domain:drmanhattan16.substack.com
Florida and many other states have a mandatory two-semester Holocaust class.
They had assets on the book worth more than their market cap
I'm sure FTX also had assets 'on the books' large enough to cover all their liabilities and even lawsuit/regulatory risk if they got targetted.
But the books were cooked, which is seemingly a common factor in crypto exchanges, with insider trading and light fraud showing up with some regularity. Ironically given that one use of crypto is keeping all books open for easy scrutiny at all times.
Mostly joking. Coinbase is, I think, the longest running American Crypto Exchange and has weathered several storms so I'd not be the one betting against them. I've had an account with them since they used to hold private keys printed on physical paper stored in physical lockboxes. But there was definitely a period of time around and after the IPO where it wasn't clear if the Gov't was going to really crack down on exchanges or liberalize the rules.
We're clearly in the liberalizing timeline.
Freddie's post sounds like ravings of exhaustion from having to fight a broad and deep set of ideological concepts that all have shared roots in 20th century social liberalism (feminism, civil rights, etc.), and his solution is to pigeonhole all those ideological concepts into a single overarching theory that can be attacked directly without having to get into the weeds and nuances of any individual ideology. But also, he says that it's not his responsibility to perform this abstraction, but that all of these separate ideologies must bring themselves under a single banner? For his convenience?
I don't see the appeal of his writing, either. This is the only snippet I've read, but I've stumbled across his name.
Edit: I've read more of his writing. This post seems to be written in an intentionally exasperated voice.
unfortunance
Had to check if this is a real word. It is not. :(
Yeah and there's constant double standards about this stuff. "Manosphere" and all that gets maligned like it's automatically slightly evil and should be combated. Bullshit harmful sources of programming for girls/women? crickets Just some concern about whether it's not quite girlbossy enough??
Ask her how she arrived at the calorie goal in the app (on what basis the target was set).
I think this betrays that the line you're trying to draw isn't really sharp.
I can make it sharper, but I'm not really painting with a moral brush here.
I'm trying to reason about the type of person that's going to be wielding governmental authority and whether they are going to do a better job than the others I might be able to choose.
The Coinbase lawsuits were a hoot. At one point the stock was down 90% from the IPO seemingly on the basis that they'd get bankrupted by federal lawsuits. It was a fun and profitable bet to take the other side of. They had assets on the book worth more than their market cap. I'd had this theory that Coinbase would execute a regulatory capture strategy for crypto in the US, as they'd always played well with regulators and generally licked their boots and did whatever they were asked to do to the consternation of /r/bitcoin. I never counted on the feds just having an insatiatable hardon that verged on a medical condition for destroying them on principle. I especially enjoyed the SEC being hoisted by their own petard when the Bitcoin ETFs were approved, effectively by court order.
Sometimes I do wonder what would have happened if the Biden administration hadn't nakedly antagonized so many interest, and driven them into the arms of Trump not out of affinity, but existential desperation. I guess I can be thankful that this once at least, evil did contain the seeds of it's own destruction.
Why?
Because it correlates with precisely zero of the skills/personality traits I would want to see in a legislative representative or public official.
And likely correlates with beliefs/personality traits I would distinctly not want in a representative or official.
Simply put: You cannot POSSIBLY convince me that they're the best alternative for the job at hand.
And to once again put my preferences out there: I'd generally want my officials to be Married with kids (i.e., a stake in the future), own some property in the area they're representing (i.e. have some skin in the game for the wellbeing of their local area), not be in large amounts of personal debt, and not be prone to making, call them 'self-destructive' decisions (hard drug abuse, highly risky behaviors with limited upside, wanton and regular gambling, and really weird sexual activity, even if you keep it within the confines of the law).
Basically, Hunter Biden is NOT someone I want representing me or having any sway over my life. I was endlessly annoyed to learn that he might be influencing decisions during Joe's term.
