Penn Wharton Budget Model

What is the impact of AI on labor markets? At the Penn Wharton Budget Model, we've examined how AI can impact economic growth, future deficits, and the finances of a large program like Social Security. While we model AI as more labor-augmenting than labor-replacing based on past innovations, a legitimate question is why we don't focus more on specific labor market outcomes, including job losses by exposed sector and the impact on wages across the income distribution. The reason is simple: the current data really do not yet support a clear labor-market conclusion on their own. We first need better theoretical modeling and more experience with AI. That assessment might be surprising to people who are lining up on both sides of this issue. On one hand, you might have seen a graph that shows how job postings suddenly fell after ChatGPT-3 was introduced. Plus, layoffs at Amazon and other technology companies are also making waves. But that evidence is not very useful. A lot of that change has nothing to do with AI. Many companies are laying off workers in response to macroeconomic outcomes and COVID hiring. Companies naturally want to project gains from AI to their investors, but other factors are more important. On the other hand, some recent studies have appeared to show little impact of AI on employment, often using variation in reported AI use by sector. Some tech companies love that message because it makes them seem less villainous. But much of that evidence already selects in favor of labor-augmenting change rather than labor replacing (survivor bias). For example, call centers that remain are not using nearly as much AI as those previously replaced by AI for a reason. The truth is that data—even a lot of it—cannot be interpreted without serious theoretical modeling that helps us construct the right tests and interpret the results. In fact, the very fact that macro effects seem to be contaminating micro conclusions should give some pause: the analysis is likely poorly identified. The good news is that stronger theoretical work is now being done that is laying the foundations for future empirical work. In case you want to read more, I am leaving some links below: Automation and Polarization https://lnkd.in/evqcbyku Artificial Intelligence in the Knowledge Economy https://lnkd.in/ej8xHehh Other studies referenced above: The Long-Term Outlook for Social Security: Baseline and Alternative Assumptions https://lnkd.in/eC3F6PPM The Projected Impact of Generative AI on Future Productivity Growth https://lnkd.in/e_Aw2Yee

Anthropic recently launched Skills. While they're not "MCP killers" as some claim, Skills are a key part of the context layer and can be used strategically to keep costs down. Here’s what I found actually works (and what’s overkill) when projects and teams get complex. MCP Servers: Think of these as broad, unified context management systems at the enterprise level that work across apps, AI models, and platforms. For example, if you want a single source of truth for JS/TS, Python, etc., then the Context7 MCP might be the way. At Penn Wharton Budget Model, we don’t use a monorepo across projects the form our integrated platform, so having a way to ensure consistency -- including with external resources -- across many active project repos became essential. We built a MCP server for our own use to keep these standards in sync. Skills (Claude Code and Web): Skills work across different project repos but with a tighter focus. They automate business logic, formatting, reporting, and internal procedures, all inside Claude’s ecosystem. They shine at standardizing tasks like reports or meeting analysis when multiple projects rely on the same AI setup. However, Skills remain internal to Claude: they’re portable across projects, but not deeply integrated with external systems. Rules (Cursor, Claude Code, and similar): These enforce requirements at the project level. For example, I’m strict about unit testing: sometimes, “no mocks, patches, skipTest, or wishful asserts—don’t hide bugs.” In other projects, limited mocking of expensive resources is fine if core systems get real tests. So, I want testing rules to vary by project. Trying to pack too much into a universal testing ruleset (such as a Skill) dilutes quality and leaves too much discretion to the AI model. Rules can be combined with Skills and MCP. Use them all at different context layers. And, be savvy to save money. Specifically, work backwards. Hit Rules and Skills first to minimize token usage. Keep those instructions lean to avoid overinvesting in broad MCP management that are overkill for narrow solutions. For example, I might run a full Context7 MCP session when upgrading libraries in a common requirements.txt file that impacts many files. But I am not spending 20K tokens to update a few lines of code where Skills and Rules are often enough. https://lnkd.in/epknBpXg

Cursor AI big update shifts the economics of the AI coding “last mile”. Cursor AI just launched Composer 1.0, a major update that allows its coding agent to operate without depending directly on Anthropic or OpenAI APIs—at least in its default autonomous mode. This is a significant strategic shift. Until now, most AI engineering copilots relied on large foundational models from the big players, effectively renting intelligence by the token. Cursor’s move toward independence could mark the beginning of a new phase: one where product companies optimize, fine-tune, or even train their own runtime systems for tighter control and lower marginal costs. In economic terms, this changes the “last mile” of AI delivery—the interface between model intelligence and user productivity. Ownership of that layer could unlock new defensibility, better margins, and faster iteration. If Composer 1.0 performs competitively while reducing reliance on external APIs, it will pressure others to develop their own inference layers too. The next wave of AI might not just be about who builds the biggest models—but who designs the smartest workflows around them. I fully expect Windsurf AI to follow. https://lnkd.in/egnFvGkB

What are the potential economic implications if the Supreme Court strikes down the IEEPA tariffs and orders a refund? Here's the data and details: As of yesterday (10/28/2025), the government collected approximately $222 billion in tariff-related revenue in 2025. However, an estimated $85 billion would have been collected based on tariffs in place before the Trump second administration. Those previous tariffs are not central to the current litigation. If a refund were mandated by the Supreme Court, it would total around $137 billion as of yesterday. Looking ahead, over the next decade, the government will forego approximately $2.8 trillion in additional revenue. However, removing the tariffs would be pro-growth even with the loss of revenue that would otherwise be used to pay down some government debt. The reason is that tariffs reduce economic activity more than other ways of raising the same amount of revenue. Unlike the undergraduate endowment model of trade, many U.S. imports (40 percent) are “intermediate goods” that lower the cost of U.S. production. So, tariffs raise these production costs and reduce U.S. capital investment inflows, including buyers of federal government debt over the long term. That is the reason that the dollar has fallen in value, unlike the undergraduate model where the dollar should have appreciated in value after the new tariffs. Put differently, if the new tariffs are put aside by the Supreme Court, the government would have to float more debt in the future, but the pool of buyers would also be larger, thereby allowing the government to sell that debt at a higher price and pay a lower return relative to baseline with tariffs. The price effect easily dominates the level effect in our calculations. Sources: - Real-time Federal Budget Tracker: https://lnkd.in/eBMsXFQf - Tariff Revenue Simulator: https://lnkd.in/enmRJf_c - Debt, Tariffs, and Capital Markets in a Dynamic Setting: https://lnkd.in/ef68g46k

I understand why many people are worried about the potential rise of AI superintelligence, including 850 experts who signed a petition arguing for a pause in model development. The mechanisms behind these models remain poorly understood, and discussions about AI's future often become abstract or metaphysical. However, game theory offers a clear answer: do NOT pause. The nuclear arms race involved only a few players with massive industrial resources that produced some hope of detection and inspection. But AI development can be done secretly by many more actors. Even large computing facilities don’t reveal what models they are running—whether beneficial ones like protein folding or potentially harmful ones. Pausing development would only hold back responsible researchers while creating opportunities for others to step in unchecked. Since verification is nearly impossible, the best strategy for responsible developers is to create superior models that outmatch those built by malicious actors. Build faster, build better—stay one step ahead, with good models designed to detect and counter the bad. It's the only rational move. https://lnkd.in/eU8DeuPd

Insights from PWBM on tariffs for the WSJ: https://lnkd.in/eBwV-PRy