Cutting Through the Noise: My AI Priors

I've been in many AI policy conversations where thoughtful people hold radically different mental models about how AI will shape our economic future. Some see it as an overhyped bubble; others as the start of a new Enlightenment. This divergence among credible thinkers makes policy discussions difficult. Midway through a conversation, someone will often say, “You don't get it,” and then have to step back to explain the entire worldview underlying their position. I've certainly done the same myself.

In a bid to ease communication, I’ve decided to write out my underlying priors—which, to be more exact, are a collection of assumptions, mental models, and values-based opinions. This is risky, because things change so fast that writing a static public document can quickly look dumb. But the knowledge space is so wide that I think it’s better to reveal them, lubricate conversation, and update quickly with added information. Perhaps I will keep a public change log. And who knows—if more of us did this, we might get to the crux of our disagreements faster. These views are also based on a near-midterm horizon.

Finally, I try to articulate some areas where my assumptions could be wrong, but I’m certain there are things I’m missing. I’m keen to hear about them.

Ok here they are:

1. Foundational AI models will likely have the capability to master most knowledge domains and digital tasks. If this is right, foundational models (think Claude, Gemini etc.) will have the ability to be legal assistants, clinician aides, and accountants out of the box, and many specialized applications in those domains could become less valuable. We seem close, but not quite there yet. Startups building applications in these domains currently rely on fine-tuning, guardrails, multi-agent orchestration, and scaffolding to cajole their AI systems to behave in specific ways their domains require. For example, a voice-based AI medical assistant may still need these techniques to remain accurate over prolonged use. Eventually, however, if model providers offer such functionality out of the box, they could “eat” many of the roles these specialty applications play.
2. Even if foundational AI models did have the capability to master most domains, it doesn't mean they will. Market failures will probably leave gaps. In the near term, even with compounding advancements in AI’s capabilities, firms likely will have to choose which domains to “eat.” Compute, energy, and data are currently scarce, so commercially valuable domains will need to be prioritized while many public-interest areas remain neglected despite being technically solvable. This is likely to be especially true in low- and middle-income countries (LMICs). For example, despite there being an estimated 500 million smallholder farmers worldwide, leading technology companies are unlikely to prioritize addressing their needs because gathering their data is hard, lots of agricultural work is done offline, and there are weak monetization opportunities. One way I could be wrong is if a recursively self-improving AI figures out a way to collect data remotely or create high-quality synthetic data, and dramatically drops the cost of prioritizing these less profitable use cases.
3. Public money should be spent on publicly beneficial domains and layers of the stack that the market will underprovide or provide too slowly. This principle may seem obvious, but is easier to state than to apply. The frontier moves quickly, and outsiders have little visibility into what AI labs are prioritizing. Just a year ago, building a pedagogically sound AI tutor required extensive fine-tuning, guardrails, and scaffolding. I did not expect major tech companies to invest in it, but today several frontier models can deliver much of this functionality with simple instructions.
4. Even if foundational models master most domains, the need to deliver interventions and public services, and manage organizational change to adopt AI, will likely remain. In healthcare, a doctor may recommend an AI agent to help patients manage medications, symptoms, and appointments. But in many cases, health systems will still need to ensure patients download it, have internet access, and use it. The same is true for workplace tools. Firms still need to figure out the right user interface, incentives, training, and workflows that capitalize on highly capable models. Change management and cracking human behavior remain hard. The intermediation between digital tools and their real-world use will change but probably still matter.
5. Organic AI use beyond formal interventions will be a major driver of impact. While many beneficial AI applications will be delivered through interventions, much of AI’s impact will also come from people independently downloading consumer models like ChatGPT, Claude, and Meta AI. These effects can be both positive and negative. We already have emerging evidence of the negative impacts: students are using generic chatbots to complete homework, harming learning outcomes at scale. My hope is that if these products are designed to promote beneficial behaviors—such as adhering to medication and seeking healthcare—they could positively impact billions of users. We need to think about how to influence and regulate these consumer products, in addition to delivering beneficial interventions.
6. In a world where most digital interventions are AI-enabled, intervention type may become a weaker predictor of impact than an implementer’s ability to rapidly iterate and improve. Traditionally, the evidence movement has evaluated categories of interventions against one another to identify which is most impactful at moving a development outcome. For example, when trying to impact student learning, we may compare the results of phone-based tutoring versus teacher training versus scholarships. I’m still forming my thoughts here, but it seems plausible that AI-enabled iteration could make in-category variation so large that intervention type becomes far less informative. Consider two AI tutoring programs: one delivers a static course, while the other achieves much larger impact because it continuously updates based on user errors, adjusts pedagogy, and runs A/B tests to improve engagement. Continuous improvement has always mattered, but AI enables much more rapid cycles of learning and improvement. As a result, what predicts impact may no longer be labels like “AI tutoring,” but an organization's ability to rapidly iterate.
7. While AI may eventually touch every digital activity, there are still many predominantly analog and offline interventions worth funding for human flourishing. Highly capable AI could have huge benefits, but people still need capital, mentorship, physical medicine to ingest, vehicles to truck those pills to that person, etc. There’s still a base level of human capital that you want to invest in to enable people to benefit from the AI economy, and the most effective ways to build this base-level capital, especially in very rural LMIC communities, might still have large offline elements.
8. Stepping back from human capital interventions, much of the productivity growth in high-income countries could come from cloud-based multi-agent systems. I’m less optimistic for LMICs. It’s hard to imagine micro entrepreneurs running a business off WhatsApp orchestrating an army of agents in the cloud. However, could a new class of small and medium enterprises source business opportunities, develop marketing material, manage supply chains, and build an entire customer service platform off these agents? There will likely be a few standouts, but I suspect the cost of doing business and growth bottlenecks are not addressed by AI enablement, no matter how powerful models become.
9. New AI-era philanthropic wealth may be largely interested in solving new problems created by AI, rather than longstanding social challenges. AI is creating new problems that may require new solutions. The Industrial Revolution sparked social unrest between workers and factory owners, but it also led to new protections such as child labor laws, workers’ compensation, public education, and public health systems. We need to consider whether today’s institutions and social protections are equipped for the structural changes AI may bring. Labor displacement is only one widely discussed possibility. For these reasons, investing in AI-related challenges is more than legitimate. However, if we believe that many major problems predate AI (item 7) and solutions exist outside the digital sphere, we should also think about how to use the wealth AI generates to address both the new challenges it creates and the longstanding ones it does not.
10. We are in a rational financial bubble, as our new board chair, Mohamed A. El-Erian, said in a recent CGD fireside chat. Mohamed explains it’s a rational bubble because the gains to the winners of the AI race are huge, but it’s hard to tell who the winners are, so capital is spread out across a wide portfolio in the hopes that when one or two investments pay off, they will more than compensate for the losses of others. This is typically done by venture capital at a relatively small scale and can look terrifying at the current scale. The presence of a bubble is sometimes used to argue that AI is mostly hype and isn’t as valuable as many claim. I think we need to believe multiple things at the same time: (i) the data center buildout is ginormous; (ii) many will not recoup their investment on the timelines needed for them to be financially viable; (iii) forecasts will swing between too much compute demand or too little supply, but there is likely a bubble regardless; (iv) lots of things could cause the bubble to burst; and (v) even if the bubble bursts, this could still be one of humankind’s most transformative inventions.

I conclude with some disclaimers: I can't claim originality for many of these ideas. If you’ve read this and thought, “That came from me,” it may very well have, but I will only express my gratitude if you turn out to be right. If you’re reading this and going “geez what a Luddite,” or “what a wacko,” I hope it at least explains where some of my policy ideas come from.