Examining the Technical Critique That May Pop the AI Bubble

Amber Sinha is a contributing editor at Tech Policy Press.

A few weeks ago, Nvidia became the first publicly traded company ever to hit the $5 trillion figure in market capitalization. Not only is this feat remarkable, what makes it staggering is that the company crossed the $1 trillion only some eight quarters ago. There is a single market phenomenon that has driven this astounding increase in market value — the AI boom.

The rise in Nvidia’s fortunes is only a small part of this story, which extends to the entire US technology sector and large parts of the US economy over the last two years. The Swiss bank UBS estimated global spending on AI worth $375 billion and $500 billion globally in 2025 and 2026 respectively. Harvard economist Jason Furnam reported that investments in data centers, the infrastructure behind AI, was responsible for 92% of GDP growth in the US in the first half of this 2025. Aside from direct spending on building AI, this wave of investment supports other sectors of the economy driven by demand for AI infrastructure, including semiconductor factories and investments in energy supply. It seems clear that the US economy is, for the moment, largely propped by AI.

As is often the case, the direct and indirect impacts of jitters in US tech equities are already being felt in other markets — stock markets in Asia and Europe fell sharply in early November in response to warnings from big financial firms that the valuations of technology companies are highly inflated, and that a correction is likely in the near future.

Over the last few months, conversation about the AI boom being a bubble and its potential fallouts have begun to occupy more space in the discourse. Two recent events made markets and investors even more nervous. Michael Burry, the well-known short-seller famous for forecasting the 2008 housing crash, placed a bet against two leading AI firms, Palantir and Nvidia. Reports about the circular deals struck by Open AI with several partners like Microsoft, CoreWeave, SoftBank, Oracle and Nvidia suggested that the risks arising from an AI bubble could be further amplified.

Financial bubbles are notoriously difficult to predict, and it would be a fool’s errand to definitively speculate about when the market boom might end. However, we can look at the broad assumptions that underlie the astronomical investments in AI and the outcomes promised to entice them, and critically analyze the state of technology. The boom requires billions to acquire and process the necessary training data across large computing infrastructures. Furthermore, simply using the models is itself expensive. Given how high both costs and expectations are, it is unlikely that paying customers can generate enough revenue to support the full costs of development and operation on the current trajectory.

So far the response from tech companies has been to double down — to add more layers of networks and more data to correct flaws in AI, leading to only more spending and more demands for resources. As The Atlantic's Matteo Wong points out, “in this tautology, the act of spending is proof that the spending is justified.” It is worth noting that even in the pitch for AI, as framed by the tech companies, it is the promise of achieving Artificial General Intelligence (AGI) that serves as the implicit or explicit justification for their business models. Ergo, if AGI is not achieved in 2026 or 2030 or some other comparable timeline, the promises to justify the financial investments will not be met.

In this piece, I will provide a short technical history of AI, explain how the current versions of AI work, and opine on how close (or distant) they are from achieving AGI, based on what technical experts say.

Sub-symbolic AI

When we look at current applications of AI that are driving market speculation, we are generally referring to a sub-discipline of AI called sub-symbolic AI, and even more specifically, its subset called neural networks.

In the early years of AI research and development, the focus was on a different approach towards building AI called symbolic AI. Symbolic AI, dominant in the early decades of AI, works by defining a current state and a desired state, then using a set of encoded rules to transition between them. This approach is based on the idea that intelligence can be entirely captured by a program processing symbols (representing real-world concepts) and implementing abstracted, logical decision-making rules. Its most common application was in expert systems, which converted specialized human knowledge into computationally processed rules.

Conversely, sub-symbolic AI draws on a rudimentary understanding of neuroscience to model the unconscious, fast thought processes (like recognizing faces) that cognitive psychologist Daniel Kahneman called "System 1" thinking, which are not discursive or interpretive. This approach resembles the famous psychologist Donald Hebb's witticism that "Neurons that fire together, wire together" — meaning repeated co-firing strengthens connections, forming "cell assemblies" that reflect learning by association. The technical groundwork for this sub-symbolic approach traces back to the 1940s, when neuropsychologist Warren McCulloch and logician Walter Pitts modeled neurons as simple logical circuits to create a flexible mathematical model for neural networks.

