Unsupervised Learning• Apr 15, 2025• 51:07Interview

Aidan Gomez: Scaling Limits Emerging, AI Use-cases with PMF & Life After Transformers

From Unsupervised Learning

Aidan Gomez•Co-author, Transformer Paper & CEO, Cohere

Executive Summary

Cohere is strategically focused on the enterprise market, offering a vertically integrated, cloud-agnostic platform (North) to accelerate AI adoption and solve complex integration challenges.
The future of AI model improvement is shifting away from pure scale towards new capabilities like reasoning and learning from experience, with synthetic data becoming the primary source for training.
High-value enterprise AI use cases with clear product-market fit include customer support automation and AI research agents that can condense a month of knowledge work into an hour.
Aidan Gomez predicts the enterprise AI market will first fragment with best-of-breed solutions for different departments before consolidating into unified platforms that can access and orchestrate data across the entire organization.

12 quotes

Concerns Raised

The potential for state-level bad actors to misuse powerful AI capabilities.
Significant privacy and security risks due to the extensive data access required by enterprise AI agents.
The near-term maintenance burden for enterprises using fragmented, best-of-breed AI solutions before market consolidation.

Opportunities Identified

Dramatically increasing knowledge worker productivity with AI research agents.
Accelerating enterprise AI roadmaps by 12-18 months using platforms like Cohere's North.
Developing a new generation of foundation models specialized for scientific domains like biology and chemistry.
Gaining a competitive edge through a vertically integrated strategy that combines model development with application building.

Key Themes

Enterprise AI Strategy and Deployment

The discussion centers on Cohere's enterprise-first strategy, positioning itself as a cloud-agnostic provider that can deploy anywhere, including a customer's private cloud. Gomez argues for a vertically integrated model, where building both the foundation model and the application (like their North platform) is a significant competitive advantage for delivering quality and customization.

This highlights the emerging business models in the AI space, contrasting the vertically integrated approach with companies that simply build applications on third-party models, and outlines the strategic choices enterprises face when selecting an AI partner.

The Evolution of Foundation Models

Gomez asserts that the 'scale is all you need' era is ending due to diminishing returns. The next frontier for model improvement lies in adding fundamental capabilities like reasoning (test-time compute) and, most importantly, the ability for models to learn from experience and user feedback over time, a feature currently missing.

This provides a forward-looking view on AI research and development, suggesting that future breakthroughs will come from architectural innovation and new learning paradigms rather than simply increasing parameter counts and training data.

Pragmatic AI Adoption and Use Cases

The conversation identifies concrete enterprise AI applications that are delivering value today. These include mature areas like customer support, specialized verticals like automating medical note-taking, and a high-demand use case in augmenting knowledge workers with AI research agents that can rapidly synthesize vast amounts of information.

This grounds the often-hypothetical discussion of AI's potential in real-world examples, offering a clear picture of where businesses are successfully deploying generative AI and realizing tangible ROI.

The Data Integration and Privacy Challenge

A core challenge for deploying effective AI agents is providing them with extensive access to a company's diverse internal systems (CRM, ERP, email, etc.). This creates significant hurdles related to privacy, security, and the technical complexity of integrating with each enterprise's unique software stack, necessitating a hybrid product-plus-services approach.

This underscores that the primary bottleneck for enterprise AI is often not the model's capability, but the practical challenges of data access, governance, and integration, which must be solved to unlock automation.

Geopolitics and National AI Priorities

The conversation touches on the global nature of AI development, including Cohere's strategic decision to be based in Canada and its deep partnerships in markets like Japan (with Fujitsu) and Korea (with LG). Gomez also expresses concern about state-level bad actors gaining access to powerful AI, advocating for liberal democracies to maintain a technological advantage.

This theme highlights the intersection of AI technology with national interests and global strategy, showing how foundation model companies are becoming key players in international tech ecosystems and geopolitical considerations.

Get started free

Topics

Enterprise AI Foundation Models Cohere Aidan Gomez AI Agents Model Training Synthetic Data AI Reasoning Vertical Integration Cloud Agnostic AI Strategy Customer Support Automation Knowledge Work AI Risk Transformer Architecture

Processed Apr 3, 2026 yt-dlp + mlx-whisper + Gemini