*The Nvidia Way*
I quite liked this new book by Tae Kim, offering a 245 pp. history of the company. Here is a useful review from the WSJ.
I can note that recently, a bit before Thanksgiving, I had the chance to visit Nvidia headquarters in Santa Clara, receive a tour, see some demos, and (a few days earlier) chat with Jensen Huang. I am pleased to report very positive impressions all around. My intuitive sense also jives with the portrait painted in this book.
As for my impressions of Nvidia, I was struck by the prevalence of large, attractive plant displays in the headquarters, and also how much care they take to ensure quietness on the main corporate floors and spaces (I notice funny things about companies). The geometric shapes and designs, for whatever reason, reminded me of the early 1970s movie Silent Running. If I visit an AI company, in the hallways many people will recognize me. At Nvidia nobody did, except those who invited me in. That too is an interesting contrast.
I am honored to have seen their lovely facilities.
Kevin Bryan and Joshua Gans have a new AI educational project
Just wanted to ping you about a tool Joshua Gans and I launched publicly today after a year of trials at universities all over the world (and just a stupid amount of work!) which I think is up your alley.
Idea is simple: AI should be 1) personalized to the student, 2) personalized to the professor’s content, and 3) structured to improve rather than degrade learning. In a perfect world, we want every student to have individual-level assistance, at any time, in any language, in the format they want (a chatbot TA, a question bank, a sample test grader, etc.). We want all assignments to be adaptive “mastery learning”. We want the professor to have insight on a weekly basis into how students are doing, and even into topics they may have taught in a somewhat confusing way. And we want to do this basically for free.
Right now, we have either raw GPT or Claude accomplishing 1 but not 2 or 3 (and some evidence it degrades learning for some students), or we have classes big enough to build custom AI-driven classes (like Khan Academy for basic algebra). For the thousands of classes where the professor’s teaching is idiosyncratic, the latter set of tools is basically “give the students a random textbook off the library shelf on the topic and have them study it” – not at all what I want my students to do!
We set up a team including proper UX designers and backend devs and built this guy here: https://www.alldayta.com/. It’s drag-and-drop for your course audio/video, slides, handouts, etc., preprocesses everything is a much deeper raw than raw OCR or LLMs, then preps a set of tools. Right now, there is a student-facing “virtual TA” and an autosummary weekly of where students are having trouble with the rest rolling out once we’re convinced the beta version is high enough quality. In my classes, I’ve had up to 10000 interactions in a term with this, and we ran trials at [redacted]. And we can do it at like a buck or two a student across a term, spun up in like 30 minutes of professor time for smaller courses.
There’s a free trial anyone can just sign up for; if your colleagues or the MR crowd would be interested, definitely send it along. I put a Twitter thread up about it as well with some examples of where we are going and why we think this is where higher ed is headed: https://x.com/Afinetheorem/status/1867632900956365307
Midnight regulations on chip access
Let us hope the Biden administration does not do too much damage on its way out the door (WSJ):
The U.S. is preparing rules that would restrict the sale of advanced artificial-intelligence chips in certain parts of the world in an attempt to limit China’s ability to access them, according to people familiar with the matter.
The rules are aimed at China, but they threaten to create conflict between the U.S. and nations that may not want their purchases of chips micromanaged from Washington.
…The purchasing caps primarily apply to regions such as Southeast Asia and the Middle East, the people said. The rules cover cutting-edge processors known as GPUs, or graphic processing units, which are used to train and run large-scale AI models.
Should we not want to bring the UAE more firmly into the American orbit? Is there not a decent chance they will have the energy supply for AI that we are unwilling to build domestically? Might not these regulations, over time, encourage foreign nations to become part of the Chinese AI network? More generally, why should an outgoing administration be making what are potentially reversible foreign policy decisions for the next regime?
Monday assorted links
1. Using LLMs to simulate public opinion research. And o1 doing the math Putnam exam.
2. New international trade survey paper by Pol Antras.
3. What Henry Oliver has been reading.
4. English-speaking Indian households. And better than most Zakir Hussain video clips.
5. Polarization is declining are you happy now?
7. Good observations on bureaucracy and the courts.
8. Comparative economic growth across Syria.
9. An interesting WSJ piece about how Elon Musk does not have access to the highest national security activities of his own company.
