Edward Conard

Top Ten New York Times Bestselling Author

  • “…a must-read for serious students of economic policy.” - Glenn Hubbard, Dean, Columbia Business School, and former Chairman of the Council of Economic Advisers
  • “…reminds us that inequality sends a signal of what society lacks most, in America’s case, entrepreneurship and risk taking.” - Lawrence Lindsey, CEO, The Lindsey Group, former Director of the National Economic Council
  • “Unintended Consequences provides a provocative interpretation of the causes of the global financial crisis and the policies needed to return to rapid growth. Whether you agree or not, this analysis is well worth reading.” - Nouriel Roubini, New York University; Chairman, Roubini Global Economics
  • “…serious thinking for serious thinkers. …a thought-provoking blueprint for growing middle- and working-class incomes.” - Mitt Romney, former Governor of Massachusetts
  • “Unintended Consequences represents the most cogent and persuasive analysis of the Financial Crisis to date.” - Andrei Shleifer, 1999 John Bates Clark Medal Winner
  • “…a fresh argument for the productive value of inequality.” - David Autor, Professor of Economics, Massachusetts Institute of Technology
  • “…challenges misconceptions that distort our economic debates.” - Arthur Brooks, President of the American Enterprise Institute
  • “Unintended Consequences is far smarter and more thought-provoking than most economics written for the general public” - Greg Mankiw, Harvard University, Former Chairman of the Council of Economic Advisors
  • “Unintended Consequences is full of substance, it is one of the must-read books of the year, and once I finish it I will be giving it a second read through right away.” - Tyler Cowen, Professor, George Mason University
  • “There are an amazing number of good ideas and interesting points made in Unintended Consequences. The thinking underlying it, and the obvious depth of understanding of the author, are very impressive.” - Steven Levitt, coauthor of Freakonomics; 2004 John Bates Clark Medal
  • “Unintended Consequences offers deep and well-argued analyses on almost every issue.” - The New York Times
  • “…a fresh argument for the productive value of inequality.” - David Autor, Professor of Economics, Massachusetts Institute of Technology
Upside of Inequality Oxford Unintended Consequences

AI Summary. Achieving 60% on MMLU, a test of academic reasoning, once required a 540bn-parameter model but was reached by a 3.8bn-parameter model two years later, a 142x reduction in parameters. Models small enough to run on phones or sensors shift inference from data centers to devices, weakening the case for paying cloud providers per token.

Paul Kedrosky Applied Complexity
Date Posted:
Is Database:
Database

Small models have increasingly closed the performance gap with frontier models. To Kedrosky, this implies “edge compute becomes real,” which in turn, at the margin, means reduced relative demand for data center-based versus local compute.

Are small AI models closing the gap with frontier models?

Core argument: Reaching 60% on MMLU, a test of academic reasoning, required a 540bn-parameter model in 2022 but only a 3.8bn-parameter model by 2024, a 142-fold parameter reduction in two years.

A quiet story in AI isn't that large models keep getting better, it's that small ones are catching up quickly. In 2022, reaching 60% on MMLU, a test of academic reasoning, required a 540-billion-parameter model. By 2024, a 3.8-billion-parameter model hit the same threshold — a 142-fold parameter reduction in two years. The direction is clear, and the implications are wide-ranging. Edge compute becomes real. Models that fit on a phone or a sensor don't need a data center. Inference shifts from hyperscaler to device, which is a different supply chain and a different competitive moat than anyone has built for. The data center API business model has a half-life. If a fine-tuned 7B model running locally can handle 80% of enterprise use cases, the case for paying per-token to a cloud provider shrinks fast. The most exposed providers are those whose moats are model quality rather than distribution, i.e., the capex-heavy frontier models. Compute becomes less scarce.

Takeaways by Macro Roundup® AI

  1. Reaching 60% on MMLU, a test of academic reasoning, required a 540bn-parameter model in 2022 but only a 3.8bn-parameter model by 2024, a 142-fold parameter reduction in two years.
  2. Models small enough to run on a phone or sensor shift inference from hyperscalers to devices, which involves a different supply chain and competitive moat than providers have built for.
  3. The case for paying per-token to a cloud provider shrinks if a fine-tuned 7B model running locally can handle 80% of enterprise use cases, leaving capex-heavy frontier providers whose moats rest on model quality rather than distribution most exposed.

Related Articles:

  • Commoditization, Orchestration, and the New AI Stack — The author argues that as incremental gains from larger AI training runs shrink, the return on additional computing power falls, weakening the case for the training spending that dominates data center demand. Survivors of model commoditization will be those owning the orchestration layer or the applications where outcomes are priced.
  • The Geopolitics of AI: Decoding the New Global Operating System — The performance gap between the top American and Chinese AI models narrowed from 9.3% to 1.7%, with Chinese firms relying on cheaper open-source models—DeepSeek prices usage 20-40x below OpenAI's ChatGPT. This price advantage could win adoption in developing countries and build dependency on Chinese AI ecosystems.
  • Misanthropic: On Mythos, Bad Human Behaviors and Systems Vulnerabilities — Anthropic reports that its Mythos model has detected thousands of high-severity cyber vulnerabilities, including ones found by chaining multiple obscure software weaknesses, and that these hacking skills are emergent rather than a designed capability. Mythos completed a corporate network attack simulation estimated to take an expert over 10 hours, which no other frontier model had solved.
Showing 77 database articles primarily about Innovation/Research