The 55-Slide Deck That Predicted AlphaFold, LLMs, and the Next Decade of AI

Demis Hassabis presented this at MIT in March 2019. Most of what he predicted came true. Here is what he saw before anyone else did.

In March 2019, Demis Hassabis stood in front of an MIT audience and walked through 55 slides that laid out DeepMind’s entire theory of intelligence.

This was before AlphaFold 2. Before the LLM era at DeepMind. Before the Google Brain merger. Before most people in the room had any idea what was coming.

Six years later, the deck reads like a roadmap.

The core argument is simple. Two approaches exist to building general-purpose AI. Expert systems rely on hardcoded knowledge and break when the unexpected happens. Learning systems learn from first principles, generalize to new tasks, and keep getting better. DeepMind picked the second. Everything that followed was proof that the bet was right.

What makes this deck worth your time in 2026 is not nostalgia. The five open problems Hassabis listed in March 2019: unsupervised learning, memory, transfer learning, imagination-based planning and language understanding, are the exact research agenda that produced Claude, ChatGPT, and every AI model running in production today.

The AI era did not start with ChatGPT. It started with the argument in this deck.

What is inside the full breakdown:

▫️ The complete 55-slide DeepMind presentation

▫️ The AlphaGo to AlphaZero lineage and what it actually proved

▫️ Why AlphaStar used a Transformer in 2019 before the rest of the industry caught up

▫️ The AlphaFold thesis Hassabis described as “still a long way to go,” one year before he solved it

▫️ What aged perfectly, what the LLM era replaced, and what it means for founders and investors building in AI today

---

The Full Presentation (downloadable) + Complete Breakdown: