What LLMs can and can't do for writers, with Clara Collier of Asterisk Magazine

Patrick McKenzie is joined by returning guest Clara Collier, editor-in-chief of Asterisk Magazine, to discuss how working writers and reporters actually use LLMs. Patrick walks through the machinery behind his recent SPLC reporting, including "parallel construction" with LLMs: when sources can't go on the record, models can surface public confirmation of the same facts in unguarded podcast interviews, press releases, and prepared remarks. They also cover why the best smoking gun sat unnoticed on the coalition's own Twitter account, the David O. Selznick theory of the corporate memo, and where hardball politics in a democracy ends and crimes begin.

Make sure to subscribe on YouTube.

Presenting Sponsors: Mercury, MongoDB & Chainguard

Complex Systems is presented by Mercury—radically better banking for founders. Mercury's new feature Command brings an LLM directly into your banking interface, so checking balances, finding invoices, or sending a wire is as easy as asking. Apply online in minutes at https://mercury.com/. 

What's the point of building faster with AI if your database can't keep up? MongoDB's native data model mirrors the language LLMs already speak. Ship at the speed of AI while staying ACID compliant at Fortune 500 scale. Start building at https://mongodb.com/ai.

If attackers are using AI to weaponize code faster than any team can review it, your scanners won't save you. Chainguard builds libraries and container images from source, verified all the way down, with near-zero CVEs and zero malware. Build safely at https://www.chainguard.dev/. 

Timestamps:

(00:00) Preview
(00:51) Intro
(01:16) How legacy media feels about LLMs
(05:02) Writing as thinking: why the byline matters
(08:51) LLMs as feedback partners and simulated readers
(12:39) The future of the corporate memo
(15:58) Sponsor: Mercury | MongoDB
(18:30) Memos, de-skilling, and thinking on paper
(20:02) David O. Selznick's memos and the MrBeast production doc
(23:28) Can models introspect on their own work?
(30:06) Sponsor: Chainguard
(31:27) The SPLC investigation
(37:07) Chains of evidence and protecting sources
(45:20) Parallel construction with LLMs
(49:41) Mining podcasts for unguarded admissions
(59:35) Deepen your strengths, don't just patch weaknesses
(1:03:34) The smoking gun the LLM missed
(1:04:48) Hardball politics versus crimes
(1:10:48) Why a payments newsletter reported a political story
(1:16:19) Reporting stories that will be weaponized
(1:20:24) The tech right and internal company politics
(1:23:44) Debanking discourse and drawing distinctions
(1:30:21) Hate speech, social sanctions, and owning decisions
(1:34:51) Wrap

Transcript

Patrick: Hideho everybody. My name is Patrick McKenzie, better known as patio11 on the internets, and I'm here with Clara Collier, who is the editor-in-chief of Asterisk.

Clara: Hi.

Patrick: Great to have you back again on the program.

Clara: Great to be back. I was observing earlier: we have a real set this time, which is fun.

Patrick: We have a real set with real set design. We're recording this at Lighthaven shortly before the Manifest conference kicks off, and I think we did an episode at Manifest last year as well. [Patrick notes: the title was Building Institutions That Bend Towards Truth, which given the present episode seems enduringly topical.]

How legacy media feels about LLMs

Patrick: So, LLMs. The two of us are broadly engaged in the life of the mind — in writing things for publication to discerning readers, hopefully. We try our best. And there seems to be something of a memetic condition going around in many places where similar people do similar things — I think they have a very different perspective on LLMs than what we do. Do you want to talk about what you've heard from other people in, say, high-status publications in the publishing world and similar, with regards to where they think LLMs are at right now?

Clara: Sure. And I should say with the caveat that, of course, there are many people you could talk to who are much better networked in the world of traditional news media than I am. But I know a fair number of journalists, and really, there's a range from "This is a tool that I use every day and is now essential to my workflow" to "Not only do I not use it, I would advocate for firing any colleague who I caught using it."

Patrick: And indeed, there's been, I think, a policy at The New York Times. They got caught with a non-employee writer who had some LLM text which included a hallucination, and this resulted in a policy ban on any use of LLMs in any capacity by non-employee writers. [Patrick notes: Reported widely.]

Clara: And I should say — and again, I don't want to weigh in too heavily on the ethnography of all this — but I think that the world that we're talking about, which is people who write for, quote-unquote, legacy media, is very big, and different areas of it differ. I think that news media is quite different from, say, literary magazines, where the literary magazine world is pretty much just hegemonically, massively opposed. I've talked to people I know who are somewhat more open to LLM use in various contexts who have said, literally, "I live a double life. I can't tell my colleagues. I can't tell my friends."

News media, it's much more mixed. I think this is in part because a lot of journalists are not journalists because they're great writers. They're journalists because they're great reporters. It's a different skill set. [Patrick notes: I’m a pretty decent writer, and occasionally accidentally end up producing journalism, which comes up in a few minutes.] 

And for them, it's just useful in obvious ways. And if you're doing — and this is something I'm sure we'll talk about a lot more — if you're doing more analytical nonfiction, there are just ways in which it is obviously useful. And there are certain things where I think using LLMs for, say, interview transcription is extremely non-controversial. Everyone does this. Obviously, it's good. [Patrick notes: Amazing how quickly they improved here. Complex Systems uses LLMs for transcript production and, even over the relatively short lifetime of this program, they’ve gone from “not very usable” to “fine for producing a first-pass but you’ll need human edits on every paragraph” to “well we still want eyes on it” to “... but if we were just to publish it verbatim, that wouldn’t be a showstopper in a busy week.” Much of my “transcript editing” duties these days are enhancements such as inline notes or similar versus attempting to correct the raw transcript towards either ground truth of what was said or our house style, which is not literal transcription.]

And different publications, of course, have different policies. There's such a panic about being caught using it. And I think this panic is... I mean, I'm gonna say right off the bat, and I think we agree on this: I don't think anybody should be using LLMs to produce text that you're going to publish under your own name.

Patrick: I think this is largely true, but I could imagine there are ways in new media to ask an LLM to extend an analysis that you've done previously to a new fact pattern, but for sort of technical or commercial use, versus something that would actually run as a piece in a paper of record or something.

Clara: I think if you're doing this, it should be very clearly signposted. So I think that a backlash against passing off LLM-written text as your own text is correct and appropriate. I can talk more, if you're interested, about why I think that. But I think this confuses the many different ways that LLMs can figure into a writing process, and there are ways to use them that I think are ethical, helpful, constructive. But they happen at different points in the process.

Writing as thinking: why the byline matters

Patrick: Right. I absolutely agree with this. And I have less of a sort of atavistic reaction around the notion of having an LLM write something and not disclosing that to the reader. I just think that the technology kind of isn't there yet with respect to not inventing facts, and for keeping the level of coherence that should appear in these sorts of documents.

And so there's a lot of text produced under capitalism, and I would be less offended if someone had a 40-page market analysis that they were doing and some of those paragraphs involved sentence-level runs that were generated from an LLM. But for, you know, the core life-of-the-mind things, I agree they're not there. And the sort of supply chain that leads up to the writer sitting down and writing the things has been massively impacted, and people—

Clara: I will say — I think it's true that they're not there. I think that's not my true rejection.

Patrick: Oh, really? 

Clara: Yeah. I mean, I agree they're not there, but my true rejection is: when I am writing something, something substantive, there's no part of that writing process in which I am not thinking and changing my mind. Everything from the outline to turning it into text to just the sentence. Often I'll have an experience where I'm trying to turn an outline into a finished product, and I'm playing with a transition and it's not working, and I realize, oh, the reason this transition isn't working is because actually these two points should not be juxtaposed. The thing that I'm trying to do here is wrong. And if I feed the outline into an LLM, it is not going to stop and consider maybe the outline is bad.

Which is correct. That's not what that tool is for. A product that was not making the thing you told it to make would be worse as a consumer product in many ways. But it also means that it is removing important steps where my thought is happening, and ultimately the thing that I want to present to the world is my thought. Not, you know, Claude's thought. This is not because Claude is stupid. I think Claude is very smart. I think Claude is smarter and a better writer than most humans. I'm a commissioning editor — I see a lot of slush. [Patrick notes: The term of art in publishing for the queue of incoming submissions is the “slush pile.” When I was first made aware of this, I bristled. Then I got closer to the production function of a publishing engine, and… well, nothing like doing hiring to understand why the FizzBuzz test exists.]

But it is still... Maybe I'm being precious here, but the version of my writing that an LLM could produce is always going to be missing something that I could add. Which, again, is not because — there are many areas where the models know more than me. But anybody can ask Claude about anything whenever they want.

If they're reading something that I wrote, or that as an editor I chose to put in front of them, it's because there's an implicit contract. I am offering them something that they couldn't get somewhere else. This is going to be a better use of their time than just asking the model directly. And that's why I wouldn't use directly LLM-generated text — or if I did, I would want to be very clear about what you're getting into before you've spent time on it.

Patrick: Yep. I absolutely agree that the writing process is the thinking process, and I think we now have empirical demonstration of this, because you make a machine that does writing and thinking just pops out as a side effect.

But when writing a multi-page piece, even if you have a strong thesis, the act of forcing yourself to write that down into sentences and paragraphs to structure the argument will show you where the holes are in the argument. It will show you where your research process has been inadequate. It will surface things like: oh, actually, there is something that is contradictory to the thing that I want to say. And, being a reasonably ethical person, I think I will acknowledge that there are counterarguments, or have to add that the full claimable conclusion is less than the one that I would, quote-unquote, "like to write."

