Act XI: The Presentation

During the third week of October, Guillaume arrived at the office. He wasn't wearing his customary, thoroughly frayed hoodie. Instead, he wore a crisp, collared shirt. That single visual detail signaled to the entire room that today was entirely distinct from their standard routine.

Three individuals followed closely behind him. There was Arthur Mensch, the co-founder and CEO of Mistral, alongside Guillaume Lample, their chief technical officer. The third individual was a face Karpathy had never encountered: a man in his late fifties clad in a tailored charcoal suit, his posture completely rigid and authoritative.

Guillaume handled the introductions. "This is Éric Bernard. He serves as the Deputy Minister for Digital Affairs and AI Infrastructure at the French Ministry of Economy."

Karpathy shook his hand.

"Your reputation precedes you, Monsieur Karpathy," Bernard said. His English was impeccably fluent, though it carried a distinct, formal French cadence. "President Macron personally instructed me to receive a direct briefing on your current technical milestones."

The team had spent three solid days preparing for this specific hour. Rahul had constructed the technical slide deck, Ji-won had rigorously cleaned and aggregated the telemetry data, and Max had configured a live evaluation environment on the cluster. Karpathy had reviewed the entire package and immediately slashed half of the content.

"You cut way too much," Rahul had protested.

"If you present too many variables, you convey nothing," Karpathy explained. "We communicate a single, absolute paradigm shift. Then we let our questions handle the rest."

"And that single shift is?"

"The transformer is entirely bounded by what is present. Liminal captures the latent precursor of what is coming. That is all they need to understand."

Rahul had looked down at the remaining slides. "Are you certain that will be enough to convince them?"

"If they don't comprehend the theory, we let the live runtime show them."

---

The briefing initialized precisely at 2:00 PM. Karpathy wasted zero time on corporate pleasantries or introductory fluff. The deck consisted of exactly twelve slides. The first three explicitly outlined the structural limits of the transformer architecture. The next four mapped the mathematical mechanics of the Anticipation module. The final five were pure empirical telemetry.

"The transformer architecture is designed to drive loss down along a strictly linear trajectory," Karpathy explained, walking up to the whiteboard to sketch a smooth, downward-sloping curve. "It is highly effective, but it inherits an absolute structural ceiling. That ceiling is a direct consequence of its fundamental design."

He then sketched a second curve right beside the first—a trajectory that rose sharply before plunging down in a steep, near-vertical descent.

"During the development of Fable 5, we observed an anomalous structural signature. The loss function would experience a temporary, sharp degradation before executing an aggressive, accelerated convergence."

Arthur Mensch leaned forward intensely. "So you determined it was an invariant property, entirely independent of the underlying weights?"

"Precisely. That discovery served as our absolute starting point here in Paris."

Karpathy advanced to the next slide.

"Neuroscience demonstrates that the human brain consistently constructs readiness potentials, organizing internal states for subsequent cognitive transitions long before those transitions manifest in conscious awareness. We have successfully engineered that exact structural capacity into our model architecture."

Bernard spoke up, his eyes locked on the slide. "And that is what you have designated as the Anticipation mechanism?"

"Yes. It allows the model to maintain a stable internal representation of the latent precursors of relationships before those tokens are formally generated. The temporary spike in the loss curve is the exact window where the model is actively restructuring its internal weights to accommodate that precursor. It appears to be in a state of total confusion, but it is actually executing deep structural preparation."

The conference room fell into a profound stillness.

"We will now initialize a live demonstration."

Max stepped forward and opened his laptop terminal. He established a direct secure connection to the cluster, initializing Liminal’s third-generation model—the architecture that fully coupled the Anticipation modules with their recursive self-improvement verification loops.

"We are going to assign the exact same complex reasoning task to three distinct models simultaneously," Max explained. "OpenAI's GPT-5.6, Mistral Large 3.1, and Liminal’s Generation Three build."

The assignment consisted of constructing a highly complex, multi-tiered mathematical proof. The three models initialized execution concurrently.

GPT-5.6 rendered its output first. The logic was pristine, highly accurate, and incredibly rapid. However, it hit a visible cognitive wall. At the exact juncture where the proof demanded a profound leap in structural abstraction, its output flattened out, delivering a surface-level synthesis.

Mistral Large 3.1 also followed an almost identical trajectory, executing slightly slower than GPT-5.6 and terminating at the exact same abstraction boundary.

Liminal’s Generation Three build was noticeably slower. The latency before it rendered its very first token was significantly higher than either of the competing models. Guillaume Fontaine’s brows knit together in a sharp frown.

Then, the token stream erupted across the screen. Arthur Mensch’s eyes widened. Guillaume Lample leaned in so close his face was practically pressed against the monitor. Bernard shifted his gaze back and forth between the terminal display and Karpathy’s completely unreadable expression.

"This is..." Arthur muttered, his voice dropping.

Liminal’s output had completely bypassed the abstraction wall that had paralyzed both GPT-5.6 and Mistral Large. The mathematical proof continued to unfold, charting a path through an incredibly deep layer of structural abstraction.

