I don’t know about the rest of you, but the bulk of my token usage is not what I type in - it’s that data the AI is processing for me. I’m shocked they even got an 8% reduction.
Classical latin is not very order constrained because it embeds so much grammar morphologically; it could aound very much like english or not at all like english, depending on the author's choices.
Vulgar latin does correspond to modern romance languages (and english) much more closely.
Using Chinese could lead to a small amount of gains, but it's nowhere near as drastic as you'd think
Putting aside the question of whether English's larger training corpus would be a dimension of quality: content is "tokenized" before the model sees it. The "~4 English chars/token" rule isn't a strategy to optimize around because in practice, many English words become one token, the same way Chinese words do.
"I love you" and "我爱你" both tokenize down to 3 tokens. So no efficiency gains from a Chinese translation. "Telephone" is just one token, and "电话" (Chinese word for telephone that spans 2 characters) also gets condensed into one token. Sometimes there are minor gains, like "经济" being one token but "economy" being two. But it doesn't end up materializing in the gain you'd hope for by increasing meaning-per-character.
You can play around with OpenAI's tokenizer here. Great for getting particular about how to save tokens in prompt and tool definitions
https://platform.openai.com/tokenizer
I cannot tell whether you were really serious but this actually could be an interesting study: same task, same procedure, only the language and writing system is different: English vs Chinese vs Korean vs ...
Since the information density is about the same in all (spoken) languages, there shouldn't be much of a difference. But maybe one writing system is better suited for LLMs.
Heh, deepseek is known to randomly switch to Chinese output for no reason at all. As I was using Claude chat as a poor man’s reviewer, I just pasted it over and Claude interpreted it without skipping a beat and responded in English.
My dad told me he had a German math professor in college that would sometimes have to pause to think while writing out equations on the board and then turn around to a room full of confused students to ask how long he had been speaking German to know how far back in the lecture to restart in English. Again. He was known for doing it. Maybe we can anthropomorphize deepseek similarly??
It would take the typical English speaker many years of dedicated study to learn Mandarin Chinese to a level of fluency required to take advantage of any token savings. "Caveman speak" can be adopted instantly, by anyone (in, presumably, ~any language) with no study or time investment required.
problem is there are thousands of pictograms you have to know with ambiguous sounds
i think korean is the best, just learn the alphabet and you can do pretty crazy compaction by dropping honorifics and abbreviation, you can also sound out foreign words ex. 짱깨 if given the right context LLMs should have no problem understanding
Isn’t spoken language in general and English in particular famously 50% redundant? Does that not provide the same (if not more) benefit to the model when narrowing the scope of intent?
Ask, what is the function of grammar in language? I would argue two major functions. Forcing an encoder to marshal ideas into a coherent structure. Preserving the integrity of that structure when communicated and decoded.
Grammar sounds off? Was the message correctly understood; is there enough left to apply error correction and regenerate the intended message?
I think it's funny this works somewhat, but that's not what the goal is, IMO. The goal is to have the model do this internally in its "thinking" stage. On the very rare occasions in the past where gpt5 leaked its true internal thinking, the "CoT" was itself kinda similar. Instead of the open source "so the user wants me to... but wait... maybe I should... blahblah...", GPT5 was using internal traces like "try x.. no.. try y... no.. from x yes then z yes...". That's probably more token saving, or faster responses, if you can get the model to remain accurate.
Say caveman, cavemen ... unheard[0] cavewoman. No good. Bad.
When Sabrewulf[1] come, some fight, some run ... some run in cave. No good. Bad.
Token cost. Seem smart less token. But already discount. No profit.
Use less token? Price go up.
What money?
You should be bothered by the persistent & popular use of the "caveman" terminology and the myth surrounding it. However, from a dedicated amateur linguistic point of view, the "caveman language" that so many people seem to know by default is kind of remarkable. Perhaps it points to (it doesn't) some deeper, ancient understanding.
[0] this is a joke because the caveman myth comes with a lot of misogynistic baggage and you should have a problem with it. Also because I made you read this. (this is not an explanation of the joke itself, that's up to you)
[1] the sabre-tooth cat is a trope that over-represents its relationship with people. I mean it's extinct, but it was around "back then", right?
