S3 E4 Cohere Founder CEO Aidan Gomez: How to build Instructable Agents

In this talk

1. 2:42

   How language models learn

   Gomez explains that Transformers can scale across large, varied text datasets and are trained on powerful computer clusters. Models first learn from cleaned web text, then receive smaller, curated datasets that steer their behavior toward tasks such as following instructions or holding dialogue.

2. 12:28

   Steering models and making them accessible

   Gomez describes prompts, examples, and fine-tuning as ways to guide a model, while hoping that future models will need less specialized prompting. He says Cohere aims to spread the costs of compute, data, and expertise across many users so more developers can build with language models.

3. 20:58

   From knowledge bases to agents

   Gomez describes assistants that retrieve current information from company or personal data sources, then use it to answer questions. He expects customer support tasks such as resetting passwords or redirecting packages to be early applications, and points to browser control as a broader but harder way for agents to take action.

4. 34:31

   Research, company building, and what comes next

   Gomez explains that Cohere for AI supports research and mentorship through its Scholars Program and research residencies. He recounts that he did not foresee the Transformer paper’s impact, hopes researchers find more efficient architectures, and advises aspiring researchers to study fundamentals and papers while product builders focus on customer needs.