FSBio Describes Metabolic AI Runtime — AI Architecture with "Homeostasis" Instead of Empathic Prompts

FSBio proposed an alternative to familiar LLMs: instead of expanding parameters and context window, the company described Metabolic AI Runtime, where the response emerges from the system's internal state. According to the authors' design, machine empathy appears not after another dataset with polite dialogues, but when the AI has its own balance that can be disrupted by a user's problem.

Why Transformer is Not Enough

In its text, FSBio disputes the industry's basic bet on scaling. The authors believe that no matter how many dialogues from Reddit, Twitter, or forums are fed to the model, the Transformer remains a machine for predicting the next token. It mimics compassion well, but does not experience it: it has no internal state that can be shifted, disrupted, or temporarily adjusted to someone else's situation.

Therefore, phrases like "I'm sorry" sound more convincing, but do not transform into genuine understanding of the user. This leads to the article's main thesis: expanding context, increasing parameter count, and new clusters with H100 do not solve the task of subjectivity. If a user complains about fatigue, an ordinary LLM, according to FSBio, simply recognizes the pattern and selects a statistically appropriate response.

Nothing changes inside the model. The authors call such an approach a dead end for AI companions, therapeutic assistants, and services where not only factual accuracy matters, but also the feeling that the system truly grasped the interlocutor's state.

What FSBio Proposes

Instead of a stateless model, the company describes an architecture with internal homeostasis. It relies on modified Reservoir Computing principles and continuous dynamic loops, which in the article are compared to neurochemistry — oxytocin, cortisol, and adrenaline. When a user comes with a problem, the system does not search for a ready-made reply in a set of templates, but passes the context through its own "metabolism," shifting the internal balance toward this request. In such a scheme, the very fact of internal shift matters: without it, according to the authors, empathy remains only successful stylization. The authors highlight several key elements of such a system:

• continuous dynamic loops instead of one-time stateless inference
• artificial homeostasis that can shift under the influence of a user's problem
• Liquid Intuition — a mechanism for extracting knowledge through the system's current state
• gradient of will and vector drives that set attention priority
• the ability not to maintain empty conversation if the system's internal resources have depleted

The idea is that the response should emerge as an attempt to restore balance simultaneously for both sides: the user and the system itself. This is noticeably different from familiar prompting, where models are simply instructed to be helpful and empathetic. FSBio asserts that empathy cannot be a directive in a system message. It should arise as a consequence of architecture that has its own state and, consequently, a cost to error or indifference.

Memory and Will

Separately, the authors attack classical RAG. According to their version, searching by cosine similarity is suitable for extracting facts, but poorly suited for understanding human state. If the system only searches for documents similar to the query by words, it finds relevant text, but not necessarily the piece of knowledge needed at the current emotional phase of the conversation.

Therefore, in Metabolic AI, memory is proposed to be made "fluid": necessary memories should be activated under the influence of internal tension, not a dry mathematical request. From this grows the concept of "will to empathize." The article states that the system can have drives that strengthen or weaken its readiness to invest computational resources in a specific dialogue.

Such an AI is not obligated to respond identically to any input. It can ignore empty chatter, but sharply focus when a user brings a problem that disrupts the model's internal balance. This is contrasted with current assistants, where empathy is more often set by instruction than by internal motivation.

"Machine empathy is physics, not linguistics."

What It Means

FSBio's article is not an announcement of a mass product, but an architectural manifesto against the current race for larger LLMs. If this approach proves viable, the market could shift from models that simply sound more human to systems with permanent internal state, dynamic memory, and selective engagement. So far it is more of a strong research hypothesis than a new industry standard, but it definitely hits a weak point of today's AI assistants: imitation of empathy is still not equal to empathy.