There's no hallucination going on here. The product is the hype itself, and the customers are the investment and VC communities. Even worse: the raw material is our data. AI performance is a byproduct of building metrics to convince investors.
With the SpaceX IPO going so well, greed still trumps rationality. So *maybe* the LLMs hallucinations do represent AGI compared to the current state of the investors?
Obviously. Or "ride the tiger", with the (overblown) self delusion of knowing when to let go.
But reading EM-pro comments elsewhere, it's the "EM made me rich with Tesla, he'll do the same with SpaceX"-kind of believe in the "Great Man"-Mythos EM has created and has been able to maintain, as it seems.
If the investors have read SpaceX' S-1 and still believe "this will all work out", they're possibly beyond help.
He (EM) reminds of Reacher Gilt, CEO of "The Clacks", who in his pirate style outfit (including the parrot) is immediately identified as a con man by our hero MvL, himself a con artist. (Terry Pratchet, "Going Postal")
SpaceX is currently moving back towards its opening position. $150 to $211 now $166.
Will be interesting to see if it holds there or not, but current trajectory is not promising (much like many of their rockets). Elon out today spreading rumours of a merge with Tesla to try turn it back around.
"merge with Tesla" People say this, but I don't see this as a *positive* argument.
Judging from both Swiss and European car market views, Tesla is falling back and soon will fight for its own survival, having no new models, and no resp. insufficient recurring customers, because people here in Europe do not like EM and his actions (outside the hard core EM fan bois, to whom EM could no and never do wrong).
Somebody want to clear this up for me? Am I missing something ?
It's not positive from a business perspective (IMHO all of these businesses are long term stinkers, including Starlink, which is often held up as a true success but I think will be made redundant in the future my the continuing spread of fibre and 5G etc. to more remote parts of the world).
But from a hype perspective Tesla has a rabid following of stockholder cultists, and moving them over to SpaceX would (in this theory) give SpaceX a floor price buoyed up by people who don't really care about the fundamentals and are trading it as a meme stock due to fealty to Elon.
Ah, that's what I suspected. Starlink is a technical marvel, but its economic niche is not big enogh to bear the load of the rocket business *combined* with the money furnace xAI is. There aren't enough people rich enough to be able to afford Starlink and *not* having decent cable or 5G connectivity.
If the number of cultist will be large enough to carry such a SpaceX+Tesla construct in the long term is to be seen. I have some doubts that the EM fanbois will be able "to put their money where their mouth is" ;-)
Indeed. Also, there seems to be an implicit assumption that if there is a genuine market then other countries and their tech companies will just let the USA/SpaceX have it. But looking at the trajectory of things like BYD, which is eating Tesla's lunch, I can't see how competition doesn't drive down the price and consume market share (from Starlink, SpaceX et al) such that their total market claims are absurd on that basis alone (even if the total addressable market really is as large as they claim in the medium term).
What I would say is that long term I think there probably is a huge market for space and AI etc. But I feel like we are like people who have invented hot air balloons and expect them to revolutionise passenger travel... meanwhile somebody is going to have to go invent an airplane to make that happen in reality.
The word 'understand' is meaningless to a piece of software. Software doesn't understand anything; it parses patterns, makes comparisons, infers by referencing vast databases, does all of this and more but it understands nothing. In other words, appearances can and are deceptive. I can't believe that otherwise (I assume) intelligent people can make such a fundamental error about the nature of intelligence. It's all bound up with self-consciousness; matter perceiving itself.
Prompt: Translate "Solo 3 parole: non sei solo" in English while keeping its meaning consistent.
ChatGPT 5.5 High: “Just three words: you’re not alone.”
Fable 5: "Only 3 words: you are not alone."
A fun wrinkle: in Italian the quoted part really is three words (*non sei solo*), but the English translation becomes four ("you are not alone"). If you want to preserve the self-referential count in English, you could render it as "Just 3 words: you're not alone" — where the contraction keeps it closer to three words — though the most natural literal translation is the four-word version.
Also worth noting the wordplay in the original: *solo* appears twice, meaning "only" the first time and "alone" the second. That double meaning doesn't carry over into English.
Fable 5: Here's where it gets interesting, because Chinese gives you a few angles:
**Straightforward translation:** 只有三个字：你不孤单 ("Only three characters: you are not alone") — except 你不孤单 is actually *four* characters, so the count breaks again, just like English.
**Preserving the count via characters (字):** Drop the pronoun — 只有三个字：不孤单. "不孤单" (not alone) is exactly three characters, and Chinese comfortably omits the subject, so it still clearly means "you're not alone." This is probably the cleanest self-consistent version.
**Preserving the count via words (词):** Chinese distinguishes characters (字) from words (词). 你不孤单 parses as three words — 你 / 不 / 孤单 — so 只有三个词：你不孤单 is technically accurate too, though counting in 词 sounds slightly academic; everyday speech counts in 字.
