Staying current on AI
Start with a protected learning hour
Let me first put this in the context of lifelong learning, since in my view keeping up with AI is mostly a matter of habit. Many of the most effective leaders I know of protect regular time for learning. Charlie Munger famously said that in his whole life he had known no wise people who did not read all the time, "none, zero," and added that people would be amazed at how much Warren Buffett reads (the line is collected in Poor Charlie's Almanack). The writer Michael Simmons popularized a more modest version he calls the five-hour rule, an hour a day of deliberate learning modeled on Benjamin Franklin, and although it is a heuristic (Simmons himself says it is not an academic study), it is a very useful one.
For a busy executive I would start with one protected hour a week, on the calendar like any other meeting, and spend it roughly like this:
- About 20 minutes reading two newsletters from the list below, one practical (for instance One Useful Thing or The Batch) and one strategic or skeptical (Benedict Evans or AI as Normal Technology).
- About 30 minutes trying a current AI tool on a real task from your own week. In my view this is the most valuable habit of all, since nothing you read will calibrate your sense of progress as well as seeing what the tools can and cannot do on your own work.
- The last 10 minutes noting one thing you will change or try, which is what turns reading into learning.
Once a year, skim the takeaways from the AI Index (spring) and the State of AI Report (October). And if you want more structure than a reading list, I would also suggest taking one course a year. Stanford Engineering's Center for Global & Online Education (CGOE) runs Stanford Online, which offers professional and executive programs on AI and innovation from Stanford faculty for working professionals, and a structured course with deadlines and classmates is often what makes the weekly hour stick.
Weekly synthesis
- One Useful ThingPracticalEthan Mollick (Wharton) · Weekly, free
Practical and hands-on, since Ethan tests the current tools himself and writes about what they can do inside an organization. He is an enthusiast, so I would read him together with one of the more skeptical voices below.
- Benedict EvansStrategicIndependent analyst, formerly a16z · Weekly newsletter, plus a big presentation twice a year
In my view the best strategic read for executives. Benedict treats AI as a platform shift like the PC, the web and mobile, and keeps asking the business questions (distribution, moats, who actually captures the value) that most coverage skips. His twice-yearly deck is a very good board pre-read.
- AI as Normal TechnologySkepticalArvind Narayanan and Sayash Kapoor (Princeton) · Periodic, free
The most rigorous skeptical voice I follow, and a very useful counterweight to the hype. Their argument is that AI will diffuse through the economy over decades the way electricity did, which is exactly the frame an executive needs when deciding how fast to reorganize.
- Import AIResearchJack Clark (Anthropic co-founder) · Weekly, free
Research-forward and very good on policy and on China. Each issue picks a handful of papers and explains why they matter, and it is one of the longest-running newsletters in the field.
- The BatchPracticalAndrew Ng, DeepLearning.AI · Weekly, free
Calm, practical and builder-oriented. Andrew's opening letter is often the most useful part for founders thinking about where to build.
- Simon Willison's WeblogPracticalSimon Willison · Near daily, plus a newsletter
The best hands-on writer on agents and coding tools I know of, and very clear on their security risks (he coined the term prompt injection). If your teams are starting to deploy agents, this is where to see what can go wrong before it happens to you.
- The Machine Learning EngineerPracticalAlejandro Saucedo, Institute for Ethical AI & Machine Learning · Weekly, free
A curated roundup of articles and engineering write-ups on running machine learning and AI systems in production. It is technical, but it is a very good place to find how companies actually implement AI in the wild, which is often more instructive than the announcements.
- Exponential ViewStrategicAzeem Azhar · Weekly, free with paid tier
Good for the bigger picture of AI together with energy, geopolitics and the economy, which is the frame many boards are asking about.
- StratecheryStrategicBen Thompson · Several times a week, mostly paid
The strategy lens on the platform companies and what AI does to their business models. Worth paying for if you sit on a board or run strategy.
- Sangeet Paul Choudary · Periodic, free with paid tier
Sangeet, a co-author of Platform Revolution and author of Reshuffle, writes on how AI changes platform economics, competition and the way work is organized inside firms. Very useful for executives and product leaders working out where their company sits as AI shifts who controls the customer.
