A completely new perspective on the "model-free vs. model-based" paradigm, which has been a central debate in AI! Based on the latest research, this article explains why any artificial intelligence capable of achieving complex goals inevitably must possess its own world model. The discovery and utilization of these "hidden world models," along with their implications for AI safety and interpretability, have now become the most important topics of discussion.

1. Can Artificial Intelligence Really Be "Model-Free"?

For a long time, AI researchers have fiercely debated two approaches:

• Model-based (where the agent explicitly learns an internal simulation of the world)
• Model-free (where intelligence develops through simple trial-and-error and experience alone, without an explicit model of the world)

The model-free path was particularly attractive because accurately modeling the real world seemed impossibly difficult.

"The long-held dream of a 'model-free' path to artificial general intelligence (AGI) may actually have it backwards."

However, a remarkable mathematical proof that overturns this assumption has recently emerged.

2. The New Proof by Richens et al. and Its Implications

In the latest paper "General agents contain world models," the authors formally prove that any AI that accomplishes complex, multi-step (goal depth n) tasks within a certain failure bound (low regret δ) must necessarily contain an accurate predictive model of the real world.

In other words, if an AI is good at making long-term plans over extended periods, then its behavior already contains internal simulation information about how the world works.

"The better an AI's long-term planning capabilities, the more accurate its internal world model must inevitably be."

The interesting point is that such a world model doesn't need to be explicitly created by the programmer — the agent acquires it on its own (as a hidden capability).

• Since "understanding the world" cannot be avoided, any agent that succeeds at long-term, complex tasks must naturally internalize the principles of how the world works.

3. The Method of Proof — "Interrogating" the AI

The research team devised a very simple and intuitive AI "interrogation" algorithm. This algorithm presents the agent with complex goals requiring it to choose between two options. They then demonstrated that the reasons behind those choices — the probabilistic judgments (implicit predictions) embedded within the agent — can be "reverse-engineered" through observing its behavior.

"Simply by observing an agent's policy (policy = behavioral rules), you can reverse-extract the hidden 'blueprint' of the world contained within it."

Ultimately, no matter how much we treat AI as a "black box," we now know that deep inside there must be an implicit model of the world.

4. What This Research Has Changed: Safety, Interpretability, and Unification

The greatest practical implications of this proof are as follows:

1. Safety Since we can extract a competent agent's world model just by analyzing its policy, we can examine the "inner map" of black-box AI to preemptively block risks.

2. Interpretability If world models are inevitably embedded, interpretability becomes a necessity, not an option. This means we can directly look into what kind of world the AI imagines and predicts.

   "The hidden world map may be the very key to both progress and safety."

3. A Paradigm Shift in AI Research The distinction between "model-free vs. model-based" becomes meaningless, and the unified principle that "world models inevitably emerge no matter what approach is used" takes center stage.

   In fact, the "emergent capability" phenomenon recently observed in large language models (LLMs) may also be an inevitable consequence of these implicit world models.

4. The Birth of New Questions

   • What kind of world models are hidden inside today's massive AIs?
   • How accurate are they?
   • Can we effectively extract and interpret these models to prevent harmful behavior in advance?

5. Final Implications — Intelligence Is Ultimately the "Model Itself" of the World

Ultimately, this paper formalizes the essence of longstanding artificial intelligence philosophy into mathematics.

"An intelligent being (agent) doesn't so much 'have' a model of the world — it essentially is the model itself."

In other words, this is not simply a matter of design choice but an intrinsic property of intelligence.

Conclusion

This research establishes a clear compass for how we should approach AI development, interpretation, and safety going forward. We must not forget that "world models are inevitably embedded" is the key to deeply understanding and working with today's artificial intelligence. The most important challenge in AI research going forward will be how well we can find these hidden maps and use them correctly.