This is already possible for Harry Potter specifically. There was a study demonstrating that Sonnet 3.7, among other models tested, could reproduce the first Harry Potter book 95.8% verbatim[1].
2,202 karma · joined October 22, 2025
[1] https://news.ycombinator.com/item?id=47057829, "Microsoft morged my diagram". It was in a discussion there that someone pointed out this article linking to full downloads of the Harry Potter novels, which I thought deserved more visibility.
It is certainly a bold choice to use this wildly misleading framing when you link to a news article that directly contradicts it. A more accurate framing would be "when a guy returned fire against a gang of thugs who were firing on random passerby from an unmarked van".
Court documents and transcripts reveal a far different story than the one officers told investigators, as well as the tales police and prosecutors offered up to the media.
Before the white, unmarked cargo van of the Minneapolis Police Department drove down Lake Street, an officer gave Sgt. Andrew Bittell his orders: “Drive down Lake Street. You see a group, call it out. OK great! Fuck ’em up, gas ’em, fuck ’em up.”
At 17th Avenue and Lake Street, around 10 p.m., the SWAT team saw a group of people outside the Stop-N-Shop gas station. Bittell told the driver to head toward the station and said, “Let ’em have it boys!”
They later learned they were shooting at the gas station owner, neighbors and relatives guarding the station from more looting, as well as bystanders, including a Vice News reporter who had his hands up and was yelling, “Press!”
About an hour later, three blocks to the west, they opened the sliding door of the van and began firing plastic rounds at people in a parking lot.
They hit Jaleel K. Stallings, 29, a St. Paul truck driver, who says he didn’t know they were cops because they were inside an unmarked white cargo van with the police lights off. [...] Stallings, an Army veteran, returned fire with his mini Draco pistol, for which he had a permit.
---
> Everything I've said is factually accurate. Hilarious how some commenters are saying I am "lying" or assume I disagree with the verdict when it could not be further from the truth. I'm only pointing out that armed fire upon police might be legal in Minnesota, and there is recent case example of that.
This is a silly way to have a conversation, but as for your response: every single word you picked was as misleading as possible. "AK" implies an assault rifle but it was a pistol. "Lit up on cops" implies that he started the conflict by attacking cops, rather than it being one of self-defense against people indistinguishable from thugs. You invoke "BLM riots", but there is no evidence he was involved in that at all. Your words are clearly chosen in such a way as to prime people towards a certain belief about the event. With the most charitable interpretation of your words possible, they might factually describe the events, but I think it crosses the line to the point where you would have to be so charitable as to actively misinterpret what words mean in order for them to remain factually accurate. At any rate, that level of charity is absolutely unwarranted given how intentionally uncharitable the selection of those words was in the first place.
> and you could not make deepseek or mistral if I gave you a big gpu cluster. You have the weights but you have no idea how they work and you couldn't recreate them.
I personally couldn't, but the team behind that startup as a whole absolutely could. We did attempt training our own models from scratch and made some progress, but the compute cost was too high to seriously pursue. It's not because we were some super special rocket scientists, either. There is a massive body of literature published about LLM architecture already, and you can replicate the results by learning from it. You keep attempting to make this out to be literal fucking magic, but it's just a computer program. I guess it helps you cope with your own complete lack of understanding to pretend that it is magical in nature and can't be understood.
It is extremely well-known that individual humans make mistakes. Therefore, any well-functioning system has guards in place to catch mistakes, such that it takes multiple people making mistakes for an individual mistake to cascade to system failure. A system that does not have these guards in place at all, and allows one individual's failure to immediately become a system failure, is a bad system, and management staff who implement bad systems are as responsible for their failure as the individual who made the mistake. Let us be grateful that you do not work in an engineering or aviation capacity, given the great lengths you are going to defend the "correctness" of a bad system.
> So we end up shipping documentation that nobody really reads
I'd note that the documentation may have been read and noticed as flawed, but some random person noticing that it's flawed is just going to sigh, shake their heads, and move on. I've certainly been frustrated by inadequate documentation before (that describes the majority of all documentation, in my experience), but I don't make a point of raising a fuss about it because I'm busy trying to figure out how to actually accomplish the goal for which I was reading documentation for rather than stopping what I'm doing to make a complaint about how bad the documentation is.
This says nothing to absolve everyone involved in publishing it, of course. The craft of software engineering is indeed in a very sorry state, and this offers just one tiny glimpse into the flimsiness of the house of cards.
Notably, the abstract of the 2015 study specifically points out that the 1986 study has frequently failed to replicate, and although it finds an effect, the 2015 study has n = 28. As always with psychology studies, we would do well not to assert their purported findings as facts, as with the statement "The brand effect is so strong that displaying a Visa logo near checkout literally increases consumer spend". Psychology as a field is far too unreliable to make such assertions with confidence.