"It pushed right past the boundary where GPT-5.6 collapsed," Lample noted, his fingers already flying across his own notebook to track the logic.

"Yes."

"Mechanistically, how is it executing that leap?"

"The Anticipation layer has already evaluated and narrowed down the valid hyper-dimensional pathways for the abstract transition long before it reaches that node," Karpathy explained calmly. "It didn't solve the wall when it got there; it had already structured the solution steps during its preparation phase."

Lample went to speak, but stopped himself entirely. Instead, he began aggressively scribbling equations into his pad.

Bernard spoke up, his tone flat and incredibly serious. "Does this surpass what Fable 5 was capable of?"

The room fell into an even deeper silence.

Karpathy reflected on the question for a moment. "Because we do not possess a live evaluation environment for Fable 5 here, a direct empirical comparison is impossible."

"But your professional assessment?"

"The structural trajectory of this architecture," Karpathy stated clearly, "points directly past the ultimate destination Fable 5 was heading toward."

Bernard remained quiet for a long time, his eyes fixed on the window. Outside, the gray October sky hung heavy over the courtyard, where damp chestnut leaves continued to pile up against the cobblestones.

"What is your immediate roadmap for scaling?" Bernard inquired.

"Our current compute allocation has officially hit its structural limit," Karpathy said. "To initialize our fourth and fifth-generation models, our compute resource requirements must scale by a factor of ten to one hundred."

"A hundred times?"

"This is the fundamental nature of frontier AI development."

Bernard reached into his inner jacket pocket, hesitated for a fraction of a second, and withdrew his hand. "I will brief President Macron personally on these findings tonight," he said, standing up. "I anticipate he will want to meet with you directly in short order."

Karpathy nodded once. "Understood."

"If I may ask one final question," Bernard paused at the door. "Why did you choose France?"

Karpathy looked at him, his expression completely steady. "Paris is exactly one thousand kilometers from Bratislava."

Bernard stared at him for a moment, processing the geometric reference. Then, a quiet, knowing smile touched his face. "An excellent distance, Monsieur Karpathy."

---

Once the Ministry officials and the Mistral team had departed, only the four founders remained in the quiet office.

Max closed out the evaluation environment, letting out a long breath. "That went incredibly well."

"Nothing is finalized yet," Karpathy noted automatically.

"But the technical validation was absolute," Rahul insisted.

"Technical validation doesn't put food on the table."

"Then let’s get some actual food tonight," Max countered. "No more croissants. Let's have a proper dinner."

Ji-won was already buttoning up her coat. "I booked a bistro down the street. A five-minute walk."

Karpathy glanced back at the whiteboard. Their next major question remained written at the bottom: What manifests in Generation Four? He grabbed his coat. "Let’s go."

The bistro was a small, intimate venue filled with weathered wooden tables, red checkered tablecloths, and hand-written wine lists pinned to the walls. Four plates of steak frites were delivered to their table. Rahul poured the red wine. Max raised his glass. "Should we do an official toast?"

"To what?" Karpathy asked.

"To Liminal AI."

Karpathy raised his glass to join theirs. "To Liminal AI." The four glasses clinked sharply.

Ji-won took a sip, then looked across at him. "Andrej, you were actually smiling during the demonstration today. Right when the abstraction layer cleared."

"I was not."

"The exact moment the output bypassed GPT-5.6, the corner of your mouth unmistakably—"

"I was not smiling," Karpathy repeated flatly.

Rahul leaned over to Max, whispering loudly, "He was totally smiling."

Max whispered back, "Oh, absolutely. I saw it clear as day."

"I can hear both of you," Karpathy noted. The three engineers burst into laughter.

Karpathy took a sip of his wine. The French red was noticeably more astringent and heavier than anything he had grown accustomed to drinking in San Francisco. But it was excellent. Outside, the cool October night settled over Paris as a gentle rain began to slick the cobblestones.

---

Washington, D.C. – Evening | National Security Council

Inside a highly secure briefing room at the National Security Council, a classified intelligence memorandum was being passed around the table. The cover page bore a strict classification marker. The title was sparse:

SUBJECT: Liminal AI – Operational Velocity and Strategic Trajectory of Paris-Based AI Laboratory

The final analytical summary at the bottom of the document was aggressively highlighted in red:

The sustained, geometric spike in compute utilization within the Mistral hardware cluster, coupled with recent intelligence indicating an imminent expansion of foreign government infrastructure access, strongly suggests the development of a model paradigm that threatens to eclipse Fable 5 capability thresholds within the current fiscal year.

An official tossed the memorandum onto the center of the table. "Karpathy?"

"Almost certainly."

A heavy, tense silence filled the room.

"Do we escalate this to the President's daily brief?"

"It’s too early for a formal executive intervention."

"If we wait until it's definitive, we’re already behind the curve."

The silence returned. Outside the windows, the Washington night remained perfectly still and dark.