The Great AI Plateau

The most telling statistic in Stanford's 2025 AI Index isn't about model performance or investment figures—it's that harmful AI incidents surged 56% to 233 cases, whilst costs plummeted 280-fold.

We've built the technological equivalent of a Ferrari with no brakes, sold at the price of a bicycle.

What's Actually Happening: The Numbers Behind the Narrative

Stanford's latest AI Index reveals a field at an inflection point, where the smallest model achieving 60% on MMLU dropped from 540 billion to 3.8 billion parameters—a 142-fold reduction that would make Moore's Law blush. Microsoft's Phi-3-mini now matches what Google's PaLM achieved two years ago, using orders of magnitude fewer resources. This isn't just efficiency; it's a fundamental shift in the economics of intelligence.

The cost dynamics are even more remarkable. Querying an AI model with GPT-3.5 equivalent performance dropped from $20 per million tokens to $0.07—a reduction that makes the traditional laws of industrial pricing look quaint. Google's Gemini-1.5-Flash-8B achieved this benchmark at pricing that represents either the greatest technological deflation in history or a race to the bottom that will define the next decade.

Meanwhile, the geopolitical chess match has intensified. US institutions produced 40 notable AI models compared to China's 15, but the performance gap has narrowed from double digits to near parity within 18 months. China's models haven't just caught up—they've done so whilst operating under export restrictions that were supposed to prevent precisely this outcome.

Corporate adoption tells its own story: 78% of organisations now report using AI, up from 55% in 2023, whilst generative AI usage in business functions doubled from 33% to 71%. Yet the productivity gains remain frustratingly elusive. Goldman Sachs warns that widespread adoption is the missing link to measurable economic impact, expected to materialise around 2027.

Why It Matters Now: The Efficiency Revolution's Dark Side

Benedict Evans would recognise this pattern: we're witnessing the classic transition from innovation to commoditisation, but at unprecedented speed. The 280-fold price reduction in AI inference costs mirrors the historical trajectory of computing power, yet it's happening in quarters, not decades.

This democratisation has profound implications. When DeepSeek's R1 model reportedly cost $6 million to build (compared to the hundreds of millions spent by US labs), it didn't just challenge American technological supremacy—it rewrote the rules of AI development economics. The Chinese approach prioritises algorithmic efficiency over brute-force computation, potentially making US export controls irrelevant.

The business reality is more complex. McKinsey estimates that AI could add $2.6-4.4 trillion annually to the global economy, yet Federal Reserve research shows workers save only 5.4% of their work hours—translating to just 1.1% aggregate productivity growth. The gap between potential and reality suggests we're still in the experimentation phase, not the transformation era.

The regulatory landscape reflects this uncertainty. US states passed 131 AI-related laws in 2024, more than doubling from 49 in 2023, whilst federal progress remains stalled. This fragmentation mirrors the early internet era, when different jurisdictions competed to define the rules of a technology they barely understood.

The Deeper Implications: Civilisation at the Crossroads

Harari would frame this moment as a new chapter in humanity's relationship with information processing. The 233 recorded AI incidents aren't merely technical failures—they're symptoms of a civilisation deploying transformative technology faster than it can understand the consequences.

The geographical divide in AI optimism reveals deeper cultural fractures. 83% of Chinese and 80% of Indonesians believe AI offers more benefits than drawbacks, compared to just 39% of Americans and 36% of Dutch respondents. This isn't just cultural difference—it's a fundamental disagreement about the relationship between technology and human agency.

China's approach represents a particular vision of technological governance: centralised, state-directed, and optimised for collective rather than individual outcomes. China installed 276,300 industrial robots in 2023—six times more than Japan and 7.3 times more than the US—whilst simultaneously deploying AI-powered surveillance systems that would horrify Western democracies.

The medical domain illustrates both AI's promise and peril. FDA approvals for AI-enabled medical devices jumped from 6 in 2015 to 223 in 2023, yet this acceleration raises profound questions about validation, accountability, and the nature of medical expertise itself. When machines can outperform doctors in pattern recognition, what happens to the human element of healing?

The investment patterns tell their own story about civilisational priorities. US private AI investment reached $109 billion—nearly 12 times China's $9.3 billion—yet this dominance may be transitory. As MIT Technology Review argues, the framing of AI development as a zero-sum competition undermines the collaborative approach needed for safe advancement.

What Comes Next: Scenarios for the Next Inflection

Scenario 1: The Great Convergence (35% probability) China continues closing the performance gap whilst reducing costs, forcing US companies to compete on efficiency rather than scale. Export controls prove ineffective as algorithmic innovation trumps hardware advantages. Global AI development becomes multipolar, with different regions pursuing distinct approaches to AI governance.

Scenario 2: The Innovation Plateau (30% probability) Current scaling laws hit fundamental limits around 2026-2027. AI winter warnings prove prescient as transformer architectures exhaust their potential. Investment shifts to other technologies, leaving AI as a powerful but specialised tool rather than a general-purpose intelligence.

Scenario 3: The Regulatory Fracture (25% probability) Rising AI incidents trigger aggressive regulatory responses in democratic countries, whilst authoritarian states embrace unrestricted development. The global AI ecosystem fragments into incompatible technological blocs, creating new forms of digital colonialism.

Scenario 4: The Breakthrough Acceleration (10% probability) New architectural innovations around 2025-2026 unlock artificial general intelligence capabilities. The productivity gains that economists have promised finally materialise, triggering the fastest economic growth since the late 1990s but also unprecedented social disruption.

The wild card remains energy consumption. As data centres could consume 10% of US electricity by 2030, the AI revolution may hit physical limits before intellectual ones. China's approach of prioritising efficiency over scale may prove more sustainable than America's brute-force strategy.

Conclusion: Intelligence as Commodity, Wisdom as Scarcity

The Stanford AI Index 2025 documents a remarkable achievement: we've made intelligence abundant. Models that required billions of parameters now deliver equivalent performance with millions. Costs that measured tens of dollars now count pennies. The technological problem of artificial intelligence is largely solved.

The human problem has just begun. As CNAS research warns, the US-China AI competition extends beyond military and economic advantages to “world-altering” questions of conflict norms, state power, and human values. The race to build more capable AI systems may be less important than the race to deploy them wisely.

Harari's observation about the 21st century—that its central challenge would be managing technological power—has crystallised around artificial intelligence. We've created tools that can think but not feel, reason but not care, optimise but not judge. The Stanford Index shows we're getting remarkably good at the first part. The second remains humanity's work alone.

The bottom line: We're approaching peak AI capability growth but valley AI wisdom implementation. The question isn't whether artificial intelligence will transform civilisation—it's whether we'll have any say in how that transformation unfolds.

Thanks for reading XTRACT! Subscribe for free to receive new posts and support my work.

Subscribed

References

  1. Stanford AI Index 2025: State of AI in 10 Charts - Official summary of key findings from Stanford HAI

  2. The 2025 AI Index Report - Full 400+ page comprehensive analysis

  3. AI costs drop 280-fold – Tom's Hardware - Technical analysis of cost reductions and safety concerns

  4. Goldman Sachs AI productivity analysis - Economic impact and timeline projections

  5. Federal Reserve productivity study - Empirical research on workplace AI impact

  6. US-China AI gap analysis – Recorded Future - Comprehensive geopolitical competition assessment

  7. MIT Technology Review on AI arms race - Critical analysis of competitive dynamics

  8. CNAS report on world-altering stakes - National security implications analysis

  9. Vanguard productivity forecast - Long-term economic projections

  10. AI Winter analysis - Historical patterns and current risks