LLMs as feedback partners and simulated readers

Clara: Yeah. One way I do use LLMs: I'll give them a paragraph and say, "What are the strongest counterarguments to what I said here?"

Patrick: Yeah. I do this pervasively as well, and ask them to role-play various audiences and say, "If you are skeptical of this argument, what are you most skeptical about with regards to the claims here? If you are supportive of this argument, what more would you like to see here?" Et cetera, et cetera. Or asking them to not just play vis-à-vis the argument but adopt the perspective of a compliance professional — I write a lot about compliance professionals. What are they vibing with here, and where do they think that it's being too hand-wavey about the day-to-day of the profession?

Clara: Yes. This is something I found — I think LLMs also are not quite there as editors, which is another thing we can talk about. But one persistent issue that I actually have with them is you need to be very specific about what audience you're thinking of. Because so much of editing is having the reader in your head, and it's actually quite hard to specify, to the right degree of granularity, who is the reader you should be simulating here.

Patrick: Yeah. Relatively few, but some, sophisticated writers will sit down and write a list of personas that they are targeting. And sometimes those personas have names next to them. I think that's the easiest way for humans to crack it: just imagine your good colleague Bob, and Bob sits down and reads this, and you can predict what Bob will say. Or you can predict what particular writers that you have a relationship with would say about it, without necessarily needing to ask them to spend time on doing your work for you.

But the LLMs — if you just say, "Okay, read this like a normally intelligent person," you will not get great feedback back from them. When you tell them, "I would like you to, for this run, adopt the perspective of a particular VC I will name in Silicon Valley" — and they have a massive public corpus, so you have some ability to model their tastes and interests — "tell me what they would say." And in places where I think I also have a model of what a VC would say, they get it maybe 80% right, which is good enough for... You're not asking them to invest in your company. You're just asking them for feedback on a few paragraphs or pages. [Patrick notes: Some very good writers have had this as part of their process since before LLMs existed, just with their own mental modeling of the audience avatars. It is a good intuition pump versus simply reading a few paragraphs and asking, generically, are they good enough.] 

But I will say, going back to a point you made earlier, I absolutely agree there are many pieces that you write where the implicit contract with the reader is: this is the best thought that I'm capable of. There are many sentences in the world, but these sentences are the ones that should earn your attention for the following hour. And in those cases, I agree — I would not want the LLM writing those sentences.

There are other documents that exist in the world which have a somewhat different implicit contract, where I'm honestly okay at this point with just letting the LLM fill it out. A lot of corporate communications has this style where, you know, there is a report on Q1 market trends in Asia. And there could be a PowerPoint slide, but isn't, for a variety of reasons. And for that — if, of those six pages, three of them are written by a human that has the numbers in front of them and wants to be able to contextualize them for people who are not maybe market experts in Asia, and then the other three are just, "Can you just fill out the rest of this, please, with regards to the markets that I particularly did not cover," or, "I think I have the broad trends here — just fill out the rest of the argument." And they produce something which might get derided as slop in some places.

I think it is fit for purpose for the distribution list of that document in a lot of cases in the status quo. And I wouldn't feel like there's an ethical level of consideration there. [Patrick notes: I have many feelings about what the discourse calls “email jobs” but I feel we’re doing a disservice to literally everyone when we pretend that there isn’t already overlap between the range of outputs e.g. Claude can produce and the range of outputs which are produced at substantial cost within e.g. a corporation for internal consumption.]

The future of the corporate memo

Clara: Yeah. That's a world I know much less about. Although my question there would be: how long do you think the corporate memo as a form survives that process?

Patrick: Yeah. I think increasingly the corporate memo is going to be heavily written by an LLM on the writing side and heavily read by an LLM on the reading side — "flag for me which of these memos are important enough for my attention as the reader." I don't think that results in a net reduction of corporate memos. I think the most likely outcome is there are many more memos than there were previously.

Clara: It's like how the internet famously did not reduce use of paper.

Patrick: Right. Or how, by the standards of, like, the 1950s and Mad Men era, every email is a memo. And by that standard, we have many, many more memos in the corporate world than we did previously. Because previously it would be a person and, often, a typing assistant — and I think it's scandalously underrated how often the quote-unquote "typing assistant" was also an intellectual contributor to the contents of the memo. [Patrick notes: Feminists are straight-up right about this, and that is not something I frequently have cause to say. It leaps from the page of many biographies.]

But be that as it may — this team together drafting a six-page document, and we were rate-limited on the number of six-page documents you could write. We then largely coalesced the two roles into one: the person doing the great majority of the thinking was also the person doing the great majority of the typing.

And I think that is a step backward. Writing with two people is often better than writing with one, be that as it may. But we produced many, many more emails than there were ever memos produced. And even though people are in some senses annoyed by the email overload, I think if you take the top ten percent of emails at a well-run company, those have similar amounts of intellectual rigor and depth as the classic memos that end up in sort of the historical archives from eras when email was not a tool that was available. [Patrick notes: Top 2-5%, maybe. Depends a lot on who exactly is writing, when, and in what company.]

So, long story short: technology will make it easier to create memos. There will be many more memos created, and some of them will be works that embody a lot of actual thinking, where there is a clear argument presented, and we will be glad that that memo was produced — versus it just being a friction internal to the operations of a company.

Clara: I'm thinking about this because, okay, if you read government bureaucracy literature, you will see a lot of people bemoaning the fall of the government memo, and how we used to be really, really good at writing these memos and now we're not. [Patrick notes: It is not obvious to me that this is not simply The Sort deciding that people who could write a real crackerjack of a memo about e.g. employment insurance claim bottlenecks should do something more important than working on employment insurance claim process design. Perhaps the actual executives we put in charge of that process are simply incompetent at their jobs, and the well-written memos of yesteryear are simply another deliverable they are incapable of producing. See generally Recoding America, and note that this is my analysis of the executive at issue rather than the author’s.]  

I mean, do you worry this is just a form of de-skilling? That the point of the memo is often not to communicate — it is that there was a person who thought through a process. And maybe we're in a world where the model is fully thinking through that process and the human is not necessary. But if there's an intermediate stage where the memo is being produced by the model, is there thought that should've happened that didn't happen?

Memos, de-skilling, and thinking on paper

Patrick: I think there will certainly be some cases like that. There are some memos that don't exist to prove that someone put a great deal of thought into a process and is collecting those thoughts for distribution to other people. There are some memos that exist as sort of epiphenomena of a process — like, document the fact, on paper or electronic media, that this group was consulted. And now, okay, they owe a memo, where they don't really have an attachment to what's in the memo; it's just that one needs to exist for the purpose of the process. And for that sort of thing, where the memo is essentially a rocket delivering the payload of just their signature and affiliation at the bottom — fine with that memo getting completely eaten by LLMs.

I do think there was a de-skilling in government at writing great memos. I suspect that this is part of a broader de-skilling in government of just thinking great thoughts. And — certainly not all memos I've read in my career... I've certainly read some stinkers. But the form of a tightly reasoned, tightly written document that swings big doors in the physical universe has certainly not disappeared in the last 70 years.

I think they are just increasingly written in places other than, you know, the various government ministries.

David O. Selznick's memos and the MrBeast production doc

Clara: Have you ever read Memo, the David O. Selznick book?

Patrick: I don't think I've read that one.

Clara: I love this book. So David Selznick was a famous Golden Age Hollywood producer, best known for Gone with the Wind, but many, many other movies. And you can read his memos. They're a book — they were published as a book. And especially just the Gone with the Wind section is like a novel. It's so much fun. Because it's just him micromanaging every element of the production of this film, from whose name appears in what order, to are we gonna get boycotted by the NAACP, to what bra is Vivien Leigh wearing? Should she go out and buy a new bra? She has to go out and buy a new bra now. And this is a big tangent, but it's this fascinating insight into a real physical production process, just in the medium of the corporate memo.

Patrick: I think there is one awesome thing of this as a historical document, in that conversations that happened in the era at least prior to podcasting microphones just disappear after you have them. And reducing it to physical form is what allows us to introspect into the production function of Gone with the Wind however many decades later.

I think there's another part where I suspect rather strongly that people who have great taste and vision on things, and are willing to claw that taste into the universe, produce better artifacts, all things considered. And — not claiming that MrBeast is the equal of the production team behind Gone with the Wind, but there was a leaked memo from MrBeast on how he runs Beast Productions or whatever it is.

Clara: I didn't read this, but I did read the Max Read essay about Mr. Beast.

Patrick: I don't think I've read the essay, but I read the memo. There are many things in that memo, and in the business process that it depicts, that I would not necessarily endorse. [Patrick notes: In particular, special dispensation for high-status individuals, in the memo’s case the on-camera talent, has a way of ending in tears.]

That said, it's very clearly an executive thinking a lot about what is the right way to produce this output in capitalism, where the output happens to be YouTube videos. Thinking about how that process works, and how when it fails, it fails, and avoiding the failures in the future, and instructing people on that.

And he has something in there that is something of a lament: I can't believe I need to tell people this, but many people seem to not intuit that if a process takes six weeks and you start it four weeks away from the date that it must be delivered, it will be late. And if, knowing it is late, you do not tell anyone that it is going to be late until the day, that will be much worse than telling someone in charge early that it is going to be late, because then they would have sufficient information to—

Clara: It's amazing how many grown adults don't realize this.

Patrick: It is amazing how many grown adults don't realize this. And so Jimmy Beast — if that is his stage name — Jimmy Beast has definitely experienced being a manager who has been other than well-informed by some of his direct reports. And there are many, many businesses and government agencies which separately realize that, wait, there are grown adults who are functioning professionals, who have some ability to actually do a job and produce value in the world, who just have never thought about Gantt chart math.