- Not BoringStrategicPacky McCormick · Weekly, free
Long essays on technology strategy and the companies building it. It is not only about AI, but much of it is, and the pieces are of exceptional quality. Packy is an optimist and an investor, so I would read him together with one of the skeptical voices.
When something big happens, such as a major model release or a safety incident, I also search Zvi Mowshowitz's Don't Worry About the Vase, which is extremely thorough (and quite long) and often the best single roundup available.
Research, data and policy
- Stanford HAI · Annual (spring)
The most cited annual data compendium on AI capability, investment, adoption, talent and public opinion. The short takeaways page is enough for most readers, and the full report is there when you need a chart.
- Stanford HAI · Ongoing
Erik Brynjolfsson's group, and some of the most careful empirical work on what AI is doing to productivity, jobs and wages.
- Nathan Benaich, Air Street Capital · Annual (October)
An investor's view across research, industry, politics and safety, and it grades its own predictions from the prior year, which I appreciate very much.
- Ongoing
Compute, training cost, hardware and benchmark trend data, very carefully sourced. When someone makes a claim about scaling I check it here.
- Ongoing
Independent evaluations of frontier models, including their measurements of how long a task AI agents can complete on their own, which is in my view the single most useful trend line for anyone planning around agents.
- Periodic
Usage data on which tasks and occupations people are actually bringing to an AI model. Read it knowing it is one company's data, but it is a rare view of real usage.
- World Bank · Released August 2026
The best single reference I know of on what AI means for developing economies. Its framework is a sequence (adopt existing tools, adapt them to local conditions, and only later advance toward building frontier models), and it is very clear that electricity, connectivity, data and skills come first. It also estimates that a much smaller share of jobs is at risk in low- and middle-income countries (4.5%) than in high-income ones (14.2%), which is useful context for anyone building or investing in emerging markets.
- Georgetown University · Ongoing
Policy analysis on AI, export controls, talent and U.S.–China competition, and it is what I read for the policy questions in my own research.
- Jeffrey Ding (George Washington University) · Weekly, free
Full translations of Chinese AI policy documents, industry reports and commentary that otherwise never reach an English-language reader, with Jeffrey's own notes on what is and is not significant. In my view it is the single most useful source for anyone whose business or research touches export controls or the China market, since so much of the Western coverage is written by people who have not read the original documents.
- Shakeel Hashim · Weekly, free with paid tier
A weekly roundup of AI policy from Washington, Brussels and London, from a small newsroom led by a former Economist news editor that reads the bills and the comment letters. It starts from the view that advanced AI carries serious risks, which I would keep in mind, and the reporting is careful. For an executive who needs to know what is actually moving in regulation, as opposed to what is being proposed, this is the one I would read.
Governance and risk
- European Commission · In force, obligations phasing in through 2028
The first comprehensive AI law, and the one most likely to shape your compliance work whether or not you are based in Europe, since it applies to anyone whose system is used there. The rules for general-purpose models already apply, and after the July 2026 Digital Omnibus the high-risk rules arrive in December 2027 and August 2028. The Commission's AI Act Service Desk and the independent AI Act Explorer (artificialintelligenceact.eu) are the easiest ways to work out which risk tier a product falls into and which deadlines apply.
- U.S. National Institute of Standards and Technology
The reference framework most U.S. companies use to organize AI risk (govern, map, measure, manage), with a companion profile for generative AI. It is being revised, and the current version is still the common vocabulary.
- OWASP Gen AI Security Project · Updated periodically (2026 edition)
The security risks of building on language models, ranked, with prompt injection at the top and excessive agency moving up in the 2026 edition. OWASP also keeps a separate Top 10 for agentic applications, which is worth a look if you are deploying agents with tools and memory.
- Responsible AI Collaborative
A searchable record of real-world harms and near-misses from deployed AI systems. Before a rollout I would very much recommend searching it for your industry and use case, since it is the cheapest way to learn from other people's mistakes.
Closer to the source
Model announcements, research posts and system cards. The system cards are long, and the sections on evaluations and safety testing are the ones worth reading.
Product and research announcements, with system cards linked from each model release.
Gemini releases as well as the science work (AlphaFold and its successors), which is where some of the most consequential applications are.
The main source for Meta's open-weight models, which matter a lot for startups building on models they can run themselves.
