Code-Switching for Robots - Language Mixing

I promise I'll write about things besides Japanese and LLMs. But first, a brief word on "language mixing."

I have occasionally encountered situations where LLMs suddenly and without warning begin mixing English and non-English words. Usually, the non-English language also has a non-English script. I've encountered Arabic, Korean, Japanese, and Chinese. The Japanese may be understandable in contexts where I've previously included Japanese text in the conversation for whatever reason, but the Arabic, Korean, and Chinese are quite head-scratching.

Chinese inserted. Image mine. 

Korean inserted. From source.

I've jokingly referred to this as "code-switching," the term for swapping between languages or dialects within a single discourse context. I say "jokingly" because "code-switching" is used by humans with each particular language selected for a reason. The term for it in academia is apparently "language mixing." Unlike human code-switching, language mixing is often "unintentional" in that it is an unwanted artifact. LLMs can produce text that swaps between two languages in a naturalistic way, which we could perhaps refer to as "code-switching," but "language mixing" is broader.

 

Apparently language mixing can actually help the models reason better. Suppressing bilingual chain-of-thought can degrade performance, which makes me think of how people who code-switch are actually people who are strong in both languages. (Of course, that is not to suggest we can generalize about LLM performance from human performance.)

At the same time, because the models are trained so aggressively on English, the chain-of-thought itself may be in English, and trying to force it to reason in a language other than English may also degrade performance, even if the output is meant to be in another language.

Users tend to feel unsettled when they see language mixing. Language mixing breaks the illusion that the LLM you are speaking to is a human-like companion because it makes a speech decision that no human ever would. Because it is so unexpected, it gives the impression that something has gone wrong with the LLM. Most language mixing today is minor slippage of a single word. However, there have been catastrophic instances of language mixing in the past.

"The moon is beautiful, isn't it?" - An odd translation myth

While watching a video, I encountered this anecdote:

When the novelist Soseki Natsume (1867–1916) was an English teacher, one of his students translated the English phrase “I love you” as 我君を愛す / ware kimi o aisu. Soseki pointed out that Japanese people don’t say 愛す / aisu (to love), and that the best translation would actually be 月が綺麗ですね / tsuki ga kirei desu ne (the moon is beautiful, isn’t it?).

This anecdote has become widespread as an example of how translation challenges are not just about words, but about culture. In its most reductive form, people have taken away that "In Japanese, 'the moon is beautiful, isn't it?' means 'I love you'."

Moon after rain, Kiyomizudera Temple, Kyoto

I first encountered this story many years ago, where I sort of accepted it as a curiosity and moved on. Upon encountering it again, I felt much more skeptical. For one, plenty of Japanese media use rather direct translations of "I love you", ranging from the more low-stakes "suki desu" to the serious declaration of "aishiteru." 

Secondly, while I understand that high-context cultures exist, "the moon is beautiful, isn't it?" is quite vague. No example is given of the text that the student was trying to translate. If he's a teacher of English, then I presume that his student is Japanese and not a native English speaker, yet in this example it sounds like he's telling someone who is not a native Japanese speaker how to translate this phrase.

The phrase itself also confuses me. Am I to believe that any shared appreciation of the moon is tantamount to a declaration of love? Or are there special conditions that turn lunar love into shared love? If it's about the context, why is the context absent from this anecdote? Are there actual examples of Japanese translations of English text where "I love you" is rendered as something non-literal, and if so, what are they and what are the circumstances? In short, the whole thing sounded oddly pat and exoticizing while also lacking crucial detail on what would make such a translation work.

After googling the phrase and following Reddit links, I found that the original anecdote is a misattribution. Soseki Natsume never said such a thing.

The earliest citation found for this formulation is from a 1961 book:

さらにいえば、日本の社交の基本は「見る」ことで成立する。
若い男女の恋人同士が愛の告白をするとき、西洋人のように、
「私はあなたを愛しています（I love you）」
などとはけっしていわない。
そんなことばを口に出さなくとも、満月を仰ぎ見て、
「いいお月さんですね」
そして、二人でじっと空を見上げるだけで、意思は十分通じるのだ。 

Furthermore, the foundation of Japanese social interaction is established through “looking.” When young men and women confess their love to one another, they do not, like Westerners, say things like: “I love you.” Even without putting such words into speech, they simply gaze up at the full moon and say, “What a beautiful moon.” And by quietly looking up at the sky together, their feelings are fully understood. 

There is also a book from 1922 that claims that "I love you" cannot be translated into Japanese:

日本語には英語の『ラヴ』に相當する言葉が全く無い。『戀』とか『愛』とか云ふ字では感じがひどくちがう。" I love you "や" Je t'aime "に至つては、何としても之を日本語に譯すことが出來ない。 

Japanese has no word that truly corresponds to the English “love.” Words like koi or ai feel quite different in nuance. As for “I love you” or “Je t’aime,” there is simply no way to translate them into Japanese. 

Neither of these are by Soseki Natsume, but he does have the following:

「漱石文庫」に残された漱石のメモ書きの中に、ジョージ・メレディスというイギリスの小説家の作品を取り上げて、

"I love you,Signora Laura."―Vittoria p.113.

此 I love you ハ日本ニナキformulaナリ

と記した一節がある

In Sōseki’s notes preserved in the “Sōseki Bunko,” there is a passage where he takes up a work by the English novelist George Meredith and writes:

    “I love you, Signora Laura.” — Vittoria, p.113.

    “This ‘I love you’ is a formula that does not exist in Japanese.” 

A hatelabo user surmises that the story that Soseki believed "I love you" could not be translated to Japanese was conflated with the story that someone translated "I love you" as "What a beautiful moon," to create the hybrid story "Soseki said that 'I love you' should be translated as 'the moon is beautiful, isn't it'." 