Publishing porn of themselves is actually decent evidence that they don't meet my preferred criteria. So my Bayesian updates are going to require some SERIOUS evidence in the other direction to conclude that they're the most suitable person among the candidates.
And there's also this little gem in the article:
“I could barely get up off the floor for about two weeks,” she said, adding that anyone reading about her account should think about how it would feel “to know that your naked body is going to be splashed all over the internet.”
She says, after intentionally splashing her naked body on the internet.
Ability to foresee possible/likely consequences of actions in the present is another feature I'd want in my officials... but I'm well aware we rarely get that.
I'd also like them to not have a two week long emotional breakdown in response to negative events, but I'd assume she exaggerates.
A. The base model already has a world model:
Pretraining on next-token prediction forces the network to internalize statistical regularities of the world. You can’t predict tomorrow’s weather report, or the rest of a physics paper, or the punchline of a joke, without implicitly modeling the world that produced those texts. Call that latent structure a “world model” if you like. It’s not symbolic, but it encodes (in superposed features) distinctions like:
What typically happens vs what usually doesn’t Numerically plausible vs crazy numbers causal chains that show up consistently vs ad-hoc one-offs
I'm going to need a citation; I have seen no research to date that suggests LLMs develop any sort of a word model. A world model is:
- An explicit internal representation of cause-effect relationships.
- Grounded reasoning about physical, social, or conceptual structures independent of linguistic statistics.
- A structured understanding of external reality beyond pure linguistic correlation.
Instead, current research strongly suggests that LLMs are primarily pattern-recognition systems that infer regularities purely from text statistics rather than internally representing the world in a structured, grounded way.
An LLM can easily write a weather report without one, will that report be correct? Depends on what you consider the "LLM" the actual text model: no, the whole engineered scaffolding and software interface, querying the weather channel and feeding it into the model: sure. But the correctness isn't emerging from the LLM's internal representation or conceptual understanding (it doesn't inherently "know" today's weather), but rather from carefully engineered pipelines and external data integration. The report it is producing was RLHF-ed to look correct
Late last year I posted a comment here asking how I could convince my girlfriend to start eating more. Now, I'm posting an update.
Frankly I was annoyed by some of the replies to my previous post that said I should "enjoy my slim girlfriend", or that implied that I was making more of a problem out of it than it actually was. She actually lost more weight and now weighs 98 lbs. It took several months of patient intervention for me to convince her that yes, I do actually want her to gain weight, and yes, I absolutely would still think she was pretty if she weighed 120 lbs. Recently she finally caved and went to her doctor for a formal medical opinion, and he backed me up on this by telling her that she was at risk of osteoporosis and anemia if she didn't change her diet and gain weight. My cause has also seen some support from her older sister, a very intelligent woman whom she trusts a lot, telling her that she needs to start eating more red meat. So in theory, at least, I've been able to convince her that her ordinary diet and habits aren't healthy or sustainable.
The problem, at least as I see it, is that even with this realization it's been hard for her to break her habits. We go out to brunch and she still eats her little vegan salads. I tell her she should add some chicken or other protein to the salads and she declines. She still consults the app on her phone that counts all her calories for the day. It's hard for me to figure out what the line is between pushing her to be healthier for her own sake, and being outright controlling over her lifestyle. Do I just put my foot down and confront her, pushing her to be serious about her health?
If videogames cause similar brain signals, aren't those basically drugs?
No. The difference is games produce natural-brain chemicals whereas drugs are alien to brain and produce much longer effect.
Although the general principle that a government official cannot coerce a private party to punish or suppress disfavored speech was well established, the law was not clearly established that the conduct alleged here -- regulatory action directed at the nonexpressive conduct of third parties -- constituted coercion or retaliation violative of the First Amendment. Accordingly, we reverse the district court's qualified immunity ruling and remand for the district court to enter judgment dismissing the remaining claims.
[Past discussions here, here, and here. Also, someone pick up the phone, because I called it.]
