This post is fantastic. Also, the last 8 months have been an eternity for LLMs. Have you been tinkering with any of the new ones since this experiment, and if so how’d they do?
ryan_b
Karma: 19
I want one, if any remain! DM sent.
This review is fantastic, well done. I am now going to go seek out books about American government written by people in Japan or Finland or something.
For other aspiring policy hobbyists:
A lot of changes at the legislative level are non-obvious, because stuff gets hidden away in huge bills ostensibly about something else. For example: The National Defense Authorization Act Contains AI Provisions
My understanding of the last 10 years or so of policy is wide agreement that this is both the easiest way to get small changes made when you want them, and the hardest to stop when you don’t, which is why the US government shuts down periodically.
Most of the executive details can actually be influenced by well informed private citizens with no special access whatsoever, because regulatory agencies have public comment periods in the US. There was one for AI standards by the National Institute of Standards and Technology, for example.
Naturally most of the public comments for this kind of thing are coordinated by lobbyists instead of actually being comments from the public, which makes a public comments a natural target for coordination effort from the Unblock.
I agree that making direct comparisons don’t make sense on their own merits; I used them as stand-ins for the previous period of American growth (which may be in the book, but were not in the link). I don’t think the frontier-vs-catch-up distinction matters to the point argued in the post, though: I strongly expect the American technical frontier 1920-1970 period looks more like the China or India catch-up 1970-2020 period than it does the American technical frontier 1970-2020.
Phrased another way, the time-price method gives us the same stagnation story as the conventional methods do. This is a separate question than what is to be done to speed up frontier progress.
The tweet summary from Philippon is very interesting—I just pulled it from NBER, where the title appears to be Additive Growth.
This is a fantastic writeup, thank you for putting in the effort!
I like this time-price comparison mechanism, because it looks like it will better for tracking human-level impact than money will. I am looking forward to the book!
Out of curiosity, what was the time price gain for the previous 40-50 year spans before the ones you mention? The stagnation claim isn’t that progress is literally zero, but that the last 50 years has shown us much less than the 50 years before that, and the 50 years before that. Thiel’s position is more specific to the United States, and I note that in the link you compare the Chinese, Indian and American cases:For the time it took to buy one unit in the three-commodity basket in 1960, they would get 27 units in 2021. The Chinese gained 7 hours and 42 minutes a day to devote to other activities.
The Indian case:
For the time it took them to buy one unit in the three-commodity basket in 1960, they got 5.35 baskets in 2021. Thus, they gained 6 hours and 30 minutes a day.
The American case:
For the time it took them to earn enough money to buy one unit in the three-commodity basket in 1960, they got 4.45 baskets in 2021. Americans gained 46.5 minutes a day to devote to other activities.
7.7 hours and 6.5 hours are very different from 0.775 hours. This makes it look like the Chinese made just shy of 10x the progress America did, by time-price comparisons. Following on the roughly 50 year chunks, this means that China made 10x the progress America did over the last 50 years, by following the path America did the 50 years before that.
This seems consistent with the claims in The Rise and Fall of American Growth, which to oversimplify amount to most of the growth after WWII being due to fully capitalizing on all the inventions from the end of the 19th century, and that there are limits to that growth.
In the context of federal government action, 5 years feels like a huge win! Out of curiosity, was there any kind of generic background on the kind of policy being worked on? For example, with ARPA-H, did the background include the founding of other ARPA-pattern agencies, the references about the relevant authority, or budgetary shenanigans?
It’s unrelated to the OP, but what I am driving at here is how much pre-work on behalf of the government is a valid optimization target. I want to make a comparison with the legislative case, where a successful strategy in lobbying is providing draft language for a bill; is there an equivalent in the executive case?
List of Book Recommendations
Do you have any methods of analysis or threads of scholarship that you think are definitely wrong, or seriously misleading, and so should be avoided?
[Question] Do you have a model of how to approach the problem of progress?