10. Chrystia Freeland steps down as finance minister of Canada.
The epistemics of drone incursions
I do not pretend to know what is going on, nor do I think it is aliens. I do read:
The sprawling Wright Patterson Air Force Base in Ohio is the latest military installation to report mysterious drones flying over its airspace, The War Zone has learned.
“I can confirm small aerial systems were spotted over Wright Patterson between Friday night and Saturday morning,” base spokesman Bob Purtiman told The War Zone on Sunday in response to our questions about the sightings. “Today leaders have determined that they did not impact base residents, facilities, or assets. The Air Force is taking all appropriate measures to safeguard our installations and residents.”
The drones “ranged in sizes and configurations,” Purtiman said. “Our units are working with local authorities to ensure the safety of base personnel, facilities, and assets.”
The airspace over the base was closed for a while. My point here is to beware self-styled “debunkers,” who often acquire excess ownership in a schtick. Most of all I mean figures such as Mick West. It is easy enough to find stupid claims about drones (especially in New Jersey?) and counter them. The better way to proceed is to confront the strongest claims head on. The head of DHS is mystified, and a classified security briefing was held for Congress. Those are puzzles we should try to figure out. There is at least a reasonable probability that something interesting is going on here. Beware debunkers, very often they are not your epistemic friends.
Addendum: Here is an update from a man who receives high-level intelligence briefings.
Technological Disruption in the US Labor Market
Deming, Ong and Summers have a good overview of long-run and very recent changes in the US labor market. Using a measure of occupational titles the authors find:
The years spanning 1990-2017 were the most stable period in the history of the US labor market, going back nearly 150 years.
It’s a bit too early to distinguish an AI revolution from a COVID shock but the last four years look to be more disruptive than any since the 1970s and over a slightly longer period there are trends including a decline in retail, as consumers shift to online shopping and delivery, and a decline in office work, the latter especially suggesting an AI effect:
There were 850,000 fewer retail sales workers in the US in 2023 compared to 2013 even though the US economy added more than 19 million jobs over this period.
There are nearly five hundred thousand fewer secretaries and administrative assistants in the US labor force now than there were a decade ago. At the same time, management and business occupations have grown very rapidly. There were four million more managers and 3.5 million more business and financial operations jobs in the US in 2023 than there were in 2013.
Keep in mind that these changes are occurring as employment and wages overall are rising.
o1 is still doing well on monetary economics
This is one of these “don’t bother reading through everything unless you already know what I am talking about” posts.
Scott Sumner has a long rebuttal to o1 on monetary theory, offering many criticisms. He does not like the quality of the answer given by the AI. But I view the AI more positively here, at least relative to the current state of the research literature, which admittedly is not so satisfying. I think Scott, a’la Kasparov, is being spooked by “the machine,” and his usual clarity of thought is not always present in this exchange. Here are a few points in response to Scott:
1. Nominal relationships seemed much clearer in the age of Milton Friedman and now they are extremely murky (and yes economists may have been wrong in Friedman’s time about this, but that is not the point). Scott seems to deny this, and I am not sure why. The Monetary History persuaded a significant batch of highly intelligent economists that there was a pretty stable relationship between the monetary aggregates and nominal income. Hardly anyone holds this view today. o1’s portrait of this change seems to me more accurate than Scott’s insistence that inflation has become easier to predict over time.
2. Scott in his criticisms is focused mainly on whether inflation prediction has become harder over time, but mainly o1 is answering why it is so hard today to explain inflation dynamics. (Revisit the question: “Please write an essay on how current macroeconomists find inflation dynamics so very difficult to predict, and why that has made them reject various forms of monetarism, even as approximations of what is going on behind price level behavior.”) So he is grading it on the wrong issues.
3. Scott writes “It [o1] mentions a bunch of irrelevant stuff like QE, and misses the key point that the payment of interest on reserves and the zero lower bound problem have made the money multiplier far more unstable.” Inflation forecasting has been a problem before and after the ZLB. And the payment of interest on reserves was a huge one-time problem for forecasting, but the literature rightly ignores or downplays this as a general issue over time, because it isn’t. So here o1 is closer to the research consensus than Scott is.