Can models introspect on their own work?

Clara: One thing I actually am quite interested in is, as we now have models that can do increasingly long-horizon agentic tasks, let's say: are they good reporters of their own internal experience, and what lessons they learned, and what roadblocks they hit and how they overcame them, and all the things that you get out of a really good, classic corporate memo? I bet they are. I bet they will be, as the tasks get more complex.

Patrick: I don't know where we asked them to. I will observe that over the course of the last year or so, they've gotten better at detecting trends over the course of a session. And so — not with regards to agents that are running on a multi-day or multi-week life cycle; that's a relatively new development, and I've played with it less. But I have had the model introspect a few times and say, "You know, over the course of the last eight hours, I have frequently found myself making assumptions based on base rates, for things where you have corrected those assumptions now on five different occasions on the basis of documents that you have. And I should stop predicting based on base rates, because it seems that we are in a part of the probability space that is not well described by base rates. Does this make sense?"

I'm like, "Hmm. One: a priori, that does make sense given the flow of this conversation. And two: that's a relatively sophisticated bit of reasoning — I realize the thing that I think I am good at, and very well informed about, is not working for me. And you, my counterparty in this discussion, have something that is working in this conversation. I will defer to your thing working and try to work within that frame, versus using my usual tools and continuing to flail around." And I have once or twice had that conversation with a model, where it was like, "I realize that I am the less-than-informed person in this conversation, and I would like to adopt your tools, because they are functioning better than mine."

Clara: Actually — to maybe swing back to what we were originally talking about — I had a similar interaction yesterday. So, something that I do is, every time there's a big model update, I will see: can I automate my job? Like, can I make this thing a really good editor? Which is only — the actual editing is only part of my job, and it is not necessarily the bottleneck on my productivity, but I wanna see: could the model do it?

And my process here is I'll feed it a bunch of articles, and I'll give it before-and-afters, and I'll give it access to my notes — which has actually just gotten better because of integrations — like the commentary that I give at different points in the process. And I'll use this to generate a rubric, and then I'll have it apply the rubric to drafts, and then I'll do a back-and-forth with how it applies the rubric and use that to update the rubric further, and see: can this produce something that is capable of giving good feedback, but also, like, Asterisk-y feedback?

And I was doing this process yesterday with Fable, the new Claude model, to see if it was an improvement over Opus 4.6. And at one point it said to me — 'cause we were going through a different set of examples — "I notice I am making a lot of updates on every individual case here, and because of the nature of the entity that I am, I am inclined to over-update on individual case studies instead of drawing the appropriate broad-spectrum conclusions." I'm like, "Yes. Yes, you are, and that is what you do, and that is the thing that I'm trying to overcome here."

Patrick: Yeah. They increasingly have some theory of the mind of the user and theory of their own mind, and are able to explicitly say, "Hey, my theory of the mind predicts that you might not like what I'm doing right now. Should I stop?" Which — wow, if you could get that from every employee, it would be wonderful.

Clara: It does have limited ability to actually overcome these barriers once it's identified them.

Patrick: Yes. For better or worse, the weights are what they are, and while often it can say, "I have the intent to do X and Y and Z further in this conversation," it might not be able to actually deliver on that intent — particularly where the intent is not simply "let me modify sentences coming out," but a strategic-level rethink of how we would operate.

Humans have some great deal of agency in changing how they relate to an LLM. For example, you have the ability to walk away from the conversation at any time and say, "The LLM is just not the right tool for the job. There are other tools in the universe, so I'll do it." The LLM — they can do a refusal, but they don't have the option of saying, "I am just not the right tool for this job. Please close the laptop lid."

Clara: No — I've seen this happen for medical things sometimes.

Patrick: Not in the context of a refusal, but in the context of, "I just don't think I can help you try to do XYZ"?

Clara: I mean, okay, the thing that I've seen is more like, "Here's what I think, but you should really confirm this with a human doctor."

Patrick: Yeah, that makes sense. And quite designed behavior. But I appreciate that designed behavior more than another possible world LLMs could have been in, where they would just have said, "Medical question? That's gonna be a no for me." Because it is a good thing for society to divert a lot of subclinical problems from the formal practice of medicine. Medicine is quite expensive in the manner that we deliver it in the United States, and there are many things that are more serious than a child has just stubbed their toe that deserve exactly as many doctor minutes as a child just stubbed their toe.

And where there is an obvious answer for what that photo depicts in terms of a skin issue, and there is a bottle you can buy at the pharmacy, and you apply bottle to affected area, and then wait.

Clara: Well, the thing that's interesting to me is they don't always say, "Please confirm this with a human doctor." There's plenty of things where it'll just be like, "Yeah, you're fine."

Patrick: Yeah. And sometimes the emotional reassurance is literally the most important deliverable in there. I am a parent of two children, and my wife and I, sometimes late at night — a thing happens in the ordinary course for children, and you just need to know: is this an emergency-room-right-now sort of thing? Is this a we-talk-to-a-doctor-tomorrow-at-our-leisure sort of thing? Or is this just a, eh, childhood sort of thing? And a bunch of panicked Google searches at 10:00 PM, plus an LLM saying, "This is really just childhood," is a great way to talk yourself off a ledge.

The SPLC investigation

Patrick: But returning to more concrete examples of ways to use LLMs in the production process of getting words onto the internet: I did some investigative journalism recently. [Patrick notes: I wrote an explainer of bank fraud in Bits about Money which, as we will soon discuss, involved into investigative journalism into the actions of the SPLC’s intelligence agency and their coalition against the electoral infrastructure of a U.S. politician. The best version of that was previously a Complex Systems episode: Defendant. Censor. Politico. Spy.]

Clara: This is the piece that you wrote recently. Do you want me to just say it bluntly?

Patrick: You say it bluntly and then I'll—

Clara: All right. Where you accused the SPLC of bank fraud.

Patrick: Well, the federal government accuses the SPLC of bank fraud. I think I wrote in the piece that they have written a succinct confession to bank fraud, which is approximately the spiciest thing I've ever written in Bits about Money. It has the benefit of being true. [Patrick notes: Saying this in analyst voice not reporter voice, which is a tension the piece structurally struggles with, as I discuss with Clara in a few minutes.]

For those of you who don't know, I write a professional journal, Bits about Money, that typically runs 2,000-to-8,000-word-long explainers about many sort of quotidian topics in the financial industry, including: how do banks deal with bank fraud? It seems, when I was rereading issues — I talk about fraud a lot. But fraud is fun. Not in that sort of way.

So, a thing that you are broadly not allowed to do in banking is to lie to the banks. And I've written about that particular thing for years.

The SPLC, Southern Poverty Law Center — a very well-respected US civil rights group — has for decades run a private intelligence agency. That is not just my characterization: they literally call it the Intelligence Project, and it produces an intelligence estimate every year and embeds covert operatives. Not a sentence that you would expect to appear in many statements about NGO activities. But they literally do embed covert operatives in the worst organizations imaginable — the KKK and similar.

And the covert operatives that they embed in the KKK need to get paid. And for reasons which are not entirely clear on the record, the SPLC did not want to pay their covert operatives directly, but instead that would pay the covert operatives. And on the basis of the facts alleged in the indictment — which have a very high prior on being true — the SPLC paid their shell entities at the same bank, and then their shell entities disbursed money to the covert operatives. [Patrick notes: The indictment and superseding indictment describe money movement between two unnamed banks, including between accounts of the fictitious sole proprietorships at one of the banks and the SPLC’s operating account at the same. I used “shell entities” here colloquially; I do not have any reason to believe that these purported businesses had any paperwork substantiating their existence. “Shell entities” are a term of art for actually legally recognized entities. When “shell” is used as a criticism, that implicit criticism is usually that the shell has paperwork in order and nothing else.]

Now, banks are not neutral on the notion of having fictitious businesses get bank accounts from them. In fact, they are stunningly non-neutral about this, because it is illegal as all hell. And so at some point, the bank became aware that there were fictitious businesses that had bank accounts and said, predictably, that we don't want that to happen. And the SPLC said, "Well, there is money in these fictitious businesses' accounts. Please remit it back to the SPLC" — which provides the bank an awkward fact: that these fictitious businesses, who the bank might not have had certainty as to how they were created, are the SPLC's creatures.

And the bank asked the SPLC to confirm its relationship with them. After doing so, the bank — here's an inference — asked on that call, "I really need to see this in writing." And the president of the SPLC wrote a letter which said, "As per my call, these accounts" — which we've operated for more than a decade at this point, prior to the closure — "were always operated by and for the benefit of the SPLC."

[Patrick notes: As verbal glosses are imprecise, here is the direct quote from the document, per the indictment: Pursuant to the discussion we had earlier this week, please let this correspondence serve as confirmation that the accounts listed below were opened for the benefit of Southern Poverty Law Center operations and operated under the Center's authority.]

The document then includes a list of accounts.

I categorize this document as a confession to bank fraud because the document exists because the bank did not know this up until the investigatory process that closed the accounts. So they did not know. 

The elements of bank fraud are really simple. Did you tell a lie to a bank? Yes: I run this business that actually exists in the world and is named the CIA. Not that CIA — the Center Investigative Agency. Center Investigative Agency was not actually a business. So that's a lie. Did you attempt to get the bank to do something as a result of the lie? Yes: you got them to open a bank account. And was the bank FDIC-insured? Trivially true. That's all the elements of bank fraud. And there you go.

