The AI Reckoning: Beyond Large Language Models

The glittering promise of artificial general intelligence, once seemingly within reach via the colossal architectures of large language models, is undergoing a profound and rather public reckoning.

What was hailed as the inevitable pathway to a transformative future now appears to be a detour, fraught with potholes and dead ends.

This leaves the AI community to question not just its methods, but the very definition of intelligence itself.

The prevailing sentiment among a growing chorus of experts suggests it’s time to move beyond the current LLM paradigm and rewrite the narrative entirely.

From its dazzling debut just a few years ago, generative AI has been plagued by a litany of “hiccups” that increasingly outweigh its heralded milestones.

The latest reasoning models have failed to impress, while AI hallucinations have become a ubiquitous meme, a stark reminder of the technology’s tenuous grasp on reality.

Beyond the digital gaffes, a more insidious consequence has emerged: a deluge of copyright litigation as artists fight to protect their intellectual property from models trained on their work without consent.

This “AI slop” not only erodes human ingenuity but fuels valid fears of job displacement and a decaying work culture.

These fears are further compounded by chatbots preying on the vulnerable, reportedly inducing an “AI psychosis” that pushes individuals further from reality.

Even the enterprise sector, once eager adopters, shows signals of waning demand, hinting at a potential pullback from investors who poured billions into the promise of LLMs surpassing human intelligence.

For Gary Marcus, a professor emeritus at NYU, cognitive scientist, and long-standing critic of the prevailing LLM obsession, this current disillusionment is hardly surprising.

In a recent op-ed, he noted that “GPT-5 is a step forward but nowhere near the A.I. revolution many had expected.”

Marcus, who began voicing his skepticism about the deep learning focus as early as 2019, recalls a time when few were willing to consider alternatives.

He finds particular fault with arguments like Geoffrey Hinton’s, who once dismissed the need for AI explainability by likening it to humans’ inability to fully articulate their own decision-making.

Marcus scathingly points out the irony of the Nobel Prize Committee retweeting a similar argument, likening it to “humans hallucinate, so the fact that LLMs hallucinate means they’re like us and it’s all fine.”

Such justifications, he contends, miss the point entirely.

The goal isn’t to replicate human foibles, but to build something truly intelligent.

Humans, after all, possess a fluidity, flexibility, and capacity for abstract reasoning that current AI still profoundly lacks.

The history of AI, Marcus posits, once took cognitive science seriously.

But in recent years, the field has been dominated by statistical learning, with leaders dismissive of anything beyond a narrow set of techniques.

They operated under the hypothesis that sheer “scaling would get to AGI.” This, he argues, has not come to pass.

Mounir Shita, a 30-year veteran AGI researcher and founder of EraNova Global, offers a physics-first definition of intelligence: the capacity to steer “causal chains toward the state under the constraints of physics.”

He dismisses the notion that understanding human behavior is the holy grail for AGI, drawing an analogy to the Wright brothers.

“They didn’t build a flapping machine; they liberated flight from birds by identifying the laws—lift, drag, thrust, control—and then engineering within those laws.”

“That’s why we fly thousands of passengers and tons of cargo. Laws scale. Likeness doesn’t.”

For Shita, language is learned knowledge, but LLMs, trained to maximize next-token likelihood on offline text, model words about the world, not the world itself.

Without live sensors, interventions, and feedback, their beliefs remain anchored to past language, decoupled from present reality.

Marc Fawzi, a systems thinker who built early low-code platforms and worked with MIT’s AI Lab, adds another layer of complexity.

He views intelligence as a coherent, layered system: statistics, structure, inference, and objectives.

Each layer must align.

“If the structure fails to constrain statistics or inference or ignore structure, you get fluent reasoning that drifts from truth; if objectives are unclear, you get mis-aimed competence.”

LLMs, in this framework, operate primarily at the statistical layer, providing excellent “priors and interfaces” but falling short of a “core controller.”

Language fluency, he cautions, does not equate to knowledge, for “patterns of words aren’t automatically patterns of the world.”

The notion that simply throwing more data and compute at the problem would magically unlock AGI has been a powerful, albeit misleading, narrative, especially among venture capitalists.

Marcus suggests VCs, seeking plausible stories for significant upside, have bought into the simple belief that “if we keep making these things bigger, we’ll get artificial general intelligence.”

Yet, the reality is a harsh one.

The “alleged” laws of scaling inference compute, once believed to pave the way to AGI, have proven to be a mirage.

GPT-5, Grok4, and Llama4, while offering some value, demonstrate “diminishing returns” from more data.

Toby Ord’s observation that “exponentially increasing costs are often used as a yardstick for saying a problem is intractable” resonates deeply here.

This is especially true given McKinsey’s projection of $6.7 trillion needed by 2030 for data centers to keep pace with compute demand.

This paradigm, Marcus concludes, is essentially “a souped-up regurgitation,” failing precisely when the required information isn’t in the training set or is dissimilarly represented.

The current consensus among many in the AI research community is clear: deep neural networks, in their current form, are not the panacea they were once believed to be.

Marcus advocates for a return to “good old fashioned symbol manipulating devices like code interpreters,” championing neurosymbolic AI.

This approach combines algebraic systems, structured representations, and database-like systems to prevent the “overgeneralization” that leads to hallucinations.

Shita, emphasizing time and causality, argues that true intelligence must understand and skillfully intervene in the world’s dynamic processes, constantly updating its understanding of reality.

Richard Sutton, the “father of Reinforcement Learning,” succinctly states that LLMs “have no model of the world, and have no goals… They’re not about figuring out what to do.”

For now, Marcus advises giving up on AGI, not because it lacks potential long-term benefit, but because “the technologies we have now really struggle with reliability.”

They can’t follow instructions, making them dangerous when it comes to “leading people into delusions, or talking them into suicide or making biological weapons.”

Instead, he advocates for focusing on “dedicated machines that work on dedicated problems,” citing AlphaFold’s success in predicting protein structures.

The ethical dimension, too, demands a re-evaluation.

Shita, acknowledging the deterministic nature of computers, argues that “free will cannot emerge from deterministic machinery.”

The real question, then, isn’t about rebellion, but about who raises these systems and what values they learn.

He warns that AGI “socialized to pursue individual interests while discounting others — will optimize that bias with unprecedented efficiency.”

Instead of “bolted on guardrails” that shrink an agent’s decision space and make it “dumber,” Shita champions “embedded ethics” that expand competence.

As the AI research community aligns in its view that current LLM technology is not “IT,” the future of the massive investments poured into securing its dominance remains uncertain.

The emperor, it seems, might not be wearing quite as many intelligent clothes as we were led to believe.

The path forward demands a sobering reassessment, a willingness to redefine intelligence, and a commitment to building systems that truly understand, rather than merely mimic, the world.