Q&A with DeepMind co-founder Mustafa Suleyman on LLMs becoming more controllable, “interactive” AI after generative AI, AI regulation being achievable, and more
DeepMind cofounder Mustafa Suleyman wants to build a chatbot that does a whole lot more than chat.
Context & Ripple Effects
This interview captures an early version of Suleyman’s argument that chat-oriented LLMs should evolve into systems that can be directed and trusted in more consequential interactions. It follows coverage describing chat LLMs as a general-purpose information-processing layer, making controllability a practical issue rather than solely a research goal.
The later coverage arc returns to the same themes through Suleyman’s discussion of Copilot and AI agents and regulation, suggesting a durable focus on moving from model outputs to managed AI behavior.
First-order effects
- Suleyman publicly makes controllability and broader interaction—not chat alone—the product criteria for the next generation of LLM-based assistants.
- His assertion that regulation is achievable strengthens the case for treating safeguards and oversight as deployable design requirements rather than a post-launch constraint.
Second-order effects
- AI developers building assistants and agents face pressure to differentiate on predictable behavior, user control, and governance alongside raw model capability.
- Organizations adopting more interactive AI will need operational processes for defining permissions, monitoring behavior, and assigning accountability when systems act beyond a conversational reply.
Third-order effects
- If this direction holds, competition shifts from standalone generative models toward controlled, task-capable AI systems whose reliability can be demonstrated in use.
- The industry’s regulatory debate may increasingly center on auditable operational controls for interactive systems, especially as their role expands beyond content generation.
The trend: This is one early data point in the shift from generative AI as a chat interface to governed, agent-like systems designed to act within defined bounds.