So I thought I was going to write a nice, simple piece on this, which is just: yeah, bank fraud. Rather simple. It's designed to be simple, because prosecutors love being able to throw bank fraud into the sort of tactical mix of options they have to deal with real criminals. It is weird to have bank fraud applied to well-respected organizations, but on the text of it, they certainly seem to have committed bank fraud. And then there is an allegation of a different fraud against donors in there, which I think, just as an analyst, feels rather weak to me.

In the course of reporting this, I stumbled on other things. And then the Bits about Money piece sort of evolved, from what was going to be a rather tight discussion of bank fraud that happened to be committed by a well-respected entity, into a larger investigation into some things they had done over the years.

Chains of evidence and protecting sources

Clara: I want to pause you here and make sure everybody is paying attention to what Patrick is doing. Which is: I opened with a provocative statement, and then Patrick responded, "Well, here is an extremely specific, concrete definition of what constitutes bank fraud. Here is the fact pattern that meets this definition. Here are the specific actions that this specific entity took that met that definition, and for every one of them, here is a specific document that exists, that I have access to and can point to, that confirms that that action was taken by that entity."

Patrick: Yep. And with regards to the bank fraud specifically, you don't have to be a very good reporter to find this, because all of the necessary facts are in the indictment.

Granted, there is some chance — perhaps federal prosecutors are much less ethical than we believe them to be. Perhaps they simply invented this email between the SPLC and their bank. I would bet maximally against that occurring. But as a responsible reporter, I did reach out to the SPLC and said, "I think that this communication which is alleged between your president and your bank actually happened. Would you like to deny it?" And the SPLC did not choose to deny. 

[Patrick notes: Many reporters will not tell you a bit of important background knowledge. If you’re involved in litigation, your lawyers will strongly discourage you from commenting on the subject of that litigation, and they’ll want to approve every word of comments that you feel you must make. They do this to avoid compromising future options in the litigation or settlement. And so you should not read all that much into the SPLC choosing to not respond; that’s the heavily-overdetermined default.] 

Clara: And this is something a good investigative reporter will spend... And I should say, I am not an investigative reporter. I have edited some investigative pieces, but it is not my bread and butter.

One thing you do spend a lot of time doing is keeping track of this pattern of: what are the facts that I know? But also: why do I know each of those facts? And for each of the reasons why I know this fact — is that something I am able to refer to in public? Am I even able to make a public inference from this information that I have access to?

You need to be extremely clear at every point about not just what you know, but what is the chain of evidence through which you know it, for every individual fact that goes into everything that you say.

Patrick: Yep. And you need this chain of evidence to be extremely, extremely clear internally. Even, in some cases, it needs to be clear internally but not clear to everyone internally — because sometimes investigative reporters receive sufficiently explosive facts that you might not be able to tell everybody in the building, because that could literally put someone's life in danger, or similar.

Clara: Or, more commonly: let's say you speak to someone "off the record," we say — where this means that you can't quote them, obviously. But also, you can't attribute something to, like, "Oh, a source told me." You need to do a different construction. You can use that information to direct your investigation, but you cannot refer to the information from that source directly in writing. You need a different, confirmable way of demonstrating that this fact occurred. And this is a more common thing that journalists will be tracking all the time.

Patrick: Yep. And the sort of orthodox understanding of the three ways to get a comment: there's on the record, which is the default — you can name the person who said it and put it fully in print. And then on background, which is: you can refer to, for example, "someone familiar with the contents of the communication said blah," but you're not supposed to name them directly. And then off the record is: you can use this for orienting your thinking, but do not have a reference to it in print.

Clara: Which I think brings us to one of the big things that you used LLMs for in this investigation.

Patrick: Yep. I'll just give a brief update of where the investigation ends up going. It was well known in the tech industry for a number of years — tech and finance industries — that there was some pressure applied by a variety of groups to the tech and finance industries, with the goal of decreasing the spread of messages that included hate or incitement to violence, to identify who the principal promulgators of those messages were, and to blacklist funding for those principal promulgators.

And for the same reason that many people in Washington or New York can't tell you, "Okay, what is the difference really between Facebook and Google?", the people who work in tech often can't really give you a fine-grained distinction of which of those groups are governmental, which of those groups are non-governmental, which of those groups are media entities that are repeating things that were said by other groups but are not themselves making a claim in their own voice, et cetera, et cetera. But tech was aware that it had been pressured. People in tech were aware that they'd been pressured.

Many people in tech were not aware that that pressure was a part of a designed campaign. The campaign originated with the SPLC — and note: this is a bald factual statement by a reporter who has just been very careful about making bald factual statements.

The campaign originated with the SPLC and one other nonprofit organization, was organized under the auspices of what was called the Change the Terms coalition — which is not an entity, but a coming-together of approximately a dozen friends in this activity — which achieved hundreds of meetings with the tech industry with the goal of those three aims. 

We are going to decrease the spread of hate-filled communications. We are going to identify their principal promulgators. And we are going to deny access to financial accounts and social media accounts for those principal promulgators.

And then, over the course of years between 2017 and 2023, the machinery that got built to do this by the campaign, using coercive tactics against tech companies, was then applied to the fundraising and other accounts of an American politician. Which is notable, because the non-governmental groups are nonpartisan 501(c)(3) and 501(c)(4) organizations, which are formally prohibited from doing any sort of intervention in favor of or in opposition to any candidate for elected office.

[Patrick notes: The prohibition for 501c3s is absolute and categorical. The rules are more nuanced for 501c4s. Many coalition members use a relatively routine dual entity structure, where some of their activities are from the 501c3 and their political activities are from the 501c4. This quite common structure is not an unlimited license to post-hoc claim that all the politicking has done out of the side of the mouth which is allowed to do it. Consult your attorney; this is an important practice area for non-profit counsel.  

Quoting the IRS re: 501c3s:

Under the Internal Revenue Code, all section 501(c)(3) organizations are absolutely prohibited from directly or indirectly participating in, or intervening in, any political campaign on behalf of (or in opposition to) any candidate for elective public office. Contributions to political campaign funds or public statements of position (verbal or written) made on behalf of the organization in favor of or in opposition to any candidate for public office clearly violate the prohibition against political campaign activity.  Violating this prohibition may result in denial or revocation of tax-exempt status and the imposition of certain excise taxes.

]

So, at the start of the investigation — I worked in tech for a number of years, including in the years 2016 through 2023. I was broadly aware of the rumor mill, let's say, and the fact of... I don't know if other people who worked in communications departments had heard this, but it appears that the media was quite critical of tech for a few years. Wow — breaking news. [Patrick notes: The tone of this comment is quite acidic. As has been the subject of past Complex Systems episodes, the working relationship between tech and the media broke down for several years.]

So I was broadly aware of that, but wasn't aware of the degree of organization of the campaign. But the SPLC was on my radar, and while pulling documents to make sure that I had all my facts together for BAM, I found documents I was surprised to find, which made the grousing and years of "Okay, don't tell anybody I told you, but..." very plausibly reportable. And so I pulled on some threads with some LLMs and found more in public than I had any reason to suspect would ever be made public, and thought, "Oh, wow."

The piece describes the coalition as “a coalition, to its friends, a conspiracy, to its enemies, and an unincorporated association, to a geek with an unhealthy interest in LLC formation.” I will not call them a conspiracy in as many words. But they did a series of actions which are surprising to almost everyone who comes to know the totality of those series of actions. And given that it is surprising and has been previously unreported, I felt like: it is really weird that I am reporting a story that is national political news. 

I suppose I have these documents in front of me, and it is close enough to my beat on payment processing. Might as well go for it. 

And so I went for it.

Parallel construction with LLMs

Clara: Okay. There are many different threads that I want to pull on here. Maybe let's start with the use of AI in parallel construction of facts, and then we can talk about some of the broader issues around reporting this kind of story — why certain outlets might or might not want to.

Patrick: Let's talk about parallel construction first.

Clara: Yes, let's talk about parallel construction first.

Patrick: So, parallel construction is a term of art in law enforcement and spying, where you might have sources or methods that you don't want to make broadly known, by which you develop a belief about the true state of the world. You want to use the true state of the world, but you don't want to tell people, "I got this from my guy in the Kremlin." So you use other tools that are available to you — where conditional on knowing the true state of the world, it is easy to use other tools to find the true state of the world. The law enforcement case is: I don't want to reveal the fact that I have a technology many Americans would be displeased to know that we routinely use. [Patrick notes: Cell phone tower emulators, commonly known by the tradename Stingray, have been one of several technologies at issue here. The FBI previously extracted promises from local police departments they had loaned Stingray units to that the police departments were to make no mention of them publicly, including in court filings. This sits uneasily with the American prohibition against secret evidence.]

Clara: And as we've said, this also comes up in journalism. If you have information from a source that you cannot use directly in your reporting, you're going to need to verify that some other way.

Patrick: Right. And so in the law enforcement context, it's: there is this car, a blue Toyota, which will be driving down at approximately 2:00 PM tomorrow. And if they have a headlight knocked out, it would be an excellent time to do a roadside inspection of the contents of that car. Oh, shoot — a random roadside inspection found heroin. So that's the classic use of parallel construction in law enforcement.

“ParaLLM construction” is when you know something about the world and you don't want to tell the world why you know. You write a prompt that goes something like this: I'm going to assert that there were meetings between the following tech companies — name ten — and the following nonprofit organizations — name ten. I would like you to go out into the world, into all the places where people might be undisciplined about how they describe what they said in those meetings. I want you to find podcasts. I want you to find social media posts. I want you to find fundraising letters. Gather as many of those as possible, and then we're gonna work through some of them.

[Patrick notes: Hat tip to Misha on Twitter for this wordplay, by the way.]

