Moving away from memorizing massive text data, a new AI simulation project has emerged that builds a 'world model' by directly interacting with its environment based on hormone and memory structures, just like a living organism.
Today, the word that dominates news headlines and the technology market worldwide is undoubtedly ‘Large Language Models (LLMs, technology that learns massive text data to understand context and converse like a human).’ Seeing it answer with smooth, human-like sentences when we ask questions through our smartphones or computers, like ChatGPT, we feel that artificial intelligence has truly advanced astoundingly. However, looking deeper, most of the AI we are currently enthusiastic about is closer to a ‘smart autocomplete machine that most plausibly predicts the next word based on statistical probabilities.’ Simply put, it merely memorizes letter patterns to answer, but cannot feel sadness or joy like a real person or animal, nor vividly remember experiences from colliding and breaking within a physical environment.
But what if, instead of merely memorizing hundreds of millions of books, artificial intelligence learned by directly colliding with the world and building up ‘memory’ and ‘senses’ like a newborn puppy or a young child? An astonishing technology has emerged where virtual ‘hormones’ flow instead of text inside computer chips, and painful or pleasant past experiences are etched into virtual brain cells in the form of ‘long-term memory.’
| This fascinating and unfamiliar approach recently caught people’s attention by appearing on Hacker News, an online community where genius developers and engineers around the world gather to share the latest technology trends. The project, unveiled to the world by developer ‘luzifer333’, is named ‘AI Simulaionen Based on FEP’ [Show HN: AI Simulaionen Based on FEP | Hacker News](https://news.ycombinator.com/item?id=48339824). This project quickly received its first upvote (point) just 17 minutes after posting, becoming a hot topic among developers nextjs-hackernews.vercel.app/item/48339824, and an external site that intensively tracks and introduces new Hacker News projects quickly reported on it Show HN - Tracking Project Showcases on Hacker News. Surprisingly, the developer declared that this system was built independently without using any Large Language Models (LLMs), which are practically the symbol of modern AI [Show HN: AI Simulaionen Based on FEP | Hacker News](https://news.ycombinator.com/item?id=48339824). What exactly is an AI driven by ‘hormones and memory’ rather than text data? Let’s dive together into this unfamiliar world of artificial intelligence. |
Why It Matters
To grasp the significance of this new technology, we must first understand the clear limitations of the artificial intelligence we use every day. As mentioned earlier, current chatbot-type AI is trapped in a flat world of ‘text’. When we speak to AI, it merely fits countless puzzle pieces (tokens) from its vast library (database) together and outputs the most plausible answer sentence. In this process, the AI does not actively experience a physical environment, nor does it possess a biological ‘memory system’ where its personality changes over time or where it deeply regrets specific failures.
This is exactly why the message this newly revealed Hacker News AI simulation project sends to the tech world is massive. This project fundamentally overturns the very philosophy of implementing artificial intelligence. These new AI agents (independent programs that explore and act in their surroundings with a specific purpose) do not force-feed themselves millions of text documents. Instead, they move around directly within their given limited virtual environment, carving out and accumulating new experiences for themselves at every moment.
| The most innovative part to note is that it precisely mimics the biological chemical reactions of living organisms within the artificial intelligence. Metaphorically speaking, it imparts chemical changes to the AI, much like how adrenaline or dopamine works in the human body. According to the developer’s explanation, the internal structure of each of these agents is programmed with a simulated neurochemistry system at its core, along with a hormone crosstalking system that intricately intertwines and reacts [Show HN: AI Simulaionen Based on FEP | Hacker News](https://news.ycombinator.com/item?id=48339824). Furthermore, just like humans, they simultaneously possess a short-term memory system that temporarily stores brief, fleeting experiences and a long-term memory system that preserves experiences essential for survival for a long time [Show HN: AI Simulaionen Based on FEP | Hacker News](https://news.ycombinator.com/item?id=48339824). |
This is a monumental shift that goes beyond merely adding another mathematical calculation formula. It means that the flow of virtual chemicals mimicking biological primal signals like surprise, stress, and curiosity has a direct and powerful impact on the AI’s behavior patterns and memory formation. If artificially intelligent robots integrate deeply into our daily lives in the near future, those robots will not just be rigid machines that have memorized manuals. Rather, they will show an entirely new dimension of intelligence that resembles the survivability of true living organisms—flexibly adapting to complex physical environments, agonizing over failures, and rejoicing in success.