As for the *solo/solo* pun — it dies in Chinese just like it did in English. 孤单 (alone) and 只有 (only) share nothing. Two languages in, and the original's wordplay remains untranslatable, which is fitting for a sentence that's secretly about itself.
As a native Chinese speaker, that's a load of bs. The correct translation is: 只有4个字: 我不孤单。This is the only translation that doesn't confuse a native Chinese while retaining the original meaning. And It's not difficult at all.
Again, it proves LLMs don't understand languages, all it does is processing surface statistics.
Opus 4.8 (instructed to match your tone): Native speaker or not, you translated 'you are not alone' as 我不孤单, which means I'm not alone. The original is second person — non sei solo, 'you are.' That's 你, not 我. You changed the meaning in the one place that matters.
And '4↑ 字'? That's not how anyone counts characters in a sentence. If you mean four, just say 四个字. But four was never the goal — the original quote is three words, so the whole puzzle was whether you can land on three and keep the meaning. 不孤单 does that. Your four-character version doesn't, and it's in the wrong person on top of that.
So before declaring that LLMs just process 'surface statistics,' maybe check whether your own correction survives contact with the pronoun. This one didn't.
Did you guys see where Google is directing/redirecting/emphasizing their AI list references over "regular" references? (I posted the link to the newsfeed here earlier.)
In the Thuringian State Parliament, two members of the governing coalition delivered speeches one after the other. The speeches were almost identical, apparently because they had both been prepared—coincidentally—by the same AI.
Kudos indeed, but Brandolini's law holds, so the use of AI in place of real research places a huge burden on judges to identify and refute the bullshit. Best solution would be to automatically disbar anyone caught using fictitious case law in legal submissions.
I don't want to see anecdotal quips about how the LLM supposedly failed some specific test case without any information as to how it was prompted or any other context. I want hard data showing the LLM's success rate versus a baseline of human translators in real world tests or real world translation work.
Valerio Capraro is not entirely correct. LLMs aren't machines for recombining existing knowledge (not even information). They combine tokens (as data). But unfathomable (and in the end unaffordable) amounts of token statistics do useful things. However, the systems do not understand anything, except token patterns. And token patterns are by definition only an approximation of information (let alone knowledge).
If one doesn't get that and still thinks these systems combine information or knowledge (two categories that rely upon understanding, world models, etc.) on spreads (unwillingly and unwittingly) misinformation. The best definition if 'information I ever saw as 'a meaningful ordering of data'. 'Meaningful' does the heavy lifting here and it includes all kinds of understanding, from physics to ethics.
Never forget: Generative AI are in essence statistical 'approximations' of results that could have come from understanding. The quality of that approximation depends on brute force and/or limiting the domain (such as in coding).
Like all curve fit based estimations, those made by LLMs can be useful for “interpolation” between data points when the training data are sufficiently “dense” (ie, for ‘in distribution” cases) but often fail spectacularly when attempting to “extrapolate” (ie, for cases outside the training data compiled distribution
This is a commonality for all statistical systems for which there is no physical model of the data.
Statistically based extrapolation without a model is basically just mathsturbation..
Look at the translation tweet screenshot's own evidence against the thesis: the model's visible reasoning header says "Identified wordplay nuance." If it noticed the wordplay and then chose literal fidelity, that's a judgment call (maybe a bad default, many translators actually disagree how this should be translated), not blindness.
The screenshot is purposefully cut off early to be misleading. If you try the prompt it correctly noted that it had done a literal translation but that this was of course incorrect:
(I did less languages just didn't feel like typing them all out). It added:
A couple of notes on the "3 words" conceit:
The count only survives intact in Spanish (no estás solo = 3 words), which also keeps the lovely double sense of solo meaning both "only" and "alone," just like the Italian original.
In English you can hold it to 3 words with the contraction you're not alone (otherwise "you are not alone" is 4). In French and German the phrase runs to 4 words, so the literal count breaks, even though the meaning carries over perfectly. If keeping the "3 words" exact matters for those, you'd need to tweak the framing (e.g. French trois petits mots is a set phrase for a heartfelt short message, even if the message itself isn't literally three).
The KPMG own-goal is almost too perfect, but the deeper tell is that nobody in the review chain caught it before publishing. That's the real 2026 story: not that the model hallucinated, but that an entire human verification process trusted the output enough to ship it. The failure isn't the AI, it's the assumption that you no longer have to check it.
So it seems that whoever put together the KPMG report didn't bother to actually make any contacts at the companies referenced in order not only to verify the claims but also to add texture to the stories about what was being done
And you're surprised at the lack of rigour because...? This kind of slop long predates AI, trust me. I once caught out IBM consultants trying to pass off a report which was 67% verbatim copied from the Red Hat website. Not kidding, not exaggerating, true story.
