Episode Summary
Executive Summary: Aiden Gomez argues AI progress is still accelerating, but the next wave will be driven less by brute-force scaling alone and more by data quality, synthetic data, retrieval, and reasoning/planning methods. He expects a crowded model market to consolidate, with value shifting toward platforms that combine models, tools, and enterprise deployment trust. He also stresses that AI will augment rather than replace humans, especially in enterprise workflows.
Main Topics: From rural Ontario to AI research (Priority: 4/5): Gomez describes growing up on a remote 100-acre maple forest in Ontario with limited internet, which pushed him toward gaming, coding, and eventually computer science. Scaling laws vs. data and methods (Priority: 5/5): He says scaling compute still works, but is inefficient and increasingly costly. Major gains now come from better data, synthetic data, and new methods such as reasoning and search. Model market structure and commoditization (Priority: 5/5): He expects both horizontal and vertical models to coexist, but believes there will be consolidation and severe margin pressure for companies selling only base models. Enterprise AI adoption and trust (Priority: 5/5): Enterprise buyers care most about security, privacy, and IP protection. Cohere emphasizes private deployments, VPC/on-prem options, and RAG to reduce hallucinations and improve usefulness. OpenAI, productization, and the application layer (Priority: 4/5): Gomez admires OpenAI’s conviction on scaling and sees it as increasingly product-led. He thinks value is currently accruing both at the application layer and the chip layer, while model margins are compressed. Agents, workforce augmentation, and future interfaces (Priority: 4/5): He is bullish on agents and voice as interfaces, arguing models will increasingly act as copilots for employees, though humans will remain central for important decisions and sales. Robotics and long-term frontier capabilities (Priority: 3/5): He expects major breakthroughs in robotics within 5-10 years as models become better planners and more robust, enabling cheap general-purpose humanoid systems.
Key Arguments: Gaming helped shape founder resilience because it teaches repetition, recovery from failure, and optimism through retries. Scaling compute still improves models, but it is the 'dumbest' and least efficient path; better data and methods now drive a growing share of gains. There is no market for last year’s model; newer generations obsolete old ones quickly, making commoditization and price pressure intense. The future will include both horizontal general-purpose models and specialized vertical models, with startups often prototyping on large models and then distilling into smaller ones. Enterprise customers will not train on their own data because it is too sensitive; private deployment, synthetic data, and RAG are key to adoption. Model-only businesses face shrinking margins due to open-source and price dumping; companies will need broader product stacks to defend value. Reasoning is hard mainly because the internet lacks visible reasoning traces; models need training data that shows problem-solving steps. Agents are justified because the core promise of AI is autonomous work over long horizons, but the best agent products will likely require deep model-level access. AI will augment humans rather than replace them; consumers and enterprises still want accountable humans for meaningful decisions. Robotics will be a major breakthrough area once models can act as robust planners in messy real-world environments.
Data Points: Cohere valuation: $5.5 billion - Media-reported valuation for Cohere's last round, as discussed by Gomez. Cohere capital raised: About $1 billion - Total amount raised by Cohere to date. GPT-4 parameter estimate: 1.7 trillion parameters - Gomez references reported size of GPT-4 to illustrate scale. Alternative model size: 13 billion parameters - He notes some smaller models are now better than GPT-4 at a much smaller scale. Enterprise POC phase: Last year was '100%' the year of the proof of concept - He says enterprise AI has shifted from experimentation to production urgency. Time horizon for robotics breakthroughs: 5-10 years - His estimate for general-purpose humanoid robotics becoming cheap and robust. Model training efficiency: 10x / 100x cheaper over time - He says previous-generation models become dramatically cheaper to build each year. Pricing reference: $20 per month - He cites ChatGPT-style consumer pricing as an example of application-layer monetization. Platform count: Multiple clouds and multiple chips - Cohere runs on NVIDIA, AMD, Google TPUs, and engages with other chip startups to preserve optionality. Funding timeline: Five years - Cohere has been around for roughly five years, helping it avoid early compute shortages.
Pivotal Quotes: "There's no market for last year's model." — Aiden Gomez: He explains why model generations become obsolete quickly and why newer models dominate the market. "The reality of the matter is, if you throw more compute at the model, if you make the model bigger, it'll get better." — Aiden Gomez: He acknowledges scaling works, but argues it is inefficient compared with data and method improvements. "It's really dangerous when you make yourself a subsidiary of your cloud provider." — Aiden Gomez: He warns that model companies can lose independence and bargaining power if they tie themselves too closely to a cloud vendor.
Implications: AI value is shifting from raw model size toward data, reasoning, enterprise trust, and product breadth. Winners will likely be platforms, not single-model vendors, and enterprises should prepare for agentic copilots, private deployments, and faster workflow automation.