How To Win The AI Race

AI Policy should prioritize getting tech off the shelves and into people's hands.

While there is a lot of energy in AI policy, strategic and political conversations often feel rudderless. Some want to regulate, others want to beat China; neither angle, however, seems to have a sense of where we are going, if these tactics are sustainable, and what the purpose of these policies even are.

What is the “why?”

In a new piece penned with Jordan Schneider in Noema Magazine  we argue for a new policy heading, which we call a “diffusion-centric AI strategy.”

“The most promising path forward is instead structural: a diffusion-centric AI policy laying the groundwork for long-term productivity growth. AI success requires breaking the tech out of the lab and putting it into people’s hands.”

Our core premise: The country that “wins” the AI race will be the country taking advantage of the economic and political benefits of actually *using* it. Churning out research papers for fleeting advantage should not be the goal. Abundance should be. While continued R&D and investment will always be needed, “Diffusion, not development, is the bottleneck for success and today’s true public policy challenge.”

Unsustainable R&D Advantage

Our emphasis on diffusion is in part a recognition of the unstable strategies of the past which have focused on getting as “large of a lead as possible” in frontier AI technology.

We argue that while OpenAI and other leading western firms hold advantage today, that lead could be eroded by:  

1.     Talent Transfers: “Unlike past technological advances, which favored government and firm control, individual AI developers enjoy unique competitive autonomy today.” In the past, firms often controlled the ‘secret sauce’ of competitive products – today, that secret sauce of AI is often readily available online. When it isn’t, core design elements can still easily diffuse to other companies through talent transfers– especially as firms in China offer top developers salaries sometimes in the millions.

2.     Industrial Espionage: While espionage is always a threat with geopolitically significant technologies, for AI technology this threat is uniquely high. AI doesn’t have the physicality or complexity of past manufacturing-based developments and algorithms can be easily smuggled through hard drives, and in some cases memorized.

3.     Open Source: Its unclear if AI labs, and government R&D, will maintain the technical frontier. Open source development is challenging, often successfully, traditional R&D structures. We argue:

“Today, the AI open-source hive mind is collectively far larger than any single AI company, rapidly developing innovations that are increasingly competitive with the latest and greatest coming out of the top private AI labs. If the best models are to come from open-source development, the best models will be equally accessible in both China and the West. The success of open source may mean the elimination of any software-related geopolitical technological gaps.”

4. Unsustainable hardware advantage: The western lead in AI hardware may also be susceptible to erosion . Despite our best efforts, Chinese actors have been able to creatively skirt our porous export controls, been able to access leading AI chips remotely through cloud computing, and can always lean on chip smuggling as a fallback. Meanwhile, Chinese chip investments are starting to yield an ever-stronger domestic industry. Even if we do continue to lead, the slowing of Moore’s law may yield a future where hardware R&D may provide diminishing productive and strategic returns.

A diffusion Centric Strategy

If R&D focused strategies offer unsustainable competitive advantage and questionable long term strategic benefit, why might diffusion succeed? We note:

“While technical diffusion may sound amorphous from a policy lens, as the Princeton scholar Helen Milner argues and the OECD has since empirically validated, the process is significantly influenced by regulatory and institutional design factors. The implication is that policymakers potentially hold the tools and levers needed to speed up the process and to come out ahead.”

If this is true, what “tools and levers” exsist and what might a diffusion centric strategy look like?

Start With A Light Regulatory Touch

“Like innovation itself, diffusion can be a protracted creative process, characterized by trial and error and experimentation. For this process to play out, regulation cannot act as an excessive burden. As a first step in any diffusion-centric strategy, regulators should aim to do no harm. That requires analyzing regulations and determining what rules and processes may inhibit AI success for little societal gain.”

A significant focus of a diffusion centric strategy should be retooling regulation to ensure process doesn’t stand in the way of use. For policymakers this should start with these basics:

Creating An AI Regulatory Map: As I argued in a previous post, current AI regulations remain hazy – we know AI is regulated under pre exsisting statutes, but we don’t know to what degree. To promote the confident innovation and risk taking essential to rapid diffusion, we must act to reduce regulatory uncertainty. To these ends, the administration should create a regulatory map - or inventory - which clarifies exsisting laws, identifies any regulatory overlaps, and notifies congress of legal contradictions.

AI-ready regulations: A second task is identifying which regulations may need to be adapted to the needs of AI. Many laws are decades old and many will not fit the dynamism and capabilities of AI. The administration should push agencies to analyze their regulations, compare them to the needs of AI, and adapt as needed.

Resisting Over Regulation: Finally, we should resist the impulse to over-regulate that is common amid such forward uncertainty. Many proposed AI regulatory schemes – such as AI licensure - could slow down AI deployment (and research) in exchange for questionable benefits. As congress embarks on its new AI legislative push, members should consider potential impacts of regulations on diffusion.

Speeding Up Diffusion

Policy should not only “do no harm” but set the table for rapid diffusion. To speed up this process, policymakers should consider the following:

Invest in AI-ready education: AI diffusion requires a talent pool that understands and embraces AI. For policymakers, this means pivoting towards not only technical education, but an education that puts emphasis on technical adaptation.

“Through proper curricular design, the workforce can be trained to flexibly adapt to uncertain technical needs and prepare to continuously pivot and adopt cutting-edge systems throughout their careers. “

Invest in AI Ready Government: AI diffusion should extend to both the public and private sectors. This likely requires changes to federal processes to ensure government systems and policy can keep pace. Policymakers should eliminate waterfall-style IT decision making, loosen the often-strict and over prescriptive rules that bind developers hands, and consider prioritizing leadership with AI or technical subject matter expertise.

Soften The Inevitable Backlash

Finally, we argue that steps to manage a backlash are needed, the alternative being an over-reaction that could stall progress and unnecessary harm. Measures to manage AI's use in elections, to improve transparency, to invest in safe systems, to shore up cyber infrastructure, and speculatively, to ease potential labor market transitions, could mitigate obvious challenges and help build the foundation of trust needed to ensure diffusion is not disrupted.

Work to be done

There is much more we discuss in the essay and I recommend giving it a read.

Still - for a diffusion-centric AI policy strategy, these ideas are only a starting point. Beyond what we’ve identified, there are no doubt countless regulatory and institutional factors that can be improved. In coming weeks, I’ll be digging into the challenge of AI diffusion, analyzing the scope of the diffusion challenge, and adding flesh to this framework’s ideas.