marginal revolution read out: Saturday assorted links · marginal revolution read out

Listen: https://blossom.buildtall.systems/947ea92a64e6401b63f2504a1aaa4d6f8e06e0676d1fd4d0ea792537526f6633.mp3

Source: https://marginalrevolution.com/marginalrevolution/2026/10/saturday-assorted-links-581.html?utm_source=rss&utm_medium=rss&utm_campaign=saturday-assorted-links-581

marginal revolution read out presents Saturday assorted links.

The first link is Rewarding failure. The author passes it along with a question mark, turning the title into an open question. The linked paper could not be retrieved, so its full argument is not part of this episode. What the title suggests is an inquiry into whether institutions, markets, or organizations end up rewarding failure. The author's comment is deliberately minimal: just that question mark.

The second link is An excellent Brian Potter explainer on how matrix algebra is used in both LLMs and robotics. The author calls it excellent, and it appears on Construction Physics by Brian Potter. Potter explains that the AI models controlling many robots are vision-language-action models, or VLAs. A large language model takes text as input and produces text as output; a VLA takes text, images, and robot information, and produces a series of robot actions. The core of the explainer is linear algebra: matrices of numbers are added, scaled, and multiplied with the dot product. Neural networks are implemented through matrix multiplication, adding bias vectors, and then applying a nonlinearity such as ReLU. Training works by running a forward pass, measuring loss, then using backpropagation and gradient descent to update weights and biases. Text is broken into tokens, turned into embedding vectors, and images are split into patches and embedded through a learned function. Potter then reaches attention and the transformer, which let a model encode how a token relates to its surrounding context. The author's framing is that this is a clear route into a technical subject; the broader point is that the same mathematical machinery behind chatbots is now moving physical systems.

The third link is Chat with Mircea Cărtărescu in the Amsterdam Review. The author shares the conversation without additional gloss. Cărtărescu, one of the most celebrated writers of contemporary fiction, discusses his new book Theodoros, forthcoming from Deep Vellum. The novel begins from a real historical detail: a servant in nineteenth-century Wallachia named Tudor, or Theodoros, was obsessed with becoming an emperor, and may have become Tewodros the Second of Ethiopia. Cărtărescu describes the book as a pseudo-historical novel, narrated by archangels in the second person, presenting Theodoros's life to God at final judgment. The conversation touches on power and corruption, the way a life can be read as tragedy or tragicomedy, and the idea that literary truth is constructed by the book itself. A key image is the Möbius strip: dream and reality seem separate, but they are one continuous surface. The author's link directs readers toward a long, thoughtful interview rather than a short note.

The fourth link is It seems the expansion of remote work reduced births? The author frames it as a tentative finding. The linked post comes from Andrew C. Johnston, who says many people hoped remote work would raise birth rates. No commute, more flexibility, and the chance to live near family all seemed likely to help. Johnston tests that idea in a new paper and finds that among married women, the 2020 expansion of remote work reduced births. That runs against the optimistic expectation. It matters because debates over remote work usually focus on productivity and location; this points instead to demographic effects in an unexpected direction.

The fifth link is Living Science uses an AI agent to reproduce key findings from seminal papers in economics, document what holds up and extend the analysis with newer data. The author highlights the project as described by Philipp Heimberger. Living Science sends an AI agent after influential economics papers. It reproduces the key findings, documents what still holds up, and then extends the analysis with newer data. The result is compared to a follow-up paper with robustness checks and extensions, and the project produces a new paper every week. The interesting implication is that replication moves from an occasional, labor-intensive craft toward a continuous, automated process, with AI probing the reliability of empirical economics.

The sixth link is Joshua Rothman on Garicano’s Messy Jobs, from The New Yorker. The author notes the New Yorker source. In the piece, Joshua Rothman explains why the jobs apocalypse hasn't arrived yet. Something else is happening instead. He discusses a book by Luis Garicano, Jin Li, and Yanhui Wu called Messy Jobs: The Work That AI Cannot Reach. The central idea is the ninety-ten production function. AI can get work to ninety percent quality very quickly, but getting from ninety to one hundred is often where it stalls. That means some tasks are commodity tasks where good enough is fine, while others are star tasks where excellence is the whole point. A further problem is that when everyone's memos and proposals are polished by AI, bosses lose access to tacit information about who is creative, trustworthy, and ready for authority. Garicano and his coauthors argue that many jobs are strong bundles of tasks; separating them destroys value. In the robots above scenario, AI gives workers expertise they otherwise lack. In the robots below scenario, AI handles routine problems while the human handles exceptions. The article reports the authors' view that the scarcer workers are not technical specialists, but rank-and-file employees who know how work is actually done. The Spanish bank BBVA is offered as a positive example: AI was given only to enthusiastic workers, and good uses were rewarded. The closing worry is whether bosses will use AI to increase worker value, or merely as a cost-cutting tool.

Across the links, a common theme appears: systems that look simple turn out to have hidden structure. Whether the subject is reward design, matrix multiplication, literary construction, remote work, replication, or the shape of jobs, the post is pointing to work that complicates first impressions.

Your signing session ended