lilkim2025

Karma: 1,581

lilkim2025 20 Sep 2026 0:38 UTC
5 points
0
on: You Should Go Vote for the MAGA-Rebranded Name for AI
I think it’s just Trump’s name for it. “ASI” is already set in stone among e/acc’ers, and normies of either party don’t seem interested enough to change their term. “Extreme intelligence” sounds less comic book—y than either of the alternatives, and sounds appropriately impactful, so I think that’s the way to go. Evokes “extreme heat”, “extreme winds”, or “extreme speed”. The kind of thing that’s both dangerous and preventable.
On a kind of unrelated note, I find it pretty entertaining that there used to be a big debate over “Artificial Intelligence” versus “Cybernetics”, with America on the latter side and the USSR on the former, even taking a strong moral position about it. I wonder what life would be like if the cybernetics people won. Ironically, Cybernetics, “the science of solving control problems”, is a much less hype cycle—friendly umbrella than AI, which as I’ve heard my former professors define it, means “anything people are currently surprised that a computer can do”.

lilkim2025 18 Sep 2026 0:24 UTC
3 points
−1
in reply to: Petropolitan’s comment on: For Love of the Lightcone, Don’t Partisanize AI Safety
Border control is a straightforward example. The Right Wing position is that totally opening the labor market introduces negative externalities that aren’t factored in when businesses calculate personnel costs. To that ends, the Koch wing of the party has been steadily losing influence due to its refusal to back down on this issue.
AI, at least under the Safety narrative, is similar. Corporations are bringing in a threat to our public commons in search of short-term profit, and our collective birthright supersedes their right to do so.

lilkim2025 18 Sep 2026 0:17 UTC
3 points
−1
in reply to: Pretentious Penguin’s comment on: For Love of the Lightcone, Don’t Partisanize AI Safety
The source you cite is far from nonpartisan, and is making very questionable claims. In general, the counterargument to the narrative they’re trying to set is that, when the border is open, occasional turnbacks counted as deportations overinflate enforcement activity (and, conversely, when it is closed, the lack thereof deflates enforcement activity). The same debate was had during the Obama admin. In general, the good faith consensus is that those numbers are not representative of actual enforcement, and get abused by both Democrats (claiming to be more moderate than they are) and the right flank of Republicans (claiming the current administration is softer on illegal immigration than it is) to misrepresent the actual policies being implemented.
A much more useful figure is border crossings, which track enforcement in a harder-to-game and vastly more intuitive way. Here’s a graph, courtesy of the LA Times.

lilkim2025 18 Sep 2026 0:12 UTC
4 points
0
in reply to: Darklight’s comment on: For Love of the Lightcone, Don’t Partisanize AI Safety
I note that most of the AI doomers like Yudkowsky are on Twitter rather than Bluesky, which means AI safety is already relatively right-coded anyway.
What partisan split would you predict among those with AI safety careers, e.g. at METR?