> looks like a vendor, and we have a group now doing a post-mortem trying to figure out how it happened. It'll be removed ASAFP
> Understood. Not trying to sweep under rugs, but I also want to point out that everything is moving very fast right now and there’s 300,000 people that work here, so there’s probably be a bunch of dumb stuff happening. There’s also probably a bunch of dumb stuff happening at other companies
> Sometimes it’s a big systemic problem and sometimes it’s just one person who screwed up
This excuse is hollow to me. In an organization of this size, it takes multiple people screwing up for a failure to reach the public, or at least it should. In either case -- no review process, or a failed review process -- the failure is definitionally systemic. If a single person can on their own whim publish not only plagiarised material, but material that is so obviously defective at a single glance that it should never see the light of day, that is in itself a failure of the system.
It actually might. Coca Cola had $48b revenue last year, or in other words, 4800 millions. Spending 7 of those millions to put your product in front of 100 million people seems like a reasonable bet. If even a couple percent of those people are (sub)consciously influenced to pick up a 12-pack the next time they stop by a store when they might otherwise not have, it would likely be a profitable endeavour given the profit margins on their sugar water.
I think there's also a longer-term status play at stake. If only one of Coca Cola or Pepsi engaged in flashy advertising to this degree, it might give them a slight edge in status perception. In the long term, even an 0.1% shift in consumer preferences between Coca Cola or Pepsi would shift significantly more than 7 million in value. So if one of them engages in this, the other is obliged to follow, in a classic prisoner's dilemma. At any rate, given that 4800 millions in annual revenue translates to 13 million in sales per day, the number paid for that advertisement is a rounding error and doesn't have to move the needle very much at all to be successful.
This is doubly true with anything that involves your real name. I don't think enough people understand that there is a genuine, non-zero risk that commenting on political matters with your real name risks getting you imprisoned or killed 5, 10, 20 years later. The internet is a permanent record for your views, and history has given us many examples of dictators who have purged millions of people deemed intellectuals, enemies of the regime, etc. This is true regardless of how much you trust your current government, because governments change and authoritarianism can happen anywhere. Even the self-proclaimed bastion of free speech is currently attempting to persecute people criticising the government online[1].
[1] https://www.nytimes.com/2026/02/13/technology/dhs-anti-ice-s...
(...that said, the LLM spam situation is so dire that even with these views on anonymity, I would be willing to create a new account where I carefully self-censor everything I say until I meet whatever proposed threshold would be necessary to submit a ShowHN, so I'm not opposed to such a limit out of necessity.)
I literally did provide a definition and my argument for it already: https://news.ycombinator.com/item?id=47051523
If you want to argue with that definition of intelligence, or argue that LLMs do meet that definition of intelligence, by all means, go ahead[1]! I would have been interested to discuss that. Instead I have to repeat myself over and over restating points I already made because people aren't even reading them.
> Not even that current models are not; you seem to be claiming that they cannot be.
As I have now stated something like three or four times in this thread, my position is that machine intelligence is possible but that LLMs are not an example of it. Perhaps you would know what position you were arguing against if you had fully read my arguments before responding.
[1] I won't be responding any further at this point, though, so you should probably not bother. My patience for people responding without reading has worn thin, and going so far as to assert I have not given an argument for the very first thing I made an argument for is quite enough for me to log off.
Good thing I didn't make that claim!
> Ignoring refutations you don't like doesn't make them wrong.
They didn't make a refutation of my points. They asserted a basic principle that I agreed with, but assume acceptance of that principle leads to their preferred conclusion. They make this assumption without providing any reasoning whatsoever for why that principle would lead to that conclusion, whereas I already provided an entire paragraph of reasoning for why I believe the principle leads to a different conclusion. A refutation would have to start from there, refuting the points I actually made. Without that you cannot call it a refutation. It is just gainsaying.
> Any definition of intelligence that does not axiomatically say "is human" or "is biological" or similar is something a machine can meet, insofar as we're also just machines made out of biology.
And here we go AGAIN! I already agree with this point!!!!!!!!!!!!!!! Please, for the love of god, read the words I have written. I think machine intelligence is possible. We are in agreement. Being in agreement that machine intelligence is possible does not automatically lead to the conclusion that the programs that make up LLMs are machine intelligence, any more than a "Hello World" program is intelligence. This is indeed, very repetitive.
They still make these errors on anything that is out of distribution. There is literally a post in this thread linking to a chat where Sonnet failed a basic arithmetic puzzle: https://news.ycombinator.com/item?id=47051286
> Good thing LLMs can handle this just fine I guess.
LLMs can match an example at exactly that trivial level because it can be predicted from context. However, if you construct a more complex example with several rules, especially with rules that have contradictions and have specified logic to resolve conflicts, they fail badly. They can't even play Chess or Poker without breaking the rules despite those being extremely well-represented in the dataset already, nevermind a made-up set of logical rules.
