Recent pieces on AI and research

A quick post to recommend a few articles I’ve enjoyed recently. They come at the AI-and-academia question from different angles, and they are all interesting.

Andy Hall’s AI is already 10x-ing academic research argues that AI tools have already accelerated research productivity by an order of magnitude — not by replacing researchers, but by pairing human expertise with AI capability. He is concrete about what getting to 100x would require: benchmarks for measurable questions, open research repositories, and reward structures that actually incentivize building on each other’s work.

One way Hall’s argument is already playing out is in the tooling around research itself. Refine.ink is a paid paper-review service that has (rightfully!) gotten a lot of attention, but there is now at least one free alternative, along with several other paid versions. The /review-paper skill I’ve posted is one example, and every “Claude Code for academics” package seems to come with its own version.

There is a second channel through which AI is changing research, alongside the direct productivity gains Hall describes. A decade ago, building a tool that reads a paper and returns structured feedback required a team. Today, one researcher with something like Claude Code can prototype it in a weekend. If we take a step back to the old times (5 months ago, just before Christmas), this is astonishing. A lot of software can be built by the person who actually needs it. And the possibilities for customization are almost endless: I changed my review-paper skill to also provide feedback on grants and pre-analysis plans with one prompt. Someone took my skills and made a version for condensed matter physics. That is mind-blowing.

Another intriguing thought in the piece is what a research paper will look like in the future. I tend to think that proprietary data and unwillingness to share all the secrets will limit some of the changes Andy is talking about. But I also think that there will be changes, and that those are probably for the better. Paul Goldsmith Pinkham has a very interesting proposal on “LLM-friendly papers”. I plan to implement that one for my own papers.

Alexander Kustov’ Academics need to wake up on AI is the third in a series, and all three are worth reading. The claim that stuck with me is that most slop is human: the complaints about AI producing low-quality writing often describe a failure mode that was already endemic in academic work and conferences. This changes the question from “will AI ruin academic writing?” to “will AI raise the floor on writing that was already bad?”

I especially agree with one of his closing points: “The best work happens when humans and AI collaborate.” It is great to have a collaborator that can answer all your questions and help you with the project. Now I don’t have to find an econometrician or bother my colleagues any more when I have a question about something. I now have lots more free time to bother them about something else.

Alex Imas What will be scarce applies some economics to AI. As AI makes commodity production cheap, the scarce sector becomes the “relational” one (education, care, arts, hospitality, universities(?)) where the human element itself is the value. This is a more productive frame than the usual debate about whether AI will take all the jobs, and it has direct implications for how we think about academic labor. A very good post.

This piece is also a nice illustration of something that I really like about economics: Economic theory is genuinely valuable, and equilibrium thinking is one of the core skills the field has developed. Alex Imas does not predict what happens to jobs by extrapolating current trends, but instead works out what has to happen in equilibrium when one sector gets drastically more productive and incomes rise. That kind of reasoning comes up constantly. In housing, which I know best, almost every important policy question — rent control, zoning reform, homeownership subsidies — turns on equilibrium effects that partial-equilibrium intuition tends to miss. Giving homebuyers more money will probably not help them buy a property, and may instead raise prices (oops). Economic theory and equilibrium thinking are very valuable.

Dylan Matthews’ The AI people have been right a lot is a reflection on his earlier dismissal of AI-safety concerns and an argument for intellectual humility toward predictions that sound extreme. Worth reading even if you are already AI-curious, as a reminder that the people who have taken AI seriously the longest have mostly been right.

Read all four. They pair well: the first two make the case that AI is already a useful collaborator for research, and the next two tell us what that means — that economic reasoning still does the work of sorting out what will be scarce, and that the people who have been taking AI seriously the longest deserve a careful hearing on where this is going.