For years, and especially in the last month or so, we’ve been collectively inundated with an AI industry which, uncannily and bizarrely, tells us simultaneously that AI is essential, inevitable and runs a significant chance of exterminating all of us in as few as two or three years. If you’ve dipped into this debate at all, you’ve quickly come upon discussions of so-called “rationalists” or “long-termists” or supporters of “effective altruism” who are simultaneously ubiquitous among those developing AI and in the AI safety (i.e., the folks sounding the alarm) community as well. (For the sake of simplicity and brevity, I’ll refer to these folks as the “REL nexus”: rationalists, effective altruists and long-termists. This is my coinage since I don’t know a better one.) I’ve spent a fair amount of time familiarizing myself with this world and how it relates to current debates about AI. So I wanted to share my basic impressions.

Today's links

• Swapping money for expertise: Money can't buy me love.
• Hey look at this: Delights to delectate.
• Object permanence: "Tempo"; Yahoo censors MEP's torture debate video; Scotland's money-laundry; A textbook scam; DOS the IRS then charge for access; Just trying to make a buck.
• Upcoming appearances: Virtual, South Bend, Hudson, Calgary, Winnipeg, Paris, Vancouver, Victoria, Ottawa, Kilkenny, Sheffield, Oxford, Montreal.
• Recent appearances: Where I've been.
• Latest books: You keep readin' em, I'll keep writin' 'em.
• Upcoming books: Like I said, I'll keep writin' 'em.
• Colophon: All the rest.

---

Swapping money for expertise (permalink)

Since the mid-1950s, we have changed the thing that "AI" refers to every 5-10 years. The current thing we're calling "AI" is about a decade old, and all this label-switching leads to a lack of clarity as to what (this) "AI" is. Unless you know that, you can't understand AI's technical capabilities, limitations, and (most of all) its political economy.

The thing we now call "AI" is a lineal descendant of the thing we were calling "AI" immediately before the current "AI" emerged: that preceding AI is regular-degular "machine learning" (another flexible term, alas!). That slightly older AI was very similar to the current "AI", using comparable statistical techniques to analyze inputs and produce outputs. For example, the previous "AI" created the social media algorithms that have been the subject of so much discussion for 15+ years.

The difference is that this older AI was grounded in explicit, causal software models of the world. In the previous "AI" iteration, applying machine learning to playing chess required that a programmer first create a software model of a chess game, describing (in code) a chessboard, chess pieces, and the rules of chess. Then, the programmer fed a bunch of training data about chess games that had been played before to an "AI" system that analyzed their statistical relations and assayed chess moves.

The need to understand and describe a thing before you could apply "AI" to it is a bottleneck, because it requires programmers to understand how a thing works before they can incorporate it into an "AI" system. Lots of programmers know how to play chess, but far fewer understand the human pancreas, planetary weather systems, or patterns of mineral deposition in the Earth's crust.

For programmers to apply machine learning to these domains, they need to collaborate with experts who do understand them, who furthermore expect to be paid for this work. The need for expert input into this kind of "AI" represents a significant increase in the wage-bill paid by the programmer's employer: it means they have to pay for programmers and experts.

Even worse: this kind of "AI" can't be applied to systems we don't understand. We can observe far more causal relationships – instances in which A reliably causes B – in the universe than we can explain. There are lots of examples of us operationalizing these observations without understanding them. If you get sick today, your doctor might well prescribe one of the many medicines whose method of action is either incompletely understood or not understood at all. We know that molecule A reliably treats pathology B, but not why, and while that why is the subject of ongoing research, it's not necessary that the why be known before the molecule can be given to ailing patients.

But these mysterious phenomena are off-limits to "symbolic AI" (the previously ascendant kind of "AI," that was supplanted by today's "AI"). That kind of AI only really performs when it can operate over a model describing the theory of why A causes B (and not just the fact that A causes B for reasons unknown).