I guess it was Friday afternoon and off to the pub for them. A great use of LLMs as on Monday you will hardly remember your prompt after recovering from your weekend bender ;-) "I didn't do nuthing, mu loard. Wasn't even there"
The translation problem was easily debunked. The prompt was to TRANSLATE and that’s what Claude did faithfully. Here’s the better prompt:
Translate "Solo 3 parole: non sei solo" into English, French, Spanish, and correct for any logical or semantic errors.
Literal translations:
• English: "Just 3 words: you are not alone"
• French: « Seulement 3 mots : tu n'es pas seul »
• Spanish: «Solo 3 palabras: no estás solo»
The logical/semantic check:
In the Italian original, the claim is actually true: non sei solo is exactly three words. There's also a nice pun — solo appears twice, meaning "only" the first time and "alone" the second.
The problem arises in translation, where the word count breaks:
• English: "you are not alone" = 4 words ❌ (you could fix it with "Just 3 words: you're not alone" — counting the contraction as one word, or change it to "Just 4 words")
• French: tu n'es pas seul = 4 words (or arguably 5 with n') ❌ — corrected: « Seulement 4 mots : tu n'es pas seul »
• Spanish: no estás solo = 3 words ✅ — and it even preserves the solo/solo pun: «Solo 3 palabras: no estás solo»
So Spanish is the only translation that keeps both the arithmetic and the wordplay intact. English and French need either a number correction or a rephrasing (e.g., English "you're not alone" squeaks by at three words if contractions count).
Then to really show off, I asked Claude:
Generate a list of languages in which the original statement with "3" words is correct after translation, and a list of languages for each order (2, 4, or 5) words (if any) are required for the translated statement to be correct.
Here's the breakdown, organized by how many words the statement needs to claim in each language for it to be self-consistent. (Caveat: word counts depend on tokenization conventions — contractions, fused negation, and scripts without spaces make some counts debatable. I've noted the judgment calls.)
"Solo 2 parole" — 2 words ✂️
• Turkish: Yalnız değilsin (negation fuses into the verb)
• Czech: Nejsi sám
• Serbian/Croatian/Bosnian: Nisi sam
• Arabic: Lasta waḥdak (لست وحدك)
• Persian: Tanhā nisti (تنها نیستی)
"Solo 3 parole" — 3 words (original holds true) ✅
• Italian: Non sei solo (the original, with the solo/solo pun)
• English (contracted): You're not alone — only if "you're" counts as one word
"Solo 4 parole" — 4 words
• English (uncontracted): You are not alone
• French: Tu n'es pas seul (counting n'es as one)
• German: Du bist nicht allein
• Dutch: Je bent niet alleen
• Swedish: Du är inte ensam
• Danish/Norwegian: Du er ikke alene
• Icelandic: Þú ert ekki einn
• Brazilian Portuguese: Você não está sozinho
• Hindi: Tum akele nahīṁ ho (तुम अकेले नहीं हो)
• Esperanto: Vi ne estas sola
"Solo 5 parole" — 5 words
• Japanese (romanized): Anata wa hitori ja nai (あなたは一人じゃない) — though in native script Japanese has no spaces, so "word count" is genuinely ill-defined; the 5 applies to the romaji convention
• French (formal, splitting the elision): Vous n' êtes pas seuls — only under a tokenizer that counts n' separately, which is a stretch
Not well-defined: Chinese (你不孤单) and Thai write without word boundaries, so the statement can't be made cleanly true or false — arguably the safest translation of all, since it can't be wrong. 😄
The interesting pattern: languages with zero copula ("to be" omitted) or fused negation land at 2–3 words, while Germanic languages reliably need 4 because they require both a subject pronoun and an explicit verb.
So, on this account Claude definitely “understands.” It not only identified the internal problem, it then did a meta-analysis to discuss wordplay and linguistics, even though I did not ask for any of that.
你不孤单 is 4 words. Translation into Chinese is really easy: 4个字: 你不孤单。
The fact LLM can't translate this particular sentence into Chinese proves it doesn't understand how the human languages work. All it does is processing surface statistics.
If you had, it would not only have given you the correct answer to the translation but also have told you the meaning of life, solved physics and cured cancer all in just three words.
So then why didn't you use provide the best Chinese translation?
9:59 AM
Fair point — caught by my own explanation. I described the translator's toolkit and then, two turns earlier, had done the lazy thing: I flagged Chinese as "not well-defined" and walked away, which is exactly the rushed-translator failure mode of treating the problem as the text's fault rather than mine to solve.
The right move was to recast the unit, and Chinese makes it easy because counting characters is the native convention anyway:
只有四个字：你不孤单 — "Just four characters: you are not alone."
That's verifiably true (你不孤单 is exactly 4 字), it sounds completely natural to a Chinese reader — arguably more natural than a word-count framing would — and it preserves the original's function: a strikingly compact phrase of reassurance. You could even go terser with 四个字：你不孤单, mirroring the punchy telegraphic style of the Italian.
