May 4, 2025. From LLMs to AGI.

The gap between current Large Language Models and true Artificial General Intelligence isn't just about scale or data. It's about fundamentally different approaches to machine intelligence. Current LLMs lack genuine causal understanding of our world. They can predict which words follow others with astonishing accuracy, but they don't truly comprehend what those words mean in relationship to physical reality.

Think about it this way: if someone memorizes thousands of Chinese phrases without ever living in China or deeply engaging with its culture, do they truly understand Chinese? They might produce grammatically correct sentences and even fool native speakers in brief conversations, but their knowledge remains superficial. They've mastered symbols without grasping their deeper meanings and contexts.

This is precisely where our current AI systems stand. They're impressive pattern matchers without true comprehension.

The models we build today mirror human knowledge rather than expanding it. They reflect what we've written and thought rather than discovering new knowledge or generating truly novel insights. This creates a significant vulnerability when these systems encounter situations beyond their training distribution. They falter precisely when we need them most - when facing the unknown.

I've been obsessing over a simple question for months: Could we create AI systems with reasoning capabilities independent of training data? Systems that don't just recombine existing knowledge but genuinely create new knowledge?

Current models don't have knowledge representation beyond statistical correlations. They don't understand physics, chemistry, or biology in any meaningful sense. They don't grasp that objects fall when dropped because of gravity - they just know that humans typically write about objects falling downward because that pattern appears consistently in their training data.

My work now focuses on building pathways toward models that can continuously learn. Not just pattern matching but genuinely understanding. Systems capable not just of using their neural architectures but of learning from them, adapting them, and modifying their own architecture to address fundamental flaws.

Models with access to internal self-modification.

We can approach this challenge in several ways. First, instead of teaching models knowledge, we must teach them how to learn and solve problems systemically. Current architectures only deliver automated versions of our existing cognitive capabilities. They can't truly learn, making them unsuitable for high-stakes environments.

Second, we need to develop reasoning capabilities transferable across domains. Not just bit-to-bit connections but bit-to-axis and bit-to-atom - bridging digital concepts with physical reality. The models need to connect abstract mathematics to physical phenomena, understanding not just correlations but causality.

Third, these systems must independently supervise their discoveries, adding to fundamental understanding while staying grounded in physical laws. They need to verify their own knowledge against reality through experimentation and observation.

Fourth, the infrastructure challenge itself is enormous. We can't rely on centralized computing forever. A truly intelligent system needs to process critical information locally. Imagine a humanoid robot that needs to connect to a central server for every decision - the latency alone would make it impractical, not to mention the security vulnerabilities.

They need to sustain themselves autonomously for decades, not hours or days.

Fifth, perhaps most importantly, we need to instill curiosity. The drive to question, to seek information, to parse feedback, to develop pattern recognition without supervision. That spark of inquiry that drives human discovery needs to be encoded somehow into these systems.

The technical challenges ahead are immense. We need new energy sources that can be collected, stored, and harvested in closed environments operating continuously. Power sources the size of a human palm that can safely store energy and self-recharge. Current battery technology won't cut it.

We need fundamentally new forms of computation. Our current approaches are too inefficient, too brittle, too limited. Humans must learn to co-optimize and synchronize software and hardware operations in ways currently unimaginable.

We need to develop operational cycles for these systems. Even human brains need sleep to function properly. Our materials science simply doesn't allow for continuous operation over decades without rest periods for recalibration and repair.

This is why studying humans comes first in my approach. Not anatomy or mechanics but cognitive processes. The thinking chains and unseen aspects of human cognition. If we can observe these processes through an engineering lens, we can potentially map them, understand them, and eventually replicate them.

Consciousness, genuine understanding, curiosity - these aren't merely philosophical concepts but engineering challenges waiting to be solved.

My background in physics gives me a particular perspective on these challenges. The fundamental constraints we'll face as computational demands grow exponentially aren't primarily algorithmic but physical. Energy efficiency, cooling technologies, material limitations, and computational boundaries will all become increasingly critical factors.

Those who understand these constraints early will shape the emerging AGI landscape. While others focus on the next incremental improvement in neural network architecture, the real breakthroughs will come from addressing the physical limitations that constrain all computing systems.

The path forward combines deep technical insight into physical constraints, precise timing in addressing these limitations, strategic positioning at key control points in emerging value chains, and disciplined execution of complex strategies.

The window for establishing this position exists now, before most participants fully recognize these fundamental constraints. The physical limitations that become binding as computational demands increase represent both the greatest challenge and the greatest opportunity in this space.

Working from Indonesia gives me a different perspective than those in traditional tech hubs. Sometimes distance from the center provides clarity. I see patterns that might be missed by those too close to existing paradigms. The constraints I face force creativity that abundance might not.

This pursuit isn't just about building better AI. It represents a pathway to abundant energy, plentiful resources, and the elimination of scarcity. It offers the chance for humanity to live fully rather than merely survive. The unseen physical limitations on computational scaling will increasingly shape the trajectory of AI development.

I understand how all this sounds. One person in a bedroom with a squeaky chair claiming to tackle problems that multi-billion dollar companies with thousands of researchers haven't solved. The history of science is full of people dismissed as odd, weird, or delusional before eventually being recognized as visionaries. The same humans who criticize can later adore. This pattern of judgment says more about human emotion than about the ideas themselves.

That's why I'm focused on the work itself, not on convincing others. The results will speak for themselves.

The foundations for this work are being laid now, beginning with a deeper understanding of human cognition that goes beyond surface-level imitation. True artificial general intelligence requires understanding the difference between correlation and causation, between pattern matching and genuine comprehension.

The models we build today are impressive shadow puppets - creating the illusion of intelligence through clever trickery. The models we need will be something else entirely - systems that understand reality at a fundamental level and can help us expand our knowledge rather than simply reflect it.

This represents the next frontier of my work - systems that genuinely expand human knowledge through continuous learning and adaptation. The technology that emerges will ultimately help us address our most pressing challenges while creating unprecedented opportunities for those positioned at critical control points in this emerging landscape.

The path from LLMs to AGI isn't about more data or bigger models. It's about fundamentally different approaches to knowledge representation, learning, and intelligence itself. That's where I'm focused now - not on incremental improvements to existing paradigms but on laying the groundwork for the paradigm shift that's coming.