ryan_b

Sequences

National Institute of Standards and Technology: AI Standards

Comments

As a Washed Up Former Data Scientist and Machine Learning Researcher What Direction Should I Go In Now?

I notice you have the following:

  • Long-term concern about the problem
  • A relevant background in several dimensions
  • Some time on your hands

Sounds to me like an excellent opportunity to firm up your analysis of the risks. With this, you can make a much more informed decision about whether to tackle the problem head on.

Also this:

One of the things I learned from doing research in industry is that experimental rigor is actually very hard to do properly, and almost everyone, people who publish papers in academia, as well as industry, cut corners to get things out ASAP so they can flag-plant on ArXiv.

I am far from the expert on the subject, but rigorous and safe toy models in code demonstrating any of the things which we discuss seem like they would be very useful.

ryan_b's Shortform

Ah, the humiliation of using the wrong ctrl-f inputs! But of course it would be lower level.

Well that's reason enough to cap my investment in the notion; we'll stick to cheap experiments if the muse descends.

Welcome to LessWrong!

Facts are independent of beliefs, which is sort of their defining characteristic. But beliefs can be in alignment with the facts, or not; the goal is the former.

What rational person doesn't? 

None. But there are no such people in the strong sense, yet. This is the ambition of the project.

Can we hold intellectuals to similar public standards as athletes?

The point about proof generation is interesting. A general proof is equivalent to collapsing the scope of predictions covered by the proof; a method of generating strong evidence effectively setting a floor for future predictions.

A simple way to score this might be to keep adding to their prediction score every time a question is found to succumb to the proof. That being said, we could also consider the specific prediction separately from the transmissibility of the prediction method.

This might be worthwhile even with no change in the overall score; it feels obvious that we would like to be able to sort predictions by [people who have used proofs] or [people who generate evidence directly].

ryan_b's Shortform

Personally I've found the biggest problem with spaced repetition for skills and habits is that it's contextless. 

Could you talk a bit more about this? My initial reaction is that I am almost exactly proposing additional value from using Anki to engage the skill sans context (in addition to whatever actual practice is happening with context).

I review Gwern's post pretty much every time I resume the habit; it doesn't look like it has been evaluated in connection with physical skills.

I suspect the likeliest difference is that the recall curve is going to be different from the practice curve for physical skills, and the curve for mental review of physical skills will probably be different again. These should be trivial to adjust if we knew what they were, but alas, I do not.

Maybe I could pillage the sports performance research? Surely they do something like this.

ryan_b's Shortform

Is spaced repetition software a good tool for skill development or good practice reinforcement?

I was recently considering using an Anki prompt to do a mental move rather than to test my recall, like tense your muscles as though you were performing a deadlift. I don't actually have access to a gym right now, so I didn't get to put it into action immediately. Visualizing the movement as vividly as possible, and tensing muscles like the movement was being performed (even when not doing it) are common tricks reported by famous weightlifters.

But I happened across an article from Runner's World today which described an experiment where all they did was tell a group of runners the obvious things that everyone already knows about preventing injury. The experimental group saw ~13% fewer injuries.

This suggests to me that my earlier idea is probably a good one, even though it isn't memory per se. The obvious hitch is that what I am after isn't actually recall - it isn't as though runners forget that overtraining leads to injury if you were to ask them, and I have never forgotten how to do a deadlift.

Rather the question is how to make it as correct and instinctive as possible.

  • This feels like a physical analogue of my earlier notion about Ankifying the elements of a problem, so as to integrate it into my perspective and notice relevant information.
  • Maybe a better way to say this is using an Anki prompt to help respond to a physical prompt, that being the task itself.
  • A physical action in response to the physical task instinctively already has a name; it is called muscle memory.
Babble challenge: 50 ways of hiding Einstein's pen for fifty years

The heuristics are pretty good within their scope, which I believe because I watched them work. That being said, the scope was limited - the explicit target of the project was something in mode of "As Seen On TV" and it had to get to a working prototype in two semesters, so the goal for heuristic 4 simultaneously became make sure no one else is doing this thing and people doing something similar is evidence of the market and investor interest. The best ones (in my opinion) were those which chose a different method for tackling a known-but-not-solved problem.

That being said, I did still sit through ENTIRELY too many coffee and/or headphone ideas. As a consequence of this experience I have concluded that solve a problem you have is pretty terrible advice when you are university student.

Babble challenge: 50 ways of hiding Einstein's pen for fifty years

What I initially had in mind was doing a babble exercise, and then later come back to prune that same babble exercise, which is what I meant by symmetric.

I agree that doing pruning is a different question. I did an entrepreneurship capstone at university, and for our projects they used a method similar to babble and prune. We got into teams, and had to generate 124 ideas, which predictably involved a lot of nonsense filler. For the prune side of the exercise, we applied a series of filters to narrow it down. These were:

  1. Is it technically feasible? Can it be done at all?
  2. Is it possible for the team to execute?
  3. Are any other companies doing it?
  4. What is the size of the potential market?

Then, of the ideas that had a reasonable market size and could be executed by the team, an option was chosen.

This leads me to think that a good prune prompt would consist of some reasonable filters with which to prune the earlier babble.

That being said the counter-babble idea is also good. I strongly recommend attempting it at least as an experiment.

The Rise and Fall of American Growth: A summary

I find myself wondering about incentives and transaction costs, here.

Is the historical record good enough to describe the people who were producing these inventions, and the environments they were working in? Do we have enough information to apply modern models to them, like cognitive load, or culture of innovation, etc?

By way of example, we've discussed several times Xerox PARC and Bell Laboratories as unusually fecund sources of innovation; but both of these were after 1920 and their chief role in the story here is launching the computing industry, which was responsible for the growth since 1970.

I have no real idea what the big sources of innovation were in 1870. There was the War of the Currents between Westinghouse and Edison, which points to companies investing in relatively ad-hoc engineering labs; but there was also a lot of acquisition of small companies in a way analogous to startup acquisition today. Whence these companies' inventions?

This was also the Reconstruction Era, so Civil War innovations might have been able to distribute. Although my inexpert impression of that war was not so much that it was a source of new discoveries in the way of WWII, but rather than it was a boom in adoption of recent inventions, and that largely on the marine and military front.

I also have the impression that patents were much weaker at the time, and the cost of doing business was lower: patents were easier to get, cheaper to keep, easier to prosecute, and there was nothing like today's patent trolls. There were a bunch of developments on that front during this period too, including international recognition of patents, switching from requiring scale models to requiring diagrams, etc.

Tangentially related is the overall question of liability and regulation, which on the aggregate I expect to have the effect of squeezing out growth of the [large harm + larger gain] type. At least legendarily, there were lots of cases where a small town or village would be destroyed or relocated in the name of a mine, or a port, or a dam. These options are all off the table now.

Following on that, this was also the period during which Westward Expansion was completed, which meant the introduction of huge quantities of land for distribution to migrants, as well as the above mechanism being applied to Native American tribes. We added...14? 15? states to the country during this period. We had only won the territory west of the Rocky Mountains <25 years prior, more than doubling the territorial expanse of the country. One complete generation working more-than-double the total natural resources seems like the kind of thing which would drive a huge growth boom. These options are also off the table in the future.

How To Fermi Model

Looks like the Internal Family Systems and multi-agent mind people are about to get a boost in accuracy!

Load More