Category: Web/Tech
Daniel Litt on AI and Math
Daniel Litt is a professor of mathematics at the University of Toronto. He has been active in evaluating AI models for many years and is generally seen as a skeptic pushing back at hype. He has a very interesting statement updating his thoughts:
In March 2025 I made a bet with Tamay Besiroglu, cofounder of RL environment company Mechanize, that AI tools would not be able to autonomously produce papers I judge to be at a level comparable to that of the best few papers published in 2025, at comparable cost to human experts, by 2030. I gave him 3:1 odds at the time; I now expect to lose this bet.
Much of what I’ll say here is not factually very different from what I’ve written before. I’ve slowly updated my timelines over the past year, but if one wants to speculate about the long-term future of math research, a difference of a few years is not so important. My trigger for writing this post is that, despite all of the above, I think I was not correctly calibrated as to the capabilities of existing models, let alone near-future models. This was more apparent in the mood of my comments than their content, which was largely cautious.
To be sure, the models are not yet as original or creative as the very best human mathematicians (who is?) but:
Can an LLM invent the notion of a scheme, or of a perfectoid space, or whatever your favorite mathematical object is? (Could I? Could you? Obviously this is a high bar, and not necessary for usefulness.) Can it come up with a new technique? Execute an argument that isn’t “routine for the right expert”? Make an interesting new definition? Ask the right question?
…I am skeptical that there is any mystical aspect of mathematics research intrinsically inaccessible to models, but it is true that human mathematics research relies on discovering analogies and philosophies, and performing other non-rigorous tasks where model performance is as yet unclear.
Podcast with Jake Sullivan and Jon Finer
Mostly about geopolitics, plenty of fresh content. And here is the transcript. Excerpt:
Jon Finer:
Should the United States be willing to take military action to defend Taiwan? It’s a thorny question for politicians to answer, but we’d be interested in your view.
Tyler Cowen:
Well, this is what economists would call a mixed strategy. Ex-ante, we should have strategic ambiguity, and not just say, we’re not going to defend Taiwan. And when Joe Biden said, “Well, we are going to defend Taiwan,” I was quite happy.
Jon Finer:
Four times. Four times.
Tyler Cowen:
Four times, yes. I know there’s different versions of how it was talked back and the like, but it should be unclear. That said, when push comes to shove, if China has made its move, you have to look at what are the terms of the deal? What are they going to do with TSMC to our best knowledge? What’s the domestic quality chip production in the United States? How do we feel about Japan and maybe South Korea getting nuclear weapons? Can South Korea remain an autonomous nation? Those are a lot of balls to juggle and they’re all hard to judge at this moment. But I think ex-ante, we should definitely create some risk that we will go to war over Taiwan, but then make the best decision ex-post. But China knows that too, right? They’re not fools. They’ve studied game theory.
Jake Sullivan:
Tyler, I’m going to put you down as that being Tyler Cowen’s version of strategic ambiguity.
Tyler Cowen:
It may not be that different from your version.
Jake Sullivan:
Exactly.
Recommended, and I also talk about my secret, unpublished China book, still pending at Tsinghua, almost certainly forever. And we cover UAPs and curling as well.
GPT as a Measurement Tool
We present the GABRIEL software package, which uses GPT to quantify attributes in qualitative data (e.g. how “pro innovation” a speech is). GPT is evaluated on classification and attribute rating performance against 1000+ human annotated tasks across a range of topics and data. We find that GPT as a measurement tool is accurate across domains and generally indistinguishable from human evaluators. Our evidence indicates that labeling results do not depend on the exact prompting strategy used, and that GPT is not relying on training data contamination or inferring attributes from other attributes. We showcase the possibilities of GABRIEL by quantifying novel and granular trends in Congressional remarks, social media toxicity, and county-level school curricula. We then apply GABRIEL to study the history of tech adoption, using it to assemble a novel dataset of 37,000 technologies. Our analysis documents a tenfold decline of time lags from invention to adoption over the industrial age, from ~50 years to ~5 years today. We quantify the increasing dominance of companies and the U.S. in innovation, alongside characteristics that explain whether a technology will be adopted slowly or speedily.
That is from a new NBER working paper by .
India AI Data MCP
The Government of India’s Ministry of Statistics and Program Implementation has created an impressive Model Context Protocol (MCP) to connect AI’s to Indian datasets. An AI connected to data via an MCP essentially knows the entire codebook and can make use of the data like an expert. Once connected one can query the data in natural language and quickly create graphs and statistical analysis. I connected Claude to the MCP and created an elegant dashboard with data from India’s Annual Survey of Industries. Check it out.
The mainstream view
Multiple studies have either shown that smartphone and social media use among teens has minimal effects on their mental health or none at all. As a 2024 review published by an American Psychological Association journal put it: “There is no evidence that time spent on social media is correlated with adolescent mental health problems.”
