This should probably be 3 posts instead of one, but for now I’m going to go through three connected but separate ideas.

Also it’s not really been edited.  Sorry.

Progress Studies and Young Scientists

I spent a bit of the last year exploring some topics and questions in Progress Studies, and I came away with a few core (to me) ideas.

The first was that, given a long time horizon (50-100 years or more), it seems like most progress is scientific and technical progress.  I expect that this is a thing a lot of people disagree with (and many have disagreed with me in person about it).  Some of the biggest objections are:

• Social progress/population growth/other progress unlocks scientific progress by making it cheaper/more likely.  I think this is a valid point
• It’s hard to figure out a metric for progress and keeping score (assigning credit to specific things) is pretty subjective.  I think this is also a valid point, and I think a lot of people disagree with my assignment.
• On a relative-to-other-existing-people metric, scientific and technical progress increases inequality/other measure of wellbeing.  I think this is the objection I push the hardest against, since it does seem like given long enough time horizons, scientific and technical progress seems to have large and broad benefits.  (It’s also worth admitting that I care more about aggregates like “total human wellbeing” than relative metrics like “difference between the top 10th and bottom 10th percentile wellbeing”)

The second point I found was that across almost every place I looked, people participating in scientific and technical progress had problems with the way we pursue and fund these.  There’s tons of stories of research labs pursing the topics they thing the grant committees want, instead of the research they want to do.  Separately, there’s stories of grant committees only giving grants to boring research since no one is going for the breakthrough research they think needs to be done.  Even for funding programs specifically targeting breakthrough research, it seems difficult to match up funding with research projects.  It seems like there is at least a few inadequate equilibria.

The third point was that scientists are getting older.  The average age of Nobel prize winning research (counting the age at publication, not at award) is getting older.  The average age of grant recipients is getting older (this could be in part, but not exclusively, people retiring later).  The average age of principal investigators founding their own labs or research organizations is getting older.

I think this one is tricky, because there are some good reasons for this trend (e.g. if scientific and technical progress is getting more difficult because of Gordon-esque factors, then we’d expect the average age of science discoveries to increase).  However, I’m not convinced that this must be the case, and in particular I think we can make a pretty big (one-time) demographic push for science to get younger.

Mentor Young Scientists

When I ask myself: “What age do I think a person could make useful scientific or technical contributions to progress?” I get some pretty young answers.  My current best bet is that people aged 14-18 could do a lot, and possibly younger than that (but I am uncertain).

An important caveat here is that I don’t think *all* 14-18 year old people could be making useful contributions to scientific and technical progress — but that’s true for older people as well.  Regardless of how you cut it, I think only a small part of the population will be involved directly in scientific and technical progress.

There was going to be a whole section here about how horrifying I find the current education system, but I’m going to skip that for now.  Instead, I will only say that as a prerequisite it seems important for people to have freedom and autonomy able to choose what they work on or study.

As I’m learning how to do scientific research, mentorship has been exceedingly useful to me.  I’ve also been a mentor a number of times, but consider myself still learning that skill, too.  In any case I think I’d be happy to sign up for 5-10 hours/week of mentoring a young scientist, and if/when I have kids, I expect some of my friends will be willing to mentor/teach as well.

I think this match is probably win-win on just the object level — the student gets mentorship, which is especially valuable when figuring out how to navigate scientific and technical problems, and the mentor gets to support someone with a steep growth trajectory, which is pretty rewarding and exciting.  These benefits are all before the possible benefits of long-term gains from improving scientific and technical progress on the margin.

If it is the case that there’s a bunch of latent supply for scientific mentorship, and young scientists with freedom, then we mostly have a matching problem.  I don’t know exactly what a solution to this would look like, but my guess is that it wouldn’t be too difficult to prototype.

I expect there also would be a lot of soft/social things to figure out, and having a bunch of mentors have access to each other (and have a bunch of the young scientists have access to each other) would be good at creating group social support systems.

Young Scientists and My Job

I think quite a lot of my job could be done by someone a lot younger — possibly even someone 14-18, given some background knowledge and skills.  My research is about understanding and aligning Language Models, a particular kind of neural network that reads and writes text.

Different people have different specific research goals, but some common themes:

• Figuring out what patterns of behavior and mis-behavior language models exhibit
• Make datasets that allow us to evaluate progress on solving tasks or problems
• Test out techniques that mitigate problems or improve evaluated metrics
• Also more things I’m leaving out for brevity.  The field has a lot of interesting directions!

We seem to be in a bit of a mini-golden era of this kind of research, since now it’s possible to study language models without needing to have cutting edge understanding of gradient descent/optimization/etc.

This happened before to image classifier models, where it used to be that producing cutting edge image classifiers was difficult technical research by itself - to having drag-and-drop interfaces allowing anyone to build their own.

I think the parts that would be a sort of bare-minimum to do this kind of research would be:

• Programming ability (basically all this work is coding-based, but doesn’t require competition-winning-levels-of-skill)
• Language Models (there are a bunch of open source ones, and additionally many industrial labs have programs to give researchers access)
• Fine-tuning (e.g. the gpt-2 fine-tuning colab notebook that was popular in the last few years)
• Zero-shot classification (mechanism for using the language model to answer multiple-choice questions)
• Few-shot tasks (mechanism for specifying a specific pattern of task to the model)
• Interfaces for giving human feedback on the model (could be rating like likert or more rich full text feedback)

This misses a bunch of advanced stuff around deep learning and optimization, but I don’t think that’s strictly necessary for the sort of research I’m doing.  I think it’s analogous to how my programs still run on assembly code, but I don’t have to know assembly to do my day-to-day work.

In Conclusion

I don’t have any concrete plans for this yet, but it seems like a space where it’s possible to try things and iterate.