Summary: More EAs should learn to speed-read. Main benefits of speed-reading: reading non-fiction books and articles faster and choosing how long to take on a report/book, rather than an open-ended amount of time. Speed-reading is not simply skimming. People often take weeks-long courses to develop all the related-subskills (I used a 36 chapter-a-day book). Specific sub-skills are: learning to not say the words aloud in your head, reading one or two lines at a time, and even using fewer eye-swipes to get the general sentiment of each sentence/paragraph. You’ll have gotten the gist of each paragraph and chapter through efficient skimming of the parts of the book that summarize the chapter’s directions (first/last paragraphs, table of contents, etc.) I use the 752-page NSCAI report as an example of speed-reading.

The views in this article belong solely to the author and do not represent those of the U.S. government.

Speed-Reading is Underutilized

I’ve seen speed-reading be a critical asset within policy circles, particularly at the higher levels, since you have so many sources of intelligence and analysis to review daily. Even as an entry and mid-level diplomat, I found that speed-reading tactics made my portfolios more manageable. Very famous policymakers who sped-read: President John F. Kennedy took a speed-reading course with his brother, promoted its use among his cabinet ministers; President Jimmy Carter and his staffers learned the skill while in the White House; President Theodore Roosevelt was a known speed-reader.

People earlier in their careers, researchers, and people with expertise in many domains should consider learning to speed-read. Learning to speed-read (via a self-paced course) has made large, important bodies of literature significantly more accessible. It allows me to decide how much time I think it’s worth spending on a book, article or report, and to get as much information as I can within that self-imposed time constraint. I.e., reading is no longer an open-ended exercise; I choose how long it will take.

You probably already use some aspects of speed-reading, like skimming through a paper where you’re familiar with the topics or skipping to the section with new information. But I am specifically discussing the comprehensive speed-reading skillset, which aims to help you:

  1. Read text faster via tactics like minimizing eye swipes, not reading aloud in your head, absorbing a paragraph as one unit as you would a sentence or a group of words.
  2. Focus on parts of the book you’d remember if you did read it more closely.
  3. Quickly find the key messages and, knowing what to expect, read rapidly since you don’t need to concentrate on the details to understand the book’s direction.

Learning to Speed-Read

I want to emphasize the different subset skills of a true speed-reading course. Many of these subsets are you gaining a new ability, similar to learning a guitar strum. To develop speed-reading skills, start by using your finger as a pacer while you read, progressively increasing your speed. Once you’re going as fast as you can, start dragging your finger over two lines at a time, and then across whole paragraphs so you’re reading an entire paragraph with fewer eye swipes. These exercises will feel uncomfortable and unnatural at first; you’re training your brain to absorb the information without slowing down by “saying” the individual words in your head or reading line by line.

Next, learn to quickly obtain the key messages. You’ll typically scan the table of contents and then the book itself to determine its structure. This pre-skim will help you read faster because you’ll know the final conclusion so won’t need to read the detail in depth. Additionally, you’ll improve your notetaking through mind mapping and prioritizing the distinct parts of the book.

I learned speed-reading from an older, physical copy of Peter Kump’s well-regarded book, Breakthrough Rapid Reading. It’s significantly easier to use a hard-copy book for finger-pacing than an e-book. After a while, you won’t need your hand to direct your gaze, so you’ll be able to achieve similar speeds on e-books.

What Speed-Reading the 752-page NSCAI Report in 2hr 45min Looks Like

I used the 752-page “National Security Commission on Artificial Intelligence (NSCAI) Final Report” as a practical example to illustrate the speed-reading skillset. My objectives were to obtain information not available in online summaries, learn how best to use it as reference material, and review the recommendations that most interested me.

I split my reading time into thirty-minute sessions, summarized below. My notes, taken while reading, were specifically for details of interest to me, new facts, and page numbers for later reference. These are included at the end of this article for anyone who’s interested.

First Thirty Minutes

I carefully read and took notes on the four-page “Executive Summary”, knowing it formed the framework for the document. I expected high reading speeds since content should align with the “Executive Summary”. I calculated I should complete the report in under three hours.