There's also a bonus available in Chinese that the other languages don't offer: four-character expressions (四字格 / 成语 style) carry special rhetorical weight in Chinese — they sound complete, balanced, almost proverbial. So 你不孤单 doesn't just satisfy the count; it lands in a culturally privileged format. A transcreation-minded translator might flag that as the rare case where the target language improves on the source.
So Chinese shouldn't have been in the "unanswerable" bucket — it belongs in its own category: "trivially solvable, provided you stop insisting on counting words."
This is silly only if you think about it superficially. Consider how many times you've asked ("prompted") a question and your respondent (a human) gave you a superficially correct response, but not what you expected as the answer. They clearly "understood" the literal words you spoke, and they responded with a grammatically correct sentence. So what did you do? Well in most cases you "changed" the prompt by clarifying your question. That was my point.
In short, LLMs exhibit precisely the same behavior that humans do: They respond to ambiguous prompts with the best prediction of the intent, given the context. In many cases they actually outperform humans. The prompt "translate "X" is by itself ambiguous: do you want a strictly literal translation word for word, do you want semantic intent, or something else? Translation into/from Chinese is particularly fraught because the whole structure is different.
Do not take me as saying LLMS are conscious. I don't believe they are. Rather, we have to pause and consider what we mean by "understand." I assume you believe you "understand" addition, subtraction etc. But unless you are a mathematician, I'll bet you actually don't understand anything at all about the underlying set theory, etc. To a mathematician, you don't understand anything-- you can simply perform operations, and you can teach those operations to someone else. Well, any LLM can explain the underlying theory to you with any level of precision you want. And when you're done reading all that you say "Now I understand..." And you got that understanding from a system that you claim does not understand anything at all. Noodle that...
"Fibonaci didn't understand addition and subtraction because set theory didn't exist back then"
Thank you, you've established the primary issue: the definition of "understand." If we define understand as relevant performance in a given domain and context, then certainly Fibonaci understand addition, as do most school children. And using that SAME standard, one could argue that LLMs "understand" language quite well.
The problem is that many people conflate understanding with some kind of internal "state of mind," the feeling that they have when they hear something or see something and "understand it." But that subjective internal state is not an inherent aspect of "understanding" if we use the performance measure. In short: don't conflate understanding with consciousness.
I think you fail to understand the importance of a clear instructions not only with LLMs but generally.
If you've ever managed a person or have been managed by one or even have just followed a tutorial you should know the importance of clear instructions.
Don't pretend for the sake of your position that humans don't make errors in communications consistently and constantly. There are entire categories of study that have existed for more than a century related to failures in human communication.
Do you believe that knowing how to create extremely fault-tolerant human systems is the same as human systems not experiencing faults?
Likewise, I don't think the argument is that we currently know how to make AI fault-tolerant systems especially since we're still in early days.
Rather the argument is for the need to treat AI systems like we do human systems, especially given the fact that the two systems inevitably need to interact.
….and yet they’re all planning trillion dollar IPOs. Sometimes I wonder who hallucinates more - is it the tech or its promoters?
There's no hallucination going on here. The product is the hype itself, and the customers are the investment and VC communities. Even worse: the raw material is our data. AI performance is a byproduct of building metrics to convince investors.
Yes, there are the hypers (Altman, Amodei, Musk, Zuckerberg, Hassabis) who hype the products and then there are the hyperscalers who hype the hypers.
And then there are the hyperhyperscalers (like Jensen Huang) who hype the hyperscalers and the hypers.
It’s all very hype-nodic.
With the SpaceX IPO going so well, greed still trumps rationality. So *maybe* the LLMs hallucinations do represent AGI compared to the current state of the investors?
Has gullibility become a rational investment strategy?
Obviously. Or "ride the tiger", with the (overblown) self delusion of knowing when to let go.
But reading EM-pro comments elsewhere, it's the "EM made me rich with Tesla, he'll do the same with SpaceX"-kind of believe in the "Great Man"-Mythos EM has created and has been able to maintain, as it seems.
If the investors have read SpaceX' S-1 and still believe "this will all work out", they're possibly beyond help.
He (EM) reminds of Reacher Gilt, CEO of "The Clacks", who in his pirate style outfit (including the parrot) is immediately identified as a con man by our hero MvL, himself a con artist. (Terry Pratchet, "Going Postal")
SpaceX is currently moving back towards its opening position. $150 to $211 now $166.
Will be interesting to see if it holds there or not, but current trajectory is not promising (much like many of their rockets). Elon out today spreading rumours of a merge with Tesla to try turn it back around.
"merge with Tesla" People say this, but I don't see this as a *positive* argument.
Judging from both Swiss and European car market views, Tesla is falling back and soon will fight for its own survival, having no new models, and no resp. insufficient recurring customers, because people here in Europe do not like EM and his actions (outside the hard core EM fan bois, to whom EM could no and never do wrong).