This is an interesting post, and the arguments make sense to me. Upvoted.
I did find one idea which is very popular in economics thinking that I want to push back on:
Some amount of time that the second group spent will have been duplicative and so wasted.
I claim none of the effort spent by the second group is wasted: all of the duplicative effort pays out as reduced time to understand (and therefore use) the discovery. In cases where two groups are very close, that amount of time is basically zero; in cases of multiple discovery it is actually zero. I strongly expect that having multiple groups with a good understanding of a discovery increases the likelihood of successfully getting to downstream discoveries, and I suspect it would disproportionately increase the likelihood of leapfrogging and branching into new areas as the multiple groups look to differentiate themselves.
Separately and not directly related to the post, I claim that situations of multiple discovery are the most valuable events for the study of progress, because they give us n>1 experiments in how much information is required to make the discovery in question. An example of what I mean here is that if the same thing was discovered three independent times, and we look at what each person or group knew when they made the discovery, then:
Stuff known by all groups tells us what is necessary
The group that knows the least stuff gives us an idea of what is sufficient
Stuff that one or more groups was wrong about can be either dismissed as irrelevant, or if another group had it right they could be compared to see what influence that part had on how far they got
The way the second point relates to the first is that I believe the analytical lens which looks at the efficiency of a single discovery—whence the duplicated effort is wasted idea—is fundamentally mistaken. A single discovery doesn’t make sense to me as a unit of analysis for this because they are not independent; they depend on the discoveries that came before them and are in turn depended on for later discoveries. If we shift from the abstract discovery level to concrete ones like steps in the chain of producing products, this becomes much more stark: what sense does it make to compare the efficiency of an automated truck in a Uranium mine to the efficiency of an additive in paint manufacturing? In order for the efficiency numbers to make sense we need the context of the process of which they are a part.
Turning at last back to the actual subject of the post—that the value of a discovery by a person or group should be considered in light of the duplicated effort—feels to me like carrying the the same frame of analysis one step farther and applying it to the groups in the research process. If we want to identify which groups we should look to for lessons on progress (which I realize was not identified in the post) then it feels like my intuitions about this point in the opposite direction of yours.
This is a cool project! I am going to spend some time poking around in there, and if I think of anything in the way of improvements I’ll let you know.
This is a fantastic write-up, and I would like to see more like it.
I always enjoy seeing love lavished on important intermediate artifacts, like a particular paper, or particular tool, or particular storytelling technique. For papers in particular I would love to see piles and piles more highlights, walkthroughs, and appreciation of great papers across all fields.
Comments in no particular order:
Marc Andreessen continues to sound just like himself. I think this is good for the piece, it feels very genuine. In the main I agree.
Markets is the biggest section. This feels telling and also kind of wasteful. It also had the clunkiest bits which were, but of course, the ones about economics. By contrast the Technology section felt a bit thin, but I could easily forgive a certain amount of c’mon, you know why you are here in the effort.
What is this for, really? I can tell who it is for, because it doesn’t seem like it would register much with people who didn’t already mostly agree. But what problem is the manifesto solving? Guessing by some of the keywords included in the bad ideas list, this feels like maybe trying to further crystallize e/acc into a broader concern?
On the flip side, this bit here under The Enemy makes it seem more like talking book:
Purely as a matter of style, I thought the “we believe” and “we had a problem” chunks were great, that’s what I want out of a manifesto. I would jettison all the quotes and argumentation, moving names and sources to footnotes or something; I mostly found it distracting, like he was so used to the argumentation side of things he had trouble letting it go (Not that I blame him, I would have the same issue were I to write a manifesto). I thought some of them were compacted too much, like compressing all the progress in agriculture into the green revolution, which sort of deprived it of emotional impact (not least because of term confusion, since green revolution shows up in advertising campaigns and slogans constantly meaning something entirely different).
All that said, I liked it and I wish more public figures like Marc would do things like this.