Session Two

30 seconds on the table of contents and 5 minutes skimming the entire book (“flash-reading” for less than 1 second per page to get book’s structure). At the 24-minute mark, I switched to speed-reading each page (2-3 seconds per page) until I finish all 590 main pages (over multiple sessions). 

After the five-minute skim, I realized that specific policy recommendations, known as “blueprints”, were concentrated at the back. I thus adjusted my general reading-pace: I’ll spend less time on the first half of the book to focus more on the second half. I bookmarked lists of concrete policy recommendations as I read.

Session Three to Five + 14 minutes

Continued speed-read at the rate of two to three seconds per page.

I allotted an extra five minutes at the end of the fifth session for the final chapter and checked the “Appendices” for nine minutes.



Below, I have included the notes I made while speed-reading the report. These show what I found interesting or noteworthy during my reading, but also illustrate that speed-reading does provide a clear understanding of an article’s content. Many of my notes remain in their relatively “raw”, original form.

Session 1

Executive Summary

  • AI will be a source of enormous power for companies and countries.
  • The People’s Republic of China (PRC) may overtake as the world's AI leader. Deepens the threat posed by cyberattacks and disinformation from Russia, the PRC, and others. Calls for AI-enabled solutions.
  • US must invest in AI resources now, not via incremental change.
  • NSCAI aims to reorganize the government and rally allies.

Part I: Defending America in the AI Area lays out the stakes and explains recommendations .

  • Defend against emerging AI-enabled threats to America's free and open society. I.e., task force and 24/7 operations center to fight disinformation, secure our databases, prioritize data security in foreign investments, supply chain risk management,
  • Prepare for future warfare by 1. Seeking widespread integration of AI by 2025, i.e., digital infrastructure, agile acquisition budget oversight processes, divesting from ill-equipped military systems. 2. Achieve a state of military AI readiness by 2025 via organizational reforms, innovative war fighting concepts, and establish AI and digital readiness performance goals; augmented DOD AI R&D portfolio.
  • Manage AI-enabled and autonomous weapons. The US doesn't find global prohibition of AI-enabled and autonomous weapon systems feasible but recommends checks. NSCAI suggests:
  • US should affirm existing US policy that only humans can authorize nuclear-weapon strikes and seek similar commitments from Russia, the PRC
  • Establish venues to discuss AI’s impact on crisis stability with competitors
  • Develop standards of practice for development, testing, and use of AI-enabled and autonomous weapons systems.
  • Transform national intelligence. IC should adopt AI-enabled capabilities.
  • Scale-up digital talent in government. Talent deficiency in DOD and IC impedes AI readiness by 2025. Government needs new talent pipelines, including a US Digital Service Academy to train current/future employees, civilian National Digital Reserve Corps to recruit people, and Digital Corp modeled on the Army Medical Corp to organize technologists already in gov.
  • Establish justified confidence in AI systems: government should ensure its AI systems are robust and reliable, including through R&D in AI security and human-AI teaming via national research labs and enhance DOD testing and evaluation capabilities. AI leads should be appointed across government to improve oversight.
  • Present a democratic model of AI use in national security such as AI-powered oversight and auditing, increasing public transparency about AI use, and building AI systems that advance privacy preservation and fairness. US government should provide redress and due process for those impacted adversely by AI tools.


