In a landscape obsessed with the sheer velocity of artificial intelligence, Brad DeLong amplifies Noah Smith's provocative counter-narrative: we are likely drowning in software that doesn't actually do anything useful. While the market celebrates an impending trillion-dollar valuation for Anthropic based on hyper-growth, the core argument here is that "tokenmaxxing"—the indiscriminate spending of AI credits to generate code—is a classic case of measuring inputs while ignoring outputs. For busy leaders trying to distinguish between genuine technological leaps and expensive theater, this piece cuts through the hype by asking the one question Wall Street refuses to: what are we actually shipping?
The Illusion of Productivity
DeLong introduces Smith's central thesis by highlighting a bizarre disconnect in the tech sector. "Not much of the kinds that we are used to", says Noah Smith, noting that while scaling laws for machine capability remain strong, the economic law connecting machine output to human value has hit rapidly diminishing returns. The author points out that companies are spending billions on AI coding agents without seeing a corresponding explosion in new features or product quality.
Smith illustrates this with a telling anecdote from an entrepreneur who ordered his staff to "spend their salary in tokens," creating code equivalent to their entire payroll. When pressed, the entrepreneur couldn't explain what was being built. This isn't just an isolated incident; it's a systemic behavior where Meta briefly ran leaderboards for token usage and one company reportedly spent half a billion dollars on Claude Code alone. As DeLong notes, this frenzy mirrors historical patterns of experimentation with general-purpose technologies like electricity or the internet, where initial chaos precedes real utility.
However, the data suggests we are stuck in the chaotic phase longer than expected. "People are spending hundreds of thousands of dollars a month on tokens? Guys, what are you shipping?" asks John Loeber, capturing the frustration of engineers watching their peers chase metrics rather than results. The evidence is stark: only 18% of AI coding spend translates into shipped products that reach real users, according to EntelligenceAI. This aligns with the concept of Goodhart's law mentioned in related deep dives—when a measure becomes a target, it ceases to be a good measure. Companies are optimizing for token consumption, not economic value.
"The link is not there yet... I think maybe implicitly there is more that is getting shipped, but it's very hard to draw a line between one of those stats and, 'Okay, now we're actually producing 25% more useful consumer features.'"
Uber COO Andrew Macdonald's admission underscores the difficulty in connecting raw AI usage to tangible business outcomes. Critics might argue that software development cycles are long and the benefits of technical debt reduction won't be immediately visible on a quarterly balance sheet. Yet, the sheer scale of spending relative to the lack of observable innovation suggests a bubble of activity rather than a surge in productivity.
The Bottleneck is Human, Not Computational
DeLong pivots to a deeper structural critique: automation hits a ceiling because it cannot solve the "weak links" in complex workflows. Even if AI can write code a million times faster, the overall system speed is constrained by the least-automated tasks—often human decision-making, integration, or physical-world constraints.
The commentary draws on historical context to explain why this matters. Just as having 100 million times more computing power than in 1970 hasn't made individual workers that much more productive, AI is not a magic wand for corporate efficiency. DeLong writes, "Within firms, task automation hits the same ceiling: even spectacular acceleration of coding does not explode total corporate productivity because downstream tasks remain human bottlenecks." This reframes the debate from "can machines code?" to "can we reorganize our businesses to let them?"
Smith suggests that the software industry may be a mature sector, similar to steelmaking or internal combustion engines. In such industries, technological improvements often lead to consolidation and cost-cutting rather than the creation of entirely new markets. The "consumer internet" frontier appears saturated; new apps are displacing old ones rather than expanding total usage. As DeLong paraphrases Smith's view, the real upside lies not in better social media or e-commerce, but in robotics and radically reconfigured business processes that control the physical world.
"The key role played by electricity in the shift from the applied science to the mass production mode of societal organization was not something that anyone could have predicted in 1890."
This historical parallel is crucial. We are currently in a phase where we know how to use the tool (natural language interfaces) but haven't yet figured out how to restructure society or industry around it. DeLong compares current AI management to driving a car with unreliable driver-assist features: "I am always terrified that the automatic braking and lane keeping... are going to turn themselves off, and driving has become a much more cognitive intensive and challenging task." The machine saves time on specific tasks but adds cognitive load in supervision.
Bottom Line
DeLong's curation of Smith's argument offers a necessary reality check for an industry intoxicated by its own potential. The strongest part of this analysis is the distinction between "task-level productivity" (writing code faster) and "economic productivity" (shipping valuable products), a nuance often lost in IPO hype. Its biggest vulnerability lies in predicting the timeline; history shows that general-purpose technologies can take decades to show their full economic impact, meaning today's waste could be tomorrow's necessary infrastructure investment. Readers should watch not for the next billion-dollar app, but for companies that successfully reorganize their entire operations around AI rather than just bolting it onto existing workflows.