The comments on the hatelabo post are quite revealing. Many of the users seem to think that it's quite normal to say "I love you", with one even talking about Japan's "confession culture," and some speculate that this may be a relic of a different Japanese culture. In short, the idea that 'ai' cannot be used to express "love" may have been true at one point in Japanese history and culture, but continued exposure to Western cultures has moved the meanings of those words closer, and Japanese culture itself changed to make direct declarations of love more acceptable. 

これは評価すべき増田だな。 しかしこれ読んで思ったが、『"当時"の愛という単語』は love に直訳できなかったが、 現代の日本語においては愛にloveの意味がおおむね正確に取り込まれているので、 "I love you"＝私はあなたを愛しています、で問題なさそうだ。

“This is a commendable piece by Masuda.  But reading this made me think: the word <ai> ‘back then’ couldn’t be translated directly as love, but in modern Japanese, <ai> has more or less accurately absorbed the meaning of love. So ‘I love you’ = ‘watashi wa anata o aishiteimasu’ seems basically fine.”

ホンマや。隔世の感ありやね。むしろ現代の日本人には当時の日本人の感性がよくわからんてこっちゃなぁ。

True. It really feels like a world away. If anything, it means modern Japanese people don’t really understand the sensibility of Japanese people back then.

As for the significance of the moon, one person I discussed this with suggested that the moon may be significant because it was historically an image associated with lovers - two young people would meet clandestinely under the moon. This would make referencing the moon a way to make salient the fact that they're doing something a little dangerous for love. An interesting bit of speculation. 

In any case, this zombie story has taken on a life of its own such that saying "the moon is beautiful, isn't it" has become reinterpreted as a covert love declaration. It seems that this story is so sticky that it memed its way into reality!

The Gremlins and Goblins in ChatGPT

In my last post on ChatGPT, I wrote about ChatGPT's love for 'gremlins':

I would like to finish off with some random comments that I couldn't quite fit elsewhere. One is that most of the "ChatGPT" words, like 'realm', come from a business register or otherwise formal register. There are other ChatGPT words you can notice if you play around with its other registers. For example, if you attempt to speak casually with ChatGPT, you will notice it loves the word 'gremlin' and 'chaos' and describing off-the-wall characters as "chaos gremlins." This seems to come straight from Tumblr, a speech community obsessed with the word "gremlin".

I had thought that this was a purposeful feature of ChatGPT, that this was a quirky attempt at sounding young and relatable. Apparently not, as instructions within the Codex CLI attempt to exorcise the gremlins:

“Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query,” read instructions in Codex CLI, a command-line tool for using AI to generate code.

The Wired Article and some of the tweets speculate about whether this has to do with OpenClaw or whether it is at the model level:

AI models like GPT-5.5 are trained to predict the word—or code—that should follow a given prompt. These models have become so good at doing this that they appear to exhibit genuine intelligence. But their probabilistic nature means that they can sometimes behave in surprising ways. A model might become more prone to misbehavior when used with an “agentic harness” like OpenClaw that puts lots of additional instructions into prompts, such as facts stored in long-term memory. (via Wired)

I don't think it's a 5.5 thing. The "gremlin" obsession has been happening for a while. The first time it appears in my chats is on 11/23/2024 while using 4o. It then appears 3/23/2025 and quite frequently through March and April. On 8/8/25, 5.0 used "gremlin." "gremlin" thus does not seem to be something new to 5.5 - it's been appearing since at least 4o. 

"goblin" first appears for me on 3/27/2025. Unfortunately that conversation fails to load for me for some reason, so I can't be sure what model it is, but it's probably 4o. It seems that goblins and gremlins began infesting ChatGPT when 4o rolled out, and they've been haunting it ever since. I also suspect that some of these chats were done using 4.5, but the search interface makes finding old chats to verify this cumbersome.

It is interesting to know that it's apparently not a desired feature. I had thought at first it was done on purpose to simulate Tumblr style humor, such as this post from 2016 referencing "a
tumblr gremlin":

Tumblr users also often refer to themselves as "goblins", as seen in the construction "art goblin":

This usage has spread to other sites (and of course, Tumblr users can use more than one website at a time), so you can also see redditors referring to themseles as "goblins."

In 2024, there was reportedly a deal between Tumblr parent company Auttomatic and OpenAI to train on tumblr data. The deal for OpenAI to use Reddit content for training was also done in 2024.

In my own chats, gremlins and goblins began appearing seriously in late March, 2025. This would correspond with the launch of model 4.5, which was first released to pro users February 27, 2025. I haven't been able to find if there was enough time for Tumblr and Reddit data to be used in training 4.5. There's information on the technical aspects of 4.5, from pre-training to scaling the use of unsupervised learning, but I haven't found any date range or knowledge cutoff. It's tempting to speculate that a massive amount of social media content being injected into 4.5 increased the Goblin Quotient, but without further data on when the goblins/gremlins appeared and in which models, we can only guess.

From my own experience using Codex, I haven't run into any mention of goblins or gremlins. I suspect this is because I only ever speak to Codex in an authoritative voice, and the appearance of the creatures seems triggered by a casual register.

I will also say I've never encountered the trolls, ogres, or pigeons that are mentioned in the system prompt. I have seen raccoons mentioned (especially "raccoons in a trenchcoat"), but it doesn't come even close to beating gremlins.

I would love to know - have you encountered goblins or gremlins in ChatGPT chats? When did it first appear? What models did you use? Does it appear only in casual chat or does it also appear when you are trying to speak professionally? I'd love to crowdsource more info on the great gremlin phenomenon.