Category: Web/Tech
Artificial Intelligence, Scientific Discovery, and Product Innovation
This paper studies the impact of artificial intelligence on innovation, exploiting the randomized introduction of a new materials discovery technology to 1,018 scientists in the R&Dlab of a large U.S. firm. AI-assisted researchers discover 44% more materials, resulting in a 39% increase in patent filings and a 17% rise in downstream product innovation. These compounds possess more novel chemical structures and lead to more radical inventions. However, the technology has strikingly disparate effects across the productivity distribution: while the bottom third of scientists see little benefit, the output of top researchers nearly doubles. Investigating the mechanisms behind these results, I show that AI automates 57% of “idea-generation” tasks, reallocating researchers to the new task of evaluating model-produced candidate materials. Top scientists leverage their domain knowledge to prioritize promising AI suggestions, while others waste significant resources testing false positives. Together, these findings demonstrate the potential of AI-augmented research and highlight the complementarity between algorithms and expertise in the innovative process. Survey evidence reveals that these gains come at a cost, however, as 82% of scientists report reduced satisfaction with their work due to decreased creativity and skill underutilization.
That is from a new paper by Aidan Toner-Rodgers. Via Kris Gulati.
Generative AI and the Nature of Work
Here is a new paper by the following set of authors: Manuel Hoffmann Harvard Business School, Sam Boysel Harvard Business School, Frank Nagle Harvard Business School, Sida Peng Microsoft Corporation, Kevin Xu GitHub, Inc. Here is the abstract:
Recent advances in artificial intelligence (AI) technology demonstrate considerable potential to complement human capital intensive activities. While an emerging literature documents wide-ranging productivity effects of AI, relatively little attention has been paid to how AI might change the nature of work itself. How do individuals, especially those in the knowledge economy, adjust how they work when they start using AI? Using the setting of open source software, we study individual level effects that AI has on task allocation. We exploit a natural experiment arising from the deployment of GitHub Copilot, a generative AI code completion tool for software developers. Leveraging millions of work activities over a two year period, we use a program eligibility threshold to investigate the impact of AI technology on the task allocation of software developers within a quasi-experimental regression discontinuity design. We find that having access to Copilot induces such individuals to shift task allocation towards their core work of coding activities and away from non-core project management activities. We identify two underlying mechanisms driving this shift – an increase in autonomous rather than collaborative work, and an increase in exploration activities rather than exploitation. The main effects are greater for individuals with relatively lower ability. Overall, our estimates point towards a large potential for AI to transform work processes and to potentially flatten organizational hierarchies in the knowledge economy.
Via the excellent Kevin Lewis.
Does the internet limit immigrant assimilation?
This paper documents the effects of new communication technologies on immigrants’ socio-economic integration, spatial and job segregation, and networking behavior. Combining data on home-country Internet expansion shocks with data on immigrants’ linguistic skill, naturalization, location choice, and employment in the US, I find that home-country Internet slows down immigrants’ social and economic integration. The effect is driven by lower-skilled and younger immigrants. On the other hand, home-country Internet decreases spatial and job segregation with co-nationals, and increases immigrants’ subjective well-being. For the mechanisms, I use the American Time Use Survey data to show that home-country Internet changes networking behavior of immigrants. I also explore the role of (i) return intentions, (ii) international phone calls, and (iii) Facebook usage. The evidence is consistent with a simple Roy model, augmented with a choice between destination- and origin-country ties. Overall, this paper shows how new ICTs transform the links between immigration, diversity, and social cohesion.
That is from the job market paper of Alexander Yarkin from Brown University.
My EconEd talk on how technology is remixing the political spectrum
Here is the video, delivered in Chicago, mostly fresh material not in my other talks. Twenty-seven mimnutes or so, recommended.
My Conversation with the excellent Christopher Kirchhoff
Here is the audio, video, and transcript. Here is the intro:
Christopher Kirchhoff is an expert in emerging technology who founded the Pentagon’s Silicon Valley office. He’s led teams for President Obama, the Chairman of the Joint Chiefs of Staff, and CEO of Google. He’s worked in worlds as far apart as weapons development and philanthropy. His pioneering efforts to link Silicon Valley technology and startups to Washington has made him responsible for $70 billion in technology acquisition by the Department of Defense. He’s penned many landmark reports, and he is the author of Unit X: How the Pentagon and Silicon Valley are Transforming the Future of War.
