Artificial intelligence is about to shift from simply predicting the next word to actually understanding how the world works. According to DeepMind CEO Demis Hassabis, the next major breakthroughs in AI won't come from just making models bigger or feeding them more data. Instead, 2026 marks a pivotal year when AI systems will gain the ability to learn continuously, remember across conversations, and simulate physics and causality the way humans do .

For years, the AI industry has focused on scaling: building larger language models (LLMs), which are AI systems trained on vast amounts of text to predict and generate language. But Hassabis and other leading researchers now argue that scaling alone won't get us to artificial general intelligence (AGI), the theoretical point where AI matches or exceeds human intelligence across all domains. Instead, the field is pivoting toward what researchers call "world models," internal simulations that let AI understand physics, causality, and how objects behave in the real world .

What Are World Models and Why Do They Matter?

Imagine an AI system that doesn't just read about gravity or watch videos of falling objects. Instead, it builds an internal mental model of how gravity works, predicts what will happen when you drop a ball, and adjusts its predictions based on new information. That's the promise of world models. Unlike current AI systems that excel at pattern-matching and next-token prediction, world models enable planning, imagination, and grounded interaction with the physical world .

Hassabis explained that current AI systems lack several critical capabilities that humans take for granted. These gaps are not just about raw computing power; they're architectural and algorithmic challenges that require fundamental rethinking of how AI learns and reasons .

What Specific Breakthroughs Does DeepMind Expect in 2026?

According to Hassabis's recent statements and DeepMind's research roadmap, 2026 will see several convergent advances that reshape how AI systems work:

• Reliable World Models: Interactive systems similar to DeepMind's Genie will enable real-time physics simulation for training embodied AI, allowing robots and agents to learn from simulated environments before acting in the real world.
• Continual Learning Prototypes: AI systems that learn continuously from new experiences without catastrophic forgetting, the problem where learning new information causes a system to forget what it previously learned.
• Persistent Memory Architectures: Nested learning and Titans-style memory systems will become standard in agentic frameworks, allowing AI to maintain context and identity across multiple conversations and sessions.
• Omni-Models: Foundation models that integrate text, vision, action, and memory into unified systems, moving beyond single-modality AI.

Hassabis allocates roughly half of DeepMind's resources to blue-sky algorithmic innovation and half to maximal scaling, reflecting his belief that both approaches are necessary . This dual strategy suggests that the path to AGI requires neither pure scaling nor pure algorithmic innovation, but rather a hybrid approach that combines both.

The research community is already moving in this direction. Yann LeCun, Chief AI Officer at Meta, heavily promotes world models through his JEPA and V-JEPA research families, arguing that predictive, grounded intelligence is more fundamental than pure language modeling . This consensus across major AI labs indicates a genuine shift in research priorities, not just speculation.

How Will These Advances Change AI Capabilities?

The practical implications are substantial. Current AI systems struggle with long-horizon tasks, consistency, and real-world adaptation. A system with world models and continual learning could handle multi-week projects, self-correct when plans fail, and personalize its behavior based on individual user interactions .

Consider robotics: today's robots rely on pre-programmed behaviors or extensive real-world training. With world models, a robot could learn physics from simulation, understand cause and effect, and adapt to novel situations without retraining from scratch. Similarly, AI agents could plan complex multi-step tasks by mentally simulating outcomes before acting, much like humans imagine consequences before making decisions.

"Scaling laws have not hit the limits. LLMs will not commoditize easily. Push scaling to the absolute maximum will be a key component of AGI. However, it's roughly 50/50 whether scaling alone suffices or if one to two more breakthroughs are needed," explained Demis Hassabis.

Demis Hassabis, CEO at DeepMind

Hassabis views AGI as plausible within the next five to ten years, roughly between 2030 and 2035, driven by these algorithmic advances plus relentless progress on models like Gemini . He characterizes the potential impact as roughly ten times larger and faster than the Industrial Revolution, underscoring the stakes of getting these breakthroughs right.

What About Pure Scaling? Is It Still Important?

Scaling isn't being abandoned; it's being reframed. DeepMind, OpenAI, Anthropic, and Google all agree that pre-training, post-training, and especially inference-time scaling have substantial headroom through at least 2027 to 2028 . Inference-time scaling refers to letting models "think longer" during problem-solving, similar to how humans spend more time reasoning through difficult questions.

However, the field is shifting emphasis toward inference scaling and multimodal data (combining text, images, and video) rather than purely expanding training data. Test-time training, where models improve their answers by reasoning through problems step-by-step, is emerging as a major source of capability gains. This approach mirrors OpenAI's o1 model, which uses extended reasoning chains to solve complex problems .

Steps to Understand How World Models Will Transform AI Development

• Recognize the Shift: The AI industry is moving from static transformer models that predict text to dynamic, memory-augmented systems that simulate and reason about the world.
• Track Algorithmic Efficiency: New architectures are delivering four to seventeen times effective performance gains over raw scaling in specific domains like memory and reasoning, indicating that innovation is compounding beyond just larger models.
• Watch for Hybrid Systems: The convergence of scaling, world models, continual learning, and planning will define the next generation of AI agents and robotics applications.
• Monitor Robotics and Scientific Discovery: These domains will serve as early testbeds for world models, with grounded simulation accelerating progress in both fields by 2027.

The timeline matters. Hassabis and the broader AI research community expect 2026 to be a watershed year for reliable world models and continual learning prototypes. By 2027, unified foundation models with persistent memory should emerge. By 2028 and beyond, if these breakthroughs compound, the field could see systems capable of autonomous self-improvement and multi-week project execution .

What makes this prediction credible is not just Hassabis's track record, but the alignment across competing labs. OpenAI, Anthropic, Meta, and Google are all pursuing similar directions: inference-time scaling, memory augmentation, world models, and agentic loops. This convergence suggests the field has identified genuine bottlenecks and is moving toward solutions in parallel.

The era of scaling-only large language models is ending. The era of memory-augmented, world-model-driven, continually learning agentic systems is beginning. For developers, researchers, and organizations building on AI, 2026 represents a critical inflection point where the capabilities and architectures of AI systems will fundamentally change.