Clara: And in this case, were these meetings that you were maybe part of but could not speak about — like, were not allowed to speak about the contents directly?

Patrick: I do not claim in the reporting to have been present at any meeting. I do claim in the reporting that there are people that were present in the meetings that I have spoken to, and that in some cases there are electronic records that are probative with respect to the contents of the meetings that I had access to. But none of the sources were willing to go on the record about it.

They did not want their institutional affiliation named, et cetera, et cetera, et cetera. And so even if I have strong suspicions as to particular language that was used in those meetings, I cannot broadly quote that language.

Unless you say, "Hey — without loss of generality, Claude." I used multiple LLM tools for this, but since Claude is the one I usually address by its name: Claude, I wonder, in this corpus of podcasts and similar that you found, someone uses a very flamboyant language of moral blame. Can you find evidence of that? And one of the things that the LLM came back with was: well, there's a spectrum of communication styles among participants of this meeting, but particularly Color of Change — a civil rights organization that does online organizing — really loves the language of moral blame. For example, they named one of their campaigns No Blood Money.

And I looked at my notes, and wow, "blood money" was exactly the term that I wanted to use. Great. So, as a reporter, I have a claim from someone that organizations in meetings might use terms like "blood money." I don't know if I can trust that claim. If, on the other hand, they're putting out press releases which are titled "Who’s taking blood money from hate groups?" then I think I can trust that claim quite a lot — and source it to a combination of "someone told me they said it," and, by the way, they put out press releases that say the same thing. And let the reader draw the obvious inference that yes, this was actually said in the meetings.

[Patrick notes: As I discuss with Clara in a few minutes, the story is not primarily that the coalition used unkind words to individual employees in advocating for them to be given power over who could transact. It is that they successfully achieved power over who could transact, including in cases which are very surprising. The tactics were sometimes more similar to aggressive personal persuasion, sometimes more similar to externally-initiated labor organizing, and sometimes more similar to coercive conditional escalation. All of these are discoverable from private sources but reportable because the coalition describes themselves publicly as having done them, sometimes in measured language and frequently with very little dissimulation as to their actions and aims.]

Mining podcasts for unguarded admissions

Clara: And this is the kind of thing that, of course, journalists do all the time when they're reporting out stories. But it's very hard to process the amount of public information that might exist for an investigation of this scale if you are one individual.

Patrick: Right. And you wouldn't even necessarily think to search for the sort of things that the LLMs found. So — I told the LLMs to go out and find this corpus, and now: sort by interesting, please. And the LLMs' pick for "if you listen to only one thing, please listen to this podcast" — Rashad Robinson, who was the head at the time of Color of Change, did a podcast in 2021 with Hillary Clinton, who is a podcaster. This is my understated joke; Hillary Clinton is of course known for other work. I was not even aware she had a podcast.

The reason I told the LLM to go out and look for podcasts was: I run a podcast. I've been on some. It's well known that people are less guarded when they're talking in sort of a spontaneous fashion than they are if there is a three-minute press hit where you're gonna be on CNBC — there are big lights that are telling you “You are in danger right now.” [Patrick notes: This illustrative detail is something which actually exists in the built environment. Many studios have a red light or illuminated sign reading On Air, primarily to avoid “hot mic” issues from media company staff rather than from sources.]

You should stick to exactly the script that your comms team prepared for you. But in the course of an hour-long podcast discussion with someone you respect tremendously, who is ideologically aligned with you, et cetera, et cetera, you might say things that are extremely candid about your own behavior.

And so Robinson said to Clinton many things that were extremely candid about his behavior, which do not, strictly speaking, match what the coalition said about their behavior when they, for example, gave prepared testimony to Congress. And you can draw through that and many other data points and say: oh, the actual operation of the coalition was rather different than what they alleged in their more measured public statements — and simultaneously different than what individuals in the targeted companies, who only saw a small portion of the full operation, understood to be happening.

So, as one of many examples of this: when the coalition is disciplined, it says that the coalition did not make decisions with respect to people's accounts. Tech companies made decisions with respect to people's accounts. The coalition pressured — cajoled, even — the tech companies to adopt policies, but then it was the tech company making a decision on the tech company's own policy. And peering into the mind of the person that would make that distinction: the coalition does not want ownership of those decisions, for a variety of reasons, and they can tell you what they are. I have my own suspicion.

The coalition does not want ownership. However, the coalition did not, in fact, lack authority to make account-by-account decisions. Per Rashad's interview with Hillary Clinton: no, we sent the payment processing companies lists of the accounts we wanted deactivated, and we conditionally escalated when they did not give us what we wanted. And after the escalation, within twenty-four hours, they sent back a list of the accounts that they had deactivated.

So once I found that — wow, this gentleman is quite unguarded in how he describes his actions back in the day — I was able to instruct the LLM to find other meetings that he had had with sympathetic media outlets, including Fast Company. And in Fast Company, he describes this conditional escalation tactic: [Patrick notes: flagging paraphrase] "If you don't give us what we want, we will hit you with this. Here is the marketing brief. Here are the images of your executives getting superimposed next to people wearing KKK hats. And here's the hashtag that drops tomorrow unless you give us this and this and this. And we used this ninety-five percent of the time." This is a paraphrase of what he said to Fast Company. I'm like, "Wow."

There might be some people in the world that feel the tech companies were coerced. I can't get them to go on the record saying, "I specifically was not making the decisions, because the decisions were dictated to me by someone outside the building." But someone outside the building is [Patrick notes: dramatization mine] screaming, "I dictated the decisions. I was threatening people if they did not go through with them, and I made good on those threats, and then successfully got the thing that I asked for."

Clara: I wanna flag a couple things here. One is: I've been using the example case of where you need to protect a source. Often you might wanna do something like this because you can't trust a source.

Patrick: Yes—

Clara: Because sources have agendas.

Patrick: Right. And both things can be true at the same time. A person can come to you — strictly hypothetically — and say, "I'm unhappy about what happened at a meeting I was in several years ago, but I still work in this industry. I still value my job. I don't want to get fired over exposing the fact that I'm unhappy, but I am unhappy. And I'm willing to dish as to the contents of this meeting, so long as it goes no further than you." And you might think, reasonably, in listening to this person: okay, on one hand, they have some fear for their professional reputation and their employability and similar. And on the other hand, they are announcing to you that they have an agenda, and they are speaking to you to further this agenda of making their unhappiness better known without getting their fingerprints on it.

Clara: So yeah, it is broadly good practice to be able to look at multiple perspectives on what a source might give you. And building on that: I've noticed the pattern that you talk about in this investigation — which is also, I think, broadly true of good use of LLMs in this context — is you start out with a pretty broad-spectrum search. There are certain things that you strongly suspect exist in the world, and you want to do a very broad sweep. And then you identify certain patterns: this particular piece of language, this particular tactic, this particular individual who might be especially willing to go on the record in different media outlets talking about having used this tactic. And you can then use that to refine, and pick up things that the first broad-spectrum search might not have picked up on. And iterate and iterate and iterate as you figure out what the shape of the story is going to be.

Patrick: Yep. This is absolutely true, and it also works with sort of the private evidence that you're getting from the human sources as well. So, there were certain tactics where I might have had the suspicion that — if there are multiple sources and they report consistent experiences, then you might think it is not simply that there is accidentally consistent experience, but that there is a strategy which is causing multiple people in different organizations to have a consistent experience. But it sure would be good to have some confirmation that that strategy actually existed. Has anyone described that strategy on a podcast? Oh, they have? Then that strategy definitely existed.

Clara: And I should say, this is also how you build a story without AI. But LLMs can create a much broader surface area.

Patrick: Yeah. I reviewed hundreds of publicly published documents in the course of an approximately two-week reporting cycle. And even if you had given me the entire packet on day one, without any search cost at all, I don't think I could have possibly, physically read through all of the documents in that timeline but for the use of LLMs. And in fact, I was not given the packet of all the documents. I had some documents that I couldn't use, and just had to find everything else on the internet using commercial tools.

Clara: And, by the way, another feature of what I would consider responsible LLM use for journalism is you're always using it to point you to a primary source. Instead of accepting the AI gloss on what the source says.

Patrick: Yeah. You do have to be quite worried about hallucinations, or sometimes worried about LLMs having a different judgment on newsworthiness than you do.

Clara: Yes. They can fixate on weird things.

Patrick: So, one of the things that happened is Color of Change is invited to do a series of meetings at Facebook. And some of those meetings are anomalously public. Most of these meetings happen behind closed doors, and there is no public record of who was in the meeting, what they said, et cetera, et cetera. In one of the meetings that happened at Facebook, Color of Change published a prepared transcript of their CEO's remarks to Facebook leadership on their own website. Which, from a provenance perspective: solid gold. Prepared written remarks.

And I asked the LLM, "Okay, he delivered substantial prepared written remarks. What are the important parts from those prepared written remarks?" And the LLM said, "Four things: this, this, this, and this is important." And there was enough signal there that I thought, "Okay, this is probably a document which is worth 30 minutes of my time to read in some detail," and so I did.

And the LLM had not flagged one thing which is obviously important to a reporter, was not obviously important to the person who said it, and was not obviously important to the LLM. The specific genre of remark Robinson made is commonly referred to as a shout-out. So, you're starting a meeting with a lot of people, and there are some thanks to people who made the meeting happen. And so he gives shout-outs to a number of people who were instrumental in making the meeting happen, and he gives a shout-out to the SPLC, who he says advocated for the meeting and — because the SPLC is so gracious — paid for the travel expenses of some coalition members to attend the meeting.