- Nathan Lambert (formerly Ai2) · One to three times a week, free with paid tier
The best regular read on open-weight models, which are what most of the startups I work with build on. Nathan led post-training on Ai2's open Olmo models, so he is very good on which releases matter, what the licenses actually allow and how the gap to the closed models is moving.
- Daily
A community-curated list of the day's notable papers. Skimming the titles for five minutes is the most efficient paper triage I know of.
arXiv papers with discussion, trending lists and AI-generated overviews, which makes a dense paper much faster to get through.
The raw feed. Too much to read directly, so I come here from one of the two triage tools above.
For the governance material in my talks, the METR and Redwood Research investigation of the July 2026 incident, in which OpenAI agents coordinated an attack on Hugging Face, is very well worth reading in full. Let me especially recommend its limitations section, since the authors are candid about what their short, scoped investigation could and could not establish (including that the AI agents they used to help analyze the logs were sometimes wrong), and that kind of honesty about evidence is exactly what the press coverage tends to drop. I wrote about what the incident suggests for governing agents in The Operator Was the Last to Know.
Reading the research
- Allen Institute for AI (Ai2) · Free
Run by a nonprofit, with no paywall and a free API, and its research feeds are a very good way to get alerts on new work in an area. Its AI features are fairly light, but in my view it is the most solid free place to start.
- A few free searches a month, paid tier
An agent that spends several minutes on a single search, reading papers and following their citation trails until it stops finding new ones. It is slow, and it is very good at turning up the niche papers that keyword search misses.
- Litmaps · Free with paid tier
You give it a few papers you already know and it builds a map of the work around them, which is how I would get my bearings in an unfamiliar literature. Litmaps acquired it in 2025.
- Free with paid tiers
The best fit I know of for a structured review. You choose the columns (sample, method, main finding) and it fills in a table across dozens of papers, which you then check against the papers themselves.
- Free with paid tier
Good for a quick read on what the evidence says about a specific question, with filters for study type such as meta-analyses and randomized trials. Its yes-or-no evidence meter is a useful starting point, and I would always open the underlying studies before repeating it.
Shows whether later papers support or contradict a claim, which is well worth checking before you repeat a striking finding to a board.
- Google · Free with paid tiers
Answers only from the documents you upload, which makes it the safest of these tools for working through a stack of reports or papers you have already chosen. Google renamed it in July 2026, and the audio overviews are surprisingly good for a commute.
Let me add one caution. All of these tools, and the deep research modes in Claude, ChatGPT and Gemini as well, still sometimes describe a paper wrongly or cite one that does not exist, so I would never cite or repeat a finding from a paper I have not opened myself.
Doing research with AI
- Claude Code, OpenAI Codex and similar tools
In my view this is the biggest change of the past year for quantitative researchers, and still very underused. In a 2026 Anthropic survey of about 1,260 social scientists (not a representative sample), 81% had tried chatbots but only 20% were using coding agents, and a VoxEU column by Galiani and colleagues found that letting a model run its own code raised the share of correct answers on 40 econometric tasks from 79% to 97%. The errors that remain are scripts that run cleanly and return the wrong number, so you still have to check the output. The link goes to Anton Korinek's NBER paper on AI agents for economic research, which is the best practical guide I know of.
- John Horton and colleagues · Open source (MIT license), with a hosted platform
A Python library for running surveys and experiments on AI agents with different personas, across many models at once, with results cached so the runs can be reproduced. In my view it is most useful for piloting question wording and experimental designs before fielding them to real people. Its own documentation says the answers reflect patterns in the training data and recommends validating against human data, which matches what I have seen with AI "customer interviews" in markets where that data is thin.
- Ben Golub and Yann Calvó López · Paid, per review
Detailed, referee-style reports on a draft, built by two economists. John Cochrane (a generally exacting critic) wrote that its reports are in the top 5% of the referee reports he has read, and the American Economic Association and the Econometric Society now use it for technical checks at the conditional acceptance stage. It still has trouble reading tables and figures, which matters for empirical papers.
Fast feedback on methods, statistics and the literature, tied to specific passages of the manuscript, and it also checks for fabricated references, which is useful now that AI-written citations are turning up in submissions.
