What makes a college course popular or unpopular? I’ve long been interested in courses for non-science majors that satisfy “general education” requirements, their aim being to foster overall scientific literacy and to convey an understanding of topics that are important to society. I often teach such courses at the University of Oregon, for example a biophysics-for-non-scientists course and one on renewable energy. Last term I again taught The Physics of Energy and the Environment, a course for non-science-majors that I’ve written about before (for example, this).

Here’s the enrollment in Physics of Energy and the Environment for the past 15 years. (See Methods for how I constructed the plot.) The datapoints with the circles are the terms in which I taught the course.

You’ll notice that there are enormous fluctuations, with the number ranging from about 40 to 140. Last term had among the lowest numbers of students. I wondered why.

Here’s enrollment data for The Physics of Light and Color, usually a popular course. Last term was particularly low, less than 50 when it’s usually over 150.

Are there “general education” Physics courses with more students, and in which enrollment last term was high? Yes: Essentials of Physics. Note the scale, 300 students last term:

These were the three general education Physics courses offered in Winter 2026. Even before the term started, I was paying attention to the enrollment, tensely checking to see if my course would cross the 20-student threshold to avoid cancellation. Here’s the graph, starting a week after enrollment opened:

300, by the way, is the maximum allowed for Essentials of Physics. The ceilings for Energy and the Environment and Light and Color were 76 and 218 respectively, indicated by dashed lines above.

What if we look at all Physics general education courses for the past 15 years?

There’s a spaghetti of lines, but it’s clear that something is unusual in recent terms.

What sets the Essentials of Physics course apart? Why is it so popular? The content is “Physics 101” for non-science-majors, i.e. not a particular theme of social or humanistic interest.

While you’re formulating a guess, I’ll note that I’ve often heard great things about the Physics of Light, Color, and Vision course.

Though I’m biased, I’ll note that students also seem fond of Physics of Energy and the Environment. I’ve had enthusiastic students tell me, sometimes even years later, that they like the course. Plus, it has a lot of real-world relevance, and we like to think our students care about this.

From this past term’s student evaluations:

“The relevance of this course content can’t be overstated. This course clearly connects to real world examples and helps explain world phenomenons.”

and

“He [i.e. me] also is very good at including active learning in his lectures by making students think first before directly stating answers.” (The relevance of this will be clear in a moment.)

I’ve posted all the student evaluations here, so you can verify that I’m not cherry-picking a few cheerful kids from an otherwise angry mob.

I have yet to hear praise of Essentials of Physics, though I haven’t specifically investigated. (We don’t have access to other courses’ evaluations.)

Modalities and the Ethics of Instruction

As you’ve likely guessed, what’s different about Essentials of Physics in Winter 2026 (and Winter 2025), is that it’s an online, asynchronous course. This means that there’s no in-person interaction; lectures are recorded. Most importantly, Most importantly, students submit all work online. In principle there could be proctored in person exams at a testing center, but this doesn’t exist for this course, or for most UO online courses. The other two courses, Light … and Energy and the Environment, like nearly all of our other Physics courses, are in person.

The University of Oregon is a residential university that makes a point of stressing in its public relations “live” interactions, student experiences, topical courses, etc. University of Oregon students, therefore, are presumably not enrolling from far away, nor enrolling with the aim of taking classes in their pajamas. The interactions enabled by actually having a room full of students, especially incorporating active learning methods that stimulate student engagement and allow a back-and-forth of questions and answers, are effective ways to enhance learning. Plus, they’re fun.

Apparently all this does not diminish the appeal (or temptation?) to students of online courses.

Obviously, one can’t think about online courses in 2026 without thinking about artificial intelligence. (This has been true since at least 2024, but in 2024 one could perhaps be unaware of AI without being professionally negligent.) Even in high-level undergraduate classes, there is nothing one can assign that can’t be answered perfectly by AI; in a general education course, perfect AI-delivered answers are trivial to obtain. We are all seeing as one of the consequences the evaporation of correlation between homework scores and (in person) exam scores, the former being generally perfect and the latter increasingly bimodal with a large fraction showing stunningly low levels of understanding.

The concern is not simply academic dishonesty, though addressing this is essential to avoiding the devaluation of higher education. Perhaps more sadly, we’re seeing students use AI as a crutch for their understanding. It’s easy to ask any modern LLM to answer and then explain a homework question, read that explanation, and think this is a substitute for thinking about the question and constructing the solution oneself. The student, then, bypasses the actual process of learning, and without meaningful assessments (like quizzes or exams), the students delude themselves about their skills.

Is the immediate filling of the 300-student Essentials of Physics really a consequence of it being online? As an additional datapoint, note the Physics Behind the Internet in the graph above. Having hovered between about 20 and 100 students, it surged to 150 two years ago, and 300 this term. What’s new about Physics Behind the Internet? Two years ago it became an online asynchronous course (ceiling 154 students in 2024, 300 now).