Tyler and Christopher cover the ascendancy of drone warfare and how it will affect tactics both off and on the battlefield, the sobering prospect of hypersonic weapons and how they will shift the balance of power, EMP attacks, AI as the new arms race (and who’s winning), the completely different technology ecosystem of an iPhone vs. an F-35, why we shouldn’t nationalize AI labs, the problem with security clearances, why the major defense contractors lost their dynamism, how to overcome the “Valley of Death” in defense acquisition, the lack of executive authority in government, how Unit X began, the most effective type of government commission, what he’ll learn next, and more.
Excerpt:
COWEN: Now, I never understand what I read about hypersonic missiles. I see in the media, “China has launched the world’s first nuclear-capable hypersonic, and it goes 10x the speed of sound.” And people are worried. If mutual assured destruction is already in place, what exactly is the nature of the worry? Is it just we don’t have enough response time?
KIRCHHOFF: It’s a number of things, and when you add them up, they really are quite frightening. Hypersonic weapons, because of the way they maneuver, don’t necessarily have to follow a ballistic trajectory. We have very sophisticated space-based systems that can detect the launch of a missile, particularly a nuclear missile, but right then you’re immediately calculating where it’s going to go based on its ballistic trajectory. Well, a hypersonic weapon can steer. It can turn left, it can turn right, it can dive up, it can dive down.
COWEN: But that’s distinct from hypersonic, right?
KIRCHHOFF: Well, ICBMs don’t have the same maneuverability. That’s one factor that makes hypersonic weapons different. Second is just speed. With an ICBM launch, you have 20 to 25 minutes or so. This is why the rule for a presidential nuclear decision conference is, you have to be able to get the president online with his national security advisers in, I think, five or seven minutes. The whole system is timed to defeat adversary threats. The whole continuity-of-government system is upended by the timeline of hypersonic weapons.
Oh, by the way, there’s no way to defend against them, so forget the fact that they’re nuclear capable — if you want to take out an aircraft carrier or a service combatant, or assassinate a world leader, a hypersonic weapon is a fantastic way to do it. Watch them very carefully because more than anything else, they will shift the balance of military power in the next five years.
COWEN: Do you think they shift the power to China in particular, or to larger nations, or nations willing to take big chances? At the conceptual level, what’s the nature of the shift, above and beyond whoever has them?
KIRCHHOFF: Well, right now, they’re incredibly hard to produce. Right now, they’re essentially in a research and development phase. The first nation that figures out how to make titanium just a little bit more heat resistant, to make the guidance systems just a little bit better, and enables manufacturing at scale — not just five or seven weapons that are test-fired every year, but 25 or 50 or 75 or 100 — that really would change the balance of power in a remarkable number of military scenarios.
COWEN: How much China has them now? Are you at liberty to address that? They just have one or two that are not really that useful, or they’re on the verge of having 300?
KIRCHHOFF: What’s in the media and what’s been discussed quite a bit publicly is that China has more successful R&D tests of hypersonic weapons. Hypersonic weapons are very difficult to make fly for long periods. They tend to self-destruct at some point during flight. China has demonstrated a much fuller flight cycle of what looks to be an almost operational weapon.
COWEN: Where is Russia in this space?
KIRCHHOFF: Russia is also trying. Russia is developing a panoply of Dr. Evil weapons. The latest one to emerge in public is this idea of putting a nuclear payload on a satellite that would effectively stop modern life as we know it by ending GPS and satellite communications. That’s really somebody sitting in a Dr. Evil lair, stroking their cat, coming up with ideas that are game-changing. They’ve come up with a number of other weapons that are quite striking — supercavitating torpedoes that could take out an entire aircraft carrier group. Advanced states are now coming up with incredibly potent weapons.
Intelligent and interesting throughout. Again, I am happy to recommend Christopher’s recent book Unit X: How the Pentagon and Silicon Valley are Transforming the Future of War, co-authored with Raj M. Shah.
Effective Altruists and finance theory
One of the most admirable and impressive things about the EA movement is how many people in it will avidly learn about other areas. Whether it be animal welfare, mosquite bed nets, asteroid risk, or the properties of various AI programs, you can find numerous EAs who really have gone out of their way to master many of the details.
They don’t quite acquire expert knowledge, but due to their general facility in the application of reason, often they can outargue the experts themselves.
Yet one thing I have never met — ever — or seen on Twitter, is an EA who understands finance at a comparable level. Never.
And that is odd, because EAs so stress the import of probabilistic thinking.
If you pose the “have you thought through being short the market?” question, one hears a variety of answers that are what I call “first-order wrong.” That is, there may well be more sophisticated defenses of those points of view, but you just hear the first-order response, designed to dispose of the question without much further thought. A few of those responses are:
1. “Why should I have to gamble?” (Given your other views, it is hedging not gambling)
2. “There is already evidence I am right. My friends and I made a lot of money buying Nvidia stock.”