Frank Rosenblatt's development of the Perceptron in the late 1950s advanced the concept of neural networks by modeling how a biological neuron functions. In this model, connections between inputs have varying weights (strengths), and the perceptron makes a binary decision (1 or 0) only if the cumulative sum of these weighted inputs exceeds a set threshold, similar to a neuron firing when input electricity is sufficient. Unlike symbolic AI, the perceptron makes decisions without relying on encoded rules, allowing it to perform perceptual tasks like image recognition if its weights are correctly tuned. The crucial step of adjusting these weights to achieve accuracy is called supervised learning: the system is trained by feeding it a labeled dataset (e.g., images correctly identified as "cat" or "not-cat"), enabling it to learn and solidify the correct internal connection strengths.

How do neural networks work?

A multi-layer neural network is, in simple terms, one that has several layers between the input and output layers, often called hidden layers. This is accomplished through a learning algorithm called back propagation. What the back-propagation approach does is to identify an error at the output stage and ‘propagate’ the blame for the error back to the hidden layers. This assignment of blame works by adjusting the weightage in the network to reduce the error. Rinse and repeat until the output error gets as close to zero as possible.

If one has to stretch the brain-AI parallel further, computers execute tasks incrementally, handling operations like addition or toggling switches one by one, requiring numerous steps to achieve significant outcomes. What sets computers apart is the speed with which transistors can switch on and off, operating at speeds of billions of times per second. But the brain excels at parallel processing, using billions of neurons to conduct many computations simultaneously, though each neuron fires much slower, at a maximum of a thousand times per second. Although computers are approaching the human brain's neuron count in terms of transistors, the brain still has vastly more connections. The efficiency of the brain is remarkable; for instance, recognizing a familiar face takes only about a tenth of a second, allowing for only around a hundred processing steps. Within these few steps, the brain swiftly searches memory, finds a match, and adjusts it to the new context (like lighting or clothing), suggesting that each processing step involves the complex integration of vast information, a concept known as distributed representation.

Proponents of symbolic AI would attempt facial recognition by pre-programming the computer with explicit rules to identify specific features (like eyes, noses, or whiskers) in an image. In contrast, sub-symbolic AI (building on the perceptron) approaches the task by using unsupervised learning: it is fed a vast number of digitized images (both cats and non-cats) and asked to compress the data. This involves identifying common features, such as uniform color or changes in brightness, across many image patches. The system then abstracts these features, repeating the process multiple times to find common features of features, a process called feature extraction. If the training data includes many cats, the system is naturally oriented towards recognizing cat-like characteristics at a higher level, which are determined by the images themselves, not by the programmer's predefined rules. The triumph of sub-symbolic AI is due to three factors: vast volumes of data available for learning (online, in dedicated databases, or from sensors), advanced computational methods for managing this data, and exceptionally fast computing systems.

What does it mean to achieve AGI

Many of the examples of big recent successes in AI come from the domain of Natural Language Processing (NLP), or at least have involved some element of it. NLP has long been a critical goal of AI. The question-answering format has been at the center of this research, as people want to be able to use human language to interact with machines, rather than code. This represents a huge leap in technological progress where not needing specialized computation knowledge lowers the threshold for using an AI service. True NLP would require not only a comprehension of the human language, but also a comprehension of the world it is used to navigate, akin to the ‘world-model’ described below. The first few decades of NLP witnessed symbolic rule-based programming, where computer systems were fed grammatical and linguistic rules. Much like other domains of AI, this approach was ultimately deemed deficient.

We are now in the throes of sub-symbolic AI. Rather than embedding explicit knowledge, we enable machines to 'comprehend' relationships in language by exposing them to massive datasets of text and teaching them to anticipate words. The outcome is what researchers refer to as a language model. Computer scientist Margaret Mitchell analyzes the example of speech recognition systems. Earlier, I discussed deep learning in the context of image recognition. Around the same time the work on this problem was being conducted, in 2012, pioneering research was carried out in the field of speech recognition in a project led by Geoffrey Hinton. In the same year, a new deep-network speech-recognition system was released on Android phones. Mitchell explains what a stunning improvement this was, going from “horribly frustrating to moderately useful,” to suddenly “very nearly perfect in some circumstances.”