It is possible, I should add, to create a meaningful, rigorous asynchronous online course. As noted above, one can have human-proctored exams, though UO doesn’t have the capacity to do this for large courses. One can schedule online video chats for presentation and assessment (oral exams or quizzes); one of my colleagues in Biology does this — it is effective. This won’t scale to classes larger than 20 or maybe 30; certainly not 300.

It seems obvious that online courses are pedagogical disasters. There are, as mentioned, ways to structure them well. (Doing so requires more work than an in person course, I think!) And, of course, there are motivated and self-aware students who will learn very well from such courses, as they would from other courses. However, for a 300-person general education course with no independent assessment or validation, there’s no way to take such courses seriously, or to be proud to offer them. We may as well just tell students to send a check in return for an “A”, and spare everyone 10 weeks of pretending. There would be considerable student demand for this, just as there is currently considerable demand for the online asynchronous courses.

At a faculty meeting, I asked our department to stop permitting online assessments, which would effectively stop our teaching online asynchronous courses. There was some agreement and some concern with details, but not enough enthusiasm to move forward. I lacked the energy to push the issue vigorously enough, especially because there’s a structural problem with “unilaterally” taking such a step:

The resources of a department, such as my Physics department, are tied to the number of students it teaches. (This connection doesn’t need to exist, but it’s understandable; even more than most public universities, the University of Oregon is dependent on student tuition, so an administrative insistence that departments carry their weight is understandable.) My analysis above suggests that our online courses are siphoning students from our other general-education courses, so canceling these courses would send students to these other courses, like Energy and the Environment, which I would argue would be an educational improvement. However, it would likely also send students to online courses in other departments. Should we hurt our own income, which helps us accomplish our many worthwhile goals, to uphold a general principle about educational validity? I’d argue yes, but I can see that this isn’t an obvious choice.

What we need to solve this dilemma is a university-wide policy about online education that is honest and forthright about what learning looks like in 2026, that considers actual teaching goals and student experiences, and that has teeth. So far, we lack such a policy. UO is not unique; this is a common problem.

On the plus side, my many conversations about AI and teaching with faculty at many institutions, and with students, show a universal agreement that online, un-proctored assessment is meaningless and that universities need to think clearly about what they’re doing. (Students, by the way, are some of the strongest voices against AI-enabled cheating and its facilitation by clueless professors and administrators.) At some point, this will have to translate into changes in how we run universities. The institutions that do this quickly and well may survive more easily than those that don’t.

Methods

Data on course enrollment over time at the University of Oregon isn’t readily available, at least for those of us without any administrative superpowers. However, all our course schedules are available online, so it’s possible to get a web page for every course offered by a given department (like Physics) in a given term, and save it as an HTML file. Reading this by eye is easy. Writing code to read the HTML is hard — the table structure isn’t simple. This is a completely uninteresting programming task and is, therefore, ideal for current AI tools! (Without this, I would not have bothered with this analysis.) I therefore downloaded the HTML files, asked Claude (Sonnet 4.6) to convert all the HTML files to more comprehensible CSVs, and then asked it to write code to extract information from the CSVs. I then read the code, made a few changes, and ran it. this works well.

I don’t use AI to write prose, and I’m witnessing the disastrous results of students offloading learning to AI, but writing routine and boring code is an ideal task for modern artificial intelligence. There’s a lot to think about with all these developments.

Today’s illustration…

I painted a whale to use in a public talk I gave in January. My wife noted that I’ve had two whale paintings on the blog before, in 2013!

— Raghuveer Parthasarathy, April 12, 2025

I have a feeling that everyone likes using AI tools to try doing someone else’s profession. They’re much less keen when someone else uses it for their profession.

— Giles Turnbull, AI and the human voice

Tags: ai-ethics, writing, ai

KIERAN GILBERT, HOST: Prime Minister Anthony Albanese, thanks for your time. What's your reaction to news of a two-week ceasefire, including the reopening, albeit temporarily, of the Strait of Hormuz?

My position on the urgency of rolling out quantum-resistant cryptography has changed compared to just a few months ago. You might have heard this privately from me in the past weeks, but it’s time to signal and justify this change of mind publicly.

There had been rumors for a while of expected and unexpected progress towards cryptographically-relevant quantum computers, but over the last week we got two public instances of it.

First, Google published a paper revising down dramatically the estimated number of logical qubits and gates required to break 256-bit elliptic curves like NIST P-256 and secp256k1, which makes the attack doable in minutes on fast-clock architectures like superconducting qubits. They weirdly¹ frame it around cryptocurrencies and mempools and salvaged goods or something, but the far more important implication are practical WebPKI MitM attacks.

