Tech • AI • Robotics • Game
An AI PhD explains to me why the bubble is going to burst
7/10TL;DR
A specialist in machine learning argues that today’s AI boom rests on powerful but limited statistical models, making a near-term AGI breakthrough unlikely while raising the risk of an AI investment bubble.
KEY POINTS
Current AI is not general intelligence
Today’s systems are largely built on machine learning, especially deep learning and large language models, which remain a subset of older statistical and neural-network methods. These models map inputs to outputs, but they do not demonstrate consciousness or broad reasoning comparable to human intelligence. Their apparent sophistication often reflects interpolation and extrapolation from training data rather than independent understanding.
The technology behind the hype is decades old
Many of the mathematical foundations behind modern neural networks date back to the 1970s, while simpler methods such as linear regression have been used far longer. What changed recently was not a scientific revolution in intelligence itself, but the combination of vast datasets and massive computing power. The surge in performance came from scale, especially the use of GPUs, rather than from a fundamentally new theory of mind.
Agents are workflows, not autonomous minds
So-called “agentic” systems are better understood as workflows wrapped around existing models. An agent can feed outputs back into inputs, call tools, process images or text, and trigger actions, but that does not make the underlying model inherently self-directed. The practical risk lies less in the model alone than in what developers allow the agent to do and how well its environment is sandboxed.
Human oversight remains the central safety issue
A core regulatory distinction emerges between models and actions. In areas such as medical imaging, an algorithm may flag a possible tumor, but a radiologist is still expected to validate the result before any intervention. The same logic applies more broadly: automated systems can assist or prioritize, but fully delegating high-stakes decisions, such as military action or surgery, remains a far more serious threshold than simply using pattern recognition in low-risk tasks.
Calls for regulation are colliding with commercial interests
Demands to regulate advanced AI are intensifying, but the debate is entangled with competitive strategy. Restricting foreign models, especially from China, could slow rivals while protecting companies that have already committed huge sums to infrastructure. That creates a tension between legitimate safety concerns and the financial need of major firms to defend valuations and justify investment ahead of further fundraising or market expansion.
Infrastructure costs are becoming a structural constraint
The AI race depends on data centers, electricity, cooling, and chip supply at extraordinary scale. Building and powering these facilities requires investments measured in millions and billions, while grid connections and water use are already becoming bottlenecks. The economics resemble earlier supercomputing cycles: demand for computing can rise faster than the physical infrastructure needed to support it.
Nvidia’s dominance highlights the concentration risk
The boom has been tied closely to Nvidia, whose hardware became the backbone of large-scale AI training. Efforts by firms such as OpenAI and Anthropic to design custom chips aim to reduce dependence and improve efficiency, but they also bind software strategy even more tightly to expensive hardware development. That concentration increases the financial fragility of the sector if expected returns fail to materialize.
Competition from China is pressuring margins
One estimate places Chinese frontier models only about four months behind the most advanced American systems. If that gap is accurate, it would be narrow enough to challenge pricing power, especially if Chinese alternatives remain significantly cheaper. That dynamic matters because the leading US firms are trying to monetize products while simultaneously financing a costly arms race in computing capacity.
Bubble concerns are gaining traction in finance
The sector’s economics are under growing scrutiny as companies chase adoption without clear profitability. One example cited is Microsoft Copilot, where a license may bring in about $200 while operating costs are alleged to be far higher, even above $10,000 per month in some estimates. Whether or not those figures generalize, they capture the broader fear that revenue growth may not keep pace with infrastructure spending.
AGI by 2026 or 2027 is seen as highly unlikely
The argument against near-term AGI rests on architecture as much as economics. Large language models remain structured feed-forward systems and do not reproduce the dense real-time feedback loops seen in biological brains. Without a major mathematical and architectural breakthrough beyond today’s transformers, expectations of human-level general intelligence within the next two to three years appear overstated.
CONCLUSION
The debate is shifting from whether AI is impressive to whether its technical limits and financial structure are being misread. The main challenge now is to separate genuine industrial utility from speculative claims about imminent AGI and limitless growth.
Ask a question