Does anyone more familiar with the technology know if these AI models treat abugidas like Thai or abjads like Arabic differently when it comes to token usage?
Specifically abjads since as far as I am aware their difference from alphabetical and abugida scripts is that they infer the vowels instead of explicitly marking them.
Why don't we just use a highly specialized domain-specific language that describes the code we want generated? Natural language is too inexact and ambiguous.
If you just want a smaller vocabulary, use French? If your goal is to communicate to an LLM, maybe saving tokens isn't the all in win, unless you like reading assert gronkHitThing(true)
French text is somewhere around 10-30% longer than the corresponding English text. I would guess much of what you save on smaller vocabulary is lost on the lengthened text.
As the test shows the real gain is somewhat marginal to most people, but still some ppl claim a wide skills portfolio is required to move faster or save tokens. Mixed feelings, I personally use superpowers as my basic setup - and only them.
alt.nerd.obsessive FAQ v1.4
In the episode where they were filming the Radioactive Man movie [1995], the comic book store guy tells Bart that he can find out the star of the RM movie. He promptly posts a message to alt.nerd.obsessive which states "Need know star RM pic". This information is relayed through the nerd world until it reaches a nerd hiding under the table at a meeting of movie moguls casting the film. He relays back the answer immediately (Rainier Wolfcastle).
Yes all the caveman spoke English like its freaking Flinstone. Like wtf is the caveman dilect, just english words with some randomly skipped? Why not try with actual other human languages and see if there is a token benefit in savings
Me mechanic not speak English. But he know what me mean when me say "car no go", and we best friends. So me think: why waste time, say lot word when few word do trick?
Also, this really should be more precise that it’s talking about neo-caveman. Legit caveman no speak English.
你不去，我也不去
You no go, I also no go
今天这里人很多
Today here person very many
下雨就不去
Fall rain then no go
Vulgar latin does correspond to modern romance languages (and english) much more closely.
Putting aside the question of whether English's larger training corpus would be a dimension of quality: content is "tokenized" before the model sees it. The "~4 English chars/token" rule isn't a strategy to optimize around because in practice, many English words become one token, the same way Chinese words do.
"I love you" and "我爱你" both tokenize down to 3 tokens. So no efficiency gains from a Chinese translation. "Telephone" is just one token, and "电话" (Chinese word for telephone that spans 2 characters) also gets condensed into one token. Sometimes there are minor gains, like "经济" being one token but "economy" being two. But it doesn't end up materializing in the gain you'd hope for by increasing meaning-per-character.
You can play around with OpenAI's tokenizer here. Great for getting particular about how to save tokens in prompt and tool definitions https://platform.openai.com/tokenizer
Since the information density is about the same in all (spoken) languages, there shouldn't be much of a difference. But maybe one writing system is better suited for LLMs.
What do you mean? A lot of us already use Hanzi (汉字) or Kanji.
Or are you asking why non-Chinese speakers do not prompt LLMs in Mandarin?
i think korean is the best, just learn the alphabet and you can do pretty crazy compaction by dropping honorifics and abbreviation, you can also sound out foreign words ex. 짱깨 if given the right context LLMs should have no problem understanding
Also it was tested on reasoning low, i'd have liked to have them tested on higher reasoning levels.
Grammar sounds off? Was the message correctly understood; is there enough left to apply error correction and regenerate the intended message?
When Sabrewulf[1] come, some fight, some run ... some run in cave. No good. Bad.
Token cost. Seem smart less token. But already discount. No profit.
Use less token? Price go up.
What money?
You should be bothered by the persistent & popular use of the "caveman" terminology and the myth surrounding it. However, from a dedicated amateur linguistic point of view, the "caveman language" that so many people seem to know by default is kind of remarkable. Perhaps it points to (it doesn't) some deeper, ancient understanding.
[0] this is a joke because the caveman myth comes with a lot of misogynistic baggage and you should have a problem with it. Also because I made you read this. (this is not an explanation of the joke itself, that's up to you)
[1] the sabre-tooth cat is a trope that over-represents its relationship with people. I mean it's extinct, but it was around "back then", right?
Specifically abjads since as far as I am aware their difference from alphabetical and abugida scripts is that they infer the vowels instead of explicitly marking them.
What are your thoughts on it?