I’ve spent the better part of a decade breaking my brain over AI. I’m not an academic researcher, and I’m not a neuroscientist. I’m a systems architect. And from an architecture perspective, the current narrative that "LLMs are AGI" is structurally impossible.
This isn't a post about code. It’s a dump of my thoughts on the logical requirements for a thinking machine—conclusions I reached by trying (and failing) to build one, and why the concepts behind Google’s recent TITANS paper suggest the industry is finally fixing the broken foundation.
Ten years ago, I tried to build an automated security brain. The goal was a system that didn't just run scripts, but decided what to run.
I designed a system of Atomic Agents. Instead of a linear script, I had small, immutable units of logic. If Input A appeared, Agent B would trigger. The results were dumped into a knowledge graph, which then triggered more agents.
It worked, but it was "dumb." It would run forever, but it had no direction.
This led to my first major realization, which addresses a question I often get: "Why does an AI need 'instincts'? Why isn't a `while(true)` loop enough?"
A `while(true)` loop is just an engine; it provides energy, not direction. A loop will process data until the heat death of the universe, but it doesn't care what it processes. To have intelligence, you need Intrinsic Motivation (in ML terms: an Objective Function). You need a function that tells the system: "Gaining new information is 'good'. Confusion is 'bad'."
Without this drive—specifically Curiosity (Information Gain)—a system is just a calculator. It waits for input. A real AI must seek input.
As I moved from that prototype to designing a "cognitive" architecture in Go (and writing my own in-memory graph database to support it), I hit the bottleneck that LLMs are currently hitting: Memory Structure.
You cannot build a thinking entity with a single database. You need two distinct architectural components:
1. Long-Term Memory (LTM): The subconscious. This is your persistent storage, the weights in a neural network, or a massive knowledge graph. It is vast, but static.
2. Working Memory (Active Inference): The conscious focus. This is a small, volatile window that holds your current context and, crucially, your perception of time.
This brings me to my most controversial take, and the one most people struggle to grasp: You cannot train AGI on text.
Text is a lossy compression of reality. If you feed an AI the sentence "The brown small dog barked," you are limiting its "thought" to that specific text representation. You are spoon-feeding it a pre-processed conclusion.
Real memory is not a string of text. It is a reconstruction of raw sensory data.
When you experience a dog barking, you don't save a text file. Your brain processes multiple streams of data simultaneously:
* Visual Neural Net: Detects "brown," "small," "mixed breed," "happy tail wag."
* Audio Neural Net: Detects "loud barking sound."
* Temporal Context: Records "Wednesday afternoon," and "Right when Event X started."
Your brain then fuses these into a custom, abstract data structure. You link the sound of the bark to the image of the dog. But—and this is crucial—you can still recall the dog without the sound, or the sound without the dog. You understand that the dog caused the sound, but also that the dog appeared at the same time as Event X (temporal correlation).
If you force an AI to store "The brown dog barked" (text/JSON), you strip away all that nuance. You lose the ability to form dynamic connections between the sound and the time independent of the dog.
A true AI must build its own internal data structures. We cannot dictate that it stores data in English or JSON. We must feed it raw sensory inputs (vision, audio, code streams) and let its own Neural Networks decide how to compress and store that information in Latent Space.
It has to build its own understanding of the world, grounded in physics and time, not in our dictionary definitions.
This is why the current hype cycle annoys me.
An LLM is a snapshot of Frozen LTM. It is a brilliant library, but it has no librarian.
* It has no Stateful Working Memory: It resets every time you hit enter.
* It has no Time Perception: It doesn't know if it answered you 5 seconds ago or 5 years ago.
* It has no Intrinsic Motivation: It only speaks when spoken to.
It is a "stateless" system. Intelligence, by definition, requires state.
I had largely given up on the industry solving this—everyone seemed content with better chatbots—until I saw the paper on Google TITANS.
I’m not saying TITANS specifically will be the winner. Implementations vary, and Google kills projects all the time. But the abstract concept behind TITANS is the critical step I’ve been waiting for.
TITANS introduces a Neural Memory Module. Instead of the current approach (static weights), it creates a dynamic, learning context. It effectively splits the architecture into "Short-term context" and "Long-term weights," but—and this is the key—it allows the system to update its understanding while processing.
Whether it’s TITANS or the next architecture that perfects it, this is the turning point. It moves us away from "brute force" text prediction and toward a system that maintains a continuous, evolving stream of thought (In-Context Learning).
We aren't there yet. To get from this concept to a thinking machine, we still need to solve:
1. Multimodal Sensor Fusion: Merging vision, audio, and physics into a central perception of reality, as described in the dog example.
2. Continual Learning: How to update the permanent brain (LTM) based on today's experiences without overwriting yesterday's lessons (Catastrophic Forgetting).
But for the first time in ten years, I’m seeing the industry move past the hype and actually look at the architecture of thought. I’m going to sit back and watch how it unfolds.
laughingman aka voodooEntity