FSBio Outlines Metabolic AI Runtime, an AI Architecture Built on 'Homeostasis' Instead of Empathetic Prompts

FSBio has proposed an alternative to conventional LLMs. Instead of adding more parameters and a bigger context window, the company describes Metabolic AI Runtime, where each response emerges from the system's internal state. In the authors' view, machine empathy won't arrive with yet another dataset of polite dialogues. It appears only when an AI has its own balance, one that a user's problem can disturb.

Why Transformers aren't enough

FSBio's paper challenges the industry's core bet on scaling. The authors argue that no matter how many dialogues from Reddit, Twitter, or forums a model is fed, a Transformer remains a next-token prediction machine. It imitates sympathy well, but it doesn't experience it: it has no state of its own that can be shifted, disrupted, or temporarily tuned to someone else's situation. That's why phrases like "I'm sorry" sound more convincing but never turn into real understanding of the user.

This leads to the paper's central claim: longer context windows, more parameters, and new H100 clusters don't solve the problem of subjecthood. According to FSBio, when a user complains about fatigue, a typical LLM simply recognizes the pattern and picks a statistically appropriate reply. Nothing changes inside the model. The authors call this a dead end for AI companions, therapeutic assistants, and services where factual accuracy isn't the only thing that matters, and users also need to feel that the system actually picked up on their state.

What FSBio proposes

Instead of a stateless model, the company describes an architecture with internal homeostasis. It builds on modified Reservoir Computing principles and continuous dynamic circuits, which the paper compares to neurochemistry: oxytocin, cortisol, and adrenaline. When a user arrives with a problem, the system doesn't look up a ready-made reply from a set of templates. Instead, it runs the context through its own "metabolism," shifting its internal balance toward the request. In this design, the internal shift itself is what counts: without it, the authors argue, empathy remains nothing more than skillful stylization.

The authors identify several core components of the system:

• continuous dynamic circuits instead of one-shot stateless inference
• artificial homeostasis that can shift in response to a user's problem
• Liquid Intuition, a mechanism that retrieves knowledge through the system's current state
• a gradient of will and vector drives that set attention priorities
• the option not to sustain empty conversation when the system's internal resources run low

The idea is that a response should emerge as an attempt to bring both sides back into balance: the user and the system itself. That's a notable departure from conventional prompting, where a model is simply told to be helpful and empathetic. FSBio argues that empathy can't be a directive in a system message. It has to emerge from an architecture that has its own state and, therefore, pays a price for mistakes or indifference.

Memory and will

The authors also take aim at classic RAG. In their view, cosine similarity search works for retrieving facts but is poorly suited to understanding a person's state. A system that only searches for documents whose wording resembles the query will find relevant text, but not necessarily the piece of knowledge needed in the current emotional phase of the conversation. That's why Metabolic AI proposes "fluid" memory: the right memories should activate under the influence of internal tension, not a dry mathematical query.

This gives rise to the concept of a "will to empathize." The paper says the system can have drives that strengthen or weaken its willingness to invest compute in a particular conversation. Such an AI isn't obligated to respond the same way to every input. It can ignore idle chatter but sharpen its focus when a user brings a problem that disrupts the model's internal equilibrium. The authors contrast this with today's assistants, where empathy is more often set by instruction than by internal motivation.

"Machine empathy is physics, not linguistics."

What it means

FSBio's paper isn't a mass-market product announcement. It's an architectural manifesto against the current race for ever-larger LLMs. If the approach proves viable, the market could shift from models that merely sound more human to systems with persistent internal state, dynamic memory, and selective engagement. For now, it's more a strong research hypothesis than a new industry standard, but it clearly hits a weak spot in today's AI assistants: imitating empathy still isn't the same as empathy.