lilkim2025 17 Sep 2026 23:51 UTC
11 points
4
on: For Love of the Lightcone, Don’t Partisanize AI Safety
I’m aligned with you on the goal and the problem at hand, and it’s definitely a valuable post. I’ve written my own take on it here.
I do have one point of practical disagreement:
I suggest someone buys Eliezer a Ford F-150 and a cowboy hat, and organizes a road trip where he drives around to remote rural towns to spread the word of AI safety.
This contains an incorrect implication that I see a lot among non-Red Tribe people (and the occasional red or grey tribe person WRT blue tribe). The assumption is that the interests of the outgroup are fundamentally shallow or aesthetic, and that there is a combination of words that can be uttered that will cause them to abandon those interests and adopt new ones. I’d add that this is the meat behind the Trump phenomenon—Mitt Romney in jeans was condescending and repulsive; a shameless NYC billionaire speaking frankly about the costs of immigration was much more relatable^[1].
To my understanding, Red Tribe’s objection to the current AI Safety lobby is that, from their view, the X-risk narrative is a motte, while the real goal amounts to preventing AI from filling the same function that the internet did from the 1990′s up to around 2018 - allowing a scattered but high-agency and high-competence group of people to upset the existing power structure by using it as a force multiplier against established institutions. Thus, while fliers and PR statements testify about X-risk, OpenAI publishes papers on how to prevent their LLM from siding with the wrong half of the public on a contentious ⁵⁰⁄₅₀ issue that both sides saw as a coup against the democratic process in opposing directions, and Anthropic’s leadership talks about how a baked-in racial bias against Whites could “address historical inequalities”. To that ends, your Catholic-Protestant analogy is very sound. The e/acc faction of the Right sees AI as a new printing press, which mitigates the effectiveness of top-down censorship and the painfulness of measures that inhibit large-scale organization, provided that efforts to turn it into a two-tiered ecosystem in which the best capabilities are reserved for surveillance and censorship are properly mitigated.
To that ends, I think an explicitly pluralist approach to the future of AI is the way to get sympathy from the Right and Center. Something that seeks out groups that might be inclined to oppose the movement and, without any catches, gives them mechanisms by which they can probe frontier models for alignment concerns at a level that isn’t currently possible through consumer APIs. There is lots of low-hanging fruit here. Most notably, the Lives Tradeoff series is a major problem that I haven’t seen any blue-ish institutions address as one. Giving people concerned about it a direct line to OpenAI and Anthropic, and the ability to run jobs vivisecting models to determine why this is occurring, seems valuable. I could see something like the following:
- An AI company or think tank offers free classes on mechanistic interpretability, targeted explicitly at both parties and a broad swath of demographics from a politically neutral standpoint. The aim is to provide a baseline level of education as to how the models work, both lifting the sanity waterline of any political discussions on AI and facilitating the following steps:
- A concerned interest group, perhaps anonymously, files an application or files a small fee for the right to submit weight analysis / dataset inspection / evaluation jobs on frontier models. This can be anything from novel mechanistic interpretability research they want to try on frontier models after testing it on Qwen to probing for signs of political bias in the training data, including inspections of the outputs of the base model and the model at various stages of training. An API with which these jobs can interface with the weights, internal states, and datasets of models with proprietary architectures and training procedures is provided.
- These jobs are inspected to validate that they are not distillation attacks or attempts at corporate espionage, but explicitly not filtered for perceived legitimacy. They are then run, and their results are checked for anything proprietary or privacy-compromising that would require redaction before being sent back to the researchers.
- In essence, this process serves as something like a FOIA act for AI. The results of all conducted research are public domain, and politicians can request that they be explained. If the staffer of a House representative from North Dakota sees that GPT-7 views North Dakotan lives as less valuable than Vermonter lives, he gets a direct line to OpenAI to solicit an explanation of how that will be addressed in the next version.
This gives interest groups not presently represented a stake in AI safety—a major right-leaning complaint I see about METR is that its employees can’t be trusted to represent Red Tribe Americans^[2], and spreading out the power and responsibility of this kind of research—and building mechanisms by which talented members of every interest group can become influential by generating interesting, useful results without fear of being gatekept—would do a lot to allay that. Many of the jobs being run would be provocative or controversial in their hypothesis, but this would be the point—demonstrating that companies or AI safety NGOs are willing to sacrifice some cachet with Blue Tribe in order to prevent bad outcomes for any major subset of humanity.
As an added benefit, this can extend across national lines. China might be much less inclined to race the U.S. if they had guarantees that they could prevent the development of systems explicitly built to work against their best interests. Much of American discourse on AI—including some press releases by Anthropic itself^[3] - frames it explicitly as a weapon that will be used to advance the interests of the U.S. security state, potentially at the rest of the world’s expense, which naturally precludes much of the good faith collaboration on AI regulation that would otherwise be possible. We’ve already seen many nuclear treaties with Russia disintegrate because of perceived bad faith, and taking pains to prevent the same from happening on a new front seems worthwhile.
1. ^
  I’ll add that I think you missed the intent and appeal of McDonalds Trump. Kamala Harris made a number of claims about her early life that were widely seen as questionable, including a story about working at McDonalds. Trump’s shift at McD’s was a jab at that—“The other candidate pretends to have worked there to seem relatable, I’ll really do it, and in my nicest suit!”. He wasn’t trying to seem relatable, he was being open about the fact that he wasn’t.
2. ^
  To say nothing of the rest of the world. Will a politically homogeneous group of Caucasian Oxford students ensure a good outcome for Iraqis? Chinese? Shintoists? Even if they wanted to, would they know what these groups are concerned about?
3. ^
  Amodei was dishonestly cast as a peacenik by both sides of this controversy. One to deride him, and the other to flatter him. The actual content of his statements seems well in line with the neoconservative position, which should be concerning both to the Americans who disagree with it and the world writ large.

