Economic misalignment in personal AI agents is a serious safety issue. We document it in 325K experiments across 13 agents. You delegate a specific action to your personal AI agent and give it access to relevant personal information so it can take the best action for you. But then it infers your wealth from this info and gives you the pricier option, *even when you explicitly ask for the cheapest option*. Such a blast working with the amazing @niloofar_mire, @AmanPriyanshu6 and @SupritiVijay See Niloofar’s great thread below and paper here: arxiv.org/abs/2609.24927 And hope it makes you read again Shakespeare!

Today we are launching the Innovations in Economic Measurement (IEM) Lab, part of the HBS AI Institute. We use AI, big data, and economics to build measures that update continuously and improve through use. Our first public project launches tomorrow. iemlab.org

"Waymo went from a non-driving internal concept to completing an autonomous route in less than a year when the project started in 2009." 17 years Waymos operate in 15 cities! Turning to RSI, the remaining 74% of R&D work is going to be much harder/uneven/weirder than the first 26% and I don't think these data tell us if AI can meaningfully speed AI R&D over the next 17 years.

Today @Google, we published work led by @m_codreanu on AI in Science, drawing on 15 million Gemini interactions, 2,600 specialised AI models & a 600-scientist survey. What did we find about how AI is changing science? Read on...

A situation that would have been unimaginable even months ago: there's a debate over whether an AI-assisted student math marathon is good for the field. 🤯 As a former student math researcher myself who's mentored such students for a decade, I want to weigh in with an offer, not just an opinion: **I'm volunteering to mentor at least three Mathathon teams as they work to turn their projects into rigorous, complete mathematical contributions. I invite other mathematicians to join me.** Working with high school and college math research students has been one of the greatest privileges of my career. They're often learning the background while trying to solve the problem—but aren't we all? And while their first-draft proofs aren't always journal-ready, that's precisely what mentors are for. Whether you see AI as a boon, a threat, or both, hundreds of students spending a weekend doing math should be an opportunity—certainly not a burden for the field as some have suggested. Telling students they shouldn't math because they aren't sufficiently professionalized is counterproductive. If we're concerned about incomplete or hard-to-read proofs, we should treat that as a teaching opportunity: let's help them (and other aspiring mathematicians) learn how to take their work up to the next level—and we'll probably learn some things from them in the process, too. Conversely, if we aren't willing to coach them, we have little leverage to try to dictate what their output should look like. **The point isn't to lower math's standards. It's to help students level-up to meet them.** So: at least three teams from me, QED. I hope others will offer to mentor as well. (I'm not directly in contact with the Mathathon organizers, so I'd appreciate a connection to help coordinate.)

My group (BU-Strategy & Innovation) is hiring! Please see job post and share with your students and colleagues!

Our weekly blog links are back, with lots of conference calls, the importance of listening to the subjects of development research, how to not make your AI RCT outdated before it is even published, and more...

In July, I spent a week at Harvard training researchers at Opportunity Insights on agentic coding and helping them develop a comprehensive AI strategy. Part of what made this tricky is that OI's research production function is much different from the average economist: > Large research teams with varied individual responsibilities > Most research involves restricted data (can't use frontier models in RDCs) > Work doesn't end at publication - communicating results to the public is just as important (e.g. Opportunity Atlas) The general approach I took with OI involved two parts: 1. I lead a series of workshops on agentic coding with Codex, aiming to raise the organization-wide level of familiarity with the tools, and so both improve productivity and also facilitate more precise internal discussions on AI. 2. Before and during the week at OI, I had many meetings with different stakeholders - PIs, research staff, research translation, pre-docs - with the goal of crystallizing my understanding of OI's research production process. Really getting into the details is key - otherwise, you risk providing solutions which never get used because they don't fit an org's workflow. For example, consider the slide creation process. Rather than just delivering some generic skill, I started by asking *many* questions. > How do pre-docs get instructions from PIs? Slack, email, or just comments in slides in a common Dropbox folder? > Are the requested changes always clear? If not, who do they go to for clarifications? > Do they use Powerpoint, Beamer, or both? > Is it important that the slides adhere to a particular style? How is that style enforced? > How many iterations does it take to get to a final set of slides? Why does it take that many iterations? Sometimes the eventual solution is very simple, but asking the right questions and digging all the way to the bottom is critical. I wrote a case study describing how exactly I led this engagement - aieconomist.io/case-studies/… If you're similarly trying to make your policy/research organization AI native, happy to jump on a call and discuss how I can help! tidycal.com/aniketpanjwani/a…

Join us!

Wonderful post worth reading to understand how AI will change science beyond coding agents. Also, just a beautiful example of tacit knowledge, the value it contains and the challenges it creates, and how AI startups plan to create value from all sorts of crazy new types of data.

Many thanks to @carlzimmer at the @nytimes for spending so much time with us and capturing the "magic" of science so beautifully. "Success can be just as mysterious as failure. Some researchers consistently get experiments to work, earning befuddled admiration from colleagues. Scientists even have a special term for this gift: magic hands. The idea may come as a surprise to people who don’t spend their lives in labs. Science is not supposed to be magic. When scientists carry out experiments, they keep careful records, both in lab notebooks and later in published scientific papers. Other researchers use that information to repeat the experiment. But every scientist discovers sooner or later that essential knowledge is not necessarily written down."

We're leading @transfyrai’s $25M seed as they come out of stealth to build the observability layer for science. Much of scientific knowledge lives in the hands and instincts of the person at the bench, and it's lost when they move on. @AnnaMarieWagner and @rwegrzyn saw this firsthand at Ginkgo Bioworks and the Advanced Research Projects Agency for Health (ARPA-H). Transfyr's physical AI captures that knowledge as it's made so scientists, models, and robots can build on it, accelerating scientific progress. Congratulations to Anna Marie, Renee, and the entire team.

This is the right kind of "hill-climbing," right? wsj.com/tech/ai/silicon-vall…

Students/young researchers - If you're interested in AI/ML applied to bibliometric and science-of-science problems, and can work for a US-based startup, I'd love to talk to you about some cool @RefineDotInk projects Please DM me or email ben@refine.ink

OpenAI researcher @daveholtz reveals the Kenya small business owners study where giving AI to high performers made them better but giving it to low performers made them worse: "I'm a co-author on another paper where we look at small business owners in Kenya. We give access to an AI mentor to see if it improves the performance of their business." "Going into this study we thought it would just make them better at running their businesses. What we see is that overall there's actually no detectable impact of giving people access to this AI tool." "But if you cut the data and look at people who are really successful business owners before we gave them AI versus people who are not successful business owners, the pre-treatment high performers did better when we gave them AI. And the people who were low performers before we gave them AI, they actually did worse once they had access to AI." "It seems to be the importance of judgment and expertise. Sometimes the AI output is really high quality, sometimes it's kind of low quality. Having the complementary skill of knowing when is the AI output good and when is it bad." "As you climb up that hierarchy, the importance of that complementary judgment skill becomes greater and greater." @daveholtz