Shortly after, a different paper came out from Oratomic showing 256-bit elliptic curves can be broken in as few as 10,000 physical qubits if you have non-local connectivity, like neutral atoms seem to offer, thanks to better error correction. This attack would be slower, but even a single broken key per month can be catastrophic.

They have this excellent graph on page 2 (Babbush et al. is the Google paper, which they presumably had preview access to):

Overall, it looks like everything is moving: the hardware is getting better, the algorithms are getting cheaper, the requirements for error correction are getting lower.

I’ll be honest, I don’t actually know what all the physics in those papers means. That’s not my job and not my expertise. My job includes risk assessment on behalf of the users that entrusted me with their safety. What I know is what at least some actual experts are telling us.

Heather Adkins and Sophie Schmieg are telling us that “quantum frontiers may be closer than they appear” and that 2029 is their deadline. That’s in 33 months, and no one had set such an aggressive timeline until this month.

Scott Aaronson tells us that the “clearest warning that [he] can offer in public right now about the urgency of migrating to post-quantum cryptosystems” is a vague parallel with how nuclear fission research stopped happening in public between 1939 and 1940.

The timelines presented at RWPQC 2026, just a few weeks ago, were much tighter than a couple years ago, and are already partially obsolete. The joke used to be that quantum computers have been 10 years out for 30 years now. Well, not true anymore, the timelines have started progressing.

If you are thinking “well, this could be bad, or it could be nothing!” I need you to recognize how immediately dispositive that is. The bet is not “are you 100% sure a CRQC will exist in 2030?”, the bet is “are you 100% sure a CRQC will NOT exist in 2030?” I simply don’t see how a non-expert can look at what the experts are saying, and decide “I know better, there is in fact < 1% chance.” Remember that you are betting with your users’ lives.²

Put another way, even if the most likely outcome was no CRQC in our lifetimes, that would be completely irrelevant, because our users don’t want just better-than-even odds³ of being secure.

Sure, papers about an abacus and a dog are funny and can make you look smart and contrarian on forums. But that’s not the job, and those arguments betray a lack of expertise. As Scott Aaronson said:

Once you understand quantum fault-tolerance, asking “so when are you going to factor 35 with Shor’s algorithm?” becomes sort of like asking the Manhattan Project physicists in 1943, “so when are you going to produce at least a small nuclear explosion?”

The job is not to be skeptical of things we’re not experts in, the job is to mitigate credible threats, and there are credible experts that are telling us about an imminent threat.

In summary, it might be that in 10 years the predictions will turn out to be wrong, but at this point they might also be right soon, and that risk is now unacceptable.

Now what

Concretely, what does this mean? It means we need to ship.

Regrettably, we’ve got to roll out what we have.⁴ That means large ML-DSA signatures shoved in places designed for small ECDSA signatures, like X.509, with the exception of Merkle Tree Certificates for the WebPKI, which is thankfully far enough along.

This is not the article I wanted to write. I’ve had a pending draft for months now explaining we should ship PQ key exchange now, but take the time we still have to adapt protocols to larger signatures, because they were all designed with the assumption that signatures are cheap. That other article is now wrong, alas: we don’t have the time if we need to be finished by 2029 instead of 2035.

For key exchange, the migration to ML-KEM is going well enough but:

1. Any non-PQ key exchange should now be considered a potential active compromise, worthy of warning the user like OpenSSH does, because it’s very hard to make sure all secrets transmitted over the connection or encrypted in the file have a shorter shelf life than three years.

2. We need to forget about non-interactive key exchanges (NIKEs) for a while; we only have KEMs (which are only unidirectionally authenticated without interactivity) in the PQ toolkit.

It makes no more sense to deploy new schemes that are not post-quantum. I know, pairings were nice. I know, everything PQ is annoyingly large. I know, we had basically just figured out how to do ECDSA over P-256 safely. I know, there might not be practical PQ equivalents for threshold signatures or identity-based encryption. Trust me, I know it stings. But it is what it is.

Hybrid classic + post-quantum authentication makes no sense to me anymore and will only slow us down; we should go straight to pure ML-DSA-44.⁶ Hybrid key exchange is reasonably easy, with ephemeral keys that don’t even need a type or wire format for the composite private key, and a couple years ago it made sense to take the hedge. Authentication is not like that, and even with draft-ietf-lamps-pq-composite-sigs-15 with its 18 composite key types nearing publication, we’d waste precious time collectively figuring out how to treat these composite keys and how to expose them to users. It’s also been two years since Kyber hybrids and we’ve gained significant confidence in the Module-Lattice schemes. Hybrid signatures cost time and complexity budget,⁵ and the only benefit is protection if ML-DSA is classically broken before the CRQCs come, which looks like the wrong tradeoff at this point.

