Edward Conard

Top Ten New York Times Bestselling Author

  • “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
  • “…a must-read for serious students of economic policy.” - Glenn Hubbard, Dean, Columbia Business School, and former Chairman of the Council of Economic Advisers
  • “A full-throated defense of economic dynamism.” - The Wall Street Journal
  • “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
  • “…a fresh argument for the productive value of inequality.” - David Autor, Professor of Economics, Massachusetts Institute of Technology
  • “…a very valuable contribution.” - Larry Summers, former Secretary of the Treasury and director of the National Economic Council, president emeritus, Harvard University
  • “Unintended Consequences offers deep and well-argued analyses on almost every issue.” - The New York Times
  • “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
  • “…a must-read for serious students of economic policy.” - Glenn Hubbard, Dean, Columbia Business School, and former Chairman of the Council of Economic Advisers
  • “…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
  • “…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
Upside of Inequality Oxford Unintended Consequences

AI Summary. DeepSeek appears to have trained its models about 45x more efficiently than other leading-edge models, combining existing techniques—activating only 37bn of 671bn parameters at a time, a smaller memory cache, lower-precision training, and multi-token prediction—alongside self-reinforcement learning. Training cost $5.6m, and its interface charges $0.14 per million tokens versus OpenAI's $7.50.

Michael Cembalest J.P. Morgan
Date Posted:

DeepSeek’s V3 model, the basis for their R1 release, may have been trained 45x more efficiently than other leading-edge models. DeepSeek provides API access at $0.14 per million tokens while OpenAI charges $7.50 per million tokens.

How did one model achieve far greater training efficiency?

Core argument: DeepSeek appears to have trained its models 45x more efficiently than other leading-edge models, at a cost of $5.6M for 2.7M GPU hours, while Llama 3.1 405B used roughly 10x more GPU hours.

DeepSeek appears to have trained its models 45x more efficiently than other leading-edge models. To be clear, most of DeepSeek’s approaches already existed. Its greatest accomplishment: deploying them all at once in the face of a chip ban, and introducing self-reinforcement learning. GPT-3.5 uses its entire model for training/inference despite only needing small parts. GPT-4 and DeepSeek are a mixture of expert (MoE) models activating only necessary parts. DeepSeek V3 has 671B parameters, but just 37B are active at any time.MLA refers to “multi-head latent attention,” jargon for how DeepSeek runs with a smaller memory cache. Other efficiency approaches: parameters stored with BF16/FP32 are reduced to FP8 during training. Multi-token prediction (MTP) reduces accuracy ~10% but doubles inference speed. Training cost $5.6M for 2.7M GPU hours; Llama 3.1 405B used ~10x more GPU hours. DeepSeek API: $0.14/million tokens vs. OpenAI’s $7.50. Performance rivals OpenAI 4o and Anthropic’s Sonnet-3.5, better than Llama’s biggest model.

Takeaways by Macro Roundup® AI

  1. DeepSeek appears to have trained its models 45x more efficiently than other leading-edge models, at a cost of $5.6M for 2.7M GPU hours, while Llama 3.1 405B used roughly 10x more GPU hours.
  2. DeepSeek V3 holds 671B parameters but activates only 37B at any time as a mixture-of-experts model, alongside multi-head latent attention for a smaller memory cache, FP8 training precision, and multi-token prediction that cuts accuracy about 10% while doubling inference speed.
  3. DeepSeek's API charges $0.14 per million tokens against OpenAI's $7.50, with performance rivaling OpenAI 4o and Anthropic's Sonnet-3.5 and bettering Llama's biggest model, though most of its approaches already existed, its accomplishment being deploying them at once under a chip ban and introducing self-reinforcement learning.

Related Articles:

  • DeepSeek FAQ — Consumers and businesses stand to gain most from a future of effectively free AI products, along with big consumer tech companies, since a world of free AI makes product and distribution the decisive advantages they already hold. China also benefits, and the United States faces a choice between expanding chip export restrictions and permission-based regulation or allowing itself to compete.
  • The Year-End Wake-Up: The ‘Chinese Sputnik’ That Demands Our Attention — Chinese AI company DeepSeek trained its Version 3 large language model for roughly $5.6m, about 100x cheaper than Meta's Llama 3, with performance comparable to Claude 3.5 Sonnet and interface access costing 12x less than Anthropic's. The author argues US chip export controls backfired by spurring leaner engineering, and that the gap between open-source and closed-source models is narrowing rapidly.
  • A Severe Case of COVIDIA: Prognosis For An AI-Driven US Equity Market — Computing cycles move from infrastructure to platforms to applications, and NVIDIA data center revenues as a share of market-wide capital spending are projected to reach levels last seen at the 1969 mainframe peak and the dot-com boom. Large cloud providers spend more than twice as much training foundational models as running them for customers.
Showing 77 database articles primarily about Innovation/Research