How Fast Are AI Models Actually Improving?
AI models are getting better at a surprisingly fast pace.
AI isn’t just improving — it’s accelerating, outpacing even the most optimistic predictions year after year.
In 2022, it took a 540-billion-parameter model to score above 60% on a key intelligence test. By 2024, a model 142 times smaller did the same job. Think of it like shrinking a school bus down to a bicycle while keeping the same speed.
Costs dropped too. Querying a powerful AI fell from $20 to just $0.07 per million tokens in two years.
Meanwhile, the gap between the best and tenth-best models shrank dramatically. Simply put, AI is becoming stronger, cheaper, and more competitive every year. China’s benchmark scores on MMLU and HumanEval went from double-digit gaps behind the U.S. in 2023 to near parity in 2024.
Research tracking AI over the past six years shows that models have gone from handling tasks taking seconds to handling tasks taking nearly an hour, with linear extrapolation suggesting multi-week task capability could arrive by late 2028.
These advances are also accelerating applications in finance, where machine learning is increasingly used to analyze market data and support trading decisions.
Is the US–China AI Gap Smaller Than You Think?
Speed and cost improvements tell only part of the story. The US–China AI gap is shrinking faster than many expected. According to Stanford’s 2026 AI Index Report the gap between top US and Chinese models narrowed to just 2.7 percentage points. Back in 2023 that gap was noticeably larger. Chinese models still trail by roughly seven months in frontier capability. Think of it like a race where one runner started late but keeps closing the distance. China achieves this despite spending far less. US AI startups received $109 billion in private investment during 2024 compared to China’s $9.3 billion. Chinese labs have made notable progress by embracing open-source models optimised for scale, compounding their gains even under significant compute constraints imposed by US export controls. Meanwhile, Chinese firms have leaned into deployment over raw capability, with 67% of Chinese industrial firms having deployed AI in production compared to just 34% for analogous US firms. This trend is reflected in faster real-world adoption of AI in industry, where production deployment often drives tangible economic impact.
The AI Risks You Actually Need to Worry About
Progress in AI brings real risks worth understanding. Bias is one big concern. Facial recognition software makes more mistakes identifying minorities. That is a serious fairness problem.
Privacy matters too. AI tools often scrape personal data without clear permission. Remember Cambridge Analytica? They used AI to secretly mine Facebook data for targeted ads.
Security threats are real as well. Bad actors can misuse AI for fraud or manipulation. Deepfakes make fake videos look surprisingly real.
Most experts think runaway killer robots are unlikely soon. The everyday risks though? Those deserve serious attention right now. Up to 30 percent of U.S. work hours could be automated by 2030, according to McKinsey, threatening widespread economic displacement. Employers and workers are encouraged to prioritize upskilling and training to adapt to AI-driven changes in the labor market. Drawdown risk from rapid economic shifts can intensify financial and employment instability for affected workers.
How AI Is Reshaping Jobs and Skills Demand
Reshaping the workforce, AI is changing which jobs exist and what skills employers want most.
After ChatGPT launched in 2022, postings for repetitive tasks dropped 13%. Meanwhile demand for analytical and technical roles grew 20%. Think less “filing spreadsheets forever” and more “teaching AI to do it better.” Diversification across roles and continuous learning help workers stay resilient.
New roles like telehealth coordinators and AI engineers blend tech knowledge with human judgment.
By 2030 millions of new jobs are expected in healthcare and high-tech manufacturing.
Employers now want workers who combine AI literacy with creativity, leadership and critical thinking. Basically humans still win at being human.
Routine roles like data entry, telemarketing, and administrative support are shrinking as AI automates those tasks at scale, with 76,440 such jobs eliminated in 2025 alone.
Occupations with high augmentation potential, such as microbiologists, financial analysts, and clinical neuropsychologists, combine AI-automatable tasks with work that still requires meaningful human involvement.
What Governments Are Spending to Lead in AI
Winning the AI race costs serious money. The U.S. federal government spent about $3.3 billion on non-defense AI research in FY25. That sounds big but falls far short of the recommended $16 billion target. Meanwhile private companies poured over $109 billion into AI in 2024 alone. Defense spending tells a different story. AI contract values jumped 1,200% in one year showing governments mean business.
Agencies like Treasury and DHS also ramped up spending markedly. Still most AI dollars flow through just ten departments leaving many agencies behind. The race is on but the spending gap remains wide. Decades of federally funded research built the foundations of modern AI, with public and private R&D acting as complements rather than substitutes, where government seeds early-stage ideas that private investment then scales into products. Governments often use futures contracts and other financial instruments to hedge spending and manage budgetary risk.
The Department of Defense alone accounted for 95% of federal AI potential contract value in the most recent year studied, with its potential award value surging from $269 million to over $4.3 billion. This dramatic concentration signals that federal AI strategy is increasingly defined by defense priorities, raising questions about whether civilian agencies and governance frameworks can keep pace.