In symmetric encryption, we don’t need to do anything, thankfully. There is a common misconception that protection from Grover requires 256-bit keys, but that is based on an exceedingly simplified understanding of the algorithm. A more accurate characterization is that with a circuit depth of 2⁶⁴ logical gates (the approximate number of gates that current classical computing architectures can perform serially in a decade) running Grover on a 128-bit key space would require a circuit size of 2¹⁰⁶. There’s been no progress on this that I am aware of, and indeed there are old proofs that Grover is optimal and its quantum speedup doesn’t parallelize. Unnecessary 256-bit key requirements are harmful when bundled with the actually urgent PQ requirements, because they muddle the interoperability targets and they risk slowing down the rollout of asymmetric PQ cryptography.

In my corner of the world, we’ll have to start thinking about what it means for half the cryptography packages in the Go standard library to be suddenly insecure, and how to balance the risk of downgrade attacks and backwards compatibility. It’s the first time in our careers we’ve faced anything like this: SHA-1 to SHA-256 was not nearly this disruptive,⁷ and even that took forever with the occasional unexpected downgrade attack.

Trusted Execution Environments (TEEs) like Intel SGX and AMD SEV-SNP and in general hardware attestation are just f***d. All their keys and roots are not PQ and I heard of no progress in rolling out PQ ones, which at hardware speeds means we are forced to accept they might not make it, and can’t be relied upon. I had to reassess a whole project because of this, and I will probably downgrade them to barely “defense in depth” in my toolkit.

Ecosystems with cryptographic identities (like atproto and, yes, cryptocurrencies) need to start migrating very soon, because if the CRQCs come before they are done, they will have to make extremely hard decisions, picking between letting users be compromised and bricking them.

File encryption is especially vulnerable to store-now-decrypt-later attacks, so we’ll probably have to start warning and then erroring out on non-PQ age recipient types soon. It’s unfortunately only been a few months since we even added PQ recipients, in version 1.3.0.⁸

Finally, this week I started teaching a PhD course in cryptography at the University of Bologna, and I’m going to mention RSA, ECDSA, and ECDH only as legacy algorithms, because that’s how those students will encounter them in their careers. I know, it feels weird. But it is what it is.

For more willing-or-not PQ migration, follow me on Bluesky at @filippo.abyssdomain.expert or on Mastodon at @filippo@abyssdomain.expert.

The picture

Traveling back from an excellent AtmosphereConf 2026, I saw my first aurora, from the north-facing window of a Boeing 747.

My work is made possible by Geomys, an organization of professional Go maintainers, which is funded by Ava Labs, Teleport, Tailscale, and Sentry. Through our retainer contracts they ensure the sustainability and reliability of our open source maintenance work and get a direct line to my expertise and that of the other Geomys maintainers. (Learn more in the Geomys announcement.) Here are a few words from some of them!

Teleport — For the past five years, attacks and compromises have been shifting from traditional malware and security breaches to identifying and compromising valid user accounts and credentials with social engineering, credential theft, or phishing. Teleport Identity is designed to eliminate weak access patterns through access monitoring, minimize attack surface with access requests, and purge unused permissions via mandatory access reviews.

Ava Labs — We at Ava Labs, maintainer of AvalancheGo (the most widely used client for interacting with the Avalanche Network), believe the sustainable maintenance and development of open source cryptographic protocols is critical to the broad adoption of blockchain technology. We are proud to support this necessary and impactful work through our ongoing sponsorship of Filippo and his team.

Japan’s election last month and the rise of the country’s newest and most innovative political party, Team Mirai, illustrates the viability of a different way to do politics.

In this model, technology is used to make democratic processes stronger, instead of undermining them. It is harnessed to root out corruption, instead of serving as a cash cow for campaign donations.

Imagine an election where every voter has the opportunity to opine directly to politicians on precisely the issues they care about. They’re not expected to spend hours becoming policy experts. Instead, an AI Interviewer walks them through the subject, answering their questions, interrogating their experience, even challenging their thinking.

Voters get immediate feedback on how their individual point of view matches—or doesn’t—a party’s platform, and they can see whether and how the party adopts their feedback. This isn’t like an opinion poll that politicians use for calculating short-term electoral tactics. It’s a deliberative reasoning process that scales, engaging voters in defining policy and helping candidates to listen deeply to their constituents.

This is happening today in Japan. Constituents have spent about eight thousand hours engaging with Mirai’s AI Interviewer since 2025. The party’s gamified volunteer mobilization app, Action Board, captured about 100,000 organizer actions per day in the runup to last week’s election.

