The AI Bubble: Advanced AI Faces A 'Complete Accuracy Collapse'

The Investment Context

The Gen AI theme has been the primary force shaping the financial markets since Microsoft Corporation (MSFT) invested in OpenAI and ChatGPT in late 2022. The tech mega-caps companies associated with the Gen AI theme, primarily NVIDIA Corporation (NVDA), but also Alphabet Inc. (GOOG) (GOOGL), and Meta Platforms, Inc. (META) count for over a third of the S&P 500 Index (SP500).

During this period, the valuations of these Gen AI mega-caps reached bubble levels, which consequently pushed the S&P 500 valuation near historic highs based on many valuation ratios. For example, the Shiller P/E ratio is over 37.

At the center of the Gen AI bubble is an enormous AI capex, so large that it's affecting the U.S. GDP. These AI companies are actually clients of each other, so this Gen AI capex is eventually reflected in their profits.

However, we still don't know what the ROR on this Gen AI capex will be - and it could all be a wasted investment if the Gen AI proves to be a fad technology.

At this point, the application of Gen AI is widely used in agentic AI, where the Gen AI performs some simple tasks. Yet, we still don't see major productivity effects on the broader economy.

How Does Gen AI Work?

A simple way of explaining how Gen AI works is basically:

• You need to have a large data set.
• You train a Large Language Model on that dataset.
• The large language model, or LLM, can then perform tasks learned from that dataset, for example creating lyrics, music, images, computer code, drawings, etc.

The early excitement about Gen AI was the ability of AI to understand human language and reply in human language with useful information, and especially the ability to generate images.

However, for Gen AI to be able to perform complex tasks, necessary to increase productivity, the AI must be able to start reasoning - beyond the data set training. In other words, the AI needs to become like a human, to think like a human. That's the higher level of AI - and the entire success of Gen AI technology depends on the ability to reach that level defined as the Artificial General Intelligence AGI.

In simple terms - that would be a robot, or a humanoid able to perform all complex tasks, like we see in sci-tech movies.

The industry moved towards AGI with Open AI o1 model in 2024, but the real breakthrough was the DeepSeek, from Hangzhou DeepSeek Artificial Intelligence Co., Ltd. (DEEPSEEK) R1 model in January 2025, which outperformed all other models, as a fraction of the cost.

Obviously, if the reasoning models progress and reach the AGI stage, the AI capex could prove to be profitable, but human society will fundamentally change forever. These reasoning AI models will replace human scientists - and all future innovations will be created by AGI.

The Apple's Angle

One major company that has been lagging in the development of Gen AI models is Apple Inc. (AAPL). Yes, Apple adopted Apple Intelligence, but even this largely failed.

So, Apple is a curious case. Why is Apply lagging in Gen AI, given the company's record of innovations?

The "Complete Accuracy Collapse" Study

Apple has just released a study titled "The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity," in which Apple finds that the Advanced Reasoning AI suffers from a "complete accuracy collapse" when facing a highly complex problem.

This study essentially finds that the current AI models are nowhere near AGI. In fact, apparently current LRMs are just good at memorizing the training dataset, but completely break down when facing a complex task that would require human-like thinking. In fact, even the best LRMs quit when unable to find a solution.

Yes, LRM models have been tested for Ph.D.-level intelligence by solving very complex math problems. But Apple claims that these tests were corrupt, where solutions have been hidden and easily accessible by AI.

The Apple study tests the LRM models using puzzles.

• The study finds that LLM models can solve a low-complexity puzzle.
• When facing a medium-level puzzle, advanced reasoning LRMs outperform basic LLMs.
• However, when facing a complex puzzle, both LLMs and advanced LRMs have a "complete collapse of accuracy". Specifically, the "reasoning" completely breaks down, not only by giving wrong answers, but also by quitting the problem.

So, what are the implications of the Apple study?

The key implication is that the advanced reasoning Gen AI is not close to the AGI, and that the progress has hit the wall. This means that the Gen AI capex is very likely a wasted investment.

Regarding Apple's lagging in Gen AI, it seems like Apple justifies a careful approach to AI by stating that AGI might not be reachable. On the other hand, some might look at this study as Apple's excuse for lagging in Gen AI.

Either way, Apple's study must be replicated, and the reasoning AI models must be retested using Apple's approach to verify the results. At this point, there is no reason to suspect the conclusion that AGI is currently unreachable is biased or wrong.

Implications

The investment implications are obvious. The Gen AI is likely a bubble, which is eventually going to burst once the investors realize that the massive Gen AI capex is mostly a wasted investment. Specifically, the Gen AI technology is unlikely to reach the AGI stage, based on Apple's study, which means that Gen AI is unlikely to significantly boost productivity.

Thus, the S&P 500 bubble is likely to burst as the Gen AI bubble bursts. At this point, investors are completely ignoring the warning from the Apple study, and the bubble could continue to inflate.

Regardless of short-term volatility, the S&P 500 is a Sell at these valuations, facing an eventual Gen AI bubble burst as more investors start questioning the validity of Apple's study, and a possible admission from related companies that AGI is currently unreachable.