An idea in progress · not (yet) a program
The idea: a few weeks of hands-on work and a mentored research project, run outside any university, with close feedback in small groups. You'd leave with work you can show and a recommendation from people who watched you do it, instead of a grade or a degree. I haven't decided whether it's worth doing, or in what form. That's what this page is for.
What I'd like from you: would you take part in, mentor, or fund something like this? If not, why not? There's a short form at the bottom.
You can also comment on any line: select it, or open the Hypothes.is panel at the top right. I read the comments and revise the page.
A few weeks of intensive, hands-on work, then a supervised project over a few months, ending in a feedback conference where people's papers are read in advance and critiqued in detail by people who know the field.
By decision modeling I mean building an explicit quantitative model to inform a specific choice: cost-effectiveness analysis, Fermi estimates, forecasts, and putting numbers on uncertainty, then defending them. Around that, the economics and statistics you need to build such models well and to read other people's critically.
You wouldn't be paying for content. Most of it is free, and AI tutors can teach it well on demand. What you'd pay for (or be sponsored for) is supervision, feedback, and the people you'd meet.
Rough numbers, all provisional. My 2020 sketch had 20–40 people doing three weeks of courses, about half staying on for a summer project, with fees around $5,000–6,000 on a sliding scale plus scholarships, and teachers paid around $5,000–10,000 for two to three weeks. A later, smaller version was six weeks for 8–10 people. Now I'd start smaller than either; see the first test below.
Statistics stays. If anything you'd learn quantitative methods more deeply than a typical course covers them, aimed at understanding what a result means and when it's wrong, and less at deriving it by hand or typing out the analysis.
Fermi estimation, cost-effectiveness analysis, forecasting and calibration: building a model of a contested question and defending the parameters. This builds on evidence and modeling workshops I've organized, and on formats I'm still developing.
Statistical and causal reasoning aimed at meaning: what an estimate and its uncertainty tell you, what a method quietly assumes, what would break an identification claim. Enough to catch the machine's mistakes and to read any study skeptically. Proving theorems isn't the point.
Directing AI tools on real research tasks, checking what they produce, and knowing what not to delegate. Also deliberate practice without AI on the fundamentals. In one large trial, high-school students who practiced math with unrestricted GPT-4 did 17% worse on a later exam taken without it than students who never had it, while a version with guardrails did no measurable harm. A newer experiment found AI helped undergraduates learn overall, but the lasting improvement in their writing went to those who used it to explain concepts; for those who let it write for them, the improvement vanished once it was taken away.
Applied microeconomics, behavioral economics, global priorities, policy analysis, and for those who want it, the economics and modeling of AI itself, one area that is clearly hiring quantitative social scientists. I'd keep the scope wide at first and see what people actually come for. If funding exists for one track and not the others, that may settle it.
A lot has appeared since I first sketched this, though less of it overlaps than you'd think. The well-funded mentored programs (MATS, Astra, ERA, the PIBBSS fellowship, and BlueDot's courses) are about AI safety and governance, some now with policy and forecasting tracks. GovAI runs excellent mentored research, but on AI governance. The AEA Summer Program is strong, but limited to US citizens and residents and built around coursework. SICSS is two weeks and methods-focused. Programs that place social-science students on thesis projects, such as Effective Thesis, are lighter-touch than what I have in mind.
For the economics of AI specifically there's now more: Anthropic's fellows program has an economics and policy track, and the CAIS AI and Society Fellowship recruits economists. Those look more like places people could go next than like competition.
The closest thing to what I'm describing was PREDOC's summer course: eight weeks of big-data social science with a real capstone project. It ran from 2021 to 2024; PREDOC now offers self-paced online courses. University-based programs for US undergraduates continue (PREDOC's EDE+ at Chicago Booth, the AEA program), but there's little for people outside that pipeline. The nearest template I know is psychology's Junior Researcher Programme: a one-week school, a 13-month team project, and a closing conference.
Feedback conferences exist inside academia: PhD conferences with discussants who read each paper in advance, and mentoring workshops like the AEA's CeMENT for junior faculty. I haven't found one open to people outside PhD programs (predocs, career-switchers, non-US students) and built around that kind of detailed, pre-read critique.
For many people that's the better route. A predoc is paid and mentored, and ends with letters from economists people know. But predoc places are scarce and competitive, and a master's takes a year, costs a lot, and is mostly coursework. Outside a university this could be shorter and cheaper, pick people on work samples instead of transcripts, take them from anywhere, and change what it teaches from year to year.
The costs are real too: no degree, no visa route, no brand. A recommendation from a program like this is only worth as much as the person writing it, so the whole thing depends on recruiting mentors whose judgment people already trust. That's probably the binding constraint, more than money or students.
I could be wrong about the gap, or right about it and wrong that anyone will pay to fill it. That's the main thing I'm trying to find out.
Before anything bigger, I'd want to try a small version, maybe in 2027:
If people come, finish, and I'd stand behind the letters I could write afterward, that's a reason to go further. If I can't find mentors, or nobody will pay or fund it, that's an answer too.
Some of the pieces already exist at a smaller scale: evidence and modeling workshops on contested quantitative questions, and supervising early-career researchers through fellowship placements. Others I'd like to try soon, like Fermi-estimation and parameter-elicitation sessions. The open question is whether they add up to something worth running as a program.
I co-direct The Unjournal, which commissions public expert evaluation of research, and those workshops ran through it. There's obvious overlap: careful reading and structured feedback. But this is a separate, personal exploration, not an Unjournal program, and nothing here speaks for that organization.
The ask
I'm collecting reactions before deciding whether to take it further. What I most want to know: