Brad DeLong delivers a rare, unvarnished assessment of modern artificial intelligence that cuts through the hype to reveal a staggering energy inefficiency. While the tech industry celebrates breakthroughs, DeLong argues we are currently burning the equivalent of all the world's computing power from 1985 just to answer a simple question about a grocery list. This piece is essential listening for anyone trying to separate genuine utility from the "absurd inefficient ignorant overkill" of the current AI boom.
A New Tool for the Classroom
DeLong begins with a personal breakthrough: he is finally ready to recommend his "SubTuringBradBot" to students as a first-line resource for answering questions about his work. This is not a generic chatbot, but a specialized system built on his actual writings. He illustrates its success with a query about the Marshall Plan, noting that the bot correctly identified the core thesis of his 1993 paper with Barry Eichengreen. The system understood that the Plan's $13 billion in aid was not primarily about capital transfer, but about resolving a balance-of-payments crisis to buy "three years of breathing room."
"The macroeconomics were the precondition for the politics, not the other way around."
DeLong writes that the bot successfully synthesized arguments from multiple sources, including his rebuttals to historian Alan Milward. Milward had argued that European recovery was already underway before the aid arrived, but DeLong's bot correctly countered that "without resolution of the balance-of-payments crisis, the recovery underway in 1947 would have strangled." This precision is vital; it mirrors the nuance required in historical analysis, much like the debates surrounding the Single Euro Payments Area where technical infrastructure often dictates political possibility. The bot didn't just retrieve text; it grasped the causal mechanism.
"The database here is grounded in your actual writing, not inference."
This distinction is the piece's most critical insight. DeLong explicitly rejects the idea of a "General-Purpose Transformer" that simulates a "typical internet s*poster," even one polished by reinforcement learning. He wants a "natural-language interface to a structured 'catechism' question-and-answer challenge-and-response database." The success of this approach suggests that for specialized knowledge, Retrieval-Augmented Generation (RAG) is far superior to the probabilistic guessing of large language models. Critics might argue that this limits the bot's ability to handle novel questions outside the training data, but DeLong's goal is accuracy, not improvisation.
The Absurdity of Energy Waste
The tone shifts dramatically as DeLong turns his attention to the infrastructure required to run these models. He describes a scenario where he asks his local computer, "What is on the grocery list?" In a traditional UNIX environment, this would be a trivial command. Instead, the modern AI approach triggers a massive, three-minute computational event.
"Ten times as much computing power as existed in the world in 1985. Devoted to making sense of the natural-language sentence: 'What is on the grocery list?'"
DeLong breaks down the energy cost with surgical precision. His brain burns about 50 watts; a human looking at a fridge uses a fraction of that. But the MacStudio, running a local large language model, ramps up all ten performance cores and the GPU to 100% utilization. The result is a power consumption 50,000 times higher than the biological equivalent, or a million times higher if the request were sent to a cloud server.
"It is the GPT LLM MAMLMs that are hopelessly inefficient at this. And it is because they are hopelessly inefficient at this that, right now, RAM and GPU prices are screamingly high."
This is a damning indictment of the current datacenter boom. DeLong suggests that the massive infrastructure build-out is not driven by immediate utility, but by executives betting on a future where these inefficiencies can be overcome. He compares the current state of AI to the Point Four Program, where the US attempted to export technical expertise to developing nations; just as that program had mixed results due to local conditions, the current AI infrastructure may be a massive over-investment in a technology that doesn't yet know how to do simple tasks efficiently.
"The datacenter infrastructure boom is the result of company executives applying truly enormous amounts of brute financial force to try to overcome the massive inefficiencies generated by our near-complete ignorance of how to do interpretation and generation of natural-language human communication at all efficiently."
The argument here is that we are in a "remarkable timeline" where we are burning vast resources to solve problems that computers were already perfectly designed to solve decades ago. The "brute financial force" is masking a fundamental lack of understanding about how to process language efficiently.
"A power budget of 50 x .01 x 1 = 0.5W-seconds. And the same brain-effort if I insisted on doing the task in the virtual rather than in the real world... But for the MacStudio? 140 watts."
DeLong's bemusement is palpable. He sees a future where the "SubTuringBradBot" might work well for office hours, but he remains deeply skeptical of the broader economic and environmental logic of the industry. The "inefficient energy-hungry resource-using mess" is not just a technical glitch; it is the defining feature of the current AI economy.
"I find myself excited, skeptical, and more than a little bemused."
Bottom Line
DeLong's most compelling contribution is the stark contrast between the precision of specialized, retrieval-based AI and the wasteful brute force of general-purpose models. While the "SubTuringBradBot" offers a genuine path to democratizing access to expert knowledge, the energy cost of the underlying technology remains a massive, unresolved vulnerability. The industry's bet on "brute financial force" to fix these inefficiencies is a high-stakes gamble that risks creating a datacenter economy built on a foundation of absurd overkill.