The World Model Gamble: Can AI CEOs Outperform LLM Scaling?

The debate over AI architecture is shifting from text-based pattern matching to the simulation of physical reality. At the center of this shift is Skyfall AI, a Toronto- and San Francisco-based venture founded by Sam Pasupalak, Sumit Pasupalak, and Kaheer Suleman. The trio previously co-founded Maluuba, a deep learning lab acquired by Microsoft in 2017, which focused on natural language understanding under the guidance of Turing Award winners Yoshua Bengio and Richard Sutton.

As BetaKit first reported, Skyfall AI is betting on "world models"—systems trained on spatial, movement, and physical data rather than the text-heavy datasets used by large language models (LLMs). The goal is to create internal 3D representations of the world that integrate cause-and-effect, moving beyond the "brute force" approaches of current AI. Sam Pasupalak argues that world models are superior to LLMs in critical business leadership areas, specifically high-stakes decision-making under uncertainty and long-horizon planning.

To prove this cognitive leap, Skyfall AI is pursuing a radical operational experiment: acquiring a "micro-B2B-SaaS" business for up to $1 million USD to install an AI CEO. Backed by Garage Capital and Inovia Capital, the firm intends to run the acquired company completely autonomously across finance, product, and marketing functions, with a target of doubling revenue within six months. Karamdeep Nijjar, a partner at Inovia Capital, noted on LinkedIn that Pasupalak is focused on real-world applicability over static benchmarks.

This push toward "enterprise super intelligence" is not an isolated trend. BetaKit notes that another world-model startup, General Intuition, recently reached a valuation of $2.3 billion USD ($3.2 billion CAD) by utilizing video-gameplay clips to train robots and physical AI. This approach stems from research into the failures of LLMs when tasked with complex simulations, such as the game RollerCoaster Tycoon.

However, the human cost of this architectural shift is significant. Pasupalak stated that for their first acquisition, any existing employees or contractors (estimated at one to five people) would receive "generous departure packages" and job search support. While Pasupalak suggests that automating monotonous operational tasks will free humans for creative work, BetaKit highlights broader economic anxieties. While Meta's Mark Zuckerberg and Anthropic's Dario Amodei have previously commented on white-collar job losses, Quartz reports that over 200 economists and researchers have signed an open letter warning of large-scale job displacement.