It’s how Team Mirai, which translates to ‘The Future Party,’ does politics. Its founder, Takahiro Anno, first ran for local office in 2024 as a 33 year old software engineer standing for Governor of Tokyo. He came in fifth out of 56 candidates, winning more than 150,000 votes as an unaffiliated political outsider. He won attention by taking a distinctive stance on the role of technology in democracy and using AI aggressively in voter engagement.

Last year, Anno ran again, this time for the Upper Chamber of the national legislature—the Diet—and won. Now the head of a new national party, Anno found himself with a platform for making his vision of a new way of doing politics a reality.

In this recent House of Representatives election, Team Mirai shot up to win nearly four million votes. In the lower chamber’s proportional representation system, that was good enough for eleven total seats—the party’s first ever representation in the Japanese House—and nearly three times what it achieved in last year’s Upper Chamber election.

Anno’s party stood for election without aligning itself on the traditional axes of left and right. Instead, Team Mirai, heavily associated with young, urban voters, sought to unite across the ideological spectrum by taking a radical position on a different axis: the status quo and the future. Anno told us that Team Mirai believes it can triple its representation in the Diet after the next elections in each chamber, an ostentatious goal that seems achievable given their rapid rise over the past year.

In the American context, the idea of a small party unifying voters across left and right sounds like a pipe dream. But there is evidence it worked in Japan. Team Mirai won an impressive 11% of proportional representation votes from unaffiliated voters, nearly twice the share of the larger electorate. The centerpiece of the party’s policy platform is not about the traditional hot button issues, it’s about democracy itself, and how it can be enhanced by embracing a futuristic vision of digital democracy.

Anno told us how his party arrived at its manifesto for this month’s elections, and why it looked different from other parties’ in important ways. Team Mirai collected more than 38,000 online questions and more than 6,000 discrete policy suggestions from voters using its AI Policy app, which is advertised as a ‘manifesto that speaks for itself.’

After factoring in all this feedback, Team Mirai maintained a contrarian position on the biggest issue of the election: the sales tax and affordability. Rather than running on a reduction of the national sales tax like the major parties, Team Mirai reviewed dozens of suggestions from the public and ultimately proposed to keep that tax level while providing support to families through a child tax credit and lowering the required contribution for social insurance. Anno described this as another future-facing strategy: less price relief in the short term, but sustained funding for essential programs.

Anno has always intended to build a different kind of party. After receiving roughly $1 million in public funding apportioned to Team Mirai based on its single seat in the Upper Chamber last year, Anno began hiring engineers to enhance his software tools for digital democracy.

Anno described Team Mirai to us as a ‘utility party;’ basic infrastructure for Japanese democracy that serves the broader polity rather than one faction. Their Gikai (‘assembly’) app illustrates the point. It provides a portal for constituents to research bills, using AI to generate summaries, to describe their impacts, to surfacing media reporting on the issue, and to answer users’ questions. Like all their software, it’s open source and free for anyone, in any party, to use.

After last week’s victory, Team Mirai now has about $5 million in public funding and ambitions to grow the influence of their digital democracy platform. Anno told us Team Mirai has secured an agreement with the LDP, Japan’s dominant ruling party, to begin using Team Mirai’s Gikai and corruption-fighting Mirumae financial transparency tool.

AI is the issue driving the most societal and economic change we will encounter in our lifetime, yet US political parties are largely silent. But AI and Big Tech companies and their owners are ramping up their political spending to influence the parties. To the extent that AI has shown up in our politics, it seems to be limited to the question of where to site the next generation of data centers and how to channel populist backlash to big tech.

Those are causes worthy of political organizing, but very few US politicians are leveraging the technology for public listening or other pro-democratic purposes. With the midterms still nine months away and with innovators like Team Mirai making products in the open for anyone to use, there is still plenty of time for an American politician to demonstrate what a new politics could look like.

This essay was written with Nathan E. Sanders, and originally appeared in Tech Policy Press.

One of the most widely cited findings in AI policy comes from a 2023 paper by Eloundou, Manning, Mishkin, and Rock titled “GPTs are GPTs.” The title is a nice double meaning: the paper studies how general-purpose technologies (GPTs) powered by large language models (also GPTs) may reshape the labor market. The headline finding is that around 80% of U.S. workers could have at least 10% of their tasks affected by LLMs, and roughly 19% may see half or more of their tasks impacted. Broadly, these exposure measures try to capture how “exposed” the occupation is to AI as a function of whether AI can augment the tasks involved in the job: direct exposure is defined as “whether access to an LLM or LLM-powered system would reduce the time required for a human to perform a specific DWA or complete a task by at least 50%.” The authors are crystal clear on this in the paper: exposure corresponds to the capacity of AI to be involved in the job, not the extent to which the job can be automated away.