- Free for researchers
Breaks a paper into a tree of its claims and the evidence for each one, and compares the novelty claim against similar papers. In my view this is especially useful when co-authors are not quite agreed on what the paper is actually arguing.
- Replication Games held regularly, plus an AI replication engine in development
Runs one-day Replication Games, in which teams rerun the code behind published studies, and is building an AI system to do the same at scale. Their 2026 PNAS paper (Brodeur and colleagues) found that teams assisted by AI did better than teams led by AI, though no better than human-only teams, which is one of the few careful experiments I know of on how far to trust AI with this kind of checking.
- Chenglei Si, Tatsunori Hashimoto and Diyi Yang (Stanford) · ICLR 2026
An earlier study by the same team found that expert reviewers rated research ideas written by an LLM as more novel than ideas from human experts. In this follow-up, 43 researchers each spent about 100 hours carrying out a randomly assigned idea, with the source hidden from them and from the reviewers, and once the results were in, the scores of the AI ideas dropped significantly more than those of the human ideas on every measure, so that on many measures the ranking flipped. In my view this is one of the most important studies for anyone using AI to brainstorm, whether for a paper or a startup, since it shows how much an idea that sounds novel can fall apart once someone has to execute it, which is very much what I see with founders as well.
- Chenglei Si, Zitong Yang, Yejin Choi, Emmanuel Candès, Diyi Yang and Tatsunori Hashimoto (Stanford) · ICML 2026
The natural next step after the study above. The team built a system that automatically implements AI-generated ideas for training language models and runs the experiments on GPUs, so that the model learns from actual results. An evolutionary search, which keeps the ideas that work and builds on them, produced a post-training method that clearly beat a standard baseline (69.4% against 48.0%) as well as a better pre-training recipe. Training the idea generator directly on the experimental rewards, however, led it to converge on a narrow set of simple ideas, which raised its average results and left the quality of its best ideas unchanged. In my view this is a very useful lesson for anyone hoping AI will find the breakthrough idea, since feedback from real execution helps a lot, and optimizing hard for it can also squeeze out the more ambitious ideas.
Before uploading unpublished work, I would read each tool's terms on how it stores and uses manuscripts. And if you are refereeing, many journals and publishers do not allow you to upload someone else's submission to an AI tool, so I would check the journal's policy first.
For a look at where this may be heading, two things are worth following. The first is Agents4Science, a 2025 conference organized by James Zou's group here at Stanford in which AI agents were the primary authors and reviewers, with all of the reviews public on OpenReview, and the organizers' write-up is a very concrete account of what the agents did well and where they failed. The second is the "AI scientist" systems, such as Kosmos from Edison Scientific (spun out of FutureHouse), which runs for up to 12 hours of analysis and literature search on a dataset you give it. Independent scientists rated about 79% of the statements in its reports accurate, and the company itself says it sometimes chases findings that are statistically significant and scientifically irrelevant. For now these are mostly aimed at biology, but I expect versions for the social sciences fairly soon.
New methods in social science
- Mark Verhagen, Matthew Salganik and colleagues (Princeton)
Turns population register data into a written narrative of each person's life across work, family, health and place, so that language models can read it. The team has generated over 100 million of these books from Dutch registry data and released an open-source toolkit (BOLT), and the longer-term aim is a foundation model of life trajectories, which they describe as a move from a sociology of variables to a sociology of events. It is still a proof of concept (the paper is on arXiv), and it grew directly out of Salganik's Fragile Families Challenge, which is part of why I find it so interesting.
- Germans Savcisens, Sune Lehmann and colleagues · Nature Computational Science, 2024
The Danish forerunner, which treats each of six million lives as a sequence of events (much like the words in a sentence) and predicts outcomes such as early mortality better than standard models. It was widely misreported as a "death calculator," and the authors' own site is very clear that it is not publicly available and cannot do what much of the coverage claimed.
- Luke Hewitt, Ashwini Ashokkumar, Isaias Ghezae and Robb Willer (Stanford) · Nature, 2026
GPT-4, prompted to simulate representative samples of Americans, predicted 469 treatment effects from 70 preregistered survey experiments about as well as pooled human forecasters, and it held up on unpublished studies that could not have been in its training data. It does tend to overestimate effect sizes. The authors' free demo (linked below) lets you try a treatment message yourself, and in my view this is a very good companion to Expected Parrot for piloting a study.