Edit: Case in point, a mere 10 minutes later we got someone making that exact argument in a sibling comment to yours! Nature is beautiful.
Intelligence is the ability to reason about logic. If 1 + 1 is 2, and 1 + 2 is 3, then 1 + 3 must be 4. This is deterministic, and it is why LLMs are not intelligent and can never be intelligent no matter how much better they get at superficially copying the form of output of intelligence. Probabilistic prediction is inherently incompatible with deterministic deduction. We're years into being told AGI is here (for whatever squirmy value of AGI the hype huckster wants to shill), and yet LLMs, as expected, still cannot do basic arithmetic that a child could do without being special-cased to invoke a tool call.
Our computer programs execute logic, but cannot reason about it. Reasoning is the ability to dynamically consider constraints we've never seen before and then determine how those constraints would lead to a final conclusion. The rules of mathematics we follow are not programmed into our DNA; we learn them and follow them while our human-programming is actively running. But we can just as easily, at any point, make up new constraints and follow them to new conclusions. What if 1 + 2 is 2 and 1 + 3 is 3? Then we can reason that under these constraints we just made up, 1 + 4 is 4, without ever having been programmed to consider these rules.[1] Although, of course, there is the question of reliability. For example, the "Boy Scouts" page says Boy Scouts have 2 mentions, but has references to 3 diary entries! Also, on further examination, Sep 1931 has broken dates (meaning my previous theory about it breaking only after Jul 1941 was wrong), and some pages appear to be out of order.
Low effort, minimal change suggestion: a link or table of contents header at the top of each month's page to jump to the diary entries.
Higher effort, bigger change suggestion: I think it would make for a significantly better reading experience if all of the diary pages and their transcriptions for a month were listed sequentially, such that you could seamlessly read them without clicking previous/next page.
I think it's a bit of a waste to have put so much effort into preserving this, but the actual ability to read it is de-prioritised relative to the ability to read an LLM summary.
- I think it would be a very large improvement if the actual diary pages/transcriptions were more accessible. I found the LLM summaries completely uncompelling, and did not particularly appreciate having to scroll through 5+ pages of LLM summary to get to the part where I could actually read the diary entries for a given month.
- The dates of the diary entries for many months are broken. For example, in the final month, all of the entries are labelled 1945-03-19. From a cursory examination, I believe the dating broke 24th July 1941 and was broken for every month from there to the end.
- The page for Nov 1941 seems entirely broken. For some reason, the dates labelling the pages are described in a different format that included the name of the month rather than a numeric representation, the pages are out of order, and then all manner of months are mixed in. The first pages are "November 1941", "April 1941", "October 2 1941", "October 3 1941", "November 4 1941", "November 12 1941", "November 7 1941" ... and so on. The LLM summary notes an "Event", a construction project that took place from 1931 to 1934, despite this being the entry for Nov 1941.
"Free speech" is some kind of terminal brain worm that begs itself to be invoked to browbeat most anybody into submission, I suppose. Now we're apparently extending "free speech" to mean "the government must publish sensitive private information about every citizen in an easily-scraped manner". Well, I don't buy your cheap rhetorical trick. I support freedom of speech in exactly the scope it was originally intended, that is, the freedom to express ideas without facing government censorship or retaliation. Trying to associate your completely unrelated argument with something that everybody is expected to agree with by default is weak.
Nor is it "an attack on journalism". If a real journalist still exists in the current year, they can do investigative work to obtain information that is relevant to the public's interests and then publish it freely. Nobody is stopping them.
> LOCK DOWN ALL INFORMATION ABOUT CRIMINALS
Notably, people who aren't criminals may find themselves in court to determine whether or not they aren't. Unfortunately, people like yourself and the entirety of the broadcast and print media then go on to presume every person who goes to court are criminals and do everything in their power to ruin their lives. Far from just "some musician", most people who are arrested on serious charges get a black mark on their record that effectively destroys their careers and denies them the ability to rent property outside of a ghetto, as employers and landlords discriminate against them baselessly even after they are acquitted of all charges against them.
Absolutely fucking crazy that you typed this out as a legitimate defense of allowing extremely sensitive personal information to be scraped.
> only system that could tell journalists what was actually happening in the criminal courts
Who cares? Journalism is a dead profession and the people who have inherited the title only care about how they can mislead the public in order to maximize profit to themselves. Famously, "journalists" drove a world-renowned musician to his death by overdose with their self-interest-motivated coverage of his trial[1]. It seems to me that cutting the media circus out of criminal trials would actually be massively beneficial to society, not detrimental.
[1] https://www.huffpost.com/entry/one-of-the-most-shameful_b_61...