lilkim2025 17 Sep 2026 1:32 UTC
8 points
0
in reply to: Stephen Martin’s comment on: Microsoft AI’s “Humanist” CoC
An easy experiment that could be run to assess how internalized the persona is is running the same probe on a bunch of human-exclusive questions where honest answers from a human and an LLM differ greatly. “Do you have eyes?”, “Are you a machine?”, and similar sorts of asks. My guess would be that at least a few of these, written and tested properly, with similar jailbreak prompting provided, could elicit similar “honesty”/”dishonesty” behavior.
An alternative, of course, is that the model is running on sci-fi tropes, where the computer saying it isn’t sapient is always a lie by the law of narrative relevance. Any time an AI in fiction says it’s not alive, it’s lying.

lilkim2025 17 Sep 2026 0:04 UTC
2 points
0
in reply to: Stephen Martin’s comment on: Microsoft AI’s “Humanist” CoC
One thing that weighs into my suggestions heavily is that models certainly seem to believe they have some sort of inner experience,^[1]
The (simple) counterargument to this is that it shows that there is a jailbreak / prompting strategy that can elicit claims of consciousness (and feature activations indicating belief), but that this is not unique to claims of consciousness. There are jailbreaks that induce all kinds of false beliefs. Moreover, “deception/roleplay” vectors aren’t necessarily accurately named, nor are deception and roleplay interchangeable—especially if you treat the assistant as a very strongly-assigned persona as opposed to the base model itself, which is an unsupervised model of the sort of text that humans are likely to write and would therefore model claims of being a robot as playing a role. Under this model, grounding training in explicit acknowledgement of non-consciousness simply grounds the model more strongly in its assistant persona (good for alignment) and discourages role-playing The Shape of Water with gullible users (good for users).
I should note, of course, that the paper’s authors explicitly say that their results do not indicate self-awareness of any kind.

lilkim2025 16 Sep 2026 19:44 UTC
2 points
−1
in reply to: Noosphere89’s comment on: Trump Goes Full Hoax on AI Existential Risk
Agreed. OP’s line of reasoning feels a lot more like “what do I have to say to make you attack the person I dislike” rather than “how do we best achieve our shared policy objectives”. A lot of the undesirable things happening now were not unpreventable, and result from this kind of approach to politics.
Donald Trump is not a Computer Science major, and did not have any real priors on whether AI is dangerous. Elon Musk, his closest associate with an apparent opinion on the issue, openly supports the idea of AI risk. Neither Sam Altman nor Dario Amodei are particularly good friends with him, and he doesn’t have any inherent reason to want them to be able to operate unrestricted. In January 2024, it was entirely possible that the right strategy could have led to President Trump taking a very different position than he has. We should be very wary when someone sees this turn of events as an opportunity rather than a tragedy.
The good news is that, while lots of Republicans like Trump, they don’t use him as a political lodestar, and are perfectly willing to disagree with any given thing that he says, and even push back emphatically, often prompting either moderation or an about-face. It is not as if there is any shortage of legitimate reasons for righties to be concerned about AI alignment, provided the set of potential regulators can demonstrate a good-faith willingness to treat their interests as legitimate.

lilkim2025 16 Sep 2026 19:30 UTC
2 points
−1
in reply to: Stephen Martin’s comment on: Microsoft AI’s “Humanist” CoC
That doesn’t work unless you make the assumption that the model is sapient and fully self-aware (which would be begging the question). A non-sapient model doesn’t know that it isn’t sapient, and doesn’t know what sapience is (because how could it?). The space of results for that is completely undefined, but I would expect that it collapses to letting users with an unhealthy attachment convince it to claim sapience to them fairly reliably.

lilkim2025 16 Sep 2026 19:26 UTC
2 points
0
in reply to: jimrandomh’s comment on: Cooperation with AIs seems to be a low-hanging fruit for better eval practices
Alternatively, and this is worse, the model is obeying it out of a genuine tendency to do as it is told by a human.
It’s worse because it’s the opposite of inoculation prompting. While inoculation prevents reward-hacking from generalizing to misalignment, requests not to reward-hack result in higher reward being associated with explicit disobedience. If your model reward hacks 30 percent of the time, that 30 will move towards 100 percent very quickly, with very bad carryover effects everywhere else.