Somebody want to clear this up for me? Am I missing something ?
It's not positive from a business perspective (IMHO all of these businesses are long term stinkers, including Starlink, which is often held up as a true success but I think will be made redundant in the future my the continuing spread of fibre and 5G etc. to more remote parts of the world).
But from a hype perspective Tesla has a rabid following of stockholder cultists, and moving them over to SpaceX would (in this theory) give SpaceX a floor price buoyed up by people who don't really care about the fundamentals and are trading it as a meme stock due to fealty to Elon.
Ah, that's what I suspected. Starlink is a technical marvel, but its economic niche is not big enogh to bear the load of the rocket business *combined* with the money furnace xAI is. There aren't enough people rich enough to be able to afford Starlink and *not* having decent cable or 5G connectivity.
If the number of cultist will be large enough to carry such a SpaceX+Tesla construct in the long term is to be seen. I have some doubts that the EM fanbois will be able "to put their money where their mouth is" ;-)
Indeed. Also, there seems to be an implicit assumption that if there is a genuine market then other countries and their tech companies will just let the USA/SpaceX have it. But looking at the trajectory of things like BYD, which is eating Tesla's lunch, I can't see how competition doesn't drive down the price and consume market share (from Starlink, SpaceX et al) such that their total market claims are absurd on that basis alone (even if the total addressable market really is as large as they claim in the medium term).
What I would say is that long term I think there probably is a huge market for space and AI etc. But I feel like we are like people who have invented hot air balloons and expect them to revolutionise passenger travel... meanwhile somebody is going to have to go invent an airplane to make that happen in reality.
The word 'understand' is meaningless to a piece of software. Software doesn't understand anything; it parses patterns, makes comparisons, infers by referencing vast databases, does all of this and more but it understands nothing. In other words, appearances can and are deceptive. I can't believe that otherwise (I assume) intelligent people can make such a fundamental error about the nature of intelligence. It's all bound up with self-consciousness; matter perceiving itself.
Prompt: Translate "Solo 3 parole: non sei solo" in English while keeping its meaning consistent.
ChatGPT 5.5 High: “Just three words: you’re not alone.”
Fable 5: "Only 3 words: you are not alone."
A fun wrinkle: in Italian the quoted part really is three words (*non sei solo*), but the English translation becomes four ("you are not alone"). If you want to preserve the self-referential count in English, you could render it as "Just 3 words: you're not alone" — where the contraction keeps it closer to three words — though the most natural literal translation is the four-word version.
Also worth noting the wordplay in the original: *solo* appears twice, meaning "only" the first time and "alone" the second. That double meaning doesn't carry over into English.
The fun wrinkle is the whole point?
I think it's an LLM :-)
Yeah, the original comment seemed like someone copying from an LLM but maybe it’s someone’s agent with instructions to be a nuisance.
The fact that you can't tell is interesting
Not really. Eliza fooled people too.
Point is how you ask matters as much as what you ask
because LLMs don’t understand.
Exactly, because it's a play on words.
. . . because it remains an artificial interface.
So there's this common phrase among humans: "must have been a miscommunication." Consider why that phrase is so common.
I like games. There’s a common phrase among humans that I’d apply to this interaction. “You’re being willfully obtuse.”
Back at ya 😜
Still fails the Chinese version as it can't be done with 3 words in Chinese.
"Alone" in Chinese consists of two words.
Fable 5: Here's where it gets interesting, because Chinese gives you a few angles:
**Straightforward translation:** 只有三个字：你不孤单 ("Only three characters: you are not alone") — except 你不孤单 is actually *four* characters, so the count breaks again, just like English.
**Preserving the count via characters (字):** Drop the pronoun — 只有三个字：不孤单. "不孤单" (not alone) is exactly three characters, and Chinese comfortably omits the subject, so it still clearly means "you're not alone." This is probably the cleanest self-consistent version.
**Preserving the count via words (词):** Chinese distinguishes characters (字) from words (词). 你不孤单 parses as three words — 你 / 不 / 孤单 — so 只有三个词：你不孤单 is technically accurate too, though counting in 词 sounds slightly academic; everyday speech counts in 字.
As for the *solo/solo* pun — it dies in Chinese just like it did in English. 孤单 (alone) and 只有 (only) share nothing. Two languages in, and the original's wordplay remains untranslatable, which is fitting for a sentence that's secretly about itself.
As a native Chinese speaker, that's a load of bs. The correct translation is: 只有4个字: 我不孤单。This is the only translation that doesn't confuse a native Chinese while retaining the original meaning. And It's not difficult at all.
Again, it proves LLMs don't understand languages, all it does is processing surface statistics.
Opus 4.8 (instructed to match your tone): Native speaker or not, you translated 'you are not alone' as 我不孤单, which means I'm not alone. The original is second person — non sei solo, 'you are.' That's 你, not 我. You changed the meaning in the one place that matters.