3. “I don’t know how to short the market.” Or “Amateur investors shoulnd’t short the market!”
4. “Did the stock market predict Hitler and WWII?”
5. “How possibly can I cash in if the world ends very suddenly? After all, the AGI has an incentive to deceive us.”
6. “But I don’t know when the world is going to end!”
7. “Why should I short the market when I can earn so much more going long on Nvidia!?”
8. “Well, I am not buying stocks!”
9. “If the world is ending soon, what do I need money for?”
10. “But if the world doesn’t end, things will be really great.”
And more. (I’ve even heard “Are you short the market?”) I will leave it as an exercise to the reader to work out what is wrong with these responses. In most cases o1 and Claude can come to your aid, if needed.
I do believe that Aella, for one, is in essence short the market. Good for her, as she is also pessimistic about AI. But here are two responses I have never ever heard, not once:
11. “I’m going to sit down and study finance and see if I can find a feasible way to short the market. If I can’t I will feel sad, but I might get back to you for further guidance.”
12. “Soon enough, AI will be good enough to tell me how to short the market intelligently. Then I am going to do this — thanks for the tip! ”
Nope never. The absence of the last one from the discourse I find especially odd. “AGI will be powerful enough to destroy us, but not good enough to help me do an effective short!” OK…
The sociology here is more indicative of what is going on than the arguments themselves. Because the EAs, rationality types, and doomsters here generally are very good at learning new things.
Of course, once shorting the market even enters serious contemplation (never mind actually doing it), you also start seeing current market prices as a kind of testing referendum on various doomster predictions. And suffice to say, market prices basically offer zero support for all of those predictions. And that is embarrassing, whether you should end up shorting the market or not. Many EAs and rationality types are also fans of prediction markets in other contexts.
I nonetheless would urge many EA, rationality, and AI doomster types to learn more basic finance. It can liberate you from various mental chains, and it will be useful for the rest of your life, no matter how long or short that may be.
Addendum: So, so many fallacies in the comments. Here is one brief response I wrote: “Just keep on buying puts with a small pct. of your wealth. You don’t have to use leverage, though of course a real pessimist should. What is hard about shorting is that the world isn’t in fact going to end! You are smuggling in categories from very different contexts. And none of this requires anything remotely like a “strong version of market efficiency.” It does require that the end of world is bearish for prices at some point! [once people recognize doom might be coming, not when doom finally arrives]”
Does the O-Ring model hold for AIs?
Let’s say you have a production process, and the AIs involved operate at IQ = 160, and the humans operate at IQ = 120. The O-Ring model, as you may know, predicts you end up with a productivity akin to IQ = 120. The model, in short, says a production process is no better than its weakest link.
More concretely, it could be the case that the superior insights of the smarter AIs are lost on the people they need to work with. Or overall reliability is lowered by the humans in the production chain. This latter problem is especially important when there is complementarity in the production function, namely that each part has to work well for the whole to work. Many safety problems have that structure.
The overall productivity may end up at a somewhat higher level than IQ = 120, if only because the AIs will work long hours very cheaply. Still, the quality of the final product may be closer to IQ = 120 than you might have wished.
This is another reason why I think AI productivity will spread in the world only slowly.
Sometimes when I read AI commentators I feel they are imagining production processes of AIs only. Eventually, but I do not see that state of affairs as coming anytime soon, if only for legal and regulatory reasons.
Furthermore, those AIs might have some other shortcomings, IQ aside. And an O-Ring logic could apply to those qualities as well, even within the circle of AIs themselves. So if say Claude and the o1 model “work together,” you might end up with the worst of both worlds rather than the best.
The uneven effects of AI on the American economy
That is the topic of my latest Bloomberg column, here is one excerpt:
To see how this is likely to play out, start with a distinction between sectors in which it is relatively easy to go out of business, and sectors in which it is not. Most firms selling computer programming services, for example, do not typically have guaranteed customers or revenue, at least for long. Employees have to deliver, or they and their company will be replaced. The same is true of most media companies: If they lose readers or customers, their revenue disappears. There is also relatively free entry into the sector in the US, due to the First Amendment.
Another set of institutions goes out of business only slowly, if at all. If a major state university does a poor job educating its students, for example, enrollment may decline. But the institution is still likely to be there for decades more. Or if a nonprofit group does a poor job pursuing its mission, donors may not learn of its failings for many years, while previous donors may pass away and include the charity in their wills. The point is, it can take a long time for all the money to dry up.
Which leads me to a prediction: Companies and institutions in the more fluid and competitive sectors of the economy will face heavy pressure to adopt AI. Those not in such sectors, will not.