And this:
Advocates of bans compare social media to alcohol or tobacco, where the harms are indisputable and the benefits are minimal. But the internet, including social media, is more analogous to books, magazines or television. I may not want my sons watching “The Texas Chain Saw Massacre” or reading “Fifty Shades of Grey,” but it would be crazy to ban books and films for kids altogether.
But that is the nature of these social media bans. Australia’s law not only restricted access to platforms such as Instagram and TikTok but also banned kids under 16 from having YouTube, X and Reddit accounts. Even Substack had to modify its practices.
Here is more from the excellent Sam Bowman. And many teens make money through “digital side hustles,” in this day and age that is what a teenage job often means.
Liberal AI
Can AI be liberal? In what sense? One answer points to the liberal insistence on freedom of choice, understood as a product of the commitment to personal autonomy and individual dignity. Mill and Hayek are of course defining figures here, emphasizing the epistemic foundations for freedom of choice. “Choice Engines,” powered by AI and authorized or required by law, might promote liberal goals (and in the process, produce significant increases in human welfare). A key reason is that they can simultaneously (1) preserve autonomy, (2) respect dignity, and (3) help people to overcome inadequate information and behavioral biases, which can produce internalities, understood as costs that people impose on their future selves, and also externalities, understood as costs that people impose on others. Different consumers care about different things, of course, which is a reason to insist on a high degree of freedom of choice, even in the presence of internalities and externalities. AI-powered Choice Engines can respect that freedom, not least through personalization. Nonetheless, AI-powered Choice Engines might be enlisted by insufficiently informed or self-interested actors, who might exploit inadequate information or behavioral biases, and thus co5mpromise liberal goals. AI-powered Choice Engines might also be deceptive or manipulative, again compromising liberal goals, and legal safeguards are necessary to reduce the relevant risks. Illiberal or antiliberal AI is not merely imaginable; it is in place. Still, liberal AI is not an oxymoron. It could make life less nasty, less brutish, less short, and less hard – and more free.
The Journal for AI Generated Papers
Science should be machine-readable
One of the leading tasks of our time:
We develop a machine-automated approach for extracting results from papers, which we assess via a comprehensive review of the entire eLife corpus. Our method facilitates a direct comparison of machine and peer review, and sheds light on key challenges that must be overcome in order to facilitate AI-assisted science. In particular, the results point the way towards a machine-readable framework for disseminating scientific information. We therefore argue that publication systems should optimize separately for the dissemination of data and results versus the conveying of novel ideas, and the former should be machine-readable.
Here is the paper by A. Sina Booeshagh, Laura Luebbert, and Lior Pachter. Via John Tierney.
Rebuilding our world, with reference to strong AI
When 2012 passed into 2013, we did not have to rebuild our world, not in most countries at least. It sufficed to make adjustments at the margin.
After the Roman Empire fell, parts of Europe had to rebuild their worlds. It took a long time, but they ended up doing pretty well.
After the American Revolution, the newly independent colonies had to rebuild their own world. They did so brutally, but with considerable success.
After WWII, Western Europe had the chance to rebuild its own world, and did a great job.
We moderns are not used to having to rebuild our world.
It is now the case that strong AI is here/coming, and we will have to rebuild our own world. Many of us are terrified at this prospect, others are just extremely pessimistic. It seems so impossible. How are all the new pieces supposed to fit together? Who amongst us can explain that process in a reassuring way?
Yet we have done it many times before. Not always with success, however. After WWI ended, Europe was supposed to rebuild its own world, but they came up with something far worse than what they had before. Nonetheless, in the broader sweep of history world rebuilding projects have had positive expected value.
And so we will rebuilding our world yet again. Or maybe you think we are simply incapable of that.
As this happens, it can be useful to distinguish “criticisms of AI” from “people who cannot imagine that world rebuilding will go well.” A lot of what parades as the former is actually the latter.
In any case, it all will be quite something to witness.
India’s AI wedding buffet
Shruti Rajagopalan surveys much of the AI policy debate in India. Excerpt:
If there is a single domain where India’s AI ambitions will succeed or fail, it is energy. And energy in India is not a technology problem. It is a political economy problem, arguably the most intractable one the country faces.
India’s peak electricity demand hit 250 GW in May 2024, up from 143 GW a decade earlier. The IEA forecasts 6.3 percent annual growth through 2027, faster than any major economy. Cooling demand alone could reach 140 GW of peak load by 2030. One number captures the trajectory. For each incremental degree in daily average temperature, peak demand now rises by more than 7 GW. In 2019 the figure was half that. India is getting hotter, richer, and more electricity-hungry simultaneously.
State-controlled distribution companies have accumulated $83.7 billion in debt because energy prices have been politically distorted for decades. Over 50 GW of renewable capacity sits underutilized. About 60 GW is stranded behind inadequate transmission. The shortage is financial and infrastructural, not resource-based. Without reforming distribution pricing, governance, and grid investment ($50 billion estimated by 2035), new renewable capacity will not become reliable electricity. It will become another line item on a DISCOM balance sheet no one wants to read.
