Cohere's Chief AI Officer, Joelle Pineau: Why Scaling Laws Will Continue & Future of Synthetic Data

Cohere CAO Joelle Pineau argues scaling laws remain robust, AI coding is where image generation was in 2015, and agent security is the next major unsolved frontier.

• Pineau argues most employees can achieve 10x productivity with AI in the next few years — she finds replacing the bottom 5% of workers a weaker and less realistic barometer.
• AI coding today mirrors image generation circa 2015: output quality is poor now but will be excellent in a decade; curation and intent, not coding, become the scarce human skill.
• Agent impersonation — agents acting on behalf of entities they don’t represent — is the unresolved security frontier; red-teaming for LLMs exists but agent vulnerabilities are largely unmapped.
• Scaling laws have been remarkably robust; Pineau would not bet against them, but algorithmic breakthroughs (transformers, Adam optimizer, reasoning) drive nonlinear gains while compute and data scale roughly linearly.
• Synthetic data degrades models only when diversity collapses; coding sits in a middle zone where diversity can be injected, making large-scale synthetic code training viable without performance collapse.
• Data is getting more expensive because useful labeling now requires domain specialists and building realistic agent simulation environments, not simple classification tasks.
• Pineau changed her mind on neural networks — she previously expected them to be surpassed at each new scale of data, as SVMs once were, but now believes they are here to stay.
• Closing AI research is a deep mistake; open circulation of ideas is essential to innovation, and she does not believe closed systems will prove effective long-term.

2025-11-03 · Watch on YouTube