Perhaps Justice Jackson's retaliation theory would have survived, if only the NRA were prescient enough to bring it --
The NRA also alleges that Vullo engaged in unlawful retaliation when she pursued investigations and enforcement actions against the NRA's business partners. But again, qualified immunity is proper here because the nexus between the alleged retaliation and the alleged eventual infringement of the NRA's First Amendment rights is too attenuated.
Nope.
Well, are there any protections for rights beyond 'don't do exactly this twice in a row', when this case had to go up to SCOTUS and took the better part of a decade to produce a case that 'clearly establish' matters in the first place?
To be clear, we do not hold that an official may infringe the statutory or constitutional rights of any person or entity if she can find some sufficiently novel way to use her office and powers to do so. Rather, we hold here that qualified immunity is proper only because no case (or set of cases) had clearly established, by the time of Vullo's conduct, that exercising regulatory power to pressure third-party regulated entities into refraining from nonexpressive activity and disassociating from a plaintiff crossed the line from persuasion into impermissible coercion and retaliation.
How low is the standard here?
Given that even this Court erred in its merits conclusion in Vullo I, it would make little sense to hold that a reasonable officer in Vullo's position should have known -- in 2017 or 2018 -- that her conduct violated the NRA's First Amendment rights.
Remember, that's for a case where SCOTUS -- in an opinion written by Sotomayor -- slapped down the Second Circuit. Here, over a year later, the Second Circuit instead motions to circuit precedent holding that "vigorous dissent" or "a petition for rehearing en banc [that] was denied over the objection of five circuit judges" is sufficient. A cynic might notice that the 2nd Circuit covers land where state judges make arguments like "Do not bring the Second Amendment into this courtroom. It doesn’t exist here.", and ponder whether even that would be sufficient.
The more damning revelation is that all of this is still the motion to dismiss phase. The NRA alleges quite a lot of bad behavior by Mrs. Vullo and the broader NYDFS. Long-extant circuit precedents "generally prefer lower courts to resolve qualified immunity defenses at the summary judgment stage"; we didn't get that. Discovery might well have established far more of the harm, or at least proven rather than merely alleged the bizarre array of Vullo's retaliatory behavior. It was also too deep a cut for the Second Circuit's vaunted interest in due process to let survive.
Steam Bends The Knee
On a slightly more lighthearted topic, Valve has now updated its acceptable content list, specifically forbidding:
Content that may violate the rules and standards set forth by Steam’s payment processors and related card networks and banks, or internet network providers. In particular, certain kinds of adult only content.
Steam has long had an awkward relationship with adult media, and not just because you might not want to tell your entire friends list when you first start up Giant Melons Simulator 3000. Valve is a big business, and a business mostly consisting of credit card transactions with a nontrivial chargeback rate. Even if it could afford a temporary disruption from a bank it operates with, or for that matter an eyebrow raised from one, it would cost the business orders of magnitude more than all adult video game sales would in a much longer period.
At the same time, Valve is an ideological company, and it has long committed to not being the taste police. While that commitment has had fuzzy edges even before it was written down, most of those exceptions and exclusions were at least arguably in the spirit of the original thesis. This one? There's not much but the payment processors.
In the short term, there's no shortage of other locales selling games, including adult-themed games, and other ways to make a living making games... though all are likely to get hit by the same pressures over time and sometimes already have the same sort of motions toward content provider rules. There are other ways to make a living making games, without necessarily selling the output. There are other ways to get a story out.
There's a fair argument that these sort of games shouldn't be monetizable, or shouldn't exist. I wouldn't agree to most of them, but I will admit few games banned so far are those I'd consider unique or unusual in their innovation or message. There's no cultural touchstones falling, yet, even among subcultures that go through astounding amounts of kleenex.
But it's also not clear that's what's actually banned, and that's the more damning bit. There is not, in fact, some clear list of what Steam's payment processors find acceptable; attempts to read through itch.io's list of payment processor guidelines point to a ban on porn entirely, and even a glance through those competitors will give seemingly-obvious differences from the letter of those rules. Looking through the list of removed games gives some common themes, but also some clear outliers (including one game that apparently wasn't adult, it just had a stupid name?).