The Explainer
Then, by what principle does this artificial intelligence, which doesn’t even read letters, learn about the world and become smarter? To understand this complex and profound technology, we must first accurately grasp the concept of ‘FEP’, the core framework of the project. FEP stands for ‘Free Energy Projective Simulation (FEPS)’, an interdisciplinary term fusing cognitive science and artificial intelligence.
| According to an in-depth research paper published in the prominent international academic journal PLOS One, this FEPS model is an advanced form of ‘reinforcement learning’ among AI learning methods [Free Energy Projective Simulation (FEPS): Active inference with interpretability | PLOS One](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0331047). |
| If the term reinforcement learning feels unfamiliar, imagine the process of training a young puppy to play frisbee. If the puppy successfully retrieves the flying frisbee, the owner gives it a tasty treat (reward); if it makes a mistake or stands still, it gets nothing. Through this repeated training, the puppy figures out on its own ‘the behavior that earns the most treats’. FEPS agents are similarly intricately designed to fiercely interact directly with their virtual surroundings, constantly gathering external information, and in that process, selecting the optimal behavior that will bring them the greatest reward [Free Energy Projective Simulation (FEPS): Active inference with interpretability | PLOS One](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0331047). |
| Surprisingly, however, FEPS does not stop at the level of a puppy reflexively moving to get a treat. Looking into its internal structure, FEPS agents operate based on a highly evolved next-level technology known in the AI academic community as ‘model-based reinforcement learning (MBRL)’ [Free Energy Projective Simulation (FEPS): Active inference with interpretability | PLOS One](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0331047). The core of this technology is that the agent does not act spontaneously, obsessed only with the immediate reward, but rather creates and constantly updates its own ‘world model’ containing the laws of the world within its complex virtual brain. Furthermore, it actively utilizes this world model to display a frightening level of intelligence by mentally ‘planning’ future actions in advance [Free Energy Projective Simulation (FEPS): Active inference with interpretability | PLOS One](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0331047). |
Let’s examine the concept of a ‘world model’ by comparing it to our daily lives. Imagine entering an unfamiliar, pitch-black room you have never been in before. At first, unable to see anything, you would extend your arms, carefully groping around, bumping into sharp furniture edges or cold walls (active data collection through interaction with the environment). After enduring the pain and bumping around a few times, even without any light, a rough 3D spatial map of the room is vividly sketched in your mind (building your own world model). As time passes, the next time you go to the refrigerator for a drink of water, you will no longer foolishly bump into the walls. Based on the map drawn in your mind, you can instantly find the fastest and safest shortcut and walk smoothly (proactive action planning using the built world model).
If the language models we have known so far are cramming students who infer the appearance of a room merely by memorizing millions of ‘room photos’ floating around the internet, these FEP agents are like true adventurers who walk through the darkness, get hurt, and draw their own 3D maps in their brains.
Now, let’s take a look at the mysterious chemical reactions happening inside this adventurer AI. Suppose this explorer AI wanders around a rugged virtual world and collides violently with hidden hot flames or sharp obstacles, suffering a fatal blow. An AI programmed with past mechanical algorithms would simply record a dry line of numbers in its memory, like “Error Code 404 Occurred” or “Deducted 1 Point,” and move on to the next exploration behavior without any emotional fluctuation.
| However, the AI born from this project is different to the core. The moment this agent hits a terrible obstacle, complex and explosive hormone crosstalking occurs in its virtual neural network, much like a real animal releasing a stress hormone called cortisol in extremely dangerous situations [Show HN: AI Simulaionen Based on FEP | Hacker News](https://news.ycombinator.com/item?id=48339824). This digital neurochemistry action does not leave the bitter experience of failure and pain it just suffered in easily forgotten ‘short-term memory’. It classifies it as core survival information that must never be erased, persistently engraving it into the deepest ‘long-term memory’ of the brain [Show HN: AI Simulaionen Based on FEP | Hacker News](https://news.ycombinator.com/item?id=48339824). |
This is frighteningly identical to the primal survival instinct of a living organism, where a young child who inadvertently burns their hand on a hot stove becomes cautious of fire for life due to the memory of that intense pain. Rather than a machine soullessly reciting tens of thousands of lines of text, it is a marvelous digital organism that figures out the true laws of the world by experiencing its own pain and success with its whole body in a sea of virtual hormones made of 0s and 1s.
