The more I read about the current AI race and potential bubble, the more concerned I become about the "brute force" approach to achieve Artificial general intelligence (AGI) - just throwing more processing power at the problem. The circular investment loop funding this approach is also troubling: Nvidia invests in OpenAI, OpenAI uses that capital to buy Nvidia chips, Nvidia's valuation rises from the extra revenue, further increasing their investment power.
The Niche Domain Problem
At its core, current AI remains pattern recognition, classification, and statistics at a massive scale - nothing more. LLMs are trained predominantly on the most popular publications, widely-discussed subjects, and commonly available data sources. This makes them excellent for mainstream use cases, but fundamentally limited for:
- Specialised industrial processes
- Local regulations and compliance requirements
- Specific scientific subdomains
- Company-specific workflows and institutional knowledge
There simply isn't enough training data for these niche cases in commercial models, and more GPUs won't solve that. These models have no memory of your interactions and no ability to learn from usage. Using ChatGPT or Claude repeatedly in your specialised domain doesn't make it smarter about your field - each query is processed against the same static training data. While organisations can train custom models on their proprietary data, this undermines the "scaling solves everything" narrative - it requires domain expertise, curated datasets, and specialised effort, not just more computational power.
The Efficiency Problem
The brute force approach is staggeringly inefficient. The human brain performs complex image recognition, reasoning, and language processing using approximately 20 watts of power [1]. Meanwhile, training a single large neural network requires megawatts of power and massive cooling infrastructure. Running inference on these models at scale requires entire data centres consuming gigawatts of power. Studies show that training GPT-4 requires between 50.000 and 62.000 megawatt-hours, which is over 40 times more than its predecessor [2].
This isn't just an environmental concern - it's evidence that we're solving the search for AGI the wrong way. We're compensating for algorithmic inefficiency and simplicity with raw computational power, which has fundamental scaling limits in terms of energy, cost, and physics. If we truly understood intelligence, we wouldn't need to burn through this much energy to approximate basic cognitive tasks.
The Naive Path to AGI
The belief that scaling LLMs will lead to AGI is deeply naive. What we have are systems that:
- Recognise patterns from training data
- Generate statistical outputs based on probability distributions
- Don't perform actual reasoning or logical deduction
- Don't understand causality or cause and effect
- Hallucinate confidently on edge cases and unfamiliar domains
- Cannot distinguish between correlation and causation
Current LLMs lack fundamental cognitive capabilities that define intelligence: they cannot form mental models of the world, understand that other agents have different beliefs (theory of mind), or reason about hypothetical scenarios. They have no metacognitive abilities—no awareness of their own knowledge limitations or capacity to seek information strategically. True AGI would require genuine abstraction, cross-domain transfer learning that goes beyond pattern matching, causal understanding, multi-step planning with goal-directed behaviour, and creativity. Scaling these systems amplifies their pattern-matching capabilities but doesn't bridge the architectural gap to actual intelligence. We have no clear path from statistical language modelling to these cognitive skills—it's not simply "more compute”.
The Financial Reality
As Hank Green's analysis shows [3], the money flow is circular and unsustainable:
- NVIDIA invests $100B in OpenAI
- OpenAI uses it to buy Nvidia chips
- Oracle signs $300B contract with OpenAI
- Oracle buys more Nvidia chips
OpenAI: $500B valuation, $12B revenue, negative profit [4]
The chips aren't the end product - AI services are. And those services generate an order of magnitude less money than hardware spending alone, not counting staff, power, water, and infrastructure. This economic model is fundamentally broken: the cost of computing infrastructure required to deliver AI services far exceeds the revenue those services generate, making profitability impossible without either dramatic hardware cost reductions or massive price increases that would eliminate the mass-market appeal.
Marketing Over Substance
The industry is saturated with AI rebranding—basic infrastructure and orchestration tools marketed as revolutionary "GenAI platforms" to secure funding. Everything needs "AI" branding now, regardless of whether it genuinely involves generative capabilities.
Sam Altman's promises that AGI will cure cancer and solve global poverty and Elon Musk's claims that AI will make humans obsolete are particularly troubling examples of hype disconnected from reality:
- We don't have AGI, and there's no clear path to achieving it
- Cancer research requires experimental validation, clinical trials, regulatory approval, and decades of work—not just computation
- Global poverty is a complex socioeconomic problem that requires policy, infrastructure, and systemic change, not pattern-matching algorithms
OpenAI's mission statement—"ensure AGI benefits all of humanity"—rings hollow while they:
- Charge $200/month for premium access
- Keep their models closed source
- Maintain exclusive partnerships with Microsoft
- Operate at massive losses, funded by speculative investment
- Use "humanity" as a marketing slogan rather than a guiding principle
The disconnect between utopian messaging and commercial reality undermines credibility and suggests the real goal is capturing market share and investor capital, not democratising transformative technology.
Conclusion
Current large language models lack the fundamental cognitive architecture required for genuine intelligence: they cannot reason causally, form mental models, or transfer learning across truly novel domains. Throwing more compute at statistical pattern matching won't bridge this gap. The path forward requires algorithmic breakthroughs, not just bigger data centres. This all feels like dotcom bubble 2.0. Not because the technology is worthless, but because valuations are completely detached from economic reality. The real value lies in specialised, efficient models combined with symbolic AI and human expertise - but that's less sexy than "$100B data centres will achieve AGI."
Member discussion