Part II: Winning the Technology Competition

  • White House should lead a strategy for technology competition, including establishing a new Vice President-led Tech Competitive Council.
  • Win Global Talent competition, cultivate domestic talent and recruit abroad talent. National Defense Education Act II to improve the American education system and improve highly skilled immigration for AI talent to study, work, and remain in the US
  • Accelerate AI innovation in the US via major investments in AI R&D, and establish AI research infrastructure to democratize access to resources that fuel AI development. 1. Double non-defense funding for AI R&D to $32 billion, establish a National Technology Foundation, and triple the number of National AI research institutes. 2 establish a national AI research infrastructure composed of cloud computing resources, test beds, large scale open training data, and open knowledge network. 3. Strengthen commercial competition by creating markets for AI and forming regional innovation clusters.
  • Implement comprehensive IP policies and regimes. US lacks comprehensive IP policies and is hindered by legal uncertainties in US patent eligibility and patentability doctrine. Needs a plan to reform IP to further national security.
  • Build a resilient domestic base for designing and fabricating microelectronics. The US is almost entirely reliant on foreign sources for the production of semiconductors. Must rebuild domestic chip manufacturing via funding and incentives to maintain multiple sources of microelectronics fabrication in the US.
  • Protect US tech advantages: somehow protect ideas, tech, and companies without hindering innovation.
    1. Modernize export controls and foreign investment screening to protect dual-use tech via building regulatory capacity, implementing recent legislative reforms, coordinating export controls on semiconductor manufacturing equipment w allies, and expanding disclosure requirements from some foreign investors.
    2. Protect US research enterprise as a national asset by providing gov, law enforcement, and research institutions resources to do a risk assessment, share info on threats, coordinate protection efforts with allies, bolster cybersecurity, and strengthen visa vetting.
  • Build favorable international technology order: to ensure emerging tech strengthen democratic norms, share practices and resources to defend against malign uses of tech. US should lead an Emerging Tech Coalition to achieve these goals and establish a multilateral AI research Institute to enhance the US position as a global research hub.
  • Win associations technologies competitions. US must develop a single, authoritative list of technologies that underpin national competitiveness in the 21st century and take bold action to catalyze the US in AI, biotech, quantum, 5G, and energy storage tech. US should invest in specific platforms to enable breakthroughs and continuously prioritize emerging tech.



  • makes recommendations directly for the President and Congress.
  • spent 15+ hours in public deliberations, released 2,500+ pages of material


Session 2

Pt I, Ch. 1-8

  • Specifies USG’s wants on proliferation.
  • “AI and the Future of National Intelligence” includes reforming the security clearance adjudication process.
  • “Technical Talent in Government” recommends government hires more part-time, permanent employees. (Great and forward-leaning!)
  • Chapter 7 is entirely on establishing justified confidence in AI systems (i.e., robustness, safety)
  • Recommends US government expand AI systems usage into border control, foreign intelligence, and domestic security.
  • Recommends AI risk assessment reports if an AI system impacts a US person, primarily to assess privacy.
  • Recommends potential third-party testing centers for national security-related AI systems.


  • Ch. 9 focuses on science and tech dialogue with the PRC.
  • The number of domestic-born students doing AI doctorates hasn’t increased since 1990!
  • On immigration:
    • Suggests making an O-1 (extraordinary ability) visa emphasizing AI.
    • Recommends advertising entrepreneur immigration opportunities.
    • Suggests green card to STEM PhDs and fix the green card backlog.
    • Seeks to double employment-based green cards, currently 140,000 a year; fewer than half issued to the principal worker (most for the worker’s spouse and children).
    • Suggests new visa categories: an entrepreneur visa and a new emerging disruptive tech visa.
  • “Accelerating AI Innovation”:
    • Brain drain from academic to private sector erodes US advantage in basic AI research.
    • Private sector is better at AI innovation right now, which weakens the teaching base. Compute costs also limit academia.
    • 82% of algorithms in use today originated from federal-funded nonprofits.
    • 90% of US innovation resides in five cities. Misses out talent in other parts of the country.
    • US relies too heavily on private sector for AI.
    • General AI as a priority area. Cool diagram of the national AI research infrastructure!
    • Asks the private sector to fund a $1-billion nonprofit to broaden AI research opportunities.
  • “Intellectual Property” shows that China has 1.4 million patent filings, US 621. Makes patent filing more onerous due to increased number of prior arts to be reviewed and language barriers. Dominance of US patents in worldwide prior art searches is expected to erode. US AI and biotech inventions have been denied patents since 2010.


