After almost every talk I give these days someone comes up afterward and asks me some version of the same question, which is how anyone is supposed to keep up with AI while also running a company or a team. It came up several times again at the Mercury Mentor Exchange Leadership Development Summit in Laguna Beach last month, and since I had been answering it one email at a time, I finally put the sources I follow, with a short note on each, on a new page on my site: Staying current on AI. Let me use this post to walk through how I think about it and how I would use the list, since a list on its own only gets you part of the way.
Let me first put this in the context of lifelong learning, because in my view keeping up with AI is mostly a question of habit. Charlie Munger used to say that in his whole life he had known no wise people who didn’t read all the time, “none, zero,” and that people would be amazed at how much Warren Buffett reads. The writer Michael Simmons has popularized a more modest version he calls the five-hour rule (an hour a day of deliberate learning, modeled on Benjamin Franklin), and although Simmons himself says it is not an academic study, I find it a very useful heuristic.
In my view there are four habits that matter. The first is simply protecting the time. For most executives I talk with an hour a day is not realistic, so I would suggest starting with one protected hour a week, put on the calendar like any other meeting, and spending about 20 minutes of it reading two newsletters, roughly half an hour trying a current AI tool on a real task from your own week, and the last few minutes writing down one thing you want to change or try as a result.
For the two newsletters I would pick one practical voice and one more strategic or skeptical one. On the practical side Ethan Mollick’s One Useful Thing is very good on what the tools can actually do inside an organization, since he tests them himself before writing, and Andrew Ng’s The Batch is a calm and builder-oriented alternative. Ethan is an enthusiast, however, so I would pair either one with Benedict Evans, who in my view is the best strategic read for executives and keeps coming back to questions about distribution, moats and who actually captures the value, or with AI as Normal Technology from Arvind Narayanan and Sayash Kapoor at Princeton, which is the most rigorous skeptical voice I follow. If you want more on research and policy, Jack Clark’s Import AI is excellent as well and has been running for many years.
Once a year I would also skim the takeaways from the Stanford HAI AI Index in the spring and the State of AI Report in October. In addition, if you want more structure than a reading list can give you, it is worth 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 taught by Stanford faculty, and having deadlines and classmates is often what keeps the weekly hour from slipping.
The second habit is going to the primary source when something matters. The labs’ own posts and system cards from Anthropic, OpenAI, Google DeepMind and Meta often say a great deal more than the press coverage does, including about limitations and safety testing, and for research papers I use Hugging Face’s Daily Papers and alphaXiv to triage in a few minutes what would otherwise be an unmanageable arXiv feed.
For instance, a good deal of the governance material in my talk drew on the METR and Redwood Research investigationof the July incident in which OpenAI agents coordinated an attack on Hugging Face. Most of the coverage focused on the headline findings, which were striking enough on their own. The part I would most encourage leaders to read, however, is the limitations section, where the authors are very candid that it was a short investigation with a narrow scope and that the AI agents they used to help analyze the logs were sometimes wrong or overconfident, which is extremely useful for anyone deploying agents to know and which most of the press coverage left out.
A third idea, and a newer one, is to use the tools themselves to do the first pass. Most of the major assistants can now search the web and run a task on a schedule, so on the page I have included a prompt you can copy, fill in with your role, your industry and two or three priorities, and set to run every Monday morning. It asks the agent to scan the past week’s newsletters, lab announcements and evaluation reports, pick no more than five items that matter for someone in your position, link each one to its primary source and say whether it is a measured result or a company claim, and then suggest one specific task from your own work to try on a current tool that week.
Let me also be clear about how I would use something like this. I treat the briefing as a way to decide what is worth reading, and for anything I plan to act on I click through and read the source myself, for the same reason the METR and Redwood team flagged. After a month or so it is also worth telling the agent which items were useful and which were noise and asking it to revise its own instructions, which is good practice in managing an agent as well.
Last is the habit I think matters most, which is to keep trying the new tools on your own real work. One way to do this that 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, since that gives you a much more concrete sense of what has changed than benchmark charts or leaderboards do. For a founder the set might include working through a backlog of customer emails, and for an executive it might be a first pass at a board memo. It takes perhaps half an hour a month, and I don’t know of a better way to calibrate your own sense of how fast things are really moving.
I talked about some of this in a short interview with Carol Yu at InnoX, including the good and bad ways I see founders working with AI and why, in my view, you want it to be your intern and not your CEO, since it has never been so fast to run in the wrong direction.
Overall, the page on my site has the full set with a note on each, including research and policy sources such as the Stanford Digital Economy Lab, Epoch AI, METR and CSET, the lab and paper feeds, a couple of model scoreboards, and a few podcasts for the commute (Latent Space, Dwarkesh Patel and No Priors). I will keep it updated as sources come and go, and if there is one you think belongs on it I would very much like to hear about it. I look forward to hearing what works for you.
Chuck