His vision centers on “human agents intelligence” — a future where networks of specialized AI and human agents collaborate to solve problems. This represents a fundamental shift from thinking about AGI as a single superintelligent system to viewing it as an emergent property of interconnected intelligence. The approach acknowledges what many AI researchers have ignored: true intelligence is inherently social and collaborative.
This perspective cuts through the hype to deliver a roadmap that makes sense.
System 1: We’ve Nearly Maxed Out “Thinking Fast”
The reactive, instinctive capabilities of large language models represent the low-hanging fruit of AI development.
Current models like Llama 4 have gotten remarkably good at generating immediate responses based on their pre-trained knowledge. The massive gains we’ve seen from scaling up parameters and training data primarily benefit these System 1 capabilities, which essentially rely on pattern recognition and memorization. However, we’re witnessing diminishing returns from this approach, suggesting we’re approaching the ceiling of what pure pre-training can accomplish.
The real breakthroughs will come from what happens next.
System 2: “Thinking Slow” Is Where the Action Is
Reasoning capabilities represent the current frontier of AI research and development.
Today’s most promising advancements come from models that can deliberate, consider multiple options, and work through problems step-by-step using “scratchpads.” These reasoning models utilize reinforcement learning to effectively utilize their working memory, thereby dramatically improving performance on complex tasks such as math and coding. The shift from pure pattern matching to explicit reasoning is an “evolutionary” step that addresses many of the limitations of current AI systems.
This is where we’ll see the biggest improvements in the next few years.
System 3: “Thinking Together” Will Define the AGI Era
The true path to AGI runs through multi-agent collaboration networks.
Creating systems where specialized agents can find each other, communicate effectively, and collaborate on complex tasks requires solving entirely new classes of problems. We’ll need to democratize agent creation, build platforms for agent discovery and interaction, develop standardized communication protocols, and implement sophisticated “theory of mind” capabilities. The ultimate challenge, beyond making smarter individual AIs, is orchestrating efficient collaboration between systems with complementary skills.
This third system represents the most ambitious and potentially revolutionary aspect of Van der Maaten’s vision.
The Twin Challenges: Scaling and Integration
Building these systems isn’t just hard — it’s “rocket science every step along the way.”
The AI community faces two fundamental challenges that span all three systems. Scaling must evolve beyond simply increasing parameters to include scaling agent networks and interactions. Integration requires harmonizing vertical capabilities (like code generation or visual understanding) with horizontal capabilities (like consistent tone and instruction-following). These challenges demand innovations that go beyond our current approaches to AI development.
We need breakthroughs in multi-agent reinforcement learning and mechanism design that AI researchers have barely begun to explore.
Why This Framework Matters
Van der Maaten’s three-system model is a pragmatic alternative to the magical thinking that dominates the discourse on AGI.
By breaking down the path to AGI into discrete, understandable systems with specific capabilities and challenges, this framework provides a roadmap that feels both ambitious and achievable. It acknowledges the social nature of intelligence and the necessity of collaboration rather than focusing solely on individual superintelligence. The approach also highlights underexplored areas where innovations are needed, pointing researchers in productive directions.
This could be the blueprint that finally moves AGI discussions from science fiction to science.
