Mustafa Suleiman on the Coming Age of AI Agents

How Mustafa Suleiman Sees the Transition to AI Agents

When a DeepMind co-founder who spent a decade in the "flat part of the exponential" tells you we're underreacting, it's worth paying attention.

Mustafa Suleiman's core thesis is clear: we're transitioning from a world of operating systems, apps, and browsers to a world of agents and companions. Every user interface will eventually be subsumed into conversational agentic form - a real assistant in your pocket that has all your context and can do anything.

What makes this conversation valuable is Suleiman's unique vantage point. He ground through the years when DeepMind was "optimizing air conditioning" and LLMs were predicting single words in sentences. He saw Lambda work at Google but couldn't ship it. He raised $1.5B for Inflection, then watched open-source models undermine his entire capital base months later.

His "modern Turing test" proposal cuts through academic benchmarks with a brutally simple metric: can an AI turn $100K into $1M? 10x return on investment by an autonomous agent. He predicts this is "within view" in the next couple of years.

The most surprising admission? The plummeting cost of inference. He expected AI would be expensive for years. Instead we got 100x cost reduction in two years, with open-source models worth billions released for free. "That bit I totally got wrong."

On AI for science, he's more cautious - it's harder than economic tasks because there's less training data for novel discovery and the human-in-the-loop is harder to implement. But the combination of logical reasoning (from math/coding training) plus creative interpolation is "a lethal combination" for scientific progress.

6 Insights From Mustafa Suleiman on AI's Future

• Apps to agents - The fundamental transition is from user interfaces to AI agents that subsume all current software paradigms
• Modern Turing test - Economic benchmarks matter more than academic ones: can AI 10x your investment?
• 2027 timeline - Agents that pass economic benchmarks are "within view" in the next couple of years
• Microsoft's bet - Being the trusted, steady platform for enterprise AI is the strategy, not racing on benchmarks
• Open source shock - Billion-dollar models released for free completely changed the competitive landscape
• Science is harder - AI for novel discovery lacks the training data and human-in-loop mechanics of business tasks

What This Means for Enterprise Software

We're transitioning from operating systems and apps to agents and companions. Every user interface will eventually be conversational. The "modern Turing test" - can an AI turn $100K into $1M? - cuts through academic benchmarks to ask what matters. A DeepMind co-founder thinks we'll pass it by 2027.