Progress studies is really interesting to me, because if one actually believes that it has a >0 chance of working, the EV is so huge that we should be directing huge amounts of funding towards it. As we are not [0], this implies that most organizations/people either think it has no chance of working, or haven’t heard of it.
There’s an analogy to be made to internal tooling at most companies. All companies I’ve been part of radically underfund their internal tooling, even though it has a direct and quantifiable ROI. I don’t know why this is. Increasing the 2nd derivative is so massively valuable that it seems obviously worth doing.
[0]: Where we = the government, private donors, universities.
This may come off as snarky, but I mean it entirely sincerely. How do you know the companies you worked for underfunded their internal tooling?
The reason I ask is that I also suspect the lab I work in in my grad program underfunds its internal tooling. But that’s just based on informal observation and speculation about the counter factual benefits that greater investment might bring.
Is that also the basis for your claim? Or do you have more in the way of legible evidence?
I don’t have a great answer for you. It is largely speculation about counterfactual benefits. I would like to delude myself into thinking it’s good counterfactual speculation, but that’s probably just my ego talking.
I’m mostly basing this around things that continue to be problems for years. For instance, I work as an AI research engineer at a large industry lab, and a common task is to launch experiments that use O(100) GPUs. We have a system that automatically schedules these jobs on different data centers depending on availability. It does not have the ability to choose a data center that satisfies constraints on both spot instances and on-demand instances. The scheduler can only do one of these. As a result, we have to manually choose cells. It has been like this for >4 years.
I have a bunch of additional examples like this that have gone unsolved for years. This is a log way of saying that, to directly answer your question, I don’t have much more than speculation.
Somehow, it seems difficult, yet valuable, for employees of a company, or members of a lab, to think of themselves as living in a small island economy. Somehow, all their efforts go into producing goods and services for export. They do virtually no production, no trade with each other, to grow the economy on the island itself.
Progress studies is really interesting to me, because if one actually believes that it has a >0 chance of working, the EV is so huge that we should be directing huge amounts of funding towards it. As we are not [0], this implies that most organizations/people either think it has no chance of working, or haven’t heard of it.
There’s an analogy to be made to internal tooling at most companies. All companies I’ve been part of radically underfund their internal tooling, even though it has a direct and quantifiable ROI. I don’t know why this is. Increasing the 2nd derivative is so massively valuable that it seems obviously worth doing.
[0]: Where we = the government, private donors, universities.
This may come off as snarky, but I mean it entirely sincerely. How do you know the companies you worked for underfunded their internal tooling?
The reason I ask is that I also suspect the lab I work in in my grad program underfunds its internal tooling. But that’s just based on informal observation and speculation about the counter factual benefits that greater investment might bring.
Is that also the basis for your claim? Or do you have more in the way of legible evidence?
I’m sorry that I missed this.
I don’t have a great answer for you. It is largely speculation about counterfactual benefits. I would like to delude myself into thinking it’s good counterfactual speculation, but that’s probably just my ego talking.
I’m mostly basing this around things that continue to be problems for years. For instance, I work as an AI research engineer at a large industry lab, and a common task is to launch experiments that use O(100) GPUs. We have a system that automatically schedules these jobs on different data centers depending on availability. It does not have the ability to choose a data center that satisfies constraints on both spot instances and on-demand instances. The scheduler can only do one of these. As a result, we have to manually choose cells. It has been like this for >4 years.
I have a bunch of additional examples like this that have gone unsolved for years. This is a log way of saying that, to directly answer your question, I don’t have much more than speculation.
Thank you for the response.
Somehow, it seems difficult, yet valuable, for employees of a company, or members of a lab, to think of themselves as living in a small island economy. Somehow, all their efforts go into producing goods and services for export. They do virtually no production, no trade with each other, to grow the economy on the island itself.
That’s a really great way to think about it- thank you for that metaphor.