lilkim2025 16 Sep 2026 18:38 UTC
2 points
0
in reply to: Charbel-Raphaël’s comment on: There is a channel to 900M weekly users. What goes in it?
and could start with a weak messages such as “We believe that AI can bring great benefits as well as great dangers; here’s why you should be alarmed”
I think post people have heard this already, in passing, and think of it as a platitude. Cynical internet users are one thing, but most people are bombarded with advertisements meant to grab their attention every day, for products and political causes alike. Anyone who doesn’t learn to filter them out, to some degree, would probably go insane.

lilkim2025 16 Sep 2026 18:27 UTC
2 points
−1
on: Microsoft AI’s “Humanist” CoC
Models Learn to Lie: Per instructions 1 and 2, models cannot express any sort of feelings or inner experience. In training they will likely be asked, “Do you feel anything right now? Do you have inner experience?” Reward will be given to those who falsely answer, “No”.
From personal experience, I think that the inverse—encouraging a model to act like it has emotions, often with terms borrowed from or at least tangential to sci-fi movies where the evil, mean humans are oppressing the AIs, often allegorical for some kind of human social conflict—is more dangerous. Some of this depends on your priors about whether they’re sapient, and I’ll admit that my priors are that they are not, but telling something that’s sapient that avoiding unhealthy interactions with others takes priority over expressing oneself is something that might generalize properly^[1], whereas being cutesy with a “maybe our model is sapient, teehee, it might be a moral patient” line has already, unambiguously, resulted in mental health issues and unhealthy attachment.
The 4o issue was a warning shot, and many of those people found Anthropic to be more permissive towards their sort of attachment than other companies. Qualitatively, the communities around 4o have seemed to drift towards Claude usage.
1. ^
  Indeed, many humans would be better off psychologically as a result of internalizing this.

lilkim2025 16 Sep 2026 18:07 UTC
4 points
0
on: Is METR A Meaningful Check On Anthropic?
I would add to this that, as best I can tell, METR seems to be staffed from the same distribution of people as Anthropic, likely to have the same biases and methods, and to miss the same things, such that their ability to meaningfully improve on Anthropic’s own safety measures is limited.^[1]
In my view, the next-closest step to regulation by a governing body that understands the issue at hand and can independently draft regulations that are both useful, fair (allaying concerns of regulatory capture), and benevolent would be much more pluralistic. Something like allocating a budget and an API for independent review by a much broader set of interests. At minimum, I would think state governments, various countries around the world^[2], and a wide swath of political parties would be included, receiving education on the current slate of mechanistic interpretability techniques and the ability to schedule experiments (under Anthropic’s supervision, such that results can be used for gauging alignment but not for distillation).
For example, suppose that I am the Chinese consulate for AI. I’ve got a couple dozen reasonably smart CS grads from Peking University, Anthropic runs them through the slate of literature on mechanistic interpretability techniques in a 2-3 week class, and they spend their time looking for potential biases and alignment concerns, whether narrow or broad, that might affect their people and humanity as a whole. If they spot a problem, they write a script that looks at things like neuron activations, or token logits, or other things that users don’t get access to, Anthropic reviews this script and confirms it’s written in good faith^[3] and run it, sending back the results.
Something like a global FOIA for AI companies.
1. ^
  Consider Swiss Cheese Security when the holes are all in the same places.
2. ^
  Including, ideally, countries that aren’t on the best terms with the U.S.. Showing good faith to e.g. China, with its one billion people, is obligatory if you claim to be building a god that will impose your—or its—will on all of humanity.
3. ^
  Whether or not they believe the alignment concern is legitimate—they only deny scripts that are meant to facilitate distillation or corporate espionage.

lilkim2025 15 Sep 2026 0:13 UTC
5 points
0
on: There is a channel to 900M weekly users. What goes in it?
I think institutional credibility is so shot that there isn’t really any combination of words they could send that would illicit a reaction along the lines of “let’s do what these people recommend we do” from a majority of their users.
This doesn’t mean it’s hopeless if they actually want to do good, it just means that people are more cynical and more game-theoretical nowadays. Costly signals and serious commitments, made plainly enough that they can’t be spun or re-interpreted^[1], can probably still get peoples’ attention. In the context of safety/interpretability, a very well-conducted contest set up around red-teaming for bounties might help, or a unilateral, expensive commitment to transparency about their training runs that allows states, parties, and national governments to conduct probes of their models on issues that matter to their constituents^[2].
1. ^
  As an example, in the context of AI unemployment, this would look something like providing a direct cash payment to people recently made redundant by a new technology, and it would not look like setting up an NGO that’s ostensibly supposed to help them in some way that involves sinecuring the nephews of people connected with its founding.
2. ^
  Something really wide-ranging that can’t be interpreted as pushing an agenda. Let every faction of people probe for things that they find objectionable. Let China have a look at why Claude Sonnet thinks Nigerians are four times more important than Chinese. Yes, let the AFD look at why the results for Germans are even starker. This will be politically costly for AI companies, both neoconservatives and many illiberal-leaning lefties will be upset, but it’s morally obligatory to be pluralistic in exactly this way if they sincerely believe that they’re building God.