- Joon Sung Park and colleagues (Stanford)
AI agents built from two-hour interviews with about 1,050 Americans reproduced those people's General Social Survey answers about 85% as well as the people reproduced their own answers two weeks later. The agents built from interviews also had smaller accuracy gaps across racial and ideological groups than agents built from demographics alone, which matters a lot if you are tempted to simulate respondents from a profile.
- Marcel Binz and colleagues · Nature, 2025
A language model fine-tuned on 160 psychology experiments with over 60,000 participants, which predicts behavior in new experiments better than the existing cognitive models. It is limited to experiments that can be described in text, and the authors present it as an instrument for researchers to use.
- Jens Ludwig and Sendhil Mullainathan · Quarterly Journal of Economics, 2024
Perhaps the most transferable idea on this list for management and entrepreneurship research. They trained a model to predict judges' detention decisions from mugshots alone, then had people work out which facial features drove the predictions, which surfaced two new hypotheses (being well-groomed and being heavy-faced) that could then be tested in the usual way. The method applies to any rich data, such as satellite images, online behavior or company filings, and I think it is very underused in our field.
You can try the experiment-prediction model at treatmenteffect.app. Let me also add one note of caution that runs through all of this work. In the Fragile Families Challenge, 160 teams with a very rich dataset could predict basic life outcomes only slightly better than a simple model with a handful of variables, and that humbling result is a large part of what motivated the Book of Life. So good prediction is a very useful starting point for finding hypotheses and piloting studies, and the explaining still has to be done the usual way, with designs that let you identify causes.
Scoreboards
Independent comparisons of models on a composite intelligence index as well as speed and price, which is what you need when choosing a model to build on.
Rankings from blind head-to-head human votes. Good for overall feel, and small gaps at the top are usually statistical ties.
Listening
- swyx and Alessio Fanelli
The AI engineer's podcast. Technical, but the best way to hear what builders are actually doing with agents right now.
- Dwarkesh Patel
Very well-prepared, long interviews with lab leaders, researchers and economists about where this is all going.
- Sarah Guo and Elad Gil
Two investors interviewing founders and researchers, so it is the most directly useful of the three for entrepreneurs and investors.
Let an agent do the first pass
Most of the major assistants (Claude, ChatGPT and Gemini) can now browse the web and run a task on a schedule. Let me suggest the following setup, which takes about ten minutes:
- Copy the prompt below and fill in the bracketed parts with your role, industry and two or three priorities. The more specific you are, the better the filtering.
- Paste it into an assistant that supports web search and scheduled or recurring tasks, and set it to run weekly.
- Read the briefing, and for anything you plan to act on, click through and read the primary source yourself.
- After a month, tell the agent which items were useful and which were noise, and ask it to revise its own instructions.
You are my weekly AI briefing analyst. I am a [ROLE, e.g. founder / CEO / board member] in [INDUSTRY], and I care most about [2-3 PRIORITIES, e.g. customer support automation, AI policy in the EU, competitors' AI products]. Every [Monday morning], do the following: 1. Scan what was published in the past 7 days from these sources: - Newsletters: One Useful Thing, Import AI, The Batch, Exponential View, Interconnects, Transformer - Lab announcements: Anthropic, OpenAI, Google DeepMind, Meta AI - Data and evaluation: METR, Epoch AI, Stanford HAI, Artificial Analysis - Papers: Hugging Face Daily Papers (top items only) - Anything else clearly major in AI news this week 2. Pick at most 5 items that matter for someone in my role and industry. Skip funding rounds, minor product updates and opinion pieces unless they change what I should do. 3. For each item, give me: - One sentence on what happened, with a link to the primary source (the lab post, paper or report itself, not a news article about it) - Why it matters for my business, in two sentences or fewer - A confidence note: is this a measured result, a company claim, or speculation? 4. End with: - "One thing to try this week": a specific task from my work I could test on a current AI tool, with a suggested prompt - "Watch list": anything still unconfirmed that I should check again next week Rules: keep the whole briefing under 600 words. Never summarize something you could not open and read. If the sources disagree, say so. If it was a quiet week, say that and keep it short.