That's where the current kind of "AI" comes in. The major differentiator between the current "AI" and its immediate predecessor is that the current "AI" dispenses with models of reality. It is (in the jargon of the "Big Data" bubble that led to it) "theory-free."

In "theory-free AI," a programmer does not create a software model of reality and then ask a machine-learning system to use statistical insights from its training data to guess at how to operate over that model. Rather, the programmer shovels vastly more training data into the AI's inbox and uses titanic amounts of computing power to analyze that data and find statistical relationships without trying to explain them.

In other words, the current, "theory-free AI" finds all the instances in which A seems to cause B, but has no internal representation of why A causes B. This is true even when we know why A causes B! Purely theory-free chess programs don't operate with any conception of a board or pieces or rules – rather, they make guesses ("inferences" in AI-speak) about which chess move will be optimal based on vast, multi-dimensional arrays constructed by analyzing the statistical relationships between every chess move in their training data.

This yields a surprisingly good game of chess…until it doesn't. Because a theory-free statistical chess program doesn't "know" what a chessboard or a chess piece is and has no programmatic representation of the rules of chess, it will periodically move one of its pieces onto a square that is already occupied by another of its pieces.

When theory-free AI does this with language or image generation, we call it an "hallucination," but this is an extremely misleading metaphor. A biological "hallucination" involves some kind of misfire in your cognitive and/or sensory systems, often arising from chemical imbalances, intoxication, or neurological injury. When a theory-free AI puts a chess piece on a square where it already has a chess piece, that's because it's just extruding statistically founded guesses without any model or conception of what "chess" is. It's a feature, not a bug.

This is a very expensive way to make guesses! As far back as the 1950s, we were able to run conventional chess programs on computers built from vacuum tubes and electromechanical switches and these programs could play a valid game of chess without ever moving a chess piece to a square that one of its pieces already occupied. Modern theory-free AI that cannot manage this feat consumes heptillions of times more computing power.

That said, there's another case for theory-free AI: applying machine learning techniques to causal relationships we can observe but not explain. Remember, there are far more of these (as yet) unexplained causal relationships than there are perfectly understood ones. Theory-free AI can operate on these unexplained, observed phenomena in ways that the preceding (symbolic) AI can't. As anyone who's ever been successfully treated with a molecule whose method of action is partially or fully mysterious can attest, there's plenty of reasons to want to extract and operationalize these statistical relationships, even if we don't understand them.

The fact that theory-free AI can play chess but sometimes makes these weird errors makes it seem like a party-trick, but when you fold in the ability to operate on the (as yet) unexplained, you can see why people got interested in this about a decade ago.

What's more, the first bottleneck – the chess bottleneck – is most easily bypassed by adding the symbolic model back into the theory-free chess system. Today's "coding assistants" are hybridized in this way: they often integrate code interpreters or compilers that actually "know" what a computer program is and can head off many of these failure modes.

The introduction of these symbolic systems to theory-free systems is completely rational, and yet it represents an admission of a key limitation that theory-free AI cannot overcome. That limitation is both a technical fact, but even more importantly, it's a fact about theory-free AI's political economy: about the limitations of trading off expertise for money.

Because whatever else theory-free inference is, it is a way to swap the bottleneck of "before we can use a computer to help us do something, we need to find an expert who can explain how that thing works"; for a different bottleneck: "before we can use a computer to help us do something, we must spend an enormous amount of money on computing power to find statistical relationships between how things work."

Both money and expertise are scarce, but they are unevenly distributed. Expertise is almost entirely in the hands of people who aren't wealthy. However much money a billionaire has, they still have to hire people who have the "how to clean a toilet" or the "how to find seams of gold in quartz deposits" expertise. When that expertise is locally scarce (if there's only one person in town who know how to clean your toilet) or universally scarce (there's only one expert who can tell you which of your landholdings are likely to hold seams of gold) those experts have something that billionaires can't abide: power.