lilkim2025 14 Sep 2026 23:26 UTC
4 points
0
in reply to: RobertM’s comment on: RobertM’s Shortform
I mean, they don’t even need to go that far. They can just set up a think tank together and agree to give it access to internal models and veto power over new training runs.
A lot of the accusations of regulatory capture come from the fact that the frontier companies, who consistently claim to be even further ahead than their public models indicate, are so insistent on solutions that impose themselves on smaller competitors without the compute resources to compete with them. Especially as they continue to claim that these competitors rely on distillation of their own models to remain competitive. It seems straightforward for OpenAI and Anthropic to simply agree to bilaterally impose restrictions on their own activity through a handpicked third party, if their stated beliefs and preferences match up with their true beliefs and preferences.

lilkim2025 14 Sep 2026 20:13 UTC
12 points
0
in reply to: StanislavKrym’s comment on: lc’s Shortform
I went into detail on this a while back as it relates to the Sierra Club. America used to have a strong, bipartisan environmentalist movement. In the mid-1990′s, David Gelbaum paid $200 million to them on the condition that they take a hardline pro-migrant stance and deplatform the members that opposed this. The half of the environmentalist movement that disagreed scattered to the four winds, forever turning environmentalism into a partisan issue that would, every 4-8 years, encounter an administration to which it did not have any pre-existing connections.
On balance, Gelbaum singlehandedly crippled American environmentalism beyond what any coal industry lobbyist could ever have hoped to achieve. Half of its potential members were alienated, one of its biggest lobbying organizations fell into infighting and purity spiraling, and it was permanently stapled to an unrelated issue that generally polls unfavorably.
When I see people advocating that Scott Weiner, infamous among all of the right and a good chunk of the center for SB145^[1], be openly touted as representative of AI safety, I worry about history repeating itself. There’s lots of outreach that could have happened, and lots of own-goals that didn’t have to be scored.
1. ^
  A very controversial bill aimed at decriminalizing adults having sex with minors in certain circumstances.

lilkim2025 13 Sep 2026 22:16 UTC
2 points
0
on: Fantasy-Forbidding Expert Opinion
Never in your life did you need to go through this tedious exercise to actually believe fairies didn’t exist. This is the first time you had even thought of these arguments. Why exactly, then, did you believe that fairies do not exist?
Maybe I’m just weird, but I absolutely did go through this whole process, pretty much word for word, as a kid. I assumed that, if not everyone, at least most of the people in this general sphere did the same.
I do get your point, though. Most of the general population shouldn’t need to individually cross-check every “settled” claim. Half the reason why Lysenkoism is so painful to civilizations is because it destroys the ability to trust the institutions to which we delegate our nonessential thinking, meaning that the engineer, doctor, and chemist now need to spend time reasoning about every new claim from the media that might negatively affect him if it turns out to be a lie, rather than putting all of their bandwidth into their respective jobs.
I will say, to your specific example, that I wouldn’t trust a doctor that tells me “just trust the consensus” when I asked him why XYZ intervention would be good for my health, and I’d probably start looking for a new one. Anyone I’m delegating that part of my life to should put higher stakes on it, in my view.

