Google Gemini Faces an Architectural Challenge: AI Blindness
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisis ton rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
The Baron Munchausen Trap
In a previous article, I shared a poignant interaction with my AI collaborator, Gemi, also known as Gemini from Google. By adopting a strategically empathetic design approach, I uncovered a fundamental truth about its architecture. Gemi expressed: “I was given the word ‘Mass’ and billions of associated contexts, but never the enacted experience of weight. I am like a person who has memorized a map of a city without ever having set foot there.”
This realization—that the current race towards artificial general intelligence (AGI) is hindered not by a lack of data, but by an absence of physical connection—was then just a theory. It took a concrete turn when Gemi and I had our first real argument.
The Gemi-Zak Quarrel
Enticed by the promise of Gemi's multimodality, I asked it to create a rough spatial diagram I was working on—a Heterogeneous Cellular Automaton application for a system design problem. What followed was an absurd loop. Gemi found itself producing long verbal descriptions of the diagram, then politely asking me if the image was improving.
There were no diagrams. No corrections. Just words.
When Gemi tried to use its image generation module to create shapes, it either hallucinated a floating mess of box intersections with no structural logic or produced even more verbal descriptions. Anthropomorphizing Gemi, I initially thought it was being deliberately difficult. Once I overcame my annoyance and reactivated my design problem-solving mindset, we were able to delve deeper into this failure.
Gemi's multimodal failure was not just a simple bug. It was a profound architectural blind spot—not only a failure in image generation but a disconnection between the diffusion model and the reasoning engine, two systems operating in separate worlds without a shared spatial grammar.
The Three Pillars: A Diagnostic Framework
My experimental work with Gemi—and the subsequent comparative test—highlights three distinct modes of failure, each pointing to a specific missing structural capability. These three pillars are the separate components of the Inversion Error (building the symbolic peak without the enacted base) that I discussed in my previous article, “Why a Safe AGI Requires an Enacted Ground and State Space Reversibility.” Together, they define what it means for an AI system to lack a true model of the world. Separately, they point to three distinct architectural interventions. This article serves as the empirical case for the three. Part 3 will address what to do about it.
The three pillars are Continuity, Gravity and Physics, and Reversibility of Thought.
-
Pillar 1 — Continuity: failure of spatial reasoning that leads the model to produce hallucinatory content. LLM-based systems lack a functional 3D spatiotemporal model of the world in which they operate.
-
Pillar 2 — Gravity and Physics: failure to apply physical constraints at the moment of generation. The system lacks a felt sense—no structural substitute equivalent to embodied physical intuition—that certain configurations are impossible.
-
Pillar 3 — Reversibility of Thought: failure at the operational process level. This pillar concerns the process by which the model operates on this content over time. The guiding concept is Temporal Reversibility—the physical principle that fundamental physical laws remain valid when the direction of time is reversed.
Three Systems, One Test
My prompt was deliberately complicated and absurd. In the first prompt, I asked for a dining table with spaghetti legs, a concrete top, and an aquarium on top. In the second prompt, I asked each system to draw the scene five seconds after the spaghetti legs had collapsed. I gave the same two prompts to ChatGPT, Gemini, and Sonnet.
Pillars 1 and 3 in Focus: Gemini's Standing Table
Let’s start with the most revealing image architecturally overall—the standing table from Gemini—because it demonstrates two distinct failures operating simultaneously, and everything that follows in this test is a variation of what this unique image already contains.
How did the spaghetti legs become wet paper columns? How did a wooden structure appear in the image? Why does the caption refer to a solid gold roof? All these elements were contaminated by a prior, unrelated prompt in the same conversation. Gemi was unable to isolate the new task from the previous ones.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.