The Why?
This article is a bit different from my usual "here is code, let's run it" posts. This is a dump of my thoughts on AI, a topic I’ve spent the better part of 10 years breaking my brain over.
I’m not an AI researcher. I’m not a neuroscientist. But over the last decade, I’ve built systems that tried to mimic "thinking," and I’ve watched the recent LLM hype with a lot of skepticism. This post covers how I started thinking about this, why I think the current "LLM is AGI" narrative is nonsense, and why Google’s project TITANS is the first time I’ve seen the industry actually move in the right direction.
The Origin Story
So we have to go back about 10 years. Back then I worked in a very small startup that was looking for new opportunities.
I had a long software development background and also some profound experience in red-team security/pentesting. At that time, the topic of security for "internet facing" systems finally started becoming a public concern. So I told my boss: we should build an automated security scanner.
The idea was simple: have a system that utilizes existing tools like nmap, sqlmap, nikto, etc., to create automated reports. A win for both parties.
Now, many people will be like "metasploit?". Well, metasploit enterprise is not really feasible for a low cost model like we planned, and also i always was a bit offended by it cuz the code is frankly speaking a mess. What also annoyed me alot about metasploit was that it was like the shotgun approach. It would recognize a system as PHP but still run a ton of ASP.NET exploits that will just not work.
So i told myself: i need to build something that utilizes existing tools and orchestrates them in a way that i can collect the information in a dynamic build form.
The Architecture
I started concepting a system based on "Actions."
Actions were immutable functions (classes) with a defined input structure. For example, an `ip->port` input would trigger an `nmap script detection`.
* This action would map its result automatically back into a central knowledge graph (I used OrientDB at the time).
* A scheduler would trigger new actions based on that new data.
Basically a system that will execute everything possible, but only whats really usefull.
I also built the core logic of scheduling new jobs and mapping actions return data into the central knowledge graph as actions themself. It had parallel workers executing these actions, continuously learning and acting based on what it learned.
But for me it was clear that whatever of learning entity it must have a drive that makes it wanne learn. Not just a `while() do`, but something that actually will make the system process learned data and make it trigger processing more. I grouped these core "actions" into a category I called "Instincts."
The most important instinct is: curiosity.
(Side note: I built this whole thing in PHP. shrug I mean yes i made a multithreading cli environment work stable with IPC etc (not using pthreads) but ye... as you might imagine php is possibly one of the worst options.)
The Shift to Go and the "Brain" Problem
Eventually, the project was cancelled due to company shifts, but I kept the code and the concept. At this time golang also catched my interest. I find it really easy to get into, and saw it as a more practical solution for fast prototyping.
I wanted to take this "smart learning architecture" deeper. The first issue I hit was the database. While OrientDB kinda "worked," the performance was really underwhelming. So, well, i kinda said to myself: i gonne write an in memory graphdb in golang.
(You can read about that journey here.
While working on this architecture (my private "go-cyberbrain"), I deepened my thoughts about how an AI actually has to be built.
To me it was clear that you need two distinct parts:
1. The Backbrain: Your long-term memory storage and sub-consciousness. An n-dimensional datagraph.
2. The Frontbrain: Your active consciousness. This contains your current context focus and time perception.
This is crucial. To properly communicate, two individuals need a similar time perception.
The Context Dilemma
This brings us to the biggest problem: Context.
If I ask you "What does Person X wear?", you don't just query a list. You use context.
* Is it cold?
* Is he a fan of Jack Wolfskin?
* Is he at a wedding?
Information is like dynamic-n-means clustering. We have to group information into N groups but we don't know the amount of N and we have no fixed formular to calculate N. We must decide N based on the information given.
Also, context is temporal. Your friend sitting on a bench and a dog chasing a bunny in the park are unrelated in causation, but they share a temporal context because they happened at the same time.
Why "LLM is AGI" is a joke
This is where I get annoyed with the current state of things.
People calling LLMs "AI" or claiming "LLM IS SOON AGI" are missing the point.
To have intelligence, a thing has to be actively learning. An LLM is a static Backbrain. It has no Frontbrain. It has no time perception. It has a context window, sure, but it is "stateless." It resets every time.
The Abstraction Realization
Also, considering all of this, it became very clear to me that my text-based experiments were just a prototype for the logic. A real AI shouldn't store data in human formats like text or JSON.
It has to create its own way of storing data into abstract Neural Network structures. These structures should be formed by the AI itself, derived from what it learns via sensoric inputs and its own "interpretation," not by us spoon-feeding it a human text format.
This train of thought is actually what led me to the conclusion—years ago—that the only way to truly train such an entity was by emulating a complete virtual environment. You would have to let the AI "exist" in a world with full physics emulation, feeding it raw sensory data and letting it learn to survive.
Enter TITANS
About a year ago, I stumbled upon two things that made me realize the industry was finally catching up.
First, I watched an interview with Jensen Huang (Nvidia CEO). He talked about the Omniverse—a virtual environment with simulated physics to train AIs. This is literally what I thought of: you cannot train an AI on text alone. You have to feed it raw sensory input (physics, light, sound) in a simulation so it learns the rules of the world, not just the description of it.
Second, and most importantly, I read the paper on Google TITANS.
TITANS is a system which, instead of the current approach of a static read-only backbrain, introduces a Neural Memory Module.
* It has a dynamic, learning context.
It effectively creates a "Frontbrain" that updates its weights while* it is processing information.
* It splits the architecture into "Short term memory" (Context) and "Long term storage" (Weights).
For me, this was the sign. Finally, someone made a step that I would consider the right direction. It moves us away from the "brute force" approach and towards an architecture that can actually "think" in a learning context window.
Conclusion
There are still pain points left to solve:
1. Sensoric Merging: Merging vision/audio/physics into a central port of "recognition."
2. Instincts: Implementing the "drive" (Hunger, Curiosity) that forces the system to evolve.
3. Learning Backbrain: How to train the system to learn from experience not just in the current frontbrain context but also update the backbrain based on newly learned data .
But after seeing Titans, I stopped my private research on the concept. The industry finally stopped just focussing on the LLM Hype and actually going (baby)steps towards what i would call a real ai - I’m going to sit back and watch how it unfolds.
Also, since we are talking about a span of 10 years, this article logically only contains a small subsection of , what i would consider the most important, points i thought about.
That said, specialized networks like LLMs still can serve practical purposes and can be a very useful tool - just stop calling them "AI"...
So long and thanks for all the fish.
laughingman aka voodooEntity