Inside the Paper That Changed AI Forever - Cohere CEO Aidan Gomez on 2025 Agents

The Genesis of the Transformers Paper

• 💡 Aidan Gomez co-authored the seminal "Attention is All You Need" Transformers paper after a cold email led to an internship at Google Brain.
• 📌 His placement was due to an administrative mistake, as he was an undergrad thought to be a PhD student.
• 🚀 The paper's development was a mad dash for the NeurIPS conference deadline, driven by academic freedom and organic team formation.

The Enduring Impact of Transformers

• 🔑 Transformers remain dominant in AI, largely unchanged in 8 years, due to extensive community infrastructure and specialized chip optimization.
• 🔬 New architectures like State Space Models (SSMs) and discrete diffusion models have emerged, but Transformers often absorb their best ideas rather than being replaced.
• 🎯 The bar for a new architecture to supersede Transformers is extraordinarily high, requiring compelling advantages to justify rewriting existing systems.

The Power of Reasoning in AI

• 🧠 Reasoning and test-time compute have been a known concept for years, addressing the need for models to spend varied energy on problems of different complexity.
• ✨ The effectiveness of reasoning models was surprisingly high, enabling significant intelligence uplift with relatively little effort and cost compared to pre-training.
• 📈 There is vast untapped potential for reasoning beyond math and computer science, particularly in pure sciences, medicine, and enterprise automation.

Cohere's Enterprise AI Vision

• 🏢 Aidan Gomez co-founded Cohere to focus on enterprise AI applications, driven by a desire to increase human productivity and solve real-world economic problems.
• ✅ Cohere's platform includes Command generative models and Search models (like embed v4 and rerank 3.5) as its backbone.
• 🛡️ The North AI agent platform allows enterprises to build agents that interact with internal software and data, with a strong emphasis on on-premise and VPC deployment for security and privacy.

The Role of Synthetic and Multilingual Data

• 📊 Synthetic data is now the majority of data used for training Cohere's models, proving more effective than human data for stylistic preferences and specific enterprise needs.
• 🌍 Cohere addresses underserved multilingual markets like Japan and Korea by partnering with regional champions to create native, enterprise-focused models.
• 🖼️ There's significant enterprise demand for multimodal AI, crucial for understanding visual data in documents (PDFs, slide decks) and enabling computer interaction.

AI's Future and Societal Impact

• ⚠️ While AI offers immense potential, sensitive use cases in medicine and finance require human oversight and a human-in-the-loop approach.
• 🚀 Early adopters of AI are gaining a staggering competitive advantage, with organizations increasingly identifying and implementing hundreds of AI use cases.
• 🌱 Aidan Gomez is optimistic about AI's role in driving GDP-impacting productivity gains, making goods and services more abundant, and integrating into everyone's workday globally.