4. Most of all, Scott goes out of his way to avoid presenting or even citing a better approach to inflation dynamics. You might look at this 2011 post from Scott on inflation dynamics. It endorses the quantity theory, which I think is true sometimes, but doesn’t go into why the evidence has gone so badly in the other direction.
4b. More seriously, Scott seems to dismiss the price level concept altogether. For instance he once wrote: “In the past, I’ve frequently argued that inflation is an almost meaningless and useless concept. I’m not even aware of any coherent definitions of the concept.” I don’t think this is a defensible point of view, and you have to compare Scott’s criticisms of the o1 model to his own approach, which is fairly nihilistic. And I think wrong. If inflation were higher and someone offered Scott an inflation-indexed contract to sign, would he be unable to evaluate such a transaction? Obviously not.
5. As a side note, I think o1 also did better than the varied observations by Krugman on inflation dynamics over the years. Most recently I recall Krugman arguing that we didn’t have a recession the year before because the initial inflation was almost entirely about supply side shocks. That view has been refuted by a number of recent research papers, some of those cited on MR, showing it was both a supply side and demand side phenomenon.
More generally, here is one recent model of price level dynamics, you can read through the model and results. Real wages matter in many of these investigations, which Scott dismisses and o1 endorses. Here is an attempt to forecast price inflation for 2024, again you can look at the model, which is not so simple and I would also say not super-impressive (I intend no criticism of the authors here, the questions are hard). It is still a puzzle why inflation rates were not even lower during the 2008-2010 period. In other words, interest on reserves may have mattered less than models would suggest.
I understand full well that “kitchen sink” approaches are unsatisfying to many economists. Yet when there is in fact a clear theoretical answer to an economics problem, usually o1 gives it to you.
If you ask o1 Scott’s oil and price theory question, it gets the right answer. The second sentence is: “Whether that quantity is higher or lower than before depends on why the price rose.” In other words, it does not reason from a price change.
So I think Scott is seriously underrating o1 as a reflection of what the profession believes on inflation dynamics. Scott has a right to disagree with that consensus, but I don’t see he has put up the evidence to establish a better view. In any case, on these issues o1 beats both Sumner and Krugman, noting that each is putting forward a fairly extreme point of view. Notably, o1 pro, a yet more advanced model, comes up with a better answer yet. Or you can ask it to spend at least 5000 logic tokens answering the question. Yes people it is worth $2500 a year.
Arnold Kling comments. And Kasparov did eventually come around.
Zakir Hussain has indeed passed away
I saw him perform maybe…a dozen times? I was set to go again this spring, and would have kept on going for as long as possible. His two concerts with Shakti I attended were among the very best of my life. No video clip can do justice to percussion, unfortunately. He also seemed so young, and so alive. He was such a magnet for the talent and efforts of others, and radiated that with every movement and word onstage. If you wanted to study the connection between charisma and talent, he would be Exhibit A. And how many other people have had a plausible claim to be the world’s greatest musician?
I am sad to hear of his passing. He was always in motion, and it seems not exactly right to wish him to “rest in peace.”
Sunday assorted links
1. NoVa Civil War NIMBY buffs vs. new data centers.
2. “We propose a mechanism by which accumulated scientific knowledge determines the capacity of nuclear reactors, and find that some 55 billion tons of CO2 emissions, 2.3 million premature deaths, and 14 trillion USD in health costs could have been avoided, had we displaced fossil fuels with nuclear power.” Link here. This was the environmental movement (and the NIMBYs), right?
3. A new paper on whether competition across AIs helps or hurts cultural diversity in outputs.
4. o1 pro on whether the Fed should subsidize ngdp futures contracts.
5. Weird story involving an economist, sort of.
6. Cass Sunstein on the AI calculation problem.
7. Yup.
Are LLMs running out of data?