Well, we can't ask Visa or Mastercard (or Wells Fargo) directly (and expect to get an answer), but can we look at the people who pushed them in turn?
Like most pornapocalypses, there's a bit of a bootleggers-and-Baptists activism involved between social conservatives and feminists, but one of the big actors is NCOSE, who spent the last week panicked over Grok's 'Ani' companion, previously wrote terrified fanfic about Frozen II porn, and tried to get a dating-sim-Bejeweled clone pulled. They're not likely to get everything they want; there are some fundamental disagreements in what content should be banned, even between NCOSE and other parts of this coalition. Indeed, it's hard to tell whether Steam's even seriously engaging with their rules here, or just trying to get them to go away. But to the entire coalition, the fight is clearly not done.
There's a common defense of these restrictions from a perspective of free speech, pointing to the difference between government and business coercion. No private merchant has to sell a specific product; no private credit card processor has to work with every possible client; a private bank gets a little more complicated legally but it's not a free speech question. It's only when the government leans on these groups to do that it becomes a clear-cut constitutional violation (or when Rehnquist likes you).
But we do know that happens, at least in some few other cases! It got lost in the rest of the politics, but Operation Choke Point specifically included smut in its list of suspect goods. And we only found out because of several whistleblowers, administrators so certain in their impunity that they spread documentation across multiple federal offices, and literal years of Congressional investigation.
But we're not going to get that, here. Congress doesn't and won't care. Even in the unlikely case that NCOSE and the modern-day-Sarkeesians get everything they want, Congress won't care. There's no insurance executive here who believed in the mission and will provide testimony to individual businesses at the sharp end of the stick, demonstrably. Without that support, or equivalents, there may be no serious way to actually find out what happened in any other case, including this one. There's no cause of action a game developer would have against Steam that would survive a motion to dismiss, and even in a world where multiple different government agencies were involved in pressuring someone here, those developers wouldn't even know who to sue, nevermind have a method with any chance of success.
Except...
Bank Pause Letters: Electric Boogaloo
[disclosure: while there are some technically interesting components to ponzigpucoins and there are some things like distributed DNS that I hope eventually work out, the entire field has enough scammers that I recoil from it. As a result, there may be (more) errors in the general discussion than I normally aim for. Corrections are appreciated, as always.]
In February of this year, Patio11 over at X Twitter commented on a court transcript:
Hoooooooooly cow what a document. It's about tussling over whether the FDIC should release so-called pause letters regarding banks offering crypto-flavored services. I've written about that general issue before, including why regulators institutionally claim privilege maximally.[...]
At one point the lawyer for the FDIC attempts to bullshit his way through a question and the judge helpfully reminds him that the robes mean he is a federal judge and the government is not allowed to bullshit in response to his questions.
The transcript is a fascinating (if dry) read, and if anything Patio11 is giving his trademark understatement. The rest of the history doesn't look much better. To summarize, Coinbase believed that a handful of government agencies had pressured a large variety of banks to stop taking in new ccoin-related business, and decrease existing accounts as a percentage of total business.
Here, there was a known set of agencies that could have been making the regulatory decisions. There was a very specific timeline that those regulations could have actually come down in. "History Associates", the org used as a FOIA plaintiff, was unusually well-funded, and had several organizational protections making both the initial FOIA and the lawsuit going after the agency violating FOIA more viable. There were very specific terms, and to the extent the exact rules weren't known to the plaintiffs, they could make some very specific guesses about the language the regulatory agencies would use.
(TBF, they have also gotten a whistleblower from the FDIC, although given that he seems to be coming across to the court as a bit of a nutjob, not sure if that's helpful or otherwise. But as much of a nut as he seems, he also has alleged that the FDIC was destroying data, and the FDIC admitted that it did not issue a command to preserve that data.)
