A recent National Bureau of Economic Research report just threw cold water on the whole idea of unchecked, exponential AI progress. It shows a 31% slowdown in model training efficiency gains between 2023 and 2025. This deceleration isn’t just a technical footnote. It’s a direct threat to the US’s ability to maintain its lead in AI, especially with intense geopolitical competition. So, how should a Trump AI policy respond to ensure America stays dominant?
Key Takeaways
- We must invest in fundamental AI research, not just applied AI, to counter the 31% slowdown in efficiency gains.
- The US has to diversify its AI talent pipeline way beyond the usual tech hubs if it wants to fill a projected 1.2 million AI jobs by 2028.
- Strategic export controls on the most advanced AI hardware and foundational models are a necessary tool for maintaining a competitive edge against rivals.
- Public-private partnerships, like the AI Grand Challenges program, are a good way to speed up the development of secure and ethical AI.
- Policymakers have to get serious about data governance frameworks that let us innovate without sacrificing privacy, because that directly affects the quality of our training data.
The Staggering Cost of Training: A 31% Drop in Efficiency Gains
That late-2025 report from the National Bureau of Economic Research (NBER) is a big deal. It reveals that the rate at which our AI models get more efficient at training themselves has dropped by 31% in two years. This isn’t a full stop on progress, but the “bang for your buck” on compute is clearly fading. For a long time, we saw models double their capabilities with just a marginal bump in processing power, but my interpretation is that we’ve already picked all the low-hanging fruit in AI optimization. We’re now in a phase where the next big gains will demand fundamental breakthroughs, not just tweaking the architectures we already have. This has serious implications for any Trump AI policy. Just letting market forces do their thing, as some people argue, is probably not going to be enough. The sheer amount of money it takes to train these monster models is astronomical, even with this slowdown, and without a focused effort to fund basic research into new algorithms or entirely different computing methods, the US is going to fall behind. This isn’t just about throwing cash at the problem. It’s about being smart with investments in the underlying science. The whole “move fast and break things” idea was fine for the early days, but it’s not a sustainable plan when you’re facing diminishing returns.
The Talent Gap Widens: A Projected 1.2 Million Unfilled AI Roles by 2028
A recent analysis from CompTIA, the IT industry association, is projecting a shortage of 1.2 million skilled AI professionals in the US by 2028. That number covers everything from pure AI researchers and machine learning engineers to data scientists and AI ethicists. This isn’t a new issue, but the projected deficit is now so big it’s alarming, and it tells me our current education and immigration policies just aren’t keeping up. To me, this data point is the most critical one. You can have the best hardware and algorithms in the world, but without the people to design, build, and deploy them, you’re going nowhere. The current administration’s focus on domestic talent is a good start, but it needs to be scaled up dramatically. We need more than university programs. We need vocational training, real apprenticeships, and strong reskilling programs for people already in the workforce. We also need a pragmatic approach to attracting and keeping the best international AI talent. While I understand the national security arguments, an overly restrictive immigration policy is just ceding intellectual leadership to countries that are more welcoming to global experts. This isn’t about open borders. It’s about smart, targeted immigration that puts STEM professionals with these specific skills at the front of the line.
US Venture Capital Investment in AI: A 15% Dip in Early-Stage Funding
Looking at the 2025 PitchBook data, I’m worried. Early-stage venture capital funding for AI startups in the US dropped 15% from the previous year. Late-stage funding held steady, but a dip in seed and Series A rounds is a canary in the coal mine for future development. This trend is a huge problem. Early-stage funding is where the truly disruptive, sometimes crazy, ideas get their first shot. A decline here tells me investors are getting risk-averse, maybe because of market saturation, regulatory uncertainty, or the simple fact that it now costs a fortune to get a real AI company off the ground. What can a Trump AI policy do? It needs to find ways to de-risk these early investments. This could mean expanding grant programs through agencies like the National Science Foundation (NSF) or DARPA, specifically aiming them at high-risk, high-reward AI research. Another option is creating tax incentives for investors who back qualified AI startups. We have to avoid a future where only the big, established tech companies with deep pockets can afford to build anything new, because that will stifle the agile and bold work that small teams have always produced.
China’s AI Patent Filings: Outpacing the US by 2.5x
The competition isn’t some abstract idea. A report from Georgetown University’s Center for Security and Emerging Technology (CSET) shows that in 2025, China filed 2.5 times more AI-related patents than the United States. You can argue about patent quality versus quantity, but you can’t just dismiss that disparity. That volume signals a clear national strategy to dominate the sector. From my professional standpoint, the US is being complacent. We need a more aggressive and coordinated national AI strategy that stops reacting and starts getting ahead. That means enforcing intellectual property rights more strictly and maybe even creating new legal frameworks to handle AI-generated inventions. On top of that, our public sector research institutions and national labs need enough funding to conduct the foundational AI research that produces real breakthroughs, not just small improvements on what’s already out there. The race isn’t just about who builds the fastest chip. It’s about who owns the foundational IP that the whole AI field will be built on.
The Unconventional Wisdom: Why More Data Isn’t Always the Answer
The prevailing wisdom in AI for a long time has been “more data, better models,” with companies collecting absolutely massive datasets assuming that volume alone leads to better performance. But recent research from places like Stanford University’s Institute for Human-Centered Artificial Intelligence (HAI) confirms what many of us have seen in practice: the quality and diversity of your data are becoming far more important than just the quantity, especially for advanced models. I disagree with the simplistic idea that data quantity is king. Piling on more unstructured, noisy, or biased data can actually degrade a model’s performance or, in a best-case scenario, just lead to diminishing returns. We’ve all seen models trained on huge, unfiltered internet datasets that absorb and amplify our biases, which then leads to terrible outcomes in facial recognition or loan applications. A Trump AI policy that wants to secure American leadership needs to change the focus from data accumulation to data curation and ethical governance. This means funding tools for creating high-quality, diverse, and ethically sourced datasets. It also means setting up clear rules around data privacy and use, which protects citizens and ensures the integrity of the AI systems we build. The future of AI isn’t about who has the biggest data lake. It’s about who has the cleanest, most representative, and most ethically sound data. Getting there is a complex job that requires strategic investment and a real grasp of how this technology is actually changing.
What is meant by an “AI slowdown” in efficiency gains?
An AI slowdown in efficiency gains means the rate of improvement is decreasing. For a given amount of computing power or data, we’re getting less of a performance boost than we used to. AI development hasn’t stopped, but the “return on investment” from more training is getting smaller.
How does a talent gap impact AI development?
A big talent gap means you don’t have enough skilled people for AI research, development, and deployment. This shortage slows down progress, pushes project timelines back, drives up costs, and makes it harder for a country to compete globally in AI.
Why is early-stage venture capital important for AI?
Early-stage VC is the seed money for new AI startups and research projects. It’s the capital that lets people explore new ideas, build prototypes, and get companies started. This is where disruptive tech comes from, the kind that bigger, more risk-averse investors usually won’t touch.
What is the significance of AI patent filings in international competition?
AI patent filings show how much a country is investing in protecting its intellectual property in artificial intelligence. A high number of filings, while not the only measure of success, suggests a deliberate national effort to own key technologies, which is a strong sign of future economic and tech power.
Why is data quality becoming more important than data quantity in AI?
As AI models get more complex, the quality, diversity, and ethical sourcing of their training data matters more than ever. Bad data, whether it’s low-quality, biased, or just irrelevant, produces flawed models that either don’t work well or reinforce social biases. Careful data curation is now more valuable than just collecting huge amounts of raw information.