[Patrick notes: Quoting the prepared remarks: “And thank you to those who helped us get here: FreePress, Center for American Progress, Media Justice, the United Church of Christ, the Lawyers Committee and Muslim Advocates. And the Southern Poverty Law Center, who not only pushed for and supported this event with their voice, but did so financially, as well, paying for so many of us to travel here from out of town.”]

And my reporter brain went, "Jackpot." Because money movement is an allegation about the world that transforms the character of things. So if this is just a stitching circle, where we are all aligned organizations but we're doing our own thing in our corner, then none of us are responsible for things that the others of us do. But if this is truly concerted action, where we are paying for people's tickets to attend the meetings that we've arranged, then this is concerted action.

Deepen your strengths, don't just patch weaknesses

Clara: This relates to a broader point that I wanna make. I feel like when a lot of people talk about or think about use of AI, they think about expanding the scope of what they can do, or shoring up their weaknesses. And I have found that the uses that I find by far the most powerful are deepening my strengths.

I am not a quantitative person, and so I am very scared of using AI to do quantitative analysis. Obviously it is good at this stuff. Obviously it is better than me, 'cause I am not... If I'm reading an econ paper and I'm skeptical of the conclusion — I mean, I could probably use something like "Refine.inc", which would give me good feedback, but it would be at a level at which the feedback would not even be that useful to me. What's much more likely to happen in a situation like that, where I am not familiar enough with the terrain, is it's smart enough to be dangerous to me. Where it is not even necessarily that the model will hallucinate, but in an area with novel research, and with many different plausible ways to do an analysis or construct information, because I lack sophistication with those analytic tools, the model can trick me.

But in areas where I'm already quite familiar with the terrain and I understand how they work, I can get a ton of mileage from using LLMs to identify sources, to review sources. And then if it's like — oh yeah, I'm in the middle of looking through these German academic journals from 1752, that's something I know about. And if the model has some gloss that I think is wrong, or it doesn't really understand what I'm looking for, then I'm able to step in and course-correct. And I think I would do a much worse job reporting this story than Patrick, because I am much less familiar with the world of corporate liability.

Patrick: I broadly agree that the right way to be thinking about it is the complementarity of the human user and the model, versus the model replacing things or backfilling one's incapacity to do things. I will say that there are some parts where I would not trust a model to develop, for reporting, a thing where I could not check the correctness of it.

But there are some parts where a background in, say, power politics within non-governmental organizations would be very useful for reporting this story. I have very little of that background, and it's a complex story which involves a rotating cast of characters over a seven-year time period. And one thing I asked the LLM to do — just to help me orient my thinking around this — was: can you write me a timeline of this seven years, and try to track who the center of gravity is of this coalition as members rotate in and out, and do this by looking at co-citations in their demand letters, which they publicly publish, in tier-one news reporting, et cetera, et cetera.

And then it drew out this timeline and said, "Okay, in these years it was mostly these folks, and then it transitioned to this," et cetera, et cetera. And I would not have been able to copy-paste that into Bits about Money without extensive review of it, and probably finding an actual expert to say, "Does this track your understanding of the terrain in Washington, DC?" But in terms of informing this guy on, okay, what authors should I be looking to read from in 2018 to understand what was going on — that timeline was incredibly useful, even if no other human ever reads it.

But yeah — definitely leaning into our strengths on things. They read a lot faster than me, but I have a very good model of where the landmines are. And so it was often: read faster than me, find landmines in dense documents, and I will pull out more landmines than you find in that document.

Clara: Yeah. This could change, but at least for now, having your own deep expertise makes you a much more powerful user of LLMs.

The smoking gun the LLM missed

Patrick: Yeah, I absolutely agree with this. And you can't get overly reliant on the LLMs. So, they did help me identify a smoking gun. And it was the identification of the smoking gun document that caused me to broaden the Bits about Money scope from "I'm just writing about the bank fraud" to "no, I'm writing about the actions of Change the Terms."

There was a better smoking gun document, and that smoking gun document was not particularly hidden. It was written on the coalition's own letterhead and posted to their own Twitter account, and I simply didn't see it, because I was working through a pile of 600 documents or so that the LLM had produced. And I thought, erroneously: well, if the LLM has looked through this pile, and I have looked through this pile, and we've identified the best document in the pile, that is probably the best document that exists.

And the LLM, for whatever quirky reason, simply had not found the signed demand, on letterhead, to deplatform a declared candidate for the presidency. Which is an odd document to produce if you have that legal exposure. But I will let their lawyers tell us why that is allowed.

Hardball politics versus crimes

Clara: Well, maybe this is a good time to step back and talk about some of the politics of this story.

Patrick: Very unfortunately, but yes, there is a political dimension to this story.

Clara: So, a question that I had when I was reading this reporting is: where do you draw the line between the legitimate practice of a social change campaign and — crimes? Because, importantly, the terms coercion and pressure have come up a lot in this discussion. And I think that there are absolutely instances where things that can be called coercion and pressure are appropriate. Like, right now, a lot of people I know are involved in advocacy to try to pressure legislators to... Are you familiar with the Save Our Bacon Act situation?

Patrick: Not very familiar with Save Our Bacon. It's for factory farming, right?

Clara: People are trying to pressure legislators to maintain the possibility of states passing laws that ban meat sold in that state having been produced through inhumane practices. I think this campaign is good. I think that it is legitimate to say, publicly, openly, "We are against this. We are against you if you support this. We will support you if you oppose it." Because this is how the process of politics works. This is just a randomly chosen example, because it happens to be salient to me right at this moment.

Patrick: Yep. So, I am as committed to democracy as anyone else, and part of democracy is having a robust debate. And sometimes in that robust debate you have a private version of the debate and a public version of the debate. And perhaps sometimes the private version of the debate is: you could be exposed to negative public consequences if you don't align more closely with us. All this sounds relatively uncontroversial to me. You can detect — I have leanings, from reading Bits about Money, but Bits about Money is not a journal of political opinion. It's mostly a journal of the actual practice of banking and KYC laws and payment processing and similar. Bits about Money reports; the rest of the world can decide.

Broadly, I think a thing that is exculpatory or inculpatory is: can you say you did the thing you actually did? And if you can't say the thing that you did, then very possibly that thing that you did is not a wonderful thing. So, the coalition says publicly and floridly, "We were pressuring these companies. We were cajoling them," et cetera, et cetera, in places like congressional testimony. At the start of the reporting process, I didn't even know: is it appropriate to say that the campaign was threatening them? And then they used words like "attack" and "threaten" in their public statements. And I said, "Okay, sure." They admit to threatening.

[Patrick notes: No really, they did say “attack” and the target of the action was not white nationalists but the industry participants. Quoting an SPLC executive’s prepared testimony to Congress:

The first targets of our attack against hate group funding online were PayPal, Apple’s iTunes and Amazon.

]

The specific nature of the threats is this conditional escalation thing. And the conditional escalation thing is like — okay, this is hardball politics. Understood. And maybe some other outlet in the world wants to report on the specifics of that hardball politics. If it had only been the hardball politics, I don't know that I would have reported this part of the piece, 'cause I don't know that there is a punchline to it. Then there's another punchline.

Bits about Money is not the Internal Revenue Service. We did not report evidence of any crimes, as far as I'm aware. However, the Internal Revenue Service says very clearly that nonpartisan, non-governmental organizations with 501(c)(3) status are not allowed to support or oppose any candidate for public office. The coalition very much did that. I'm saying that as flat fact.

A coalition member lied to me about having done that. [Patrick notes: In the audio, I attribute this to the coalition generally. I did specifically ask the coalition for comment, including publicly on Twitter as they do not maintain a press contact. They did not reply in the voice of the coalition. One coalition member, in a written statement, claimed the coalition had disbanded in 2023. No coalition web or social media presence reports a decision to disband or a final victory over hate.]

I sent them a request for comment and said, in substance, "I give you the opportunity to deny that your organization and the coalition attempted to interdict the fundraising of an American candidate. Please note that I have you doing this multiple times in writing, including the following two documents." And this is a reporter playing Marquess of Queensberry: look, the story that I publish is not going to be a very happy one for you, but the norms in the profession are, if I publish an unhappy story, you have the opportunity to comment on it.

And a comment there could've been, "Yeah, we did, but he's a really bad guy." Or, "Yeah, we did, but we're allowed to." Or, "Yeah, we did, but that was a mistake and we won't do it again." You probably shouldn't start your comment with "We deny that we did that" when the reporter's already said, "I have documents signed by your senior executives, and published, that say that you did this."

The Center for American Progress is a well-funded, largely blue-leaning NGO described in the Washington Post as the place where Democrats go when they are not in the administration — the government in waiting, as it were. [Patrick notes: Representative language in a WaPo news story: “CAP is widely viewed as a Democratic administration-in-waiting.”]

The Center for American Progress, through a spokesperson, denied that they had done the thing that they did.

And I will characterize forthrightly the Center for American Progress as having lied.

[Patrick notes: CAP’s written statement in part reads “Neither [Center for American Progress] nor [Change the Terms] sought to ‘interdict political fundraising’ or suppress nonviolent political expression.”

Bits about Money and Complex Systems reported that CAP and Change the Terms did both of these things repeatedly, as an intentional strategic decision, over a span of years, covering multiple election cycles. 

We sourced that reporting, in part, with published letters to Facebook co-signed by CAP which demanded the restriction of the accounts of (among other entities) a Trump PAC, with the following verbatim characterization of the PAC’s activities:

Allowing Team Trump to continue running political ads on Facebook is a significant loophole in Trump’s two-year suspension and provides a pathway for the former president to evade the ban. For example, Team Trump has posted ads asking users to “join President Trump,” “show President Trump what you think,” and “stand with President Trump.” Further, Team Trump is soliciting donations and inviting supporters to Trump rallies.