If all of that helped, it wasn't sufficient. Even before the January 2025 court hearing, the FDIC has been abusing FOIA exceptions to illegitimately hide the extent these organizations were targetted and had rewritten the FOIA request in a way that we now know changed it from dozens of documents to a mere two. The SEC, involved in determining whether ETH was a security or not, had its chair publicly claim that the change from proof-of-work to proof-of-stake could impact something somewhat, and then never publicly saying how, avoided having any court hearings turn as embarrassing as the FDIC's. But it's not done much better -- starting out by declaring no responsive documents existed, then that over a hundred thousand did but were temporarily blanket-exempted, and then ten months after the initial rejection announcing that not only did they not prepare for after that exemption inevitably ended, but that they wouldn't even know what the timeline for a Vaughn index to explain what documents could be exempted for other reasons would be.
But the results, damning as they are, don't actually mean anything. It looks like the claims of a 15% rule weren't probably made up, but it's also not explicitly in the documents, and even if it were in one it wouldn't mean it was actually being applied as a consistent rule. Even today, I wouldn't want to bet monopoly money on it. In the unlikely case that these particular lawsuits ever actually resolve, they're still FOIA lawsuits; the best "History Associates" can hope for are some damning documents and some relatively-small monetary awards. Actually achieving any concrete policy changes would require either overtures toward the public (good luck) or toward senators (gfl) or filing further lawsuits (hah).
Thankfully, we have at least one alternative. There's been a long-running joke about coup-complete problems, as a disparaging way to describe issues that can't be resolved short of having complete control over a government. Unusually for such jokes, we have what the current president's political opponents might call a test case. For various reasons, that President does not like (some) debanking. There has been some regulatory movement, and even some legislative woolgathering (uh, if you can avoid laughing at the name). Whatever results from that might not be worth the toilet paper it was written on, but lawsuits have the same problem and didn't have to filed in 2018 to get a no today, unlike Vullo.
Of course, in neither case will we ever find out what the actual rules are.
Italy voted a porn starlet into parliament:
https://en.wikipedia.org/wiki/Ilona_Staller
In 1985, she switched to the Radical Party, campaigning on a libertarian platform against nuclear energy and NATO membership, as well as for human rights. She was elected to the Italian Parliament in 1987, with approximately 20,000 votes. While in office and before the outset of the Gulf War, she offered to have sex with Iraqi leader Saddam Hussein in return for peace in the region.
I suspect the U.S. siding with USSR rather than Germany is a symptom of this phenomenon rather than the cause of it. Senior people in the Roosevelt administration were already giving Stalin a pass.
to ogle up the pages of the high-class magazine with the naked women, AND to be a full-on gooner who consumes hours of porn
I think this betrays that the line you're trying to draw isn't really sharp. There is, IMO a slope (maybe even a frequently slippery one, not intending to claim otherwise here) between nudity in "high art" (enjoys fine wine and Rubenesque paintings), moderate-but-titilating scenes in Hollywood cinema, something like Playboy, and whatever depth of depravity you want to stop at. The debate on the appropriateness of nudity in art goes back literally centuries --- fig leaves are not a modern invention.
I'm not sure I would judge "modeled for the figure drawing course in college, maybe even for reasonable payment" in isolation harshly in general. But that's just my opinion.
They're expensive broken ones that snapped at the apex of the curve where it's not conducive to a satisfactory mending job.
The original band is nylon and after exhausting the possibility of finding official spares or eBay parts donors I initially started by looking at 3D printing. I found an stl for this model of headphones on Thingiverse and got some quotes for nylon prints (FJM process if I remember right) but while the quotes were reasonable I wasn't totally happy with the model. It wasn't a 1:1 identical match, and even if it had been the originals were a little tight fitting so I thought it would be a chance to modify the model to my own spec. Turns out that 3D modelling isn't like image editing and I would need the "step" file to edit it, like having the layers for editing a photoshop file instead of printing out a finished jpg. Sadly the Thingiverse model didn't include these. Okay, it's not a hugely complex shape. Maybe this is my motivation to finally learn a bit of 3D modelling.