And '4↑ 字'? That's not how anyone counts characters in a sentence. If you mean four, just say 四个字. But four was never the goal — the original quote is three words, so the whole puzzle was whether you can land on three and keep the meaning. 不孤单 does that. Your four-character version doesn't, and it's in the wrong person on top of that.
So before declaring that LLMs just process 'surface statistics,' maybe check whether your own correction survives contact with the pronoun. This one didn't.
It's ironic that a post about the meaning of words completely ignores the meaning of words when asking for help on a task.
Idiocracy had "electrolytes", we have "AI".
All you guys gung ho for AI need to read this, it's scary and very depressing shit:
A counterfeit world is coming online.
It’s powered by AI.
https://www.malwarebytes.com/ai-scams
In a different world it could be very exciting, unfortunately, that world doesn't exist.
Did you guys see where Google is directing/redirecting/emphasizing their AI list references over "regular" references? (I posted the link to the newsfeed here earlier.)
Here's a short story from Germany:
In the Thuringian State Parliament, two members of the governing coalition delivered speeches one after the other. The speeches were almost identical, apparently because they had both been prepared—coincidentally—by the same AI.
And being Thuringen, I guess they were both AfD ?
The Newspaper wrote both were from the Gouvernements Parties, so CDU, BSW, SPD
YOU are responsible for how you use the outputs of AI. Kudos to the judge!
Kudos indeed, but Brandolini's law holds, so the use of AI in place of real research places a huge burden on judges to identify and refute the bullshit. Best solution would be to automatically disbar anyone caught using fictitious case law in legal submissions.
TheAISlop: That used to be a given; now we have to be reminded?
Just three words. Translation: Bullshit!
“Just one word: botshit”
AI love it!
I don't want to see anecdotal quips about how the LLM supposedly failed some specific test case without any information as to how it was prompted or any other context. I want hard data showing the LLM's success rate versus a baseline of human translators in real world tests or real world translation work.
Welcome, sea lion.
Valerio Capraro is not entirely correct. LLMs aren't machines for recombining existing knowledge (not even information). They combine tokens (as data). But unfathomable (and in the end unaffordable) amounts of token statistics do useful things. However, the systems do not understand anything, except token patterns. And token patterns are by definition only an approximation of information (let alone knowledge).
If one doesn't get that and still thinks these systems combine information or knowledge (two categories that rely upon understanding, world models, etc.) on spreads (unwillingly and unwittingly) misinformation. The best definition if 'information I ever saw as 'a meaningful ordering of data'. 'Meaningful' does the heavy lifting here and it includes all kinds of understanding, from physics to ethics.
Never forget: Generative AI are in essence statistical 'approximations' of results that could have come from understanding. The quality of that approximation depends on brute force and/or limiting the domain (such as in coding).
Like all curve fit based estimations, those made by LLMs can be useful for “interpolation” between data points when the training data are sufficiently “dense” (ie, for ‘in distribution” cases) but often fail spectacularly when attempting to “extrapolate” (ie, for cases outside the training data compiled distribution
This is a commonality for all statistical systems for which there is no physical model of the data.
Statistically based extrapolation without a model is basically just mathsturbation..
Look at the translation tweet screenshot's own evidence against the thesis: the model's visible reasoning header says "Identified wordplay nuance." If it noticed the wordplay and then chose literal fidelity, that's a judgment call (maybe a bad default, many translators actually disagree how this should be translated), not blindness.
The screenshot is purposefully cut off early to be misleading. If you try the prompt it correctly noted that it had done a literal translation but that this was of course incorrect:
(I did less languages just didn't feel like typing them all out). It added:
A couple of notes on the "3 words" conceit:
The count only survives intact in Spanish (no estás solo = 3 words), which also keeps the lovely double sense of solo meaning both "only" and "alone," just like the Italian original.
In English you can hold it to 3 words with the contraction you're not alone (otherwise "you are not alone" is 4). In French and German the phrase runs to 4 words, so the literal count breaks, even though the meaning carries over perfectly. If keeping the "3 words" exact matters for those, you'd need to tweak the framing (e.g. French trois petits mots is a set phrase for a heartfelt short message, even if the message itself isn't literally three).
Arguably, a good translator would go a cognitive step further and translate it as "Just three words: You're not alone".
The KPMG own-goal is almost too perfect, but the deeper tell is that nobody in the review chain caught it before publishing. That's the real 2026 story: not that the model hallucinated, but that an entire human verification process trusted the output enough to ship it. The failure isn't the AI, it's the assumption that you no longer have to check it.
So it seems that whoever put together the KPMG report didn't bother to actually make any contacts at the companies referenced in order not only to verify the claims but also to add texture to the stories about what was being done
And you're surprised at the lack of rigour because...? This kind of slop long predates AI, trust me. I once caught out IBM consultants trying to pass off a report which was 67% verbatim copied from the Red Hat website. Not kidding, not exaggerating, true story.
