Yep, known as Bloom's Two Sigma Problem[1]. Like most hard problems we know the solution, but lack the appetite to implement.

1. https://en.wikipedia.org/wiki/Bloom%27s_2_sigma_problem

Hey Bloom’s 2 sigma problem. So far, (nearly) all conversations about education on HN I’ve seen, have had a naturally point at which Bloom’s 2s should be introduced.

Humanity is now preparing students with a 20 year time horizon, while tech changes much faster. If this was agriculture, the industry would be doomed by that horizon mismatch.

We really need more teachers, if we want the median citizen to be better off.

The study picked an artificial and useless proxy.

What did tutored students go on to do? Were they over represented in Nobel Prizes, hedge fund billionaires, heads of state?

Or did they do well on a meaningless test and then forget all about whatever they “learned” just like everyone else?

The entire field is absolutely littered with this problem. Everyone is targeting cholesterol and not all cause mortality.

What evidence would change your mind?

Strong correlations to actual outcomes. That can be career outcomes or otherwise but actual outcomes people care about. Not test scores.

We have a zillion natural experiments because there are so many schools in the world and within many of them teachers given a lot of leeway.

Does any of it make a damn bit of difference controlling for everything else? We have Waldorf schools teaching woodworking and cram schools shoving AP courses at eighth graders—-that’s a large difference, what do life outcomes show?

That's fair. It helps knowing what would change things. Thank you for sharing! appreciate it

With AIs as most staff, this should be very reachable.

With AIs as teachers, I disagree. But with AIs assisting routine grading, filling in the university's assessment_framework_draft_v3_final_FINAL.docx, and otherwise freeing up time to actually focus on students - maybe? Although I fear that any productivity gains will be swallowed up by further reductions in lecturer headcount...

AI lacks both the reasoning and insight needed to teach anybody that isn't already immensely interested in the topic, and even then might leave large knowledge gaps, not to mention how often it hallucinates wrong knowledge. Especially with topics that already have a lot of bad information floating around.

Perhaps. For now, one of my one-on-one tutoring sessions (in real analysis) this semester consisted mostly on un-teaching a student a bunch of wrong crap they "learned" from ChatGPT.

I'm thinking of custom built Teacher AIs, trained in how to teach different kinds of students, and with well defined curriculums.

Even if they're not better than the best human teachers, I think being able to "personally" interact with each student 24/7 will be a huge improvement.

There are some early products around:

- https://www.khanmigo.ai/

- https://squirrelai.com/