent on a cloud provider that supplies capital as well as compute. Asked about investor returns, he predicts Cohere’s backers will make money and praises early investor Radical Ventures. This is a founder’s forecast involving his company and investors, not an achieved return.

Discussing a reported valuation above $5.5 billion, Gomez acknowledges pressure to grow into it and says Cohere is better positioned than some peers because its valuation is lower. He expects margin pressure from free and low-priced models to change and Cohere’s products to evolve. The valuation is a report cited in the interview, and the comparison and projections are his.

He admires OpenAI’s commitment to scaling and recalls Ilya Sutskever’s early conviction. Asked whether OpenAI prioritizes long-term AGI over products, Gomez thinks its consumer product has become central and speculates that AGI work may have taken a back seat. His reading of another company’s priorities is not independently verified.

Stebbings raises the cost of adding AI features without charging more. Gomez describes two approaches: include AI at the existing price to attract and retain users, or charge separately for the feature. He considers either reasonable and expects falling costs to help companies choosing the first; this is a prediction, not a margin guarantee.

Enterprise adoption faces trust and security concerns, Gomez says, including fear that a provider might train on confidential data or expose intellectual property. He says Cohere offers private deployments in a customer’s environment or on its hardware. These are the company’s stated offerings and should not be read as an independent security audit.

Asked about a move back from cloud to on-premises systems, Gomez says the answer varies by sector. He observes some financial firms investing in their own infrastructure while other industries continue moving to cloud. That is a limited observation, not a quantified market trend.

Gomez acknowledges models hallucinate and says error rates had been falling on some benchmarks. He compares this with human error and argues systems can be designed to tolerate mistakes. His claims about rates are time-specific; neither benchmarks nor comparison with people make model outputs automatically reliable.

He explains retrieval-augmented generation as giving a model access to a knowledge base, retrieving relevant material and citing sources. He says it can reduce unsupported answers and provide access to private information such as a user’s documents. Retrieval and citations can aid verification, but his discussion does not establish that they eliminate errors.

Gomez observes a shift from enterprise experiments toward production deployment and thinks employee augmentation is a common use case. He praises the assistant concept but argues that enterprises need tools spanning software from multiple vendors. This aligns with Cohere’s business strategy and is his market assessment.

He is enthusiastic about agents that might work independently over long periods, such as sourcing sales leads. He argues model builders may have an advantage because they can change the underlying model for the task. These are predictions about future capabilities and competitive structure, not demonstrations of reliable six-month autonomy.

Gomez says enterprise software is sticky and difficult to displace, while consumers can switch more readily to substantially better experiences. He praises Salesforce’s leadership and sees the possibility of transformative consumer AI products. These are competitive judgments, not proven market outcomes.

On Google’s research, he says its teams began training language models soon after the Transformer paper and were not behind technologically. He distinguishes research from product vision and an organization’s ability to execute it. He also calls the decision to spend dramatically more on model scaling a risky bet, and credits OpenAI for pursuing it.

He rejects the idea that AI progress has plateaued, predicting further work on planning, reasoning and recovery from failure. He expects new capabilities to reach production, but gives no independent evidence of their timing or reliability. He sees application-layer investment opportunities as well as opportunities for model providers.

Stebbings worries about children spending more time with AI than people. Gomez suggests a safe, patient and knowledgeable system could have educational value, but says it cannot replace human relationships. This is a speculative discussion; the safety and effects of any actual system are not established here.

Gomez expects AI to augment workers rather than cause aggregate mass unemployment, while acknowledging localized displacement and fewer people needed in some customer-support roles. He sees automation of distressing conversations as potentially helpful. These are forecasts about employment and job quality, not established outcomes.

He predicts progress in robotics as more flexible model-based planning improves on narrowly programmed systems. He speculates that inexpensive, robust general-purpose humanoid robots could emerge in five to ten years. He presents a technological expectation, not evidence that such systems were already ready.

In quick fire, Gomez says he has raised his estimate of the importance of data quality after internal Cohere experiments. He says Cohere had raised about $1 billion and describes choosing independence rather than a closer tie to a large provider. These figures and strategic claims are the CEO’s account at the time.

He praises board members Mike Volpi and Jordan Jacobs and says he is closer personally to Geoffrey Hinton but more aligned with Yann LeCun’s optimism about AI. His comments on AI risk are personal views, and his praise concerns people with whom he has close professional relationships.

Gomez contrasts what he sees as technology optimism in the UK with a more regulation-oriented European environment, while noting ties to both. He says in-person work is more productive for Cohere. These are broad personal judgments; the interview supplies no comparative productivity measurement.

Asked what future he wants, Gomez emphasizes productivity, increased supply and abundance. The speakers illustrate how a hypothetical productivity improvement in healthcare could matter, and Gomez worries about stagnating real GDP per capita in Canada. Their example and policy priorities are arguments, not measured results of deploying AI.