The illusion of “easy” AI: Meta and xAI collide with the limits of technology

Financial markets continue to reward the promises of Artificial Intelligence, celebrating each announcement as an imminent revolution, but the industrial reality is presenting a very different picture. Mark Zuckerberg and Elon Musk, two of the key players in the global technology race, are currently navigating a technical stalemate that exposes the complexity of moving from marketing hype to the operational reality of computational models.

Implementing Artificial Intelligence is not simply a cost issue, but a major infrastructural and mathematical puzzle.

Meta's "smoothie": Huge investments, but disappointing benchmarks

Mark Zuckerberg has bet his company's future on AI supremacy, but this year's harvest appears to be plagued by programming flaws. According to recent reports, Meta has had to delay the launch of Avocado , its next-generation entry-level model, from March to at least May.

The reason is technical: in internal tests, the model showed disappointing performance in logical reasoning, writing, and, above all, coding. While outperforming previous Meta models and Google's version 2.5, Avocado trails the more recent Gemini 3.0 by a long shot. This delay comes despite a plan of investment in superconductors that could be described as extreme , with the development of four new dedicated chips.

• Capital expenditure (Capex) for this year is estimated at between $115 billion and $135 billion , nearly double the previous year.
• Long-term commitments approaching $600 billion in investments in the United States.
• A $14.3 billion stake in Scale AI, a move that led to Alexandr Wang becoming Meta's head of AI.

The newly formed "TBD Lab" was supposed to churn out fruit-themed innovations (Avocado as the base model, Mango for images/videos, and the impressive Watermelon for the future), but the division's leaders are reportedly even considering licensing models from rival Google to keep their products competitive. An admission of weakness that demonstrates how capital alone doesn't automatically generate superintelligence.

Elon Musk's "shock therapy" on xAI

On the other side of the fence, Elon Musk isn't faring any better, but he's reacting with his usual radical approach. xAI , his startup founded just two years ago, is undergoing a brutal reorganization. Its flagship programming product, Grok, is lagging behind competitors like Anthropic's Claude Code and OpenAI's Codex.

Musk's reaction was relentless, applying the method already seen in Tesla and SpaceX:

• Executive Board Purge: Key technical leaders, including Zihang Dai and Guodong Zhang, were removed, blamed for shortcomings in the pre-training phase and coding. Today, of the original 11 co-founders, only two remain.
• Integration and “Fixers”: Musk brought in trusted managers from SpaceX and Tesla to oversee the work, focusing obsessively on the quality of the training data .
• The Macrohard Project: To address these problems, xAI is developing digital agents (“Digital Optimus”) capable of observing and simulating business functions to automate entire enterprises.

All this is happening in the shadow of a titanic infrastructure: the Memphis supercluster , which already has 200,000 GPUs and is expanding towards a million units.

Why is AI implementation so difficult?

The analysis of these two corporate crises allows us to draw technical conclusions about the real difficulties of implementing Artificial Intelligence, far from the mainstream narrative:

1. The limit of brute force: Until recently, it was believed that simply adding computing power (GPU) and raw data was enough to obtain better models. Today, we discover that the "law of scaling" has diminishing returns. Quality is needed, not just quantity; in fact, adding quantity provides limited advantages, just as being more muscular doesn't mean being more intelligent.
2. The data bottleneck: As Musk discovered, the quality of training data is everything. Indiscriminately ingesting data from the web produces confusing models, incapable of rigorous logic like that required in computer programming.
3. The wear and tear on human capital: Developing these technologies requires herculean efforts. The pressure to meet unrealistic deadlines is causing a severe burnout rate among the world's top engineers.

Agency	Model in difficulty	Main cause of delay	Proposed (or ongoing) solution
Half	Avocado	Poor logical and coding performance compared to competitors	Possible licensing of third-party models (Gemini)
xAI	Grok (Coding)	Poor training data, gap with Anthropic/OpenAI	Targeted layoffs, data review, talent acquisition

Deficit spending on AI infrastructure is propping up the hardware industry, but a true return on investment and real-world application in the productive economy require a qualitative algorithmic leap that, at the moment, eludes even the Silicon Valley giants. AI, therefore, turns out to be much more elusive than previously thought.

The article The illusion of “easy” Artificial Intelligence: Meta and xAI collide with the limits of technology comes from Scenari Economici .

---

This is a machine translation of a post published on Scenari Economici at the URL https://scenarieconomici.it/lillusione-dellintelligenza-artificiale-facile-meta-e-xai-si-scontrano-con-i-limiti-della-tecnologia/ on Fri, 13 Mar 2026 21:07:44 +0000.