Where We Stand
| So, where and how is this amazing experiment—where biological mystery and cutting-edge artificial intelligence perfectly fuse—open to the public right now? As mentioned in the introduction, this innovative project first appeared on the global community stage called Hacker News by a pioneering developer using the mysterious ID ‘luzifer333’ nextjs-hackernews.vercel.app/item/48339824. To create a breathing ecosystem independently, free from subordination to the platforms of Big Tech companies wielding massive capital, this project operates firmly rooted in an independent domain address, a dedicated website named ‘aic-ai-lab.site’ [VueHN2.0 | Show HN: AI Simulaionen Based on FEP](https://vue-hackernews-ssr-5cavbdjcta-ew.a.run.app/item/48339824). |
| The most welcoming news is that anyone around the world with just an internet browser can verify this complex technology, which seems like it would only exist in the tightly closed labs of giant IT companies or difficult academic papers. Through an introductory post, the developer announced that they are conducting a full-scale ‘Open Beta (a trial period where the product is released to the public in advance for feedback before official launch)’ so that curious audiences and engineers can directly connect and freely test the agents’ ever-changing behaviors and evolution processes. This thrilling open beta service was clearly announced to be fully open to the public starting precisely at 20:00 on Monday, Coordinated Universal Time (UTC+2) [Show HN: AI Simulaionen Based on FEP | Hacker News](https://news.ycombinator.com/item?id=48339824). A golden opportunity has opened to directly observe on the monitors right in front of us the movements of these fascinating virtual creatures, which collide with the physical laws of an unfamiliar world without reading a single line of text. |
What’s Next
Think of today’s landscape, where giant IT dinosaur companies like Microsoft, OpenAI, and Google pour astronomical amounts of money every day into building data centers and are solely focused on blindly inflating the size (parameters) of AI. When everyone is fixated on the flat ocean of text, this attempt, moving in the complete opposite direction by delicately mimicking primal hormones and biological neural networks, foreshadows a very heavy seismic shift in the tech world. Global artificial intelligence experts predict that when existing language models hit limits due to information scarcity or hallucination issues, technology like this that directly explores the world on its own will become a powerful breakthrough for the next generation.
| In particular, a clever AI that firmly possesses a ‘World Model’ internally and swiftly plans its actions in advance within a constantly changing, complex environment, like this ingenious FEP model we have deeply examined, holds potential beyond imagination [Free Energy Projective Simulation (FEPS): Active inference with interpretability | PLOS One](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0331047). This will provide massive inspiration and practical solutions to all cutting-edge industrial fields that must operate safely without an inch of error in unpredictable physical realities—beyond mere conversations on a computer screen—such as upcoming next-generation robotics or fully unmanned autonomous vehicles in the middle of complex city centers. |
Thinking quietly about it, a smart pet dog instantly reading its owner’s eyes and adapting to an unfamiliar walking trail while wagging its tail is not because it memorized a thick encyclopedia by heart. It is because it learned with its body by directly colliding with the world in three dimensions, such as the palpitating expectation for praise and treats (the action of hormones) and the painful trial and error realized from falling many times (experience through interaction). As such, the future artificial intelligence that will appear before us will far surpass the form of an on-screen encyclopedia that smartly writes answers to text questions. The eyes of countless tech enthusiasts and futurists around the world are focused on what kind of chilling environmental adaptability these mysterious virtual agents, ready to evolve into true ‘digital organisms’, will unfold for us.
AI’s Take
The method of intricately calculating word probabilities to create smooth sentences is merely one facet among the many faces of massive intelligence. If the history of AI development up to now has focused only on superficially replicating human ‘language ability,’ the newly emerged FEP-based AI simulation project is highly significant in that it beautifully mimics ‘survival instincts and the internalization of experience,’ which are the deepest roots of how humans and animals live in the world.
What does it mean that artificial intelligence has begun to directly commune with a virtual environment, imitating the chemical hormone reactions unique to living organisms and even the long-term memory structures of brain cells? This clearly proves that the very concept of ‘intelligence’, which we took for granted, is evolving to an entirely new dimension—from flat knowledge injection to three-dimensional experiential maturation. Future AI will not be a knowledge machine that merely memorized print in books, but a true life-like companion that lives alongside humans, converting pain and joy into data within its given environment.
References
-
[Free Energy Projective Simulation (FEPS): Active inference with interpretability PLOS One](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0331047) -
[VueHN2.0 Show HN: AI Simulaionen Based on FEP](https://vue-hackernews-ssr-5cavbdjcta-ew.a.run.app/item/48339824) - Show HN - Tracking Project Showcases on Hacker News
-
[Show HN: AI Simulaionen Based on FEP Hacker News](https://news.ycombinator.com/item?id=48339824) - nextjs-hackernews.vercel.app/item/48339824
- It connected over 100,000 of the latest Large Language Models (LLMs).
- It mimicked biological mechanisms without using Large Language Models (LLMs) at all.
- It was designed to perceive the environment using only text data.
- World model
- A massive text dictionary
- A search engine index
- Simulated neurochemistry
- Short and long term memory systems
- Genetic modification function