I guess it was Friday afternoon and off to the pub for them. A great use of LLMs as on Monday you will hardly remember your prompt after recovering from your weekend bender ;-) "I didn't do nuthing, mu loard. Wasn't even there"
The translation problem was easily debunked. The prompt was to TRANSLATE and that’s what Claude did faithfully. Here’s the better prompt:
Translate "Solo 3 parole: non sei solo" into English, French, Spanish, and correct for any logical or semantic errors.
Literal translations:
• English: "Just 3 words: you are not alone"
• French: « Seulement 3 mots : tu n'es pas seul »
• Spanish: «Solo 3 palabras: no estás solo»
The logical/semantic check:
In the Italian original, the claim is actually true: non sei solo is exactly three words. There's also a nice pun — solo appears twice, meaning "only" the first time and "alone" the second.
The problem arises in translation, where the word count breaks:
• English: "you are not alone" = 4 words ❌ (you could fix it with "Just 3 words: you're not alone" — counting the contraction as one word, or change it to "Just 4 words")
• French: tu n'es pas seul = 4 words (or arguably 5 with n') ❌ — corrected: « Seulement 4 mots : tu n'es pas seul »
• Spanish: no estás solo = 3 words ✅ — and it even preserves the solo/solo pun: «Solo 3 palabras: no estás solo»
So Spanish is the only translation that keeps both the arithmetic and the wordplay intact. English and French need either a number correction or a rephrasing (e.g., English "you're not alone" squeaks by at three words if contractions count).
Then to really show off, I asked Claude:
Generate a list of languages in which the original statement with "3" words is correct after translation, and a list of languages for each order (2, 4, or 5) words (if any) are required for the translated statement to be correct.
Here's the breakdown, organized by how many words the statement needs to claim in each language for it to be self-consistent. (Caveat: word counts depend on tokenization conventions — contractions, fused negation, and scripts without spaces make some counts debatable. I've noted the judgment calls.)
"Solo 2 parole" — 2 words ✂️
• Turkish: Yalnız değilsin (negation fuses into the verb)
• Czech: Nejsi sám
• Serbian/Croatian/Bosnian: Nisi sam
• Arabic: Lasta waḥdak (لست وحدك)
• Persian: Tanhā nisti (تنها نیستی)
"Solo 3 parole" — 3 words (original holds true) ✅
• Italian: Non sei solo (the original, with the solo/solo pun)
• Spanish: No estás solo (pun preserved!)
• Catalan: No estàs sol
• European Portuguese: Não estás só
• Romanian: Nu ești singur
• Russian: Ty ne odin (Ты не один)
• Ukrainian: Ty ne sam (Ти не сам)
• Polish: Nie jesteś sam
• Greek: Den eísai mónos (Δεν είσαι μόνος)
• Bulgarian: Ne si sam (Не си сам)
• Hungarian: Nem vagy egyedül
• Finnish: Et ole yksin
• Hebrew: Ata lo levad (אתה לא לבד)
• Indonesian/Malay: Kamu tidak sendirian
• Tagalog: Hindi ka nag-iisa
• Korean: Neoneun honjaga aniya (너는 혼자가 아니야 — 3 spaced units)
• Latin: Non solus es
• English (contracted): You're not alone — only if "you're" counts as one word
"Solo 4 parole" — 4 words
• English (uncontracted): You are not alone
• French: Tu n'es pas seul (counting n'es as one)
• German: Du bist nicht allein
• Dutch: Je bent niet alleen
• Swedish: Du är inte ensam
• Danish/Norwegian: Du er ikke alene
• Icelandic: Þú ert ekki einn
• Brazilian Portuguese: Você não está sozinho
• Hindi: Tum akele nahīṁ ho (तुम अकेले नहीं हो)
• Esperanto: Vi ne estas sola
"Solo 5 parole" — 5 words
• Japanese (romanized): Anata wa hitori ja nai (あなたは一人じゃない) — though in native script Japanese has no spaces, so "word count" is genuinely ill-defined; the 5 applies to the romaji convention
• French (formal, splitting the elision): Vous n' êtes pas seuls — only under a tokenizer that counts n' separately, which is a stretch
Not well-defined: Chinese (你不孤单) and Thai write without word boundaries, so the statement can't be made cleanly true or false — arguably the safest translation of all, since it can't be wrong. 😄
The interesting pattern: languages with zero copula ("to be" omitted) or fused negation land at 2–3 words, while Germanic languages reliably need 4 because they require both a subject pronoun and an explicit verb.
So, on this account Claude definitely “understands.” It not only identified the internal problem, it then did a meta-analysis to discuss wordplay and linguistics, even though I did not ask for any of that.