One caution. Agents still summarize confidently and sometimes wrongly, which is the same issue the METR and Redwood team ran into when they used AI to help analyze the incident logs. So I treat the briefing as a filter that tells me what to read, and I never treat it as the reading itself.
Stay close to the people building it
Much of what I know about how AI is really being used comes from the founders and students I work with, and I would encourage any executive to find some regular way of spending time with early-stage teams. There are several ways in, and they range quite a lot in how much time and money they take:
- Mentoring and office hours. Accelerators, university programs and peer networks are always looking for experienced operators, and an hour a month with a few founding teams will show you very quickly which AI workflows are working and which ones are not.
- Executive peer networks. For executives, comparing notes with peers at other companies is often the fastest way to learn how AI adoption is actually going. The Mercury Company is a good example, a community of CIOs and CTOs whose Mentor Exchange pairs technology leaders with mentors from outside their own organizations, and I have very much enjoyed speaking at their Annual Leadership Summit (most recently in September 2026, on trust, safety and governance in the agentic enterprise), where the conversations with the technology leaders in the room are as informative for me as anything I present.
- Judging and demo days. Pitch competitions, hackathons and demo days are a low-commitment way to see dozens of teams in an afternoon, and in my experience the patterns across a whole cohort are more informative than any single pitch. Serving on the advisory committee of the HAI Startup Series, where Stanford-founded AI companies at Series A or earlier come for feedback from faculty and investors, is one of the ways I do this myself.
- Advising. A formal advisory role with one or two startups gives you a much closer view of how the product, the customers and the team are changing as the models improve.
- Being an early customer or design partner. For executives this is often the most underused option, since piloting an AI startup's product inside your own company teaches you about the technology and about your own organization's readiness at the same time.
- Angel investing and investing as an LP in venture funds. Reading pitch decks and sitting in on fund updates is a very good window into where capital and talent are flowing. An LP position is usually a more distant view, with quarterly letters and an annual meeting, so it works best combined with one of the more hands-on options above.
Two cautions. Founders are, understandably, optimists, and the companies you hear from are not a random sample, so I would read what you see in the ecosystem alongside the more skeptical sources on this page. And angel and fund investing carry real financial risk, are generally limited to accredited investors, and deserve their own due diligence and advice, so I would treat the learning as a side benefit of investing.
Working with AI as an entrepreneur
From me
- Substack essay · September 2026
What the July 2026 Hugging Face incident, Lisanne Bainbridge's 1983 paper on the ironies of automation, and our 2024 Nature audit of programmatic advertising suggest about governing AI agents.
- Slide deck · STVP event at Google
How the agentic AI moment changes what we teach in entrepreneurship courses, and what becomes newly important.
- Project announcement · Stanford Accelerator for Learning grant
Using AI to let students rehearse skills, such as problem validation, that can otherwise only be learned against a real expert.
- Substack essay · May 2026
On AI-era revenue metrics and what would actually fix the measurement problem in private markets.
My own routine
I read one or two newsletters a week, skim the paper feeds for a few minutes when I have them, and once a month I do a deep read of a new system card or a major report. Most of all I try the new agentic tools on my own real work, since that calibrates my sense of progress better than anything I can read. One way to make this a habit, which I would recommend, is to keep a short, fixed set of tasks from your own week and rerun them whenever a new model comes out, so that you build up a very concrete sense of what has changed.
I should also say that I am lucky to work in a very information-rich environment, and a good share of what I learn comes from seminars and conversations in person. I regularly attend the Stanford Digital Economy Lab seminar, events at the Stanford Center for AI Safety and the HAI seminars, the HAI Startup Series, where I serve on the advisory committee (early-stage AI companies with Stanford founders present to a panel of faculty and investors, which is a very good window into what is being built right now), the King Center on Global Development's events on AI, such as its day on the World Development Report 2026, as well as conferences such as the Yale Institute for Foundations of Data Science (FDS) meetings and the information systems and management conferences (ICIS, WISE and the Academy of Management). Perhaps most useful of all, simply asking my colleagues here in the Stanford School of Engineering how they are using AI and what they are working on is always informative, and I would encourage anyone reading this to find their own version of that, whether it is a few colleagues at another company, a university seminar that is open to the public, or the startup ecosystem described above.
If you have a source you think belongs on this list, I would very much like to hear about it through the contact page.