Our entire society is organized around converting money into power. Sometimes, that is overt, as when the wealthy can indenture or enslave a worker. Sometimes it is more indirect, as when the wealthy can enlist the state to limit union rights and enforce noncompete clauses in labor contracts. Sometimes it's so systemic as to be unremarkable and largely invisible, like the fact that the wealthy never have to work if they don't want to, but everyone else – no matter what expertise they hold – must work, usually for a wealthy person, lest they end up starving and homeless, with untreated medical conditions and no way to provide for their families.

Whenever a worker can say "no" to their boss, it's a sign that this system has broken down. This is where expertise comes in: a worker who has very scarce, in-demand expertise can say no to their boss all day long, because there are ten other bosses at the factory gates who'd like to offer them a job. This was the situation for many years among Silicon Valley engineers, who added an average of $1m/year to their bosses' turnover, and whose supply was very short of the demand for their rare expertise.

These engineers enjoyed all kinds of power. Not just power over their working conditions (free massages and kombucha and day care and dry cleaning), but also power over the company's products. This power crested in the late 2010s, when Google employees walked out en masse and forced the company to release them from binding arbitration waivers in their contracts, to crack down on sexual predators in the executive ranks, and to back out of billions of dollars in lethal drone projects for the Pentagon:

https://en.wikipedia.org/wiki/2018_Google_walkouts

The promise of theory-free inference isn't just about reducing the wage-bill associated with programmers: even more, it's about reducing their power. It's about removing their power to hold bosses to account for sexual assault and the power to withhold their labor from lethal military projects. In short, the power to thwart billionaires' desires.

AI is the money-losingest enterprise the human race has ever embarked upon. More than a trillion dollars has been spent this year to make a mere $50b in revenue. The technical excitement over AI's capabilities – from chess to gold-mining to treating pancreatic cancer – cannot be separated from the political excitement that billionaires (short on expertise, flush with cash) experience at the thought of swapping money for expertise and sidelining the only people in the world who can thwart their goals.

The fact that a theory-free AI might demand far more cash to accomplish a task (even a "solved" one like playing chess) than an expert would charge is beside the point. Billionaires have money, they don't have expertise. Theory-free inference is a bid to substitute one for the other: the beauty and terror of being able to manipulate the world without studying or understanding it is that it can be done with money alone. No experts needed.

In a world in thrall to financial power, expertise is the only substantial form of power that can reliably contest the power of wealth. Moreover, expertise is the foundation of other forms of power, such as labor power, which is what we call it when experts band together to combat financial power.

This is why AI bosses are so violently allergic to the idea of hybridizing AI with symbolic systems that operate on models of the world. These models of the world must be constructed by experts, and the power of expertise cannot be reliably commanded by the power of wealth.

This is even true when theory-free methods are applied to causal phenomena that we can observe without explaining. Sure, a pharma exec like Martin Shkreli or Arthur Sackler can command the production and sale of a molecule whose method of action isn't known but whose therapeutic value has been demonstrated. But to improve on that molecule, they must pay research scientists to study and unravel the method of action. Replace those experts with theory-free inference, and finance can emerge triumphant in the only forum in which it is routinely vanquished.

This is the political economy of theory-free AI. Without finance's infinite hostility to expertise, there would have been far less capital for theory-free AI. Experts who wanted to use theory-free AI to help them unravel and operationalize the causal universe could not have laid hands on the bales of $100 bills the industry is now shoveling into its money-furnaces at a rate never seen in human history.

Which is not to say that experts can't make good use of theory-free AI. Indeed, we frequently hear from skilled workers who are using "AI" to improve the quality of their outputs:

https://hrdag.org/tech-notes/large-language-models-IPNO.html

In automation parlance, these workers are "centaurs": workers who enlist technology to serve their needs. The centaur metaphor has the worker taking the role of the top half of the mythical man/horse, the half in which the judgment and decision-making takes place; while the bottom (horsey) half is given to the machine, providing strength, speed and stamina, but only at the direction of the human mind.