But the word “exposure” turned out to bring on all sorts of anxieties about exactly that—displacement. And perhaps for this reason, these AI exposure measures have routinely gone viral on social media over the last couple of months.

A recent example is by Andrej Karpathy, one of the co-founders of OpenAI and a leader in how to think about AI more generally (e.g., he coined both the terms “jagged intelligence” and “vibe coding”). His dashboard, which he described as a “vibe-coded” weekend project, was a ranking of how exposed major occupations are to AI-driven automation. It quickly went viral on X, as it fed all of the already-existing narratives about rapid job loss due to AI.

After seeing the dashboard sensationalized and spread like wildfire, Karpathy clarified that his “exposure” scorecard was based on a quick, LLM-generated measure of how digital a job is, and was never meant to be a serious forecast of which occupations will shrink or disappear. While his own project website made the same caveat, it was largely ignored on X. To butcher the well-known phrase: “A vibe coded weekend project will travel twice around the world before the caveat has time to put its pants on.”

What this recent episode illustrates, however, is that such exposure measures have caught the public eye but are routinely misread (with some proposing a moratorium on the term “exposure” altogether). When people hear that a job is “80% exposed” to AI, they picture 80% of that job disappearing. The actual economics of AI exposure and job loss are pretty far from that characterization.

What is a “job”?

A job is a set of tasks; a person typically gets paid based on how well they complete all of the tasks associated with the job. So let’s say you’re a project manager. Your job involves a bunch of tasks like generating ideas, outlining those ideas succinctly and getting feedback from team members, putting together presentations, and a bunch of rote work (e.g., approving time sheets, fielding logistics). As the AI models become better, you’ve realized that you can automate many of these things: AI can do a lot of the rote work for you, and can even help you put together presentations. According to the exposure measure, your job is now “exposed” to AI. What happens to your job and what happens to your wage? Well, if automating some of the tasks frees up time to generate better ideas, your overall productivity goes up—you become even more valuable to the firm. Humans are still employed and if anything the wages go up.

On the other hand, if AI automates all of the tasks—let’s say your job only involves two tasks and they both get automated—then yes, human labor will get displaced. Importantly, the fewer the number of tasks (what we call the dimensionality of a job), the greater the incentive of the company to automate it in the first place. This is the part much of the analysis on automation misses: adopting AI into an existing organization is costly, so the firm will be more likely to invest if it can automate the job, not just the task. “Exposure” and risk of automation is not just a function of model capabilities, it also depends on firm incentives. And this is not a hypothetical: we now have plenty of evidence that such incentives matter greatly for what gets automated and when (e.g., firms are much more likely to automate when the cost of human labor increases).

Lastly, even if AI makes people more productive and yields higher wages, there can still be massive layoffs in that sector if consumers do not “absorb” the increased productivity: if productivity-driven price drops do not increase demand for the product, then fewer workers will be needed in that sector.

More generally, a task being exposed to AI—even if that exposure corresponds to full automation of that task—can potentially lead to higher wages and more hiring for that occupation. Or it can lead to layoffs and even full displacement. Whether exposure leads to better or worse labor market outcomes for workers depends on two key variables: the elasticity of consumer demand in that sector (how much more of the product people buy as prices decrease), and the dimensionality of the job (how many tasks are involved in that job). As we hope to convince you by the end of the piece, we should be a lot more worried about jobs like trucking and warehousing than we currently are.

The standard approach to automation

Let us start with the “standard” approach to thinking about automation. First, we decompose jobs into tasks using a taxonomy like O*NET, then evaluate how many of those tasks can be automated or augmented by AI. The total impact on the job is a weighted average of how much each task was improved, which means you can build an “exposure index”—typically defined as what share of a job’s tasks can AI do?—and that index maps linearly into how much the job is affected (see, e.g., Michael Webb’s already-classic paper). This approach has been enormously useful for mapping the landscape of AI’s potential reach. But it contains an assumption that is almost certainly wrong for most real-world jobs: it assumes tasks are separable. That is, automating task A has no effect on the productivity of task B, and the overall impact is just the sum of parts.

Consider the jobs that you know. There are many out there where the output consists of doing many different things right, not just some of them. You can’t have a cook who follows most of the steps of a recipe, a drummer who is mostly on the beat, a programmer whose code only partially works (or, for that matter, a professor who only does the research half of the job…though some have tested this requirement). These are jobs where each task needs to be completed successfully for the output to be acceptable.

Put differently, the tasks are not separable; they are complements, i.e., doing one task right or wrong affects how well you can do others in the job in order to complete it. That tasks within a job are complements rather than substitutes seems quite plausible for most real-world production. And this has a wide range of important implications for how AI will actually affect jobs.

