Nick Frosst on Cohere’s Enterprise AI Strategy and the Limits of Hype

Cohere co-founder Nick Frosst joins Harry Stebbings to discuss enterprise AI, scaling, model evaluations, jobs and sovereignty. Frosst’s account reflects both his research experience and his commercial interest in Cohere, which sells enterprise AI products.

Frosst recalls working with Geoffrey Hinton at Google Brain and being surprised by his playful, intuitive approach to research. They often reasoned about algorithms through physical analogies rather than beginning with equations alone, asking what might happen under different conditions.

Asked whether Google missed the consumer AI moment, he notes that the transformer was published at Google in 2017 but was not rapidly commercialized or scaled there. He does not claim to know why, and stresses that many excellent researchers remain at Google DeepMind.

Frosst describes Cohere as a foundation-model company focused on enterprise use. It trains models for tasks such as using business tools, accessing permitted internal data and assisting employees. That differentiation is his description of his own company, not an independent comparison of model performance.

Rather than optimizing a model to keep a consumer engaged in conversation, he says Cohere trains for workplace utility. One approach uses synthetic environments: fictitious companies, messages and APIs allow a model to practice helping with business tasks.

Synthetic data does not make real data irrelevant, Frosst says. High-quality examples are still needed to begin and guide the process, and Cohere also uses human annotators. In his view, quality data remains a meaningful constraint on making models useful.

He describes a sequence of training stages: base language modeling, supervised examples and reinforcement learning using feedback. He believes training data is a more important practical bottleneck for usefulness than the underlying algorithm, while noting that training methods have changed.

Stebbings asks whether more compute will continue to produce progress and says he personally found GPT-5 less convenient than GPT-4 because of model selection and latency. That is his product experience, not a controlled comparison of model capability.

Frosst distinguishes model scaling from the work needed to deliver useful products. He imagines an assistant filing expenses by finding receipts, reading a policy, checking an API and obtaining approval. Better models, connectors and secure data access are all needed; the complete workflow is a goal rather than a verified existing deployment.

He says he does not think current language-model technology reaches AGI as he defines it: a computer people treat and expect to behave like a person. This is his definition and contested scientific judgment, not a settled finding about the future of AI.

Frosst sees foundation models and applications as closely related. A broadly capable language model still needs refinement for the interface and use cases it serves; his example is Cohere’s work on enterprise tools and documentation.

He contrasts older task-specific neural networks with language models trained broadly before being refined. A summarization system benefits from general language training, he says, while a coding or enterprise model can then be adapted to a wider family of relevant tasks.

Stebbings asks whether Cohere is weaker on public evaluations. Frosst says the company prioritizes whether customers can deploy a model and use it successfully, rather than leaderboard position. His critique should be understood alongside Cohere’s interest in enterprise-specific evaluation.

Frosst recalls benchmarks changing over time, from language completion to tests of reasoning and other capabilities. He argues that some widely discussed tests do not resemble the tasks his customers perform. He also says benchmarks can be gamed, while acknowledging he does not know whether particular competitors do so.

Model releases can appear rapid even though training takes months and the hardware cycle is slower. Frosst says the core transformer and sequence-model approach has persisted while training procedures have evolved; this is his broad characterization of the technology.

On competition for AI talent, Frosst says he cannot verify headline compensation figures. Skilled researchers can create substantial value, he argues, but people also care about stability, purpose and alignment, not compensation alone.

Would Cohere pay $5 million for a researcher? Frosst says yes if the person creates sufficient value, and notes that equity can make employees’ stakes valuable. He does not disclose a particular person’s compensation or verify another company’s offers.

He calls AI transformative while criticizing what he sees as misleading hype. In particular, he argues that claims of imminent existential catastrophe diverted attention from concrete questions about employment and inequality. This is his disputed assessment of AI risk, not a consensus conclusion.

Stebbings says employees may be nervous about AI and predicts that some junior marketing and sales roles could be replaced quickly. Frosst says his enterprise customer experience has been more positive and believes the technology primarily augments workers. They offer conflicting predictions rather than measured employment outcomes.

Frosst argues that language models are excellent at some tasks but that human judgment, relationships and cultural context remain important. He also asserts that sequence models cannot independently produce certain breakthroughs; that is a substantive, contested research claim, not a proven limitation.

When Stebbings describes generating marketing ideas with an AI system, Frosst sees a useful starting point rather than the end of the work. He says people still choose, revise and contextualize the suggestions. The practical balance between automation and human work remains open to debate.

Frosst predicts that in five to ten years many workers will use language to ask computers to perform routine tasks, then spend more time evaluating results and interacting with people. He acknowledges the transition may be disruptive and says societies should consider workforce resilience and inequality.

He argues that whether AI increases or reduces inequality depends in part on labor policy. He looks to the industrial revolution as an example of technological change accompanied by labor protections, while recognizing that inequality was a concern before modern language models.

On open and closed models, Frosst says Cohere releases model weights for noncommercial use and negotiates commercial arrangements with business users. He calls this a middle ground that supports research access while sustaining a for-profit company; license terms depend on the specific model.

