The week the AI map stopped making sense

For roughly a decade, two of the most credible voices in AI have been telling anyone who would listen that large language models, however impressive, are a detour. Yann LeCun and Fei-Fei Li have been making the same argument in different registers: the world isn't text, and you can't capture gravity, fluid dynamics, or cause and effect through language alone. The next frontier, they argued, is spatial intelligence — machines that don't just describe the world but simulate it.

When Fei-Fei Li published "From Words to Worlds" last November, the timeline still felt long. Her own company, World Labs, had Marble in limited preview. Google DeepMind had Genie 2. The argument was sound, the products felt 2028.

This week the timeline collapsed. Marble is publicly available. Genie 3 is generating navigable, photorealistic 720p worlds at 20–24 frames per second. Gemini Omni Flash rolled out to the Gemini app, Flow, and YouTube Shorts — free — with multi-turn natural-language video editing where the physics actually holds. Each of these used to be a production budget and a six-week timeline. Now it's a text prompt.

The interesting thing for education isn't that AI makes nicer visuals now. It's that simulation as pedagogy — the thing that used to be locked behind expert production capability — has just been democratised. Anything inaccessible (ancient sites, cellular biology, dangerous industrial processes), anything where the dynamics are the hard part (disease spread, structural load), anything that needs to be specific to a region or a cohort. All of it is now in scope for any educator who can write a sentence.

Sitting against that is Mira Murati's counter-move. Her lab, Thinking Machines, released its first technical paper this week: interaction models, built around full-duplex communication and 0.4-second response latency. Her position is the opposite of the agent narrative racing through the rest of the field. The most important variable, she argues, isn't raw capability. It's the quality of the collaboration. The deliberate design choice is to keep a person inside the loop for as long as possible.

Andrej Karpathy's May 19, 2026 announcement that he has joined Anthropic to work on Claude pre-training.

And almost as an aside, Andrej Karpathy joined Anthropic, running a team that uses Claude to accelerate Claude's own pre-training research.

Three trajectories. Three different bets on what education becomes. All shipping this week.