India’s electricity reaches consumers through 72 distribution companies, 44 of them state-owned, collectively the most financially distressed utilities in the world. Accumulated losses stood at ₹6.92 trillion ($76.89 billion) as of March 2024, rising every year despite five government bailouts since 2002.
Substantive throughout.
Natural and Artificial Ice
Excellent Veritasium video on the 19th century ice industry. Shipping ice from America to India would hardly seem like a wise idea—it’s hard to imagine ever getting a committee to approve such a venture—but entrepreneurs are free to try wacky ideas all the time, and sometimes they pay off, resulting in great riches. That’s the story of the “Ice King,” Frederic Tudor, who lost money for years before figuring out the insulation and logistics needed to make the trade profitable.
What I hadn’t fully appreciated is how the ice trade reshaped shipping, diet, and city design before the invention of mechanical refrigeration. Ice created the cold chain, and the cold chain made it possible to move fresh meat, fish, and produce over long distances. That in turn enabled cities to grow far beyond what local agriculture could support and shifted the American diet from salted and smoked provisions toward fresh food.
The profits of the ice trade encouraged investment in artificial ice which initially was met with resistance—natural ice is created by God!—a classic example of incumbents wrapping their economic interests in moral language, a pattern we see repeated with every disruptive technology from margarine to ridesharing.
Lots of lessons in the video about option value, permissionless innovation, and creative destruction. New technologies destroy old industries and create new ones that no one could have foreseen. The moral panic over artificial ice replacing the natural kind is no doubt familiar.
Hat tip: Naveen Nvn
Minimum wage hikes and robots
This paper studies how minimum wage policy affects firms’ adoption of automation technologies. Using both state-level measures of robot exposure and novel plant-level data on industrial robot imports linked to U.S. Census microdata from 1992-2021, we show that increases in minimum wages raise the likelihood of robot adoption in manufacturing. Our preferred identification exploits discontinuities at state borders, comparing otherwise similar firms exposed to different wage floors. Across specifications, a 10 percent increase in the minimum wage increases robot adoption by roughly 8 percent relative to the mean.
That is from Erik Brynjolfsson, et.al., including Andrew Wang. Via the excellent Kevin Lewis.
By the way, a photo from our textbook Modern Principles of Economics:
Those new service sector jobs
Who needs to be a programmer, be hired to close the doors on Waymo vehicles:
Or see here. And more text from TechCrunch. Via Glenn Mercer, Tom McCarthy, and also Air Genius Gary Leff.
The import of cross-task productivity
Given that LLMs seem to be able to automate so many small tasks, why don’t we see large productivity effects?
I drafted a short paper recently exploring the possibility that it’s for the same reason (or at least one of the reasons) that labor is typically bundled into multi-task jobs, instead of transacted by the task, in the first place: because performing a task increases one’s productivity not only at the task itself but at related tasks.
For example, say you used to spend half your time coding and half your time debugging, and the LLM can automate the coding but you still have to do the debugging. If you’re more productive at debugging code you write yourself, this (1) explains why “coder” and “debugger” aren’t separate jobs, and (2) predicts that the LLM won’t save half your time. If you’re half as productive at debugging code you didn’t write, or less, the LLM saves you no time at all.
So I was excited to see @judyhshen and @alextamkin’s paper from a week or two ago finding basically just that!
At least the way I’m thinking about it, “cross-task learning” should make the productivity impacts of automating tasks more convex: – Automating the second half of a job should be expected to have much more of an impact than automating the first half; and – If the machines can learn from their and each others’ experience, as a worker learns by doing from her own experience, then automating two jobs will have more than twice the impact of automating one.
That is from Philip Trammell. Here is his short piece. Here is the Shen and Tamkin paper. This is all very important work for why the AI growth take-off will be much slower than the power of the models themselves might otherwise indicate. The phrase “…and then all at once” nonetheless applies. But when?
These short pieces and observations are likely among the most important outputs economists will produce this year. But are they being suitably rewarded?
Optimal timing for superintelligence
There is a new paper by Nick Bostrom with that title:
Developing superintelligence is not like playing Russian roulette; it is more like undergoing risky surgery for a condition that will otherwise prove fatal. We examine optimal timing from a person-affecting stance (and set aside simulation hypotheses and other arcane considerations). Models incorporating safety progress, temporal discounting, quality-of-life differentials, and concave QALY utilities suggest that even high catastrophe probabilities are often worth accepting. Prioritarian weighting further shortens timelines. For many parameter settings, the optimal strategy would involve moving quickly to AGI capability, then pausing briefly before full deployment: swift to harbor, slow to berth. But poorly implemented pauses could do more harm than good.
Via Nabeel.