He says founders should know enough about competitors without becoming absorbed by every headline. For Cohere, he argues, the more important questions are whether the company solves real customer problems and delivers value.

Frosst expects language prompts to remain useful, but thinks specialized prompt-writing tricks will become less important as models are trained to follow ordinary instructions. People will still need to understand what the tools can and cannot do; this is a prediction about the interface.

Reflecting on fundraising, he says investor conversations have become more specific as the industry matured. Instead of explaining what language models are, Cohere can discuss its work with customers such as RBC, Fujitsu and LG; these are company-reported relationships.

Training requires people, compute and data. Frosst says Cohere’s Command A family was designed to fit on two GPUs, which he sees as useful for customers with limited hardware. His efficiency comparisons and customer benefits are company claims, not independent benchmarks.

He recalls early research into training across fragmented GPU capacity when the company had limited funds. It was possible but slow, he says, whereas renting a large data center is easier. The story illustrates a trade-off between resource availability and training time.

Asked how Cohere competes with companies raising far more money, Frosst points to its narrow enterprise focus. His shorthand is “ROI, not AGI”: the company wants to help customers deploy systems and receive value rather than market a consumer subscription.

He says enterprise deployment differs from a large consumer service because customers need privacy, customization, cost control and integration. He describes Cohere’s North product as privately deployable and tailored to knowledge work; these are the founder’s commercial claims.

Pricing depends on the customer and the work, Frosst says. Cohere uses forward-deployed engineers to help set up and customize systems. He rejects the idea that such support necessarily means poor technology, arguing enterprise software often has to fit a customer’s existing environment.

Stebbings asks whether slower enterprise growth results in a valuation discount. Frosst says perhaps and acknowledges the reported $6.8 billion valuation is hard for an individual to grasp. This is the round valuation discussed on the show, not a current market price.

Frosst confirms that Cohere has received acquisition approaches but says the founders want to build a company that lasts beyond their own involvement. He sees the ambition to make something larger than oneself as rewarding, even while recognizing no institution lasts forever.

Stebbings and Frosst disagree about the value of different jobs. Frosst recalls physically demanding work as a cook and argues that low pay does not mean the work lacked value. Stebbings emphasizes market compensation; neither presents the disagreement as an empirical measurement of a person’s worth.

Frosst says he and co-founders Aidan Gomez and Ivan Zhang disagree over specific product and operating decisions, including past API design, but share a strong working relationship. He describes wanting to create something that endures across generations.

Asked whether Cohere should tell its story more publicly, Frosst distinguishes enterprise selling from consumer marketing. He enjoys explaining the company but says product quality and solving customer problems are more important to its strategy.

Frosst compares locally controlled AI models to infrastructure and argues countries may benefit from systems that understand their languages and cultural context. He is unsure what the final form will be, and his Canadian company has a commercial interest in demand for alternatives to US providers.

He says Cohere’s Canadian location can help in conversations with customers concerned about access and geopolitics. His view that political developments affect those purchasing choices is based on his experience and interpretation, not a quantified survey.

Looking at future devices, Frosst expects language to play a larger role but says graphical interfaces will remain useful for some tasks. He has not yet seen a dedicated AI device get the interaction right; that is an assessment rather than a prediction that phones will disappear.

He recalls initial enthusiasm for Google Glass and virtual reality, followed by a realization that he did not want technology to distance him from the people around him. He wants tools that improve connection to the world, and says playing music is valuable to him partly because it creates immediate shared experiences.

Frosst worries about loneliness and weakened community but cautions that people in many eras have believed their own generation was uniquely troubled. He argues technology makers should examine whether their products connect or isolate people, while retaining historical perspective.

In a quick-fire exchange, Frosst criticizes Sam Altman’s public rhetoric about AGI and existential risk, calling it misleading and unhelpful. These are Frosst’s contested characterizations of a competitor’s statements; the interview does not establish the precise predictions he attributes to Altman.

Stebbings asks whether provocative AI claims are connected to fundraising needs. Frosst says he does not know the strategy behind another company’s statements and notes that Cohere also raises venture capital without making the same claims. He remains convinced the technology can be transformative.

Frosst worries that regulation based on a poorly chosen benchmark could misjudge what models can do or be trained to perform differently on that test. His example of a regulator shutting down development based on one benchmark is a hypothetical policy concern.

He says Chinese companies make useful models but is unsure how future competition will develop, and does not describe himself as worried by it. His view of which models lead is a snapshot from the interview, not a definitive ranking.

For 2026, Frosst predicts that a user will be able to ask a workplace assistant to file expenses and have it find relevant information and follow policy. He emphasizes that making the workflow reliable and widespread remains difficult; this is a forecast, not a confirmed capability.

He uses Cohere’s North and the Cursor coding tool, and says the company does not mandate a single coding assistant. If prices rose dramatically, a business would reassess its options; he acknowledges that simply substituting Cohere’s model in an editor would not currently reproduce Cursor’s product experience.

Frosst describes curiosity and contrarianism as both strengths and potential liabilities. He recalls believing in language models early, but says he was wrong to assume social progress would always be monotonic and initially underestimated the data efficiency of reinforcement learning from human feedback.