www.forbes.com
Every AI mystery deserves a logical plausible explanation, including the latest one about OpenAI's ... [+] ChatGPT o1 advanced model.gettyIn todays column, I aim to resolve the AI mystery floating around on social media and in the mainstream news regarding OpenAIs ChatGPT o1 advanced AI model suddenly switching momentarily from working in English to working in Chinese. In case you havent heard about this surprising aspect, users have been posting tweets showcasing o1 doing just that. The AI is solving a user-entered prompt and while presenting the logical steps the language shifts from English to Chinese. This occurs for a line or two and then reverts back to English.Is it some kind of tomfoolery? Hacking? Maybe the AI is going off the deep end? Lots of postulated theories and wild conjectures have been touted.Lets talk about it.This analysis of an innovative AI breakthrough is part of my ongoing Forbes column coverage on the latest in AI including identifying and explaining various impactful AI complexities (see the link here). For my coverage of the top-of-the-line ChatGPT o1 model and its advanced functionality, see the link here and the link here.Whats Going On With o1Allow me to set the stage for revealing the known facts concerning the mystery that is afoot.ChatGPT o1 generative AI is a large language model (LLM) that generally rates as being at or quite near the pinnacle of modern-day AI models. There are plentiful advances jammed into o1. When you use o1, you can immediately discern that the AI has something special going on. Happy face.MORE FOR YOUTo be clear, o1 and none of the current time AI is sentient, nor have we reached artificial general intelligence (AGI). If you are interested in where we are related to achieving AGI and also the vaunted artificial superintelligence (ASI), see my analysis at the link here. At this time, generative AI and LLMs are based on human-devised mathematical and computational pattern matching that in large do an amazing job of mimicking human writing and conversation.Those who have been using o1 now for several months would likely say that they relish doing so. It does its job. You enter a prompt; you get a reply. One nice twist to o1 is that the reply customarily includes a listing of the steps that the AI took to arrive at the answer presented. This is commonly known as chain-of-thought (CoT), see my detailed explanation at the link here, consisting of a series of delineated steps of the internal processing by the AI.So far, so good.Now for the mystery. Various users have indicated that from time to time the o1 suddenly switches from English to Chinese when displaying the chain-of-thought steps that are being undertaken. Just as quickly, the portrayal shifts again back to English. It is almost like seeing a mirage, except that it really does happen, and printouts or screen snapshots bear this out.Are peoples eyes deceiving them?Nope, the accounts of this happening are verifiable and not merely fancy.Explanations Are Over-The-TopOpenAI seems to have remained mum and isnt telling us what is at the root of this oddity. Their AI is considered proprietary, and they dont allow others to poke around into the internals, nor do they make publicly available the internal design and mechanisms at play. This means that everyone can only guess what the heck might be happening inside o1.Into this vacuum has rushed a slew of quite wild suggestions.Some of the nuttiest conjecture postulates that the Chinese have taken over o1 or perhaps are secretly running OpenAI. Another equally outlandish idea is that a Chinese hacker has planted something into o1 or has accessed a secret backdoor. On and on these conspiracy-oriented theories go. Social media has an overworked imagination, indubitably.I am going to categorically reject those zany schemes.Why so?Know this the same overall issue of switching out of English has been documented by others and includes instances of switching to German, French, Portuguese, and so on. The gist is that the Chinese language is not the sole purveyor of the grand switcheroo. Other languages are momentarily displayed, beyond just Chinese.Perhaps I find myself out on a limb, but I seriously doubt that an entire cabal of earthly hackers or numerous countries across the globe are all sneakily putting their hands into the internals of o1. My viewpoint is that there is something more straightforward that can explain the multitude of sudden language appearances.Laying Out A Reasonable GuessI will share with you my theory or educated guess at what might be occurring. Dont take this to the bank. There are lots of technical reasons that something like this can take place. Lets go with one that I think is plausible, makes abundant sense, and fits with the reported facts.Is it the winner-winner chicken dinner?I cant say for sure since the AI is proprietary and the AI isnt open for inspection.Put on your Sherlock Holmes cap and go along for a fascinating journey into the guts of contemporary generative AI and LLMs.Leaning Into The CoreWhen generative AI and LLMs are initially put together, the big first step entails scanning lots of data to do pattern-matching on how humans write. All kinds of essays, narratives, stories, poems, and the like are examined. Complex mathematical and computational mechanisms try to identify how words relate to other words.This is coined as a large language model due to being undertaken in the large, such as scanning millions upon millions of materials on the Internet. Without the largeness, we wouldnt have the fluency that is currently exhibited by LLMs (for those interested in SLMs, small language models, I showcase how they differ from LLMs, at the link here).Ill use a simple example that will gradually aid in unraveling the mystery.The Word Dog Comes To MindConsider the word dog as a commonplace word that readily would be scanned when examining content on the Internet. We can assume that dog is immensely across the web as a word that people use. Thats a no-brainer assumption. Everyone has a beloved dog, or a story about dogs, or has something to say about dogs. Humankind pretty much loves dogs.If you were to detect which other words appear to be associated with the word dog what comes to mind?Some obvious ones might be fluffy, four-legged, tail-wagging, etc.From the perspective of what is taking place inside the AI, the word dog is associated mathematically and computationally with the words fluffy, four-legged, tail-wagging and so on. The words themselves have no meaning. They are each a jumble of letters. The word dog consists of the letter d followed by the letter o and followed by the letter g.You should think of the word dog as just a bunch of letters, collected together, and we will treat that collection of letters as a kind of blob. The blob of the letters in dog is statistically associated with the blobs of the word consisting of the letters fluffy.My aim here is to have you disassociate in your mind that the word dog has any meaning, such as images in your head of this or that favored dog. Instead, the word dog is a collection of letters and is associated with lots of other collections of letters that form other words.The French Word For DogShifting gears, I will pick a language other than English to set up the reveal that will be momentarily discussed.I lived in France for a while and love the French language, though I admit I am extremely rusty and would never even attempt to speak French aloud. Anyway, if its Ok with you all, I will envision that we are interested in the French word for dog (which is going to be easier as a language choice than picking a Chinese word, due to the symbols used