The O-ring model of jobs

The idea that complementary tasks create nonlinear productivity goes back to Michael Kremer’s classic 1993 paper, “The O-Ring Theory of Economic Development”. The name comes from the tragic Challenger disaster: a single faulty O-ring caused the catastrophic failure of the entire system. Kremer’s insight was that if production requires many steps, and each step needs to be done well for the final product to have value, then productivity becomes a multiplicative rather than a linear function of skill. A worker who makes slightly fewer errors per task will be dramatically more productive overall, because those small quality gains compound across every step.

This task-based model of jobs has gained fresh relevance with a recent paper by Joshua Gans and Avi Goldfarb, “O-Ring Automation,” which applies Kremer’s framework directly to AI-driven automation. While their model might appear simple at first glance, its implications are far-reaching and profound. At least one of us (Alex) has been obsessed with this paper for months (see here, here, and here).

Gans and Goldfarb build a model of a firm where each worker’s job is composed of n tasks. The job’s output is multiplicative in the quality of each task—this is the O-ring production function:

A worker has a time endowment h and allocates it across the n tasks. If task s is performed manually, the worker spends h_s hours on it and generates quality:

where a is labor productivity, assumed constant across tasks (a simplifying assumption). The worker's time constraint is:

The firm can also choose to automate any task by renting a piece of capital that delivers a fixed quality θ at cost r per task. This is the key part to pay attention to: whether firms invest in automating a task depends on the trade-offs embedded in this problem. Once a task is automated, the worker no longer needs to spend any time on it.

So far the setup is quite simple. The interesting part is what the multiplicative structure of the production function implies once automation enters the picture.

How can automation raise wages?

Now suppose a firm chooses to automate k out of n tasks. What happens to the worker, and how does that affect the wage?

Before automation, the worker allocates time evenly across all n tasks, which is optimal given the symmetric structure. Each manual task therefore receives h/n hours and has quality a · h/n. Total output is:

After k tasks are automated at quality θ, the worker now has all h hours to allocate across only n - k remaining manual tasks. Each manual task now gets h/(n-k) hours, producing quality a · h/(n-k). Total output becomes:

So output rises after partial automation if and only if:

This is an important condition which states that if the automated task quality θ is at least as good as the worker’s original pre-automation manual quality on those tasks, then the output increases for sure. Output does not automatically rise just because some tasks are automated; it rises when the quality of automation is high enough.

But here is the key insight: because automation also frees the worker to concentrate more time on the remaining tasks, output can increase even if the automated tasks are performed at slightly lower quality than the worker originally achieved before automation. Automation lets the worker concentrate on fewer tasks, raising the quality of each one. This is the “focus effect.” Because of the functional form of the production function, higher quality on the remaining manual tasks doesn’t just add to output—it multiplies through the production function. The worker becomes more productive precisely because they’re doing fewer things.

When the automation quality is sufficiently high relative to what the worker was producing manually on those tasks, the worker’s marginal product rises—and so (typically) does their wage. Partial automation, in the O-ring world, is often a complement to human labor rather than a substitute for it, which increases the worker’s wage.

But this is not necessarily good news for labor

Higher worker productivity is good for wages, but does it lead to more jobs or fewer? This depends on consumer demand. Each worker makes one calculator a day and the firm has 10 workers. All calculators are sold at the prevailing price. Now imagine each worker becomes much more productive so that each worker can make 10 calculators. The price of each calculator falls (costs fall), but consumers still demand roughly the same number of calculators. This is the case of inelastic demand—one that does not respond much to prices. Now the firm will fire 9 of the workers. But what if consumers buy way more calculators at lower prices, i.e., demand is very elastic. Then the firm will actually end up hiring more workers to meet the new demand, despite the fact that they’re more productive.

More generally, if demand is elastic (elasticity > 1), then a price decrease leads to a more-than-proportional increase in quantity demanded. Output expands a lot. The firm needs more workers to produce this higher output, even though each worker is now more productive. Net effect: more hiring.

If demand is inelastic (elasticity < 1), a price decrease leads to a less-than-proportional increase in quantity demanded. Output does not expand much and the firm can produce the same (or slightly more) output with fewer workers since each one is more productive. Net effect: displacement.

This is closely related to a popular idea commonly referred to as Jevons’ paradox: when a resource becomes more efficient to use, total consumption of that resource often increases rather than decreases. When the steam engine made coal more efficient, coal consumption skyrocketed because so many new applications became economically viable. The same logic applies to labor: if AI makes a worker dramatically more productive, and demand for that product is elastic, one may end up with more workers in that occupation, not fewer.

Why job dimensionality matters: The case of firm incentives

The relationship between tasks and the elasticity of consumer demand is an important dimension for predicting AI-driven displacement, but one variable that is often overlooked is the number of tasks in the job itself, i.e., its dimensionality. A job’s dimensionality matters for two reasons.