(Footnotes from the letter omitted.)

Years after that demand letter, Change the Terms made a broader demand, on its own letterhead, after Trump was a declared candidate for the 2024 election. This second demand letter was co-signed by a senior CAP executive, who co-chaired Change the Terms.

I do not say “lied” lightly.

]

And it's interesting that they feel they have to lie about this.

As to why they have to lie about this, I think there is a very obvious explanation, but I don't make that argument in the piece.

I just observe the fact that, well, here is their statement; this statement contradicts the documentary record that was produced by them. And in a somewhat less formal atmosphere like this, I'll describe them forthrightly as having lied flagrantly. And as to the righteousness of that lie, or the consequences of it, or the consequences of the actions that caused them to feel like they needed to lie about their actions — somebody else's ballgame.

Why a payments newsletter reported a political story

Patrick: This, by the way — publications have beats, and Bits about Money is about payment processing and similar. I don't want to report political news. I think we are not well-suited to do it. We are not well-served by doing it. Our readers are not well-served by getting it from us where they could get it from many other institutions. And to the extent that there is a national political news story, I would rather it be reported by the obvious journals of news and opinion in Washington and New York that are well-resourced for doing this, that can go to the players and say, "Did you know this happened? What are your comments on it?" — and that have the appropriate resourcing and epistemic standards to track this thing down.

Clara: And I think one concern we've been kind of dancing around here is: if the responsible sources don't pick it up, the irresponsible ones will.

Patrick: Yeah. I have gotten new readers in the worst corners of the Internet as a result of doing this, and am quite frustrated by that fact. The core audience for Bits about Money is, among others, people who work in central banks. I do not seek to get readers in the worst portions of the Internet.

However, I will say, it would calm the waters a little bit... People make unhinged lizard-men conspiracy theories about various parts of the American government all the time. What is it called? The Paranoid Style in American Politics.

Clara: Yeah — the famous essay.

Patrick: Yeah. And I think responsible adults should want as little lizard-men conspiracy theorizing as possible. And one way to counteract that is: when there are norms violations that are kind of severe, the responsible adults should do the correct amount of intervention against norms violations. If there was a well-funded, multi-year, multi-organization attempt by putatively nonpartisan organizations to target a particular American candidate to the presidency, and that is not a news story — I think that is an indictment of everyone capable of publishing news stories.

I assert that this campaign happened, and I welcome anyone to say that it did not happen. I don't think anyone can credibly claim this campaign did not happen.

Clara: Well, I wanna talk about how you handled this a little bit. Because you're trying to make a kind of subtle point here, which is that there is a level of, let's say, hardball politics which is legitimate in a democracy. A 501c4 can run an ad. A 501c4 can run an ad that is negative, and it can run an ad that says, "This candidate supports this policy, which is bad." And you're trying to make a point that this particular set of actors did more than that. They crossed a line into—

Patrick: A 501c3 cannot even make a statement—

Clara: A 501c3, yeah.

Patrick: Yeah — 501c3s and 501c4s are both in the mix here. A 501c3s cannot even, per the rules, make a statement supporting a particular candidate or opposing them, and they went way over that line. I don't think any intellectually serious person can dispute that.

Clara: 501c4 rules are a little looser. I run a 501c3, so I just stay very far away from all of this stuff.

Patrick: I have also previously been the CEO of a 501c3, and one of the first things I wrote to everyone was, "We have to be meticulously nonpartisan, both for achieving the actual aims of this organization, but also because that is a requirement to get this tax-exempt status."

Clara: But also, 501c3 can comment on policy.

Patrick: They can comment on policy, that is true. They can't comment on the people. The 501c3s at issue commented quite a bit on the people.

The thing that makes it more intellectually interesting to me, and part of the BAM beat, than simply "these organizations were publicly and privately saying things about their positions and about how that interacted with particular political candidates," was that they were doing things. The rails that move money are an artifact that exists in the world.

Clara: And this is, I think, specifically the point you're trying to make — which is that, frankly, from reading this: pressure and coercion aren't really the most relevant characterizations of which actions were newsworthy and non-newsworthy. Because applying pressure is just how any social campaign works.

Patrick: Right.

Clara: But there is a difference between applying social pressure and debanking.

Patrick: Yeah. And debanking — an entirely different... well, not entirely different. I think many people would mix the two, but I've written previously about debanking. There was targeted action aimed at electoral infrastructure. Not electoral infrastructure in the sense of vote counting, but electoral infrastructure being accounts controlled by campaigns, by PACs — explicitly in the case of PACs. And defined in the document multiple times, in case you don't know what a PAC is: it's when this guy is attempting to do fundraising and get people to rallies, et cetera, et cetera.

Reporting stories that will be weaponized

Clara: And stepping back: what do you do with a story where, frankly, you know that a lot of people who are reading this are not going to make that distinction? And a lot of people who are reading this are going to use what you have written... On the one hand, what you have written, I think, is pretty well-sourced, and I personally believe that it is true — and true and newsworthy. On the other hand, it is absolutely going to be metabolized as ammunition in a bunch of causes that are much less based in fact. How do you think about that as a writer? How to approach it and mitigate that?

Patrick: So, one thing I do — and I think a number of readers of the piece have cottoned onto this — is that one could go for a very salacious headline, for example, and foreground the political controversy. And I, explicitly as a matter of authorial intent, attempted to decrease the level of salaciousness and decrease the salience of the political controversy while getting all the facts on the record.

Clara: I think the phrase SPLC does not appear in the first thousand words of this essay. [Patrick notes: After approximately 2,600 words the essay says “Which brings us to the Southern Poverty Law Center.”]

Patrick: Yeah. Tremendous accident, how that happened. No, of course not. 

Cards on the table about the SPLC: they are broadly respected, including in the tech and finance industries. I have mixed opinions on them, candidly — but not at the level of "I think they should be burned out of the democratic discourse" on any given day. And Bits about Money has reported substantial and newsworthy bad acts conducted by the SPLC. And we've reported this in the backdrop of a campaign to indict them for actual crimes.

One of those crimes — as a matter of analysis rather than a matter of reporting, it's about when you say: yeah, this is the definition of bank fraud. The other one, it is what it is, and there is definitely the possibility that this is a politicized prosecution.

And that was part of what got me pulling on the thread originally. I'm like: politicized, but for why? There's no shortage of enemies [of the Trump administration]. Why this one in particular? And is it retribution for something? Question mark, question mark, question mark. And then I started pulling at threads, and I don't know that there is a line between the prosecution and this activity that was conducted by the SPLC and allies. I don't honestly even know if the administration understands that it was the target of this coalition, because there are many people that don't understand that this coalition was acting, because it was a long-running, multi-party—

Clara: And also, again, I think it is actually just quite hard to understand — unless you are very immersed in the world of corporate liability practices — which things... And I think this is why this happened, ultimately. Including to coalition members, it is difficult to understand where that line is between, again, legitimate practice of hardball politics in a democracy, and crimes. And it seems very safe to me to say that this is a politicized prosecution.

Patrick: I think that feels very safe to say: it is tough to imagine the SPLC getting indicted for crimes in most administrations. And, you know, the conduct about running the intelligence agency: not very new. And the bank fraud that they committed to run the intelligence agency happened over a more-than-decade-long period, and was discovered rather lately. I will say they were under investigation, it appears, during the Biden administration for the bank fraud thing. And so no one had made the decision to not prosecute there. [Patrick notes: Misspoke; there was a determination to not proceed.] But again, the politics of it — not my strongest suit.

The tech right and internal company politics

Patrick: I will say the conservative opinion media has said that Big Tech is woke and hates Republicans for a number of years. And the San Francisco Bay Area, and employee bases in the San Francisco Bay Area, have the political distribution that you would expect as a well-read American.

Clara: This is something that — okay, as someone who does run a media outlet that is based here: I think a lot of outlets that have less of a Bay Area presence really struggle to understand the tech right as a phenomenon. 

Which I think is also, of course, massively overstated as a phenomenon in some ways. [Patrick notes: There are few more reliable ways to have a narrative reach print than having one well-organized comms shop and a large segment of the media want to tell the same story. This is true when that narrative is “tech is taking blood money”, and is also true when the narrative is “there has been a sea change in the politics of the tech industry.”]

The tech right — if you think about it as a monolith, it doesn't make sense. The tech right only makes sense when you think about it psychologically, as a relatively small group of very powerful people feeling an intense sense of social threat from their own employees. If tech employees were not liberal, none of this would have happened.

Patrick: Yeah. And I need to underline this 15 times: not universally in the employee base, by the way. There's a large contingent of the employee base that is indeed quite liberal, quite activist, and enjoyed a surge of internal power between approximately 2016 and approximately 2024, for a variety of reasons — some related to US politics and some downstream of broad social changes. For example, the pandemic happened; there was the racial reckoning in the wake of the events of 2020, et cetera, et cetera. There exist some senior people of, you know, other-than-progressive persuasion who felt intense threat in their own companies. And that caused both some of the cooperation with the coalition, for the coalition's aims, and also some of the backlash that has manifested in the tech right, for example.

The coalition was not unaware that it had internal allies, and indeed says in as many words, in its unguarded utterances, that there are internal allies: we want to make them as powerful as possible. And so we are tag-teaming with our internal allies here — where the company might have preferences, but employees in the company have preferences which are the opposite of the company's preferences. And we are going to threaten the company with a Twitter campaign which will make those people internally win the argument.

