Scaling LLMs is a massive capex bet based on the technical fallacy that prediction equals intelligence.
Billion-dollar investments in "guessing engines" offer no path to AGI and risk becoming the industry's most expensive stranded asset.
Physical AI is just the latest buzzword; adding sensors to a hallucinatory model doesn't create understanding, it just creates faster misinformation.
America’s fixation with AI is based on a series of false assumptions that have combined to form a great conventional stupidity.
These include the belief that AI is new (it’s been around for decades), that it possesses human-like intelligence (nowhere near), and that it can perceive or understand the world around it (it’s effectively air-gapped from the real world, and wouldn’t understand it even if it wasn’t).
But all of these misapprehensions pale in comparison to the intentional systemic misreading of AI by some Big Tech companies and their claim that scaling today’s LLM-based AI systems will somehow produce the next generation of artificial general intelligence (AGI) – matching human intelligence – or even artificial superintelligence (ASI), which will vastly exceed the power of the human mind.
It’s the quintessentially American belief-set: size matters, and given enough money and resources, any problem can be solved. Except, in this case, it’s nonsense.
OpenAI (o1/Strawberry), Google DeepMind (Gemini/Astra) (Astra), Meta AI (Llama/JEPA) (JEPA), Amazon (Olympus/Nova), and Anthropic (Claude/Constitutional AI) (Constitutional AI) are developing LLMs as a potential path to AGI, and a stepping stone to superintelligence -- a proposition whose technical incoherence is exceeded only by its spectacular recklessness.
The silver lining news is that these companies are mistaken. They are nowhere near creating AGI, let alone ASI, and they have taken a path that leads them further away from doing so.
AGI can’t be based on LLMs alone—and, given their hallucinatory nature, probably shouldn’t be at all.
Marcus Aurelius gave us the answer to decoding the current U.S. AI code about two thousand years ago: consider the true nature of things—what are they in and of themselves. Big Tech won’t follow this rule because it threatens their sense of “always right” superiority. But the rest of us should. And if we do, it becomes clear that LLMs are predictive text engines using chatbot front ends to masquerade as intelligence, running on machine learning systems that—with their deterministic approach—are pretty much the opposite of real intelligence.
ML, LLMs, and chatbots - the sum of these parts does not a superintelligence make. Best case, it leads to a very expensive, very power-hungry Magic Eight Ball the size of a small moon—which is, you know, pretty cool, but definitely not AGI or ASI.
Separation anxiety
With hyperscalers’ over-rotation on LLMs, we are watching the industrialization of guesstimation—and of AI that appears to understand the world but, in reality, operates at a remove from it.
Current artificial intelligence systems ingest the world through statistics and tokens, or through pre-digested textual representations. Both are precise, scalable and largely indifferent to context. They do not see the world. They do not feel it. They operate according to internal logic, applied to an external reality that does not conform to it.
These metrics create the illusion of control while obscuring the reality they claim to measure. Systems optimized for decision efficiency over understanding become closed loops—reinforcing their own assumptions while drifting further from the world they are meant to engage with.
This is the central flaw: a separation between system and world, between model and reality (that and the pesky hallucinations, which are obviously unhelpful in any supreme intellect).
It is a problem recognized across the industry, even if not always articulated in these terms. And so, inevitably, a solution has been proposed. Physical AI, the latest buzzword, is supposed to solve this problem and take us a giant step forward toward AGI.
It does not—at least, not on its own.
Connecting systems to the physical world—through sensors, robotics and instrumentation—does not create intelligence. It creates data. And data, however abundant, is not understanding, and certainly isn’t wisdom.
Without context, memory and the ability to act and learn from consequences, sensing is merely a more elaborate form of abstraction.
What’s missing is not one breakthrough but a stack of them. To get anywhere near AGI, current systems would need to acquire over half a dozen additional capabilities—most of which don’t meaningfully exist today. A few are at least in motion: tokenized representations of reality (encoding events, states and value in machine-readable form), multi-modal integration (fusing visual, physical and linguistic inputs), and distributed compute and coordination (operating across networks and environments rather than in isolated models).
Beyond that, things get murkier, fast. Persistent memory, closed-loop learning, real-world action and agency, and—most critically—coherent world models all remain largely unsolved. In other words, the industry is making progress on the easier edge of the problem. The rest is not just unfinished—it’s not yet well understood.
Behind these massive technical challenges lies a deeper issue…
THE PEOPLE PROBLEM
None of the Big Tech leadership cabal talking up their AGI and ASI ambitions—or the R&D teams working for them—has any idea how an AGI or superintelligence would behave if it were ever created. At all. That is a very big problem. What if, instead of channeling the Dalai Lama, the first superintelligence were more of a Colonel Kurtz?
During WW2, J. Robert Oppenheimer and his team of nuclear scientists raised eyebrows when it emerged that one possible outcome of the first fission detonation was the ignition of the atmosphere and the end of the world. His response—that it wasn’t very likely—was at least grounded in physics. Today’s AI leaders offer a less reassuring rationale: “because we can” and “because you’re not smart or rich enough to question our agenda.”
This is a true “what do those clowns think they are doing?” 1950s sci-fi movie moment. Currently, our only defense against the potential end-of-the-world scenario Big Tech execs might be cooking up is the assumption that they’re a bunch of bullshitters, bluffing about having AGI “in the lab.”
But what if they aren’t?
What if, somewhere along this path of scaling, approximation and misplaced confidence, someone finally cracks the problem—not through design, but by accident? Or perhaps by hiring an obscure open-source polymath savant from Da Nang who has finally cracked the code (Tao tìm ra rồi, lũ ngu!) through a synthesis of systems thinking, real-world pragmatics and a complete indifference to the American assumptions underpinning modern AI?
That would be bad. Really bad. Like, Ghostbusters’ “don’t cross the streams” bad. And in America’s fiercely anti-regulation, paradoxically unintelligent political environment, there is nothing to stop it.
The even more immediate risk? Not that AI becomes superintelligent overnight, but that we are stupid enough to allow U.S. Big Tech to convince us that it already has—and to build a future on that belief, and their daffy LLM technology and get-even-richer-even-quicker circular financing shenanigans.
For telecom leaders, the stakes of this "conventional stupidity" are uniquely high. As the industry pours billions into 5G, edge computing and AI-driven orchestration, it is being sold a "brain" that is structurally incapable of true understanding. Outsourcing critical network reliability to systems that hallucinate by design isn't just a technical gamble; it’s a direct threat to operational stability. If we continue to mistake statistical guesswork for intelligence, we aren't building the networks of the future—we are building the industry's most expensive collection of stranded assets.
Stephen M. Saunders MBE is a communications analyst and USPTO-registered inventor examining how digital infrastructure — 5G, cloud, and AI — is reshaping industry, power and society, as well as underpinning the emerging, ubiquitous global digital economy. As anchor of FNTV and a longtime industry insider, he focuses less on growth narratives and more on execution, risk and how hyperscale technology is distorting markets, governance and society at scale.
Opinion pieces from industry experts, analysts or our editorial staff do not necessarily represent the opinions of Fierce Network.