DeepMind co-founder sticks to his 50% odds of AGI by 2028

DeepMind co-founder Shane Legg is sticking to a prediction he made over a decade ago: there is a 50-50 chance researchers will achieve artificial general intelligence (AGI) by 2028.

Speaking on a podcast hosted by Dwarkesh Patel, Legg reaffirmed the forecast he first made publicly on his blog at the tail end of 2011. His timeline actually traces back to 2001, right after reading Ray Kurzweil’s 1999 book, The Age of Spiritual Machines.

Legg explained that Kurzweil convinced him of two underlying truths: computational power and global data would both grow exponentially for decades. Combined with deep learning algorithms designed to process information like human brains, Legg concluded that AGI was entirely achievable in the coming decades, provided nothing catastrophic like a nuclear war broke out.

The checklist problem and scaling limits

With only five years left until 2028, Legg outlined major caveats. The primary obstacle remains defining and measuring human-level intelligence. Human thought is notoriously intricate, and creating a complete checklist of human capabilities is virtually impossible.

Key gaps still exist in modern AI, such as developing episodic memory—the capacity to recall full episodes from the past—and processing streaming video in real time. Legg pushed back against Patel's suggestion that a single benchmark like beating Minecraft could serve as a litmus test. True general intelligence requires an AI system to perform a vast battery of varied tasks without significant blind spots.

Another hurdle is scaling training runs even further. Given the massive energy footprints already required for models like OpenAI's GPT-4, Legg argued that researchers urgently need more scalable algorithms to effectively harness computing resources and data.

As for where the industry stands right now, Legg believes hardware power is finally sufficient. The immediate milestone is training models on data volumes that surpass what an individual human could experience in an entire lifetime, a threshold he thinks the industry is primed to cross.

Why it matters

Unlike many tech executives who adjusted their timelines only after the generative AI boom, Legg has held the exact same 2028 benchmark for over a decade. His reasoning relies on persistent scaling laws rather than sudden breakthroughs, which gives his perspective far more weight than casual hype.

For developers and enterprises, his evaluation framework is the real takeaway. Legg explicitly dismissed single-task triumphs, emphasizing that true AGI cannot have massive capability gaps. Today’s models still struggle with episodic memory and reliable agency, meaning current AI wrappers and workflows are operating on fundamentally incomplete foundations.

We think his balanced stance—giving it exactly a 50 percent shot—is surprisingly realistic. It acknowledges that computational scaling is on track, but refuses to ignore algorithmic inefficiencies and physical energy constraints. If half the coin lands on AGI within five years, preparing for systemic economic shifts is no longer a fringe thought.