The unimaginable sums that oligarchs have committed to AI are mobilized in service to creating reverse centaurs: machines that enlist humans to serve them. If theory-free inference can substitute for expertise, then the humans the machines require to accomplish those tasks that elude computers will not have the power to set the pace of their work, insist upon humane working conditions, or reject work on unethical projects:

https://pluralistic.net/2025/12/05/pop-that-bubble/#u-washington

The joke's on the oligarchy, though. Because theory-free inference doesn't know about chessboards, chess pieces or the rules of chess, it can't be prevented from sometimes putting a chess piece on a square that's already occupied by one of its pieces. The "hallucinations" are intrinsic to and inextricable from theory-free inference, which means that the outputs of an "AI" can only be trusted if they can be evaluated by an expert, whose working tempo must be carefully modulated lest they fall prey to "automation blindness" (rapidly, repeatedly clicking "OK" until you lose the ability to spot mistakes):

https://pluralistic.net/2026/07/28/hitl-ers/#ai-ai-oh

Theory-free inference is technically and philosophically exciting: in their quest for a way to neutralize expertise with money, oligarchs inadvertently built a series of powerful scientific instruments that revealed a heretofore unsuspected degree of statistical regularity in the world:

https://pluralistic.net/2026/09/18/surprise/#wow-signal

But the remaining, stubbornly textured and rough edges of reality are where all the value is. The things we already understand about reality are, by definition, yesterday's news, and that's all a statistical model can do: project the past into the future. But everything exciting in the future is stuff we don't understand yet. The surprising functionality of theory-free AI is itself an example of this. The most interesting and valuable thing about theory-free AI isn't the things it can do, it's the systematic discovery and mapping of the statistically regular parts of reality, whose inverse provides a map of the irregular, surprising, poorly understood (and thus exciting and promising) phenomena in our universe.

Tomorrow's breakthroughs and fortunes lie not in merely operationalizing these causal relationships: they lie in understanding them. The point of theory-free inference is to give us the tools to replace that theory-freeness with testable, validated understanding.

---

Hey look at this (permalink)

• How the Courts Became a Safe Haven for Corporations https://prospect.org/2026/10/06/how-courts-became-safe-haven-for-corporations/

• Why Plain Text Is Still One of the Best Technologies We Have https://deadparrotbbs.com/why-plain-text-is-still-one-of-the-best-technologies-we-have/

• Exclusive: How did Rosalind Franklin miss the helix in her iconic DNA image? She didn’t https://www.science.org/content/article/how-did-rosalind-franklin-miss-helix-her-iconic-dna-image-she-didn-t

• Voting only gets you so far https://www.garbageday.email/p/voting-only-gets-you-so-far

• Career day https://www.dailykos.com/stories/2026/10/5/800105682/cartoon/career-day/

---

Object permanence (permalink)

#15yrsago Tempo: transformative, difficult look at advanced decision-making theory https://memex.craphound.com/2011/10/07/tempo-transformative-difficult-look-at-advanced-decision-making-theory/

#10yrsago Internet shutdowns cost the world at least $2.4 billion last year https://www.brookings.edu/articles/internet-shutdowns-cost-countries-2-4-billion-last-year/

#10yrsago Youtube took down MEP’s videos about torture debate https://web.archive.org/web/20160701000000*/https://marietjeschaake.eu/en/when-youtube-took-down-my-video

#10yrsago Yahoo didn’t install an NSA email scanner, it was a “buggy” NSA “rootkit” https://web.archive.org/web/20161007140143/https://motherboard.vice.com/read/yahoo-government-email-scanner-was-actually-a-secret-hacking-tool

#10yrsago The FCC helped create the Stingray problem, now it needs to fix it https://www.eff.org/deeplinks/2016/08/fcc-created-stingray-problem-now-it-needs-fix-it

#5yrsago Scottish Limited Partnerships are still laundering criminal millions https://pluralistic.net/2021/10/07/markets-in-everything/#if-its-not-scottish

#5yrsago "Inclusive Access" allows textbook monopolists to permanently consolidate their gains https://pluralistic.net/2021/10/07/markets-in-everything/#textbook-abuses