It is debatable how much of the US economy falls into each category, and of course it is a matter of degree. But significant parts of government, education, health care and the nonprofit sector can go out of business very slowly or not at all. That is a large part of the US economy — large enough to slow down AI adoption and economic growth.
As AI progresses, the parts of the economy with rapid exit and free entry will change quickly.
Recommended, read the whole thing.
Scott Alexander on the Progress Studies conference
Here is one excerpt:
Over-regulation was the enemy at many presentations, but this wasn’t a libertarian conference. Everyone agreed that safety, quality, the environment, etc, were important and should be regulated for. They just thought existing regulations were colossally stupid, so much so that they made everything worse including safety, the environment, etc. With enough political will, it would be easy to draft regulations that improved innovation, price, safety, the environment, and everything else.
For example, consider supersonic flight. Supersonic aircraft create “sonic booms”, minor explosions that rattle windows and disturb people underneath their path. Annoyed with these booms, Congress banned supersonic flight over land in 1973. Now we’ve invented better aircraft whose booms are barely noticeable, or not noticeable at all. But because Congress banned supersonic flight – rather than sonic booms themselves – we’re stuck with normal boring 6-hour coast-to-coast flights. If aircraft progress had continued at the same rate it was going before the supersonic ban, we’d be up to 2,500 mph now (coast-to-coast in ~2 hours). Can Congress change the regulation so it bans booms and not speed? Yes, but Congress is busy, and doing it through the FAA and other agencies would take 10-15 years of environmental impact reports.
Or consider solar power. The average large solar project is delayed 5-10 years by bureaucracy. Part of the problem is NEPA, the infamous environmental protection law saying that anyone can sue any project for any reason if they object on environmental grounds. If a fossil fuel company worries about a competition from solar, they can sue upcoming solar plants on the grounds that some ants might get crushed beneath the solar panels; even in the best-case where the solar company fights and wins, they’ve suffered years of delay and lost millions of dollars. Meanwhile, fossil fuel companies have it easier; they’ve had good lobbyists for decades, and accrued a nice collection of formal and informal NEPA exemptions.
Even if a solar project survives court challenges, it has to get connected to the grid. This poses its own layer of bureaucracy and potential pitfalls.
Do read the whole thing. And congratulations to Jason Crawford and Heike Larson for pulling off this event.
Metascience podcast on science and safety
From the Institute for Progress. There are four of us, namely Dylan Matthews, Matt Clancy, and Jacob Trefethen as well. There is a transcript, and here is one very brief excerpt:
Tyler Cowen: I see the longer run risks of economic growth as primarily centered around warfare. There is lots of literature on the Industrial Revolution. People were displaced. Some parts of the country did worse. Those are a bit overstated.
But the more productive power you have, you can quite easily – and almost always do – have more destructive power. The next time there’s a major war, which could be many decades later, more people will be killed, there’ll be higher risks, more political disorder. That’s the other end of the balance sheet. Now, you always hope that the next time we go through this we’ll do a better job. We all hope that, but I don’t know.
And:
Tyler Cowen: But the puzzle is why we don’t have more terror attacks than we do, right? You could imagine people dumping basic poisons into the reservoir or showing up at suburban shopping malls with submachine guns, but it really doesn’t happen much. I’m not sure what the binding constraint is, but since I don’t think it’s science, that’s one factor that makes me more optimistic than many other people in this area.
Dylan Matthews: I’m curious what people’s theories are, since I often think of things that seem like they would have a lot of potential for terrorist attacks. I don’t Google them because after Edward Snowden, that doesn’t seem safe.
I live in DC, and I keep seeing large groups of very powerful people. I ask myself, “Why does everyone feel so safe? Why, given the current state of things, do we not see much more of this?” Tyler, you said you didn’t know what the binding constraint was. Jacob, do you have a theory about what the binding constraint is?
Jacob Trefethen: I don’t think I have a theory that explains the basis.
Tyler Cowen: Management would be mine. For instance, it’d be weird if the greatest risk of GPT models was that they helped terrorists have better management, just giving them basic management tips like those you would get out of a very cheap best-selling management book. That’s my best guess.
I would note that this was recorded some while ago, and on some of the AI safety issues I would put things differently now. Maybe some of that is having changed my mind, but most of all I simply would present the points in a very different context.
Does it matter who Satoshi was?
That is the topic of my latest Bloomberg column, here is one excerpt:
It also matters if Satoshi was a single person or a small team. If a single person, that might mean future innovations are more likely than generally thought: If Satoshi is a lone individual, then maybe there there are more unknown geniuses out there. On the other hand, the Satoshi-as-a-team theory would mean that secrets are easier to keep than people think. If that’s the case, then maybe conspiracy theories are more true than most of us would care to admit.
