Mustafa Suleyman: A Decade on the Flat Exponential

What Mustafa Suleyman Learned at DeepMind

Mustafa Suleyman offers a uniquely grounded perspective on AI's trajectory. Having spent a decade at DeepMind during what he calls "the flat part of the exponential" — when breakthrough papers were celebrated but commercial applications remained elusive — he provides crucial context for understanding where we're headed. His revelation that Google's $650M acquisition of DeepMind was initially justified by optimizing data center air conditioning speaks volumes about how transformative technologies sneak up on us.

The conversation's most striking moment comes when Suleyman admits what he got wrong: the democratization of AI through open-source models and plummeting inference costs. When Inflection AI raised $1.5B to build one of the largest H100 clusters, the emergence of Llama and accessible APIs fundamentally changed the competitive landscape overnight. This humility from someone who correctly predicted the scaling laws trajectory lends credibility to his other forecasts.

His "Modern Turing Test" proposal — measuring AI capability by whether an agent can turn $100K into $1M — cuts through academic benchmark theater to ask what actually matters. Suleyman suggests we'll see agents pass this test within 2 years (by 2027), while acknowledging that AI for science will take longer because novel discovery lacks the training data and human-in-the-loop opportunities that business tasks provide.

4 Insights on AI's Trajectory From Microsoft's AI Chief

• Microsoft's strategic thesis: The transition from operating systems, search engines, and apps to AI agents and companions is the defining paradigm shift — and Microsoft is positioning all 250,000 employees around this transition
• The "Modern Turing Test" measures economic capability: can an AI agent 10x a $100K investment? Suleyman predicts this will be achieved within 2 years
• AI inference costs have dropped 100-1000x in two years — a pace Suleyman admits he completely failed to predict, especially the impact of open-source models
• The Lambda moment at Google was Suleyman's most recent "mind-blown" experience — watching conversational AI emerge was when he knew the paradigm had shifted, and Google's failure to ship it led to the exodus that spawned Character AI, Adept, and Inflection

What the Modern Turing Test Means for AI Agents

The man who spent a decade on "the flat part of the exponential" at DeepMind now leads Microsoft's AI strategy. His Modern Turing Test - can an agent turn $100K into $1M? - cuts through benchmark theater to ask what matters. He thinks we'll pass it by 2027.