Ilya and many other experts say yes. I would not dare to disagree with them about AI per se, but I cannot say I am entirely convinced. Supply is elastic! That is a time-honored economic truth. So perhaps the future for traditional scaling is brighter than many of the experts currently are suggesting. We outsiders don’t know exactly which sources of data have been fed into the current top models, but surely there is plenty left? And for some price perhaps it can be mobilized. That is without even getting into data generated by mobile high-quality, life-sampling robots, which admittedly are still some number of years away.
Tax arbitrage through your business
That is the topic of my latest Bloomberg column, here is one bit:
This phenomenon is one reason that many office jobs in Nordic countries seem so pleasant. The workers have nice lunches and the use of comfortable and stylish furniture, which they are not taxed on, though of course their take-home pay may be less.
If you think that such workplace comforts make people happier than cash, then you may approve of such arrangements. And it is one vision for how to make society marginally less competitive.
An alternative model is that, with a proliferation of workplace perks and a diminution of earning power, workers become somewhat less ambitious on the earnings front. Peer norms may change, and the dynamism and innovation of the economy can decline accordingly. There are, in fact, signs of these problems in current-day Europe.
And this:
A recent study looked at some comparable effects in Portugal where the in-kind benefits accrue to a firm’s owners rather than its workers. When people own enough of a firm to control its behavior, they charge some of their personal consumption to the firm. Or, to put it another way: They draw more in-kind income from the firm, and take less cash. That lowers their total tax burden.
For the top quintile of the Portuguese income distribution, once those people are able to control a business, about 20% to 30% of their consumption expenditures are switched to benefits reaped within the firm. For the top 1% of earners, attaining a position of business manager is associated with an almost 18% drop in monthly expenditures. And lest there be any doubt about what’s happening here, the paper notes that “business expenditures on hotels and restaurants significantly increase by 9.8% in the birthday month of the owner-manager and by 6.1% in the birthday month of the owner-manager’s spouse.”
Worth a ponder.
Marriage markets in everything (tax arbitrage!)
Looking to marry someone with $1m+ of short-term capital gains (LA California) for tax savings (I have $1m+ in losses) and split the savings
I (unfortunately) lost a bunch of money this year with some risky gambles and have ~$1.2m of context of capital losses.
I would like to marry someone with very large ($1m+) short-term capital gains and split the difference on the tax savings.
I am proposing keeping ~40c for every $1 of capital losses I provided for myself and offering you the remainder (~10c or so, $120k context if you are at the highest tax bracket). The formal agreement can be formalized with a lawyer in relation to the marriage
Slight preference for females but open to males too (preference is just to avoid having to explain why I (straight male) married a man in the future).
Prefer if you are in the LA / Socal Area as that’s where I’m located.
That is from Reddit, via Stephen J.
Very good sentences
…Corbet can be just as critical of indie movies as he is of studio-backed ones. “Art-house cinema and big tentpole releases are equally algorithmic,” he said. “I’ve seen ‘4 Months, 3 Weeks and 2 Days’ remade forty-five times. I know why; it’s an extraordinary film.” Still, he went on, “There’s this kind of faux subtlety, and an allegiance to good taste, that I find really frustrating. It’s the same recipe, regurgitated over and over.”
“American indies have been conditioned to think small,” Dennis Lim, the artistic director of the New York Film Festival, told me. With “The Brutalist,” Corbet has gone full maximalist. The film, which takes place over a span of thirteen years and ends with a coda set two decades later still, promises, from its first moments, to be a capital-“E” Event. To the sounds of orchestral rumbling, a title card announces the “overture.”
That is from an excellent Alexandra Schwartz New Yorker article about the new movie The Brutalist.
Saturday assorted links
1. On taste.
2. The dangers of mirrored life? Eli Dourado is not so worried.
3. How are things going with Fermat’s Last Theorem? And are we documenting mathematics properly?
4. Claims made by some Syrians about free enterprise (speculative).
5. What is the evidence on the five-second rule?
6. Where will the Trump crowd go eat in DC?
7. Civitas Outlook, a new publication from UT Austin and the Civitas Institute.
8. Toward a measure of anecdotal value.
9. Dylan Field: “Still doing evals but feels like AGI is basically here with o1 pro mode”
Ilya’s talk
Twenty-four minutes, sixteen minutes for the core talk, self-recommending.