Session 3

Pt II continued

  • “Technology Protection” /export control chapter:
    • Recommends aligning US, Netherlands, and Japan on semiconductor manufacturing equipment.
    • Pg. 234 details cybersecurity resources for research institutions. Recommends that the government broker commercial credits to universities for secure data storage.
  • “A Favorable International Technology Order”: envisions a coordinating body of US National AI Research Institutes and multilateral initiatives to form a Multilateral AI Research Institute (MAIRI) providing formal research support to GPAI and OECD.
    • Mentions role of State’s new Cyber Bureau; D/MR bureau becomes an Under Secretary for Science, Research and Technology (State/Q)
  • No US city in the top-10 of smart-city connectedness. US internet, telecom lags. Pg. 259 shows the various “critical tech” lists used across different parts of government. AI is listed in all. Proposes world-class biobank, noting that US NIH GenBank database is poorly curated, underfunded, and used. China's similar database cost $117 million in initial funding.

Blueprints for Action: Part 1

  • “Emerging Threats in the AI Era”:
    • Requests Joint Interagency Task Force $30 million to direct top State Public Diplomacy official to lead in malign info.
    • Specific recommendations for Congress, DARPA, etc., to combat emerging cyberthreats from multiple angles.
    • Tasks CISA to create a modern-day “Cash for Clunkers” program, but for electronics!
    • Requests $125 million in FY 2023 appropriations to develop enterprise-wide data sets and $100 million for procurement and integration of commercial AI solutions for DOD business functions.
      • DOD's budget process requires funds to be requested two years in advance. Seeks to modernize. Recommends a dedicated AI fund under USD R&D, $200 million to start.
  • “Foundations of Future Defense” assigns JAIC multiple responsibilities.
    • For DOD, the AI delivery team at each Combatant command. 3.4% of the annual DOD budget to S&T and $8 billion for R&D of core AI.
    • Congress: additional $10 million to USD R&D for tech fellows and AI scouting tools and data.


Session 4

Blueprints for Action: Part 1 continued

  • Tech annex to national defense strategy is on page 317
  • “AI and Warfare” recommends:
    • Integrating AI-enabled applications into all major joint and services exercises, including war games and tabletop exercises.
    • Fostering “thinking red” via a $2.5 million, 10-week, 10-game series led by JAIC.
    • Integrate AI into systems whenever possible.
    • Coordination with Five Eyes countries, assessing gaps in AI-related tech, stress test for supply chains in critical industries, and common standards for Test, Evaluation, Verification, and Validation (TEVV).
    • Coordination with others (NATO, Australia, South Korea, Israel, etc.) for AI-related defense issues.
    • Allowing more DOD cooperative projects with private companies, academic research centers, and major non-NATO partners.
  • “Technical Talent in Government” states that there are 430k Computer Science (CS) jobs in the US but only 71k CS grads yearly.
  • “Establishing Justified Confidence in AI Systems”:
    • Calls for federal R&D on AI robustness and declares it a priority research area.
    • All departments and agencies should create an AI assurance framework.
    • DOD, ODNI red teams for adversarial testing.
    • Meet baseline criteria for robust and reliable AI through consulting outside experts, improving documentation practice, building systems architectures to limit the consequences of system failure.
    • Asks NIST to set standards
    • Lots on TEVV pg. 383
    • National AI Office should create a standing body of multi-disciplinary experts to be called upon for advice on responsible AI.


Blueprints for Action: Part 2

  • PRC tech dialogue on page 411.
  • “The Talent Competition” promotes STEM/AI learning programs, e.g., after school programs, teacher recruitment, scholarships, and asks state legislatures to require Statistics in middle school and Computer Science principles in high school. Only 47% of high schools offer CS courses.
  • “Accelerating AI Innovation” has some big asks:
    • Set up National Tech Foundation as an independent federal agency and sister to NSF, focusing on critical tech. NTF would run prizes and give grants modeled on DARPA.
      • Asks Congress to double non-defense AI R&D to $32 billion by 2026, . (done by doubling investments annually from a baseline of $1 billion in 2020)
    • Allot 1% of GDP on federally funded R&D, set up a National AI Research Resource (NAIRR) roadmap and fund teams of engineers to unlock public data currently by the government for us by the AI research community.
    • Asks OSTP to prioritize critical AI research areas, e.g., novel ML directions, TEVV, robust ML, etc.