According to many speculations, Satoshi came out of a movement obsessed with e-cash and e-gold mechanisms, dating to the 1980s. People from those movements who have been identified as potential Satoshi candidates include Nick Szabo, Hal Finney, Wei Dai, David Chaum and Douglas Jackson, among others. At the time, those movements were considered failures because their products did not prove sustainable. The lesson here would be that movements do not truly and permanently fail. It is worth experimenting in unusual directions because something useful might come out of those efforts.
If Peter Todd is Satoshi, then it’s appropriate to upgrade any estimates of the ability of very young people to get things done. Todd would have been working on Bitcoin and the associated white paper as a student in his early 20s. At the same time, if the more mainstream Adam Back is involved, then maybe the takeaway is that rebellious young people should seek out older mentors on matters of process and marketing.
I believe that in less than two years we will know who Satoshi is.
The new Claude can run and direct your computer for you
Watch here. (It’s happening.)
Acemoglu interview with Times of India
Here is part of the segment on AI:
Given the potential for AI to exacerbate inequality, how can we redirect technology?
We need to actively steer technological development in a direction that benefits broader swathes of humanity. This require a pro-human approach that prioritises enhancing worker productivity and autonomy, supporting democracy and citizen empowerment, and fostering creativity and innovation.
To achieve this, we need to: a) Change the narrative around technology, emphasising societal control and a focus on human well-being. b) Build strong countervailing powers, such as labour unions and civil society organisations, to balance the power of tech companies, and c) Implement policies that level the playing field, including tax reforms that discourage automation and promote labour, data rights for individuals and creative workers, and regulations on manipulative digital advertising practices.
Here is the full interview.
How should strong AI alter philanthropy?
That is the theme of my latest Bloomberg column, and here is one bit:
One big change is that AI will enable individuals, or very small groups, to run large projects. By directing AIs, they will be able to create entire think tanks, research centers or businesses. The productivity of small groups of people who are very good at directing AIs will go up by an order of magnitude.
Philanthropists ought to consider giving more support to such people. Of course that is difficult, because right now there are no simple or obvious ways to measure those skills. But that is precisely why philanthropy might play a useful role. More commercially oriented businesses may shy away from making such investments, both because of risk and because the returns are uncertain. Philanthropists do not have such financial requirements.
And this oft-neglected point:
Strong AI capabilities also mean that the world might be much better over some very long time horizon, say 40 years hence. Perhaps there will be amazing new medicines that otherwise would not have come to pass, and as a result people might live 10 years longer. That increases the return — today — to fixing childhood maladies that are hard to reverse. One example would be lead poisoning in children, which can lead to permanent intellectual deficits. Another would be malnutrition. Addressing those problems was already a very good investment, but the brighter the world’s future looks, and the better the prospects for our health, the higher those returns.
The flip side is that reversible problems should probably decline in importance. If we can fix a particular problem today for $10 billion, maybe in 10 years’ time — due to AI — we will be able to fix it for a mere $5 billion. So it will become more important to figure out which problems are truly irreversible. Philanthropists ought to be focused on long time horizons anyway, so they need not be too concerned about how long it will take AI to make our world a fundamentally different place.
Recommended, interesting throughout.
Reflections on Palantir
Here is a new essay by Nabeel Qureshi, excerpt:
The combo of intellectual grandiosity and intense competitiveness was a perfect fit for me. It’s still hard to find today, in fact – many people have copied the ‘hardcore’ working culture and the ‘this is the Marines’ vibe, but few have the intellectual atmosphere, the sense of being involved in a rich set of ideas. This is hard to LARP – your founders and early employees have to be genuinely interesting intellectual thinkers. The main companies that come to mind which have nailed this combination today are OpenAI and Anthropic. It’s no surprise they’re talent magnets.
And this:
Eventually, you had a damn good set of tools clustered around the loose theme of ‘integrate data and make it useful somehow’.
At the time, it was seen as a radical step to give customers access to these tools — they weren’t in a state for that — but now this drives 50%+ of the company’s revenue, and it’s called Foundry. Viewed this way, Palantir pulled off a rare services company → product company pivot: in 2016, descriptions of it as a Silicon Valley services company were not totally off the mark, but in 2024 they are deeply off the mark, because the company successfully built an enterprise data platform using the lessons from those early years, and it shows in the gross margins – 80% gross margins in 2023. These are software margins. Compare to Accenture: 32%.
The rest is interesting throughout. As Nabeel and a few others have noted, there should be many more pieces trying to communicate what various businesses and institutions really are like.