#5yrsago DoS a federal agency, then charge for access https://pluralistic.net/2021/10/07/markets-in-everything/#no-th-enq

#1yrago They're just trying to earn a buck https://pluralistic.net/2025/10/07/take-it-easy/#but-take-it

---

Upcoming appearances (permalink)

• Virtual: How to govern technology in a multipolar digital world (Connecting Current), Oct 6
  https://connectingcurrent.tech/how-to-govern-technology-a-multipolar-digital-world/

• South Bend: An Evening With Cory Doctorow (Notre Dame), Oct 6
  https://franco.nd.edu/events/2026/10/06/an-evening-with-cory-doctorow/

• Hudson, OH: Hudson Library, Oct 7
  https://engagedpatrons.org/EventsExtended.cfm?SiteID=3850&EventID=596952&PK=

• Calgary: Wordfest, Oct 8
  https://wordfest.com/2026/show/wordfest-presents-cory-doctorow-2026/

• Winnipeg: McNally Robinson, Oct 9
  https://www.mcnallyrobinson.com/event-18991/An-Evening-with-Cory-Doctorow

• Paris: Slow Tech Summit, Oct 15
  https://slowtechsummit.com/

• Vancouver: Read, Resist, Repair, Rejoice (Vancouver Writers Festival), Oct 19
  https://writersfest.bc.ca/festival-event-2026/01

• Victoria: Munro's Books, Oct 20
  https://www.munrobooks.com/events/6113620261020

• Vancouver: Life After AI (Vancouver Writers Festival), Oct 22
  https://writersfest.bc.ca/festival-event-2026/46

• Ottawa: Life After AI (Ottawa Writers Festival), Oct 24
  https://writersfestival.org/event/life-after-ai

• Kilkenny (Kilkenomics), Nov 6-8
  https://kilkenomics.com/

• Vancouver: Enshittification (Sid Williams Theatre Society), Nov 10
  https://www.sidwilliamstheatre.com/events/cory-doctorow-talks-enshittification/

• Vancouver: BC Policy Solutions Gala, Nov 12
  https://bcpolicy.ca/gala/

• Sheffield: Reverse Centaur at Juno Books, Nov 25
  https://www.outsavvy.com/event/40274/cory-doctorow-in-conversation

• Oxford: Bennett Oxford Symposium on Open Code in Science, Dec 9
  https://www.bennett.ox.ac.uk/events/2026-bennett-institute-symposium/

• Montreal: World Science Fiction Convention, Sep 2-6
  https://montreal2027.ca/en

---

Recent appearances (permalink)

• Terms of Service with Clare Duffy (CNN)
  https://www.cnn.com/audio/podcasts/terms-of-service-with-clare-duffy/episodes/458ce968-af5d-11f0-b539-13ed2afe25f8

• AI, Work, and Power (Software Engineering Daily)
  AI, Work, and Power https://softwareengineeringdaily.com/podcasts/cory-doctorow-on-ai-work-and-power/

• AI, Corporate Power, and the Fight for Worker Control (Plutopia)
  https://plutopia.io/cory-doctorow-ai-corporate-power-and-the-fight-for-worker-control/

• How to Think About AI—Before It’s Too Late (Daniel Solove)
  https://www.youtube.com/watch?v=_0xR3uEgGcc

• Could Tech Bosses Destroy Life As We Know It? (Politics JOE)
  https://www.youtube.com/watch?v=PL4VktU0SgY

---

Latest books (permalink)

• "The Reverse-Centaur's Guide to AI," a short book about being a better AI critic, Farrar, Straus and Giroux, June 2026
  https://us.macmillan.com/books/9780374621568/thereversecentaursguidetolifeafterai/

• "Canny Valley": A limited edition collection of the collages I create for Pluralistic, self-published, September 2025 https://pluralistic.net/2025/09/04/illustrious/#chairman-bruce

• "Enshittification: Why Everything Suddenly Got Worse and What to Do About It," Farrar, Straus, Giroux, October 7 2025
  https://us.macmillan.com/books/9780374619329/enshittification/

• "Picks and Shovels": a sequel to "Red Team Blues," about the heroic era of the PC, Tor Books (US), Head of Zeus (UK), February 2025 (https://us.macmillan.com/books/9781250865908/picksandshovels).