So after installing FreeCAD, uninstalling and writing it off as a nonstarter I got an account on SketchUp and started drilling the same YouTube tutorial each evening for a week until I could replicate the example bedside drawers without needing to follow the video. Equiped with the basics I started on modelling the headphone band, building up the shape as a flat linear form that I would then curve into shape as the final step. Turns out the curving function is locked behind a premium plugin. Sigh. I could pirate the software but that's a whole other tangent of a tangent.
Anyway, I was sitting here thinking about how I could DIY it when I remembered I have about 10m of beech veneer edge banding gathering dust. It's flexible, it's easy to work, it's near enough the right width, it's preglued and I can build it up to achieve the right amount of flex to rigidity. I had all the measurements I'd taken from the original product that I was going to use for the 3D model so I marked out a form on some crappy old pine board from the scraps box and got the jigsaw out. After that I looked up soaking and boiling times for bending wood (only 2 minutes for material this thin) and progressively clamped more strips to the form after allowing the previous one to dry and set. The strips needed softening with heat and water because the inside half of them have to bend counter to the way they'd been stored rolled up, and all of the layers have a tight dog-leg bend at the ears to accommodate the drivers.
Now I have six layers clamped in the form and I'm ready to heat the glue and stick them together. After that if the flex feels right there's only a little cutting and drilling to do. I have a feeling I'll have mixed results at best, but I'll be content just to have made something usable while I work on a better solution. Apparently oak is one of the woods that is both amenable to bending and also widely available here and reasonably affordable meaning I can buy a 3' strip of suitable dimensions for a little over £10 in town without having to mail order from a specialist supplier or buy a large board that I only need a fraction of.
I also think people should be free to associate or not associate with whoever they want to, so I can't really point to them doing anything exactly wrong when they do that.
I think that people should be free to do a lot of things which are bad, even specifically bad for social cohesion. You're free to say "I don't like Jews, so I won't associate with this Jew". I would still call it wrong.
We need to distinguish "break up friendships for a bad reason" and "break up friendships for something we can prove is a bad reason". If someone refuses to associate with you because you're a Jew, that's wrong. If someone refuses to associate with you because he doesn't like your attitude, that's fine. If someone refuses to associate with you because you're a Jew, and he lies and says "I just don't like his attitude", this is still wrong, even if you have no way to prove he's lying without looking into his head.
I'd probably limit wokeism to identitarian politics as well. Nobody gets cancelled for saying they don't think the workers need to own the means of production.
I think it’s even simpler than that.
• Soviet Union: US ally in World War II •Nazi Germany: US enemy in World War II
In the parallel universe where 400,000 brave American boys died holding back the red tide of Bolshevism from overrunning Western Europe, and there was an awkward alliance and later Cold War with Nazi Germany, the t-shirt situation would probably be reversed.
There have been various attempts at defining "wokism", but for me the distinguishing characteristic is the set of tactics it employs and not just its goals.
Standard "wokism" is progressive or left-wing politics (often, but not always identitarian) plus anti-liberal tactics like deplatforming, ostracism, cancelling, refusing to explain oneself or try to convince one's opponents, etc.
There's two issues with this definition.
The first, is that certain woke tactics are technically within the realm of acceptable behavior. I dislike the tendency of "woke" people to break up friendships when they learn someone is a Trump supporter, but I also think people should be free to associate or not associate with whoever they want to, so I can't really point to them doing anything exactly wrong when they do that.
The second, is that such tactics aren't unique to progressives or the left. That's why I've seen a few failed attempts to rebrand some aspects of Trumpism as "right-wing wokism." Certainly, I'm no fan of Trumpism, and some of it is that it is not the committed, little-l liberal alternative to wokism, but just another anti-liberal form of identitarianism.