First, conditional on a task being automated, a low-dimensional job is more likely to be fully displaced. If a job has 20 tasks and one gets automated, a human worker is still required to do the other 19 tasks. But if a job has one task and one task gets automated, that job is gone. Second—and this dimension is perhaps overlooked the most—organizations have a stronger incentive to automate tasks the fewer non-automated tasks are left in the job. Imagine that automating a task requires a $10 million dollar investment (buying the software, onboarding, connecting it to the rest of the system, etc.). In one case, this task is the only non-automated task left in a job; in the other case, if this task is automated, there are 19 other non-automated tasks left. The firm has a much higher incentive to automate the task in the first case than the second because it can then replace the worker and reap the cost savings involved.1

Because of this, firms have a stronger incentive to invest in technology to automate low dimensional jobs. In a low-dimensional job, automating all or most of the core tasks can eliminate the position and the wage bill altogether. That makes the return to automation much larger. In other words, not all “unexposed” tasks matter equally: in some jobs the remaining tasks still keep the existing worker at the firm; in others they do not.

This gives a clear prediction: even if a job is not currently “exposed” to AI, in the sense that AI is not being used for the tasks involved, if it is low dimensional and the technology is getting close to automating the tasks, it should be considered at risk. Firms will work harder and invest more to automate the task(s) involved than in the case where jobs have many non-automated tasks.

Trucking and warehousing, the overlooked canaries in the coal mine

This is why we think people should be more worried about jobs like trucking and warehousing.

Roughly 3 million Americans drive trucks for a living. Many are in their 50s, have been driving for decades, and live in communities where trucking is an economic backbone. Trucking is one of the best jobs one can get without a college degree. The actual work of a long-haul truck driver is dominated by a few core functions: moving the truck safely from point A to point B. The logistics, loading/unloading, etc. are all done by others. If autonomous driving becomes reliable on long-haul routes, the job of a truck driver is not just being augmented; it is fundamentally threatened and may even be displaced entirely. And that possibility is no longer theoretical. Companies such as Aurora Innovation and Kodiak Robotics are already running large-scale autonomous trucking pilots and commercial deployments on constrained routes. Warehousing tells a similar story. Warehousing employs millions of U.S. workers, and many warehouse jobs—picking, packing, sorting, pallet movement—are relatively narrow and increasingly automatable. Abroad, firms are already operating highly automated “dark warehouses” that run around the clock with minimal human labor. These warehouses look nothing like what we see today: they are designed from the ground up to be run by machines.

Now compare that to a knowledge worker, say, a management consultant. The job combines research, data analysis, client communication, presentation design, strategic reasoning, team coordination, and relationship management. That’s at least seven or eight distinct complementary tasks. Claude or Codex might automate the first pass on the data analysis and slide deck creation, but the consultant is still needed for everything. In O-ring terms, automating some tasks can make the remaining ones more valuable by allowing the worker to allocate more time to them—the consultant can spend more time talking to the client and making them comfortable with the implementation, getting buy-in from the various units, etc. As a consequence, wages may rise, and employment may rise too if better output and lower prices expand client demand.

You can see the same logic in many high-stakes professions such as medicine and academia. There are now over 870 FDA-approved radiology AI tools, and 66% of doctors use at least one AI tool, mainly for note dictation and diagnostic support. But these tools are augmenting radiologists and physicians, not replacing them. AI typically handles the routine pattern recognition aspect of the job, freeing doctors to focus on complex cases, patient communication, and clinical judgment. Likewise, academics have been debating whether advances in AI make research assistants more or less valuable. As AI automates routine analytical tasks, both professors and RAs can concentrate more on ideas and judgement, thereby expanding output and demand for skilled research labor. This is yet again the O-ring focus effect in practice.

What do exposure indices capture?

Let us bring this back to the exposure framework. In the standard approach, a management consultant is highly “exposed” to AI whereas a truck driver is not. But does this mean that the consultant is at higher displacement risk than the truck driver? Not necessarily. The consultant’s high exposure may actually be good news because it means AI will augment many of their complementary tasks, triggering the focus effect and potentially raising wages. On the other hand, the truck driver’s moderate exposure on a single critical task is much more dangerous because trucking companies have a much higher incentive to automate the task of driving, and once that’s done, the job is gone as well. These incentives are already playing out in practice:

The relevant object therefore is not average task exposure, but the structure of bottlenecks and how automation reshapes worker time around them. Two jobs with identical exposure scores can have completely opposite displacement risks depending on whether their tasks are complements, whether demand for their output is elastic or inelastic, and the incentives of the firm to invest in automation. The workers at greatest risk are not necessarily those with the highest average exposure, but those whose jobs are built around a small number of core tasks that AI can automate.