Episode Summary
Executive Summary: Aiden Gomez frames Cohere as an enterprise-first AI company built to help organizations adopt language models safely and productively. He argues the biggest hurdles are trust, data sensitivity, and implementation know-how—not raw model access—and that enterprise AI is moving from hype to practical production use. He also says inference-time compute, reasoning models, and targeted customization are reshaping the economics and product design of AI.
Main Topics: Cohere’s origin and Aiden’s path into AI (Priority: 4/5): Gomez describes growing up in Canada, entering the University of Toronto AI ecosystem shaped by Geoff Hinton, and landing a Google Brain internship that led to co-authoring Attention Is All You Need. He portrays his trajectory as a mix of luck, timing, and immersion in a strong AI culture. Cohere’s enterprise-focused mission (Priority: 5/5): Cohere is positioned not as a consumer chatbot company but as an enterprise platform for deploying AI across organizations. The goal is to improve worker productivity and transform products/services for businesses. Why enterprise AI fails and how product can fix it (Priority: 5/5): Gomez emphasizes that many enterprise RAG and chatbot projects fail because people overestimate model robustness and miss details like prompt formatting, data presentation, and database structure. Cohere is investing in stronger models plus more structured APIs, reliability, and support to reduce deployment failures. Use cases and customer value (Priority: 5/5): He highlights broad enterprise use cases including document QA, internal knowledge assistants, healthcare record summarization, and specialized research tools such as an insurance RFP assistant. The common value is faster, more accurate decision-making and higher throughput. Customization stack: prompting, fine-tuning, post-training, pre-training (Priority: 4/5): Gomez outlines a gradient of specialization from cheap/easy fine-tuning to more expensive pre-training. He argues large enterprises with enough proprietary data can benefit from continuation pre-training, while startups and SMBs generally should not. Reasoning models and inference-time compute (Priority: 5/5): He argues the industry is entering a new phase where intelligence can be improved at inference time by spending more compute during use, rather than only through bigger training runs. He sees reasoning as unlocking multi-step problem-solving and changing the capex-to-consumption model of AI products. Scaling limits, market misconceptions, and AGI (Priority: 4/5): Gomez says scaling is flattening because models are approaching the limits of human knowledge and high-quality expert data. He rejects the idea of an imminent runaway superintelligence while still believing highly capable general-purpose AI will keep improving over time.
Key Arguments: Enterprise AI adoption is constrained more by trust, security, and implementation expertise than by lack of model capability. Many RAG systems fail because users mishandle data formatting, retrieval presentation, and prompt sensitivity. Off-the-shelf tools cover the bottom of the enterprise AI pyramid, but high-value differentiating use cases must often be built in-house. Large enterprises with massive proprietary datasets can justify continuation pre-training; small firms usually should not. Vertical integration in model training gives Cohere more leverage and better control than relying only on open-source base models. Inference-time compute is a major underpriced lever that lets customers buy more intelligence without waiting for a new training cycle. Reasoning models expand what AI can solve by enabling multi-step problem solving rather than simple input-output memorization. Scaling is slowing because the frontier is moving from internet-scale data to scarce expert knowledge and harder evaluation domains. AI is not commoditized in the true sense; price declines are often dumping, not permanent commoditization. Cohere believes the next decade will be dominated by enterprise integration and productivity gains, even if frontier model gains slow.
Data Points: Cohere valuation: more than $5 billion - The company is described as valued at more than $5 billion in 2024. Founded: 2019 - Aiden Gomez says he founded Cohere in 2019. Year of Attention Is All You Need: 2017 - Gomez notes he was a co-author on the landmark paper while interning at Google Brain. Enterprise adoption timeline: 2-3 years - He says it will likely take another two or three years for the technology to really permeate enterprise development. Production runway without new model training: half decade - He claims there is roughly half a decade of integration work even if no new language models were trained. Training cost for competitive model: $10M-$20M - He says a model as good as GPT-4 for enterprise purposes can be built for tens of millions, far below frontier-lab costs. Continuation pre-training budget: about $5M - He cites $5 million training runs as a plausible enterprise continuation pre-training effort. Frontier-lab spend: $3B-$7B per year - He contrasts Cohere’s strategy with frontier labs that may spend billions annually to stay at the front. Enterprise model behavior: 100 million+ users / hundreds of millions - He says the technology is already in the hands of hundreds of millions of people and in production. Market repricing example: Haiku price 4x cheaper - He references a recent price drop of Haiku as an example of market price pressure.
Pivotal Quotes: "We are not going to build a chat GPT competitor." — Aiden Gomez: He explains Cohere’s enterprise-first mission and positioning. "There's a half decade of just resurfacing to go do." — Aiden Gomez: He argues that even without better frontier models, there is substantial work to integrate AI into the economy. "It hasn't been clocked. People haven't really priced in the impact of inference time compute delivering intelligence." — Aiden Gomez: He describes inference-time compute as an underappreciated shift in AI economics and capability.
Implications: Enterprise AI winners will likely be those that combine strong models, secure deployment, and productized workflows. Expect more reasoning, more customization, and more spending at inference time, while the biggest opportunity remains practical adoption rather than sci-fi AGI.