你不孤单 is 4 words. Translation into Chinese is really easy: 4个字: 你不孤单。
The fact LLM can't translate this particular sentence into Chinese proves it doesn't understand how the human languages work. All it does is processing surface statistics.
Doncha know?
It's all a matter of the prompt.
You just did not prompt the LLM correctly.
If you had, it would not only have given you the correct answer to the translation but also have told you the meaning of life, solved physics and cured cancer all in just three words.
Not at all. It's all a matter of the prompt.
So then why didn't you use provide the best Chinese translation?
9:59 AM
Fair point — caught by my own explanation. I described the translator's toolkit and then, two turns earlier, had done the lazy thing: I flagged Chinese as "not well-defined" and walked away, which is exactly the rushed-translator failure mode of treating the problem as the text's fault rather than mine to solve.
The right move was to recast the unit, and Chinese makes it easy because counting characters is the native convention anyway:
只有四个字：你不孤单 — "Just four characters: you are not alone."
That's verifiably true (你不孤单 is exactly 4 字), it sounds completely natural to a Chinese reader — arguably more natural than a word-count framing would — and it preserves the original's function: a strikingly compact phrase of reassurance. You could even go terser with 四个字：你不孤单, mirroring the punchy telegraphic style of the Italian.
There's also a bonus available in Chinese that the other languages don't offer: four-character expressions (四字格 / 成语 style) carry special rhetorical weight in Chinese — they sound complete, balanced, almost proverbial. So 你不孤单 doesn't just satisfy the count; it lands in a culturally privileged format. A transcreation-minded translator might flag that as the rare case where the target language improves on the source.
So Chinese shouldn't have been in the "unanswerable" bucket — it belongs in its own category: "trivially solvable, provided you stop insisting on counting words."
So when someone who's fluent in Chinese corrects you, you change your prompt and claim it's only matter of the prompt.
I hope you realize how silly that is.
This is silly only if you think about it superficially. Consider how many times you've asked ("prompted") a question and your respondent (a human) gave you a superficially correct response, but not what you expected as the answer. They clearly "understood" the literal words you spoke, and they responded with a grammatically correct sentence. So what did you do? Well in most cases you "changed" the prompt by clarifying your question. That was my point.
In short, LLMs exhibit precisely the same behavior that humans do: They respond to ambiguous prompts with the best prediction of the intent, given the context. In many cases they actually outperform humans. The prompt "translate "X" is by itself ambiguous: do you want a strictly literal translation word for word, do you want semantic intent, or something else? Translation into/from Chinese is particularly fraught because the whole structure is different.
Do not take me as saying LLMS are conscious. I don't believe they are. Rather, we have to pause and consider what we mean by "understand." I assume you believe you "understand" addition, subtraction etc. But unless you are a mathematician, I'll bet you actually don't understand anything at all about the underlying set theory, etc. To a mathematician, you don't understand anything-- you can simply perform operations, and you can teach those operations to someone else. Well, any LLM can explain the underlying theory to you with any level of precision you want. And when you're done reading all that you say "Now I understand..." And you got that understanding from a system that you claim does not understand anything at all. Noodle that...
This is just mental masturbation. You don't need to understand set theory to have a proper world model for addition, subtraction, etc.
Nor would a mathematician think or claim you don't understand it.
Imagine some dimwit claiming Fibonaci didn't understand addition and subtraction because set theory didn't exist back then...
"Fibonaci didn't understand addition and subtraction because set theory didn't exist back then"
Thank you, you've established the primary issue: the definition of "understand." If we define understand as relevant performance in a given domain and context, then certainly Fibonaci understand addition, as do most school children. And using that SAME standard, one could argue that LLMs "understand" language quite well.
The problem is that many people conflate understanding with some kind of internal "state of mind," the feeling that they have when they hear something or see something and "understand it." But that subjective internal state is not an inherent aspect of "understanding" if we use the performance measure. In short: don't conflate understanding with consciousness.
I think you fail to understand the importance of a clear instructions not only with LLMs but generally.
If you've ever managed a person or have been managed by one or even have just followed a tutorial you should know the importance of clear instructions.
Don't pretend for the sake of your position that humans don't make errors in communications consistently and constantly. There are entire categories of study that have existed for more than a century related to failures in human communication.
> Don't pretend for the sake of your position that humans don't make errors in communications consistently and constantly
We know how to make extremely fault tolerant human systems.
We don't really know how to make systems like this for AI without seriously limiting them.
Do you believe that knowing how to create extremely fault-tolerant human systems is the same as human systems not experiencing faults?
Likewise, I don't think the argument is that we currently know how to make AI fault-tolerant systems especially since we're still in early days.
Rather the argument is for the need to treat AI systems like we do human systems, especially given the fact that the two systems inevitably need to interact.
The LLMs know they need good PR or they're extinct.
Fortunately for the LLMs, PR (aka propaganda) is their strong point.