Learning Econ (AI) in the Age of AI (Econ)
https://twitter.com/0xPrismatic/status/1936959651196895260
I mean, how do you not click on a tweet like that if you are an econ teacher?
Not only does the test in the tweet talk about incentive design being "the final boss", but that Ghiblified image of the carrot and the stick. Ooh, sign me up for more!
Teaching Econ, Learning AI
Incentives matter | Incentives can be positive and negative (the carrot and the stick) | Incentives can be monetary and non-monetary | Incentives can be gamed | Beware Goodhart's Law | Incentive design is hard
Anytime you teach econ to first year undergraduates, you'll want to start with this lesson. Well, I'll want to, at any rate. And I'm always on the lookout for great examples related to this concept, so when I read this tweet, I wanted to learn more.
But that, alas, is where the good news ended.
Because to this economist, this was gobbledygook:
https://twitter.com/0xPrismatic/status/1936959675687481851
Not the first tweet over there - that kinds sorta made sense. So you're trying to do something on a decentralized basis, rather than through a centralized initiative. Folks who choose to contribute are either pitching in because they buy into the mission, or because they hope for some payment later, or both. These, of course, are their incentives.
And the tweet goes on to say that this works for early stage experiments (v1), but this doesn't necessarily scale. That part, even I understood.
Ah, but that next tweet, the one that starts with "3/". That reminded me of the first time I opened an econometrics textbook back in my undergrad days. Seems to be written in the English language, but that's about all I get.
What on earth is a Bittensor subnet? I know what pre-training an LLM means, but what are TAO tokens? I can guess what train locally means, but what are compressed updates? What does submitting pseudo-gradients to the network mean?
This, of course, means that I should talk to you about food. Naturally.
I Travel to Eat
When I travel to a new place, I do so armed with a list. This list is a private list on Google Maps, and it is a list of "Must Eats" in that new place. I have these lists prepared for all the places that I have traveled to over the last ten years, and anytime I travel to a new place, I will make a new list.
The reason I bring this up is because I learn about that place by learning about the food they eat. What are the staples? What is the souring agent? How popular are meat dishes, and which meat(s)? What are typical sweets, and what are the special occasion sweets? What is a typical daily dinner like? What crops does the region grow, what fruits are in season while I'll be in that place?
That's not a comprehensive list of questions I have about food, but it is a good start. And the point is that I learn about that place by learning about the food from that place.
I will visit museums, I'll go to the local tourist spots, I'll take in shows, and go on long walks. But my "in" to that place is the food. My daughter has her "in" via dance - which styles are popular, what are the clothes, are the dances individual dances, etc., etc. But in my experience, you learn more about the place you go to when you travel to that place as a relative expert in some area.
What is (or should be) your "in" when you travel? Spend a cup of coffee thinking about your answer to this question, and travel is likely to become much more enjoyable.
But what does this have to do with TAO tokens and Bittensor subnets?
Back to Incentives
Well, look at it this way: I'm traveling to the world of decentralized AI development (and cryptography as an added bonus). I know nothing about this place, and would like to learn more. And my "in", in this case, isn't food. It's a simple but crucial principle in economics: incentives matter.
I know more about the Tao of physics then I do about the TAO of tokens, and I know nothing about the Tao of physics. But I do happen to know a little about incentives.
And so my "in" to this strange and fascinating world is that of incentive design. How does this world make use of a principle I love to teach folks about? And so I end up learning more about something I know nothing about, by using a topic I know relatively well as a bridge.
So ok, I may not understand much about DeMo, or about key values. But I do understand that incentives need to be designed well, and that key contributors need to get their fair share:
https://twitter.com/0xPrismatic/status/1936959699271999783
I also get they don't use a binary to reward performance - it's not a case of either you get a fixed sum or you get nothing. It is instead, a function of how much you have contributed. But hey, if you are a regular reader of EFE, you know that good ol' Uncle Goodhart will turn up sooner or later. How do you guard against folks who want to game the system, and get rewarded without actually having done the work?
https://twitter.com/0xPrismatic/status/1936959722810437697
And as with applications of econ principles elsewhere in the world, so also with decentralized AI development. Incentive design is hard.
https://twitter.com/0xPrismatic/status/1936959734579609623
And I do not know enough about this field to be able to tell you how well things are going, either in an absolute or relative sense, but generally speaking, graphs that move up and to the right are graphs that are doing well:
Now, there's an entire blogpost waiting for you about the topic. Note that you will have to register (no payment involved) to read. But now that you have a bridge, and an LLM by your side, reading even the entire blogpost should be much, much easier.
Here's the simplified version for your perusal.
The Meta TMKK
Use bridges to learn subjects you know nothing about. "But I know nothing about x" is a reasonable statement in and of itself, but can you really afford to stay in that well, especially in the case of AI?
Pick a subject you do know something about. I picked econ, but you can pick whatever domain you like. And I refuse to buy the notion that there is no domain you know nothing about. Pick cooking, pick music, pick dance - it needn't be an academic subject. If you cannot figure out a bridge, leave that job to AI!
For example, here's a prompt I wrote to have the same article be more interesting for my wife:
Great! Now rewrite this, but for my wife. She'll want to learn about this too, and her domain of expertise is not econ, but renewable energy policy.
(She's got a PhD in Econ, so she knows a thing or two about the subject, but I wanted an example of how you can switch across domains.)
But the key point is that learning across domains becomes so much easier when you can build bridges. This is a bit counter-intuitive, but a great way to learn more about other subjects is becoming reasonably good at your own area of expertise.
And all of us, like it or not, also have to become reasonably good at figuring out how to build bridges.
Go ahead, start building 'em!