France truly is the 4chan of philosophy. Everybody likes its memes, but few can stomach the environment which was necessary to produce them.
This is a notable quotable. Nicely done.
Uh, where did you arrive at the idea that it isn’t? It’s a paired activity, kinda inherently a relationship there.
Wasn't there a (state? California?) legislator who was found to have had an OnlyFans? I'm not going to search for it at the moment, but I remember this happening within the last few years.
Boring practical reasons are high minded spiritual ones. There is no difference except maintenance through time. It's a mistake to think that virtues are the concerns of some untouched far away realm and not the common lives of common men of the common time.
What similar problems do drugs have?
They make reasonable men do unreasonable things.
Other vectors of addiction can do the same, but nowhere near at the same rates. Few men have killed because of a video game. Legions have for sex and drugs.
When and how did you arrive at the idea that sex is a relationship
Arguably it started when I read Plato as a teenager, but that's not what convinced me of it. I think it's about when I had friends of mine get back together for the third time after swearing each other off. That's about when I knew my mother was right both to say that sex makes people retarded and that it connects people on a special level other things do not.
But time has done nothing but confirm this for me. I've hung out with many a people of little virtue (I've probably had more candid conversations with sex workers of more genders than anyone on this website), and their existential angst seemed to be proportional to the amount of casual sex they had.
A decent amount of them openly whined to me about feeling desensitized and frustrated at how unable to feel anything for anyone they were.
Now I refuse to conclude something too specific from my life experience, because it's not really generally applicable, but I will certainly stand by the ancient idea that sex is special and that carnal knowledge is not like other acts in its implications on the psyche of the participants. That alone seems undeniable to me.
did you already believe that porn was a bad thing beforehand and this was just one more piece of supporting evidence?
I do not actually believe pornography to be evil. At least not inherently. I've stood by its artistic merits and associated freedoms here before on numerous occasions.
I used to have far more liberal views on this particular matter, but these days I think it probably needs some effective regulation or at least social framework, like we do for tobacco or alcohol.
The complete free for all of hyperreal stimuli seems a bit unwise, seeing as though I've had multiple friends make stupid life ruining decisions on account of it.
How exactly does an LLM know that Mozart wasn't a fan of hip hop without some kind of world model? Do you think that fact was explicitly hand-coded in?
Anyway:
This is the core of our disagreement. I'd argue this is a false dichotomy. How does one become a master pattern-matcher of text that describes the world? The most parsimonious way to predict what comes next in a story about balls falling or characters moving between cities is not to memorize every possible story, but to learn an implicit model of physics and geography.
Which we know happens:
Large Language Models develop structured internal representations of both space and time
There's a whole heap of mechanistic interpretability research out there, which finds well-ordered concepts out there, inside LLMs.
You can find more, this Substack has a good roundup.
You say: “The LLM cannot know today’s weather, only the scaffolding can.” True. That does not bear on whether the base model holds a world model in the predictive-processing sense. The base model’s “world” is the distribution of texts generated by humans who live in the physical world. To predict them well, it must compress latent generators: seasons, cities, typical temperatures, stylistic tropes. When we bolt on retrieval, we let it update those latents with fresh data. Lack of online weight updates does not negate the latent model, it just limits plasticity.
RLHF shapes behavior. It does not build the base competence. The internal “truth detectors” found by multiple groups are present before RLHF, though RLHF can suppress or amplify their influence on the final token choice. The fact that we can linearly read out “lying vs truthful” features means the base network distinguished them. A policy can still choose to ignore a feature, but the feature exists.
On your definition of a world model:
By insisting on “explicit, grounded, structured” you are smuggling in “symbolic, human-inspectable, modular”. That is a research preference, not a metaphysical requirement. Cognitive science moved past demanding explicit symbol tables for humans decades ago. We allow humans to count as having world models with distributed cortical encodings. We should use the same standard here.
More options
Context Copy link