Yet, she notes that, importantly, even the best speech recognition continues to be confused or misled by unfamiliar words or phrases, thus underscoring their limited comprehension of the speech they are transcribing. Would adding more layers of networks and more data correct these flaws, or are these flaws a clear and unremovable symptom of a programming system which accomplishes tasks without understanding what they are?

The holy grail of a ‘world model’

Increasingly, one of the standards for ‘intelligence’ against which emerging LLMs are being evaluated is whether they have what is colloquially called a ‘world model’ — “compressed, simulatable models of how aspects of the world work, ones that capture causal structure and can yield predictions.” In June 2024, a group that was part of a training program led by DeepMind's Neel Nanda published “OthelloGPT Learned a Bag of Heuristics.” It assessed Othello-GPT, a transformer trained to play and predict legal moves in the game Othello. This work represents another step towards resolving a larger debate about the nature of 'intelligence' possessed by large language models (LLMs) such as ChatGPT, Gemini, Bard and Llama. It disputes the findings of a 2023 paper by Kenneth Li, perhaps the most widely cited evidence for LLMs developing something like a human's mental model of the world, an internal 'world model.'

One cautionary response to the hype around LLMs is that they are 'stochastic parrots' capable of only learning superficial statistics about syntax, not meaning or semantics. What this implies, in very general terms, is that they merely predict the relationships between material in their training data while formulating responses to prompts to give an illusion of intelligence, without understanding the concepts that these responses describe. Li's paper sought to defy these doubts, by representing the internal decision-making of Othello-GPT. An LLM demonstrating ‘world model’ capabilities suits the purposes of an industry that dangles AI’s promise to achieve human-level intelligence as a carrot to aggressively fundraise. However, Karvonen et al provisionally show that LLMs likely remain a 'bag of heuristics,' based on large sets of statistical associations incapable of capturing causal and abstract knowledge.

Skeptics like Mitchell argue that most examples of successful AI employ ‘surface heuristics’ calling upon shortcuts very specific to the training data and not any abstract, causal understanding the training dataset was meant to teach. Meta’s Chief Scientist (for now), Yann LeCun, goes so far as to say that a “system trained on language alone will never approximate human intelligence,” as “LLMs have no stable body or abiding world to be sentient of—so their knowledge begins and ends with more words and their common-sense is always skin-deep.”

In the last 18 months, Large Reasoning Models (LRMs) have been portrayed as closer to have achieved close to human-level thinking. These models, trained to perform specific kinds of reasoning, have emerged from all major companies — OpenAI’s o1 and o3 models, Anthropic’s “extended thinking mode,” Google’s Gemini 2.0 Flash Thinking Experimental, and DeepSeek’s R1 model. They create a 'chain of thought' by writing out a step-by-step, natural-language description of a logical reasoning process. However, a paper published by Apple’s researcher earlier this year tested the LLM and LRM versions of DeepSeek, R1, OpenAI’s o3 and Claude 3.7 Sonnet, and found they perform very poorly against complex problems. The cognitive scientist Gary Marcus, who has been tirelessly sounding the alarm against the false promises of generative AI for some years, provides an excellent background to this argument here and here.

Despite all their storied sophistication and dazzling aura, AI systems can be fooled by simple things — small, irrelevant material in text documents, or changes in lighting. Even minor ‘noise’ can disrupt state of the art image recognition systems. If small modifications are made to the rules of games that AI has mastered, it can often fail to adapt. To be truly useful — and by extension, profitable — AI needs to provide reliable advice in complex environments that people encounter in the real world. Their failure to do so points to limitations that highlight the AI's lack of understanding of the inputs they process or the outputs they produce. This lack of understanding may be an unremovable feature of the way sub-symbolic AI is trained, not a bug that more data or more computational power can solve for.

The question is whether this surface understanding that AI systems can now demonstrate is itself sufficient to ensure a profitable return on the mammoth investment that is propping up the economy. My guess is if that was the case, the technology companies would not be promising AGI to raise more money, but focusing more on how the technology in its current state can begin to justify the investment.