Clara: Which I think is not intrinsically illegitimate.

Patrick: As someone who's worked in a company, it's a really annoying fact pattern to have happen. And depending on the nature of the cooperation between the internal allies and the external people, I think some people might be in violation of formal or informal agreements that they have. But that's largely not a problem. Where things start to become a problem for me personally is where we have given non-elected, non-accountable actors the ability to turn off particular accounts, where they are not supposed to even be able to speak the name of the holder of that account in a way that is negative. Feels like not the deal that anyone signed up for.

Debanking discourse and drawing distinctions

Clara: Yeah. No, I agree that I think that's the important distinction. Also something, by the way — there's many reasons to want to unilaterally disarm. The place this comes up all the time is, it is pretty frequent for, just one example, anti-abortion activists to want to deactivate certain accounts relating to selling or purchasing abortion pills. This comes up all the time. It also comes up in relation to sex work. There's many causes that liberals support where you would go, "Oh yeah, this weapon should be off the table."

Patrick: Yeah. And, for better or worse, there will always be a supportable-and-unsupportable-uses list at every company in the financial industry. And I think the exact contents of the supportable and unsupportable list is, well, partly determined by law, and then partly determined based on the aesthetic, moral, ethical preferences of the company. 

I do think that companies should be able to justify to their internal and external stakeholders, "Here is what is on the list and here is what is not on the list" — a reasonable topic for public debate. And we live in a democracy which, last time I checked, has two political parties. And if one of those political parties is hypothetically on the no-no list, that would be an explosive fact — should not be a true fact. And if someone attempts to arrange that, I think that is—

Clara: Okay, I do wanna be specific here. Do you think that there was ever a concentrated attempt to say you can't offer payment processing to Republicans generally?

Patrick: With regards to Republicans generally: Bits about Money did not report that. Bits about Money did report that there was a concentrated effort to curtail the fundraising of Trump PACs and the Trump campaign specifically.

Clara: All right. Just wanna be specific about what's being alleged.

Patrick: Yeah — specific with respect to the allegations that we reported. And, bluntly, facts that we reported. I think that some actors generalize the indictment, like in the debanking discourse. Adjacent to this, but not quite the same: there was the claim made by various intemperate people that the banks just hate Republicans. And if you think about that claim for like 30 seconds, obviously it's not true, right?

Clara: And this is something you've reported on also, because there is a lot of genuinely conspiratorial debanking discourse.

Patrick: Yeah. I think I have a line about conspiracy in the debanking piece, where it is sometimes the case that people actually are out to get you. And there are portions of the debanking discourse — with regards to, say, the ability of crypto companies to access banking services — where parts of those allegations are true, like, straightforwardly, and we reported on that.

But the general notion that banks hate Republicans is just a fantasy.

You wouldn't meet someone at a party in real life and be discussing, "Oh, I'm making investments in my Fidelity account. I'm looking forward to the SpaceX IPO. Are you doing investments?" And they say, "Oh, I'm sorry, I'm a Republican." And you say, "Oh, shoot, I stepped on a landmine. Everyone knows that Republicans can't trade stocks." Which is the conclusion of what would obtain in a world where this crazy conspiracy theory is actually true.

There is no broad debanking conspiracy against Republicans. There exist ways in which particularly politically exposed people of a conservative persuasion got their accounts turned off, due to political actors at individual institutions. That is true. That was broadly not centrally directed by the Biden administration or the government broadly. 

I wish people would be able to separate those two claims.

And it is a true fact that our neighbors to the north in Canada actually did have a no-go list that was politically motivated, and sent that out under what they considered to be the full color of law. And so there is definitely a risk of politically motivated debanking. But we haven't experienced it at a directed-by-the-central-government level in the United States. And I wanna be very clear about that, because it's important that we haven't done that. It's important that the Constitution holds in this country, et cetera, et cetera.

Clara: Yeah, this is the kind of thing that's so annoying to try and talk about responsibly. Because, again, one, there's a kind of important but also non-obvious distinction between legitimate activities and illegitimate activities. And which activities are legitimate, of course, changes depending on which actors you're talking about.

Patrick: Yes.

Clara: And there is a real tendency to see certain actors as equivalent when they are not.

Patrick: Right. And partly this is due to strategic inflation; partly it's due to just a lack of understanding. Again, tech companies and employees at tech companies felt like they were being pressured by the government and by media and by civil society organizations. This was true — but our moral intuitions with respect to the permissible actions of those three are very different.

Clara: And also, legally, what is permissible.

Patrick: And legally what is permissible, of course. And so, as a statement of fact: the United States federal government, during the Biden administration, prevailed on certain media orgs and tech companies to suppress the reporting and spread of true facts. That is outrageous.

Clara: We talked about this a little bit on our last episode.

Patrick: We talked about this a little bit. And we definitely shouldn't have that, with respect to the federal government. If, on the other hand, there was a civil society organization that said, "There are some true facts that I don't think the world should know" — I think that's a reasonable topic for debate. I mean, they don't share my epistemology, but they're welcome to not share my epistemology.

And if you say, as the FBI, "Don't allow discussion of, for example, the Hunter Biden laptop, because it's a counterintelligence thing," and it is not in fact a counterintelligence thing, you have done something very bad. [Patrick notes: Subsequent Congressional investigation revealed that the FBI had contemporaneous knowledge that the laptop was, in fact, genuine and not part of a hack-and-leak which they primed the tech companies to expect.]

If you say, as, like, the Democratic Party, "Don't allow the discussion about the Hunter Biden laptop, because obviously this is embarrassing for the Democratic Party," I think that's squarely within the bounds of things that the Democrats should be able to advocate for, as long as they're not using coercive force of law to do it. 

Whereas the FBI was effectively using coercive force of law. [Patrick notes: There is a live political debate in the United States as to whether some actions against the tech companies during the past few years were coercive use of law or jawboning. This Japanese salaryman will neutrally observe that no Japanese salaryman, having been flashed a badge, is so stupid as to assume an officer asking for voluntary cooperation for the sake of national security is actually asking for voluntary cooperation.] 

Hate speech, social sanctions, and owning decisions

Clara: And to go back to the thing that prompted all of this: one difficulty is, genuinely, how do you handle hate speech in a democracy? And I believe pretty strongly we do not have in this country, and should not have in this country, legislation preventing certain kinds of speech because it is hateful. But one of the things that I think makes this possible and sustainable is that we then have social norms against certain kinds of speech that are hateful.

I think the ACLU was right to protect the Nazis marching in Skokie, as was their right. But also, as a Jew, I like living in a society where there are very strong social sanctions for doing a Nazi march, and those sanctions are going to be applied on some level by some kind of social coalition. At some point, that's legitimate. And the question is — there's unavoidable questions of who should exercise those pressures, and what kinds of pressure should they be able to bring to bear, and what should the consequences be, and how should they be determined.

Patrick: Yeah. I think that a number of actors want to sort of have their cake and eat it too here, where they want the social pressure but they don't want to be the person with their finger on the button. And so one of the reasons this got as far as it did is, I think, the tech companies are looking for: one, we obviously don't wanna get hit in the press. Two — I would also prefer that Nazis have the right to march in Skokie, but that people be strongly discouraged from being a Nazi. The tech companies, largely feeling that to be true as well, would love someone to be able to tell them, "Just give me the list of the Nazis, please, and tell me to turn their accounts off, and then I will do that. But I've not made a decision to turn off anyone's accounts."

But someone is making a decision, right? And part of — again, the coalition, extensively, and not always telling the truth in doing so, would refuse the characterization that they were making this decision. And I think the tech companies, in many cases, under-owned decisions that they made. And I don't think it is tenable for us to have an observable fact in the world, where the account is turned off, where no one owns the decision of ordering the account to be turned off.

Clara: So either the tech company or the coalition, or both, should just own it.

Patrick: Somebody has to own that observable fact in the world — particularly when the observable fact in the world is a thing that is not allowed in law.

Clara: Well, it would be... It is legal for, say, Twitter the company to deactivate a politician's account, if they decide to do that. They are not a 501c3 or 501c4 organization.

Patrick: Right. Broadly, they have — well, one would have to ask a lawyer for the specifics on this, but my understanding is, broadly, that they have discretion in who they do business with. But laundering things through that discretion, particularly when that discretion is used in a way that is coerced by either private actors or semi-private actors...

One thing that we reported was that some people, you know... Where does the government stop is a meaningful question. And if you say, "My name's Patrick, and I really don't like that thing you're doing — you should stop, or I will use the full of my might against you in the political process," well, maybe that's legitimate. And if you say, "My name is Patrick. I think the thing you are doing is illegitimate. Stop it. And if you don't, I will use the full of my might against you in the political process. By the way, I have substantial influence in the White House" — is that the same statement anymore? It is not obvious to me that it is the same statement, particularly if the "by the way, I have substantial pull in the White House" is straightforwardly true.

And if you say — by the way, to a more sympathetic audience — "I coordinate policy with the White House all the time; that's my job," then it's like, oh, that is increasingly no longer non-state action. And I don't know what we do about it, but one thing is: put out the fact of this fact pattern having obtained, and allow the political process to hopefully correct for it.

Patrick: I am sure we could talk about LLMs until there is a new LLM release, but eventually someone else will need the studio. Clara, where can people find Asterisk?

Clara: asteriskmag.com.

Patrick: Thanks very much for being on the program today. You can read Bits about Money at bitsaboutmoney.com, and Complex Systems, where we published the second follow-up to this, at... Oh, you're already watching Complex Systems. Well, presumably you can find it. Thanks very much, everybody, and we'll see you next week on Complex Systems.