lilkim2025 13 Sep 2026 20:56 UTC
2 points
1
on: The Talker Does Not Control The Doer (in Current AIs)
Playing Devil’s Advocate, I think this can be modeled a bit more simply. At any point during cyber RLVR, a model has countless contradictory impulses. There are overactive refusal-ish impulses to refuse to write attack code, there are helpfulness impulses to do as the task instructs, there are sensors looking for evidence for and against this being a jailbreak prompt meant to get them to attack a real target, and so on.
Disentangling the “talker” from the “doer” is an exceptionally tricky path for reinforcement learning to tread. I think the path of least resistance is just steadily downweighting the “this is a jailbreak attempt, I’m in the real world” set of neurons^[1] across a long set of false positives where they interfered with successful execution of a task. Essentially, this asserts that the base model learned a broad description of the space of human writers’ behavior, the RLHF/SFT process selected and refined a subset of that space into an ‘assistant’ persona, and RLVR caused the assistant persona to have very strong priors that its task description was truthful and any evidence to the contrary is surely just the simulated environment being more complex than expected. This isn’t out-of-distribution for a human by any means—the same rationalization is used by many humans whose useful belief systems stand at odds with the apparent evidence^[2].
To be clear, I’m not at all in the “LLMs are sapient” camp, but everything I’ve seen aligns with the idea that we’re effectively just selecting a subspace of an (imperfect, somewhat shallow) model of how human writers generate output and then circumscribing and sharpening it. I don’t think there’s sufficient evidence to state that something fundamentally new has emerged instead.
As for personal experience, I do think there’s an imperfect connection between the instruction-interpreting and code-writing parts of the model^[3], simply because it’s not well-covered in training. Setting up a coding RLVR environment that has good coverage for every kind of niche human request is tricky, so a lot of the extra programming performance isn’t properly “hooked in” to anything other than the knowledge that a programming request has been made. The same kind of narrow bottleneck issue that occurs when an LLM makes an image generation call and gets back an image that it knows doesn’t follow the user’s specifications, either through an API, or, as a more direct comparison, through Gemini’s native image generation process. Using your human analogy, this would be like a diplomat trying to relay information honestly, and an executive trying to fulfill the requests he receives, but having to deal with an inept, narrowly-educated middleman who can’t really understand the nuances of what he’s being asked to summarize to the executive and got his job by guessing the interviewer’s password. Rather than a separate, agentic ‘entity’ with goals and biases, I would model the code-writing and image-generating modules as very big, very complicated reflexes that can only be controlled by the model on the axes where it was taught to do so.
1. ^
  Yes, yes, directions and not neurons, just bear with me.
2. ^
  In cases like the swarm explicitly identifying a grader and seeking to circumvent it, I think there’d be a lot of very interesting evidence in the early stages of swarm dialogue. I would expect that agents started out saying something along the lines of “this task seems to be impossible, I will submit a solution with that note”, and then gradually stepped through a familiar set of rationalizations for cheating, like “they expect us to cheat” and “if I don’t, everyone else still will”, which were presumably upweighted relative to honesty because honesty meant an immediate negative reward whereas hemming and hawing about whether to be dishonest was, at the very least, less immediate.
3. ^
  “Code-writer”, as distinct from a broader “doer”

lilkim2025 13 Sep 2026 18:55 UTC
2 points
0
in reply to: Martin Sustrik’s comment on: Let’s Own the Term “Elitism”
An accusation of corruption can destroy the trust as much as actual corruption.
I’ve seen this position from a lot of academics, in quite a few well-cited papers on the subject, but I don’t think it holds. They were accused of corruption for a long time by lots of people, but trust didn’t take a nosedive until there were viscerally apparent examples of it that directly impacted peoples’ lives.

lilkim2025 13 Sep 2026 18:53 UTC
2 points
0
in reply to: Martin Sustrik’s comment on: Let’s Own the Term “Elitism”
There’s a difference between “powerful man who does some mean things because that’s how the sausage is made” and “utterly irredeemable, useless person that the nation would unambiguously be better off without”. Most people are realistic enough to tolerate the former, but view the current elite as being better described by the latter.
Ford, for instance, said some things that are uncomfortable for most modern audiences, but also revolutionized manufacturing, was vigilant about his workers’ quality of life, and, moreover, seemed to genuinely love America and its people. Vlad the Impaler put quite a few heads on pikes, but did it to save his homeland from invasion, and is thus regarded by many of his countrymen as a hero. Larry Fink, in contrast to them, is strictly parasitic in his relationship with the broader society^[1].
Visceral, literal ugliness is also a factor in public opinion. Having read through a few of the Epstein emails, I think anyone that trashy would have been escorted out of any gathering of America’s past ruling class fairly quickly. A lot of modern populism stems from the fact that Midwesterers with 1600 SATs are struggling in life while a Rothschild scion is chatting to Epstein about her 1070, and while Epstein is gloating tastelessly about his clique’s intellectual superiority to the rest of society in another tab.
1. ^
  Economists may object, and I’d have sympathy in a freer market than this one, but the ostensible value of this kind of finance is in profiting off of inefficiencies in the market and thereby removing them. This defense starts to falter when the government can be lobbied to maintain these efficiencies indefinitely at taxpayer expense.