Research training outside the university. Is it worth doing?

The basic shape

A few weeks of intensive, hands-on work, then a supervised research project, ending in a feedback conference. There's no degree or accreditation at the end. What you leave with is the work itself, and the word of the people who supervised you.

Students and career-switchers work in small groups with active researchers. The closing conference matters most to me: people present, and their work gets read carefully and critiqued in detail by people who know the field. One track has researchers read each other's papers in advance, instead of sitting through slides.

You wouldn't be paying for content. It's basically free now, and AI tutors deliver it well and on demand. What you'd be paying for (or be sponsored for) is the supervision, the honest feedback, and the people you'd meet.

Why I wouldn't just run the original sketch

• 01 AI commoditized the mechanics I'd have taught. Language models now write the code, clean the data, run standard analyses, and draft the literature review, well and cheaply. Teaching those mechanics as the core offering would train people for work that's evaporating. Understanding what the methods mean is a different thing, and it's the part that got more valuable.
• 02 The entry-level quant job market contracted. Employment for young workers in AI-exposed roles fell sharply from 2022 while older workers in the same occupations grew; data-analytics postings dropped more than almost any other tech category. How much of this is AI versus a broader hiring slump is contested. Either way, "learn data science, get a job" is no longer an honest pitch.
• 03 Polished output no longer signals skill. When anyone can produce clean code and a tidy writeup, those artifacts stop telling employers or PhD programs anything. What still signals: a detailed letter from someone who watched you think, and a project whose judgment calls were visibly yours.
• 04 What got scarcer is judgment. Choosing questions worth answering, designing analyses that identify something, catching the confident errors in AI output, knowing when a model is wrong. There's decent evidence AI tutors teach content well, and evidence that leaning on AI without guardrails makes you worse. Judgment is the thing a program like this could still teach.

What might be worth teaching now

None of this means dropping statistics. If anything you'd learn quantitative methods more deeply than a typical course teaches them, but for a different purpose: understanding what a result means and when it's wrong. Deriving it by hand or typing out the analysis is the machine's job now.

What already exists, honestly

A lot has appeared since I first sketched this, though less of it overlaps than you'd think. The well-funded programs (MATS, Astra, BlueDot, ERA, PIBBSS) do technical AI safety and machine learning. That's a different discipline from quantitative social science. GovAI runs excellent mentored research, but only on AI governance. The AEA Summer Program is strong, but US-only and built around coursework. SICSS is two weeks and methods-focused. The programs that do place social-science students, Effective Thesis among them, tend to be light-touch, which I know first-hand.

The closest thing to what I'm describing was PREDOC's summer course: big-data social science with a real capstone. It was discontinued after 2024 and replaced with self-paced online courses. So the in-person, high-feedback, mentored model for quantitative social science is quietly contracting, just as the tools get better and the reasons to want it grow. Nobody runs the feedback conference at all.

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. It's also why I'd keep the scope broad: more ways to find something that works.

Who might want this

• People weighing a research career: late undergrad through early PhD, who want a mentored project and a real reference before committing years.
• Domain experts who want to model: scientists and practitioners who hold the knowledge a good model needs and want the quantitative toolkit to use it.
• Quantitative people aiming at high-stakes questions: economists and modelers working on policy, global priorities, or the questions around advanced AI.
• Strong candidates outside the usual pipelines: including people from under-resourced countries, chosen by careful screening instead of credential filters.
• Researchers who just want the feedback conference: including PhD students after finishing-school comments on the paper that matters.

Where this comes from

This wouldn't be starting from scratch. Some of the pieces already exist at a smaller scale: modeling workshops on contested quantitative questions, and supervising early-career researchers through fellowship placements. Others I'm hoping to try soon, like Fermi-estimation and parameter-elicitation sessions. The open question is whether they add up to something worth doing as a program.

I co-direct The Unjournal, which commissions public expert evaluation of research. The overlap in spirit is obvious: careful reading, structured feedback, open science. But this would be a separate, personal exploration, not an Unjournal program, and nothing here speaks for that organization.