Session 5

Blueprints for Action: Part 2 

  • (continued)
    • Requests 1-to-1 fund matching for up to 10 tech clusters across the US.
    • Seeks $1 billion from private sector donations. Government wants a private sector-led consortium and more industry-academia exchanges.
  • “Microelectronics” asks Congress for a 40% refundable investment tax credit for semiconductor manufacturing.
  • “Technology Protection”:
    • Asks Congress to establish an independent entity focused on research integrity to maintain open-source material and investigate research integrity.
    • US grant-making agencies should require cybersecurity standards, i.e., critical data is encrypted and audited.
    • Government info sharing. Asks FBI to share real-time actionable threat info with research institutions, and DHS to support research sector’s cybersecurity, similar to that of critical infrastructure
  • “A Favorable International Technology Order” chapter:
    • Asks Commerce to create a federal advisory committee to inform strategy on international standards.
    • Asks Congress to create a grant program for small AI companies to participate in international standardization efforts.
    • Seeks $300 million annually for State Department emerging tech programs.
    • The Multilateral AI Research Institute (MAIRI) plan is very cross-cutting involving many countries, physically based in the US.
    • Recommends State Department set up a Deputy Assistant Secretary for Science and Tech in each regional bureau and re-establish state representation in Silicon Valley.
    • Page 539 has a great graph of key multilateral tech initiatives – Five Eyes, GPAI, D10.
    • Pandemic preparedness is discussed on pg. 557.
    • Cooperation on space collision scenarios, AI R&D.
  • “Associated Technologies” promotes adding AI to biodefense, biotech leadership.
  • China's genomic center, BGI, has US partners and bought a US firm; may be global collection mechanism for the Chinese government.
  • US military owns most of the US mid-band, with slow 5G infrastructure expansion compared to the PRC.



  • Appendix D has segments of legislative text drafts, e.g., giving authority to DOD for cooperative research with intelligence partners.
  • Appendix E is a table of funding requests, making it very easy to see each cabinet/agency.


New comment
4 comments, sorted by Click to highlight new comments since: Today at 3:17 PM

My impression of the literature was that reading speeds above 500-600 wpm had drastic effects on comprehension, and there being general scepticism about 'speed-reading'.

You say 'speed-reading is not simply skimming', but I think it basically is? But strategically skimming and making notes of a huge report like you did would definitely allow you to comprehend a lot (most?) of the text with massive reductions in time. It's definitely a valuable skill that should be promoted. I just think speed-reading is a bit of a false moniker.

Rapid serial visual presentation also has drawbacks because a reader can't easily pause or re-read sentences, which is often important in understanding difficult text.

Do you know of any compelling peer-reviewed evidence of speed-reading? I couldn't find any systematic reviews, but the few recent studies I found were in line with my prior view.

There's some things here that definitely seem like a losing strategy for reading, like reading multiple lines at once. This post itself is definitely not something to speedread, not the least because some skepticism is occasionally required.

However, on balance, the recommendations are extraordinarily positive and seriously worth consideration. Especially the choice of the NSCAI final report as practice material, which is a very very good thing to practice on and I highly recommend trying it.

Another great thing is the Sinocism newsletter, which basically allows anyone to be an amateur China Watcher (or at least an assistant) but it's pretty long.

After skimming (or should I say speed reading? :) ) some skepticism from Scott Young (who used to promote speed reading after reading the same book as you, and certainly cares about learning quickly - he did MIT's 4-year CS curriculum in 1 year) and LessWrong, I think the summary for this post desperately needs some caveats. Is there any evidence for these claims besides personal experience and Kump's guidebook?

That being said, it would be wildly fortunate if people naturally acquired perfect reading techniques as children; there are certainly ways to read better, and some of this post's techniques might work. I've found a lot of success with the ideas from this Farnam Street article: quit books, assess why you're reading things and apply effort accordingly, read slowly when you find something really good, take notes.

Have you ever used Rapid Serial Visual Presentation (RSVP) for speed reading?  Depending on the content I can get close to 900 wpm!   An example online is:

I use the chrome plugin swift reader