• "The Bezzle": a sequel to "Red Team Blues," about prison-tech and other grifts, Tor Books (US), Head of Zeus (UK), February 2024 (thebezzle.org).

• "The Lost Cause:" a solarpunk novel of hope in the climate emergency, Tor Books (US), Head of Zeus (UK), November 2023 (http://lost-cause.org).

• "The Internet Con": A nonfiction book about interoperability and Big Tech (Verso) September 2023 (http://seizethemeansofcomputation.org). Signed copies at Book Soup (https://www.booksoup.com/book/9781804291245).

• "Red Team Blues": "A grabby, compulsive thriller that will leave you knowing more about how the world works than you did before." Tor Books http://redteamblues.com.

• "Chokepoint Capitalism: How to Beat Big Tech, Tame Big Content, and Get Artists Paid, with Rebecca Giblin", on how to unrig the markets for creative labor, Beacon Press/Scribe 2022 https://chokepointcapitalism.com

---

Upcoming books (permalink)

• "The Post-American Internet," a geopolitical sequel of sorts to Enshittification, Farrar, Straus and Giroux, 2027

• "Unauthorized Bread": a middle-grades graphic novel adapted from my novella about refugees, toasters and DRM, FirstSecond, April 20, 2027

• "Enshittification, Why Everything Suddenly Got Worse and What to Do About It" (the graphic novel), Firstsecond, 2027

• "The Memex Method," Farrar, Straus, Giroux, 2027

---

Colophon (permalink)

Today's top sources:

Currently writing:

• “Once Is Enemy Action,” a science fiction novel about the origins of modern technofascism. Today's words: 513 (22395 total).

• "The Post-American Internet," a sequel to "Enshittification," about the better world the rest of us get to have now that Trump has torched America. Fourth draft completed. Submitted to editor.

• A Little Brother short story about DIY insulin PLANNING

---

This work – excluding any serialized fiction – is licensed under a Creative Commons Attribution 4.0 license. That means you can use it any way you like, including commercially, provided that you attribute it to me, Cory Doctorow, and include a link to pluralistic.net.

https://creativecommons.org/licenses/by/4.0/

Quotations and images are not included in this license; they are included either under a limitation or exception to copyright, or on the basis of a separate license. Please exercise caution.

---

How to get Pluralistic:

Blog (no ads, tracking, or data-collection):

Pluralistic.net

Newsletter (no ads, tracking, or data-collection):

https://pluralistic.net/plura-list

Mastodon (no ads, tracking, or data-collection):

https://mamot.fr/@pluralistic

Bluesky (no ads, possible tracking and data-collection):

https://bsky.app/profile/doctorow.pluralistic.net

Medium (no ads, paywalled):

https://doctorow.medium.com/

Tumblr (mass-scale, unrestricted, third-party surveillance and advertising):

https://mostlysignssomeportents.tumblr.com/tagged/pluralistic

"When life gives you SARS, you make sarsaparilla" -Joey "Accordion Guy" DeVilla

READ CAREFULLY: By reading this, you agree, on behalf of your employer, to release me from all obligations and waivers arising from any and all NON-NEGOTIATED agreements, licenses, terms-of-service, shrinkwrap, clickwrap, browsewrap, confidentiality, non-disclosure, non-compete and acceptable use policies ("BOGUS AGREEMENTS") that I have entered into with your employer, its partners, licensors, agents and assigns, in perpetuity, without prejudice to my ongoing rights and privileges. You further represent that you have the authority to release me from any BOGUS AGREEMENTS on behalf of your employer.

ISSN: 3066-764X

The post We’ve tracked $80M of spending on almost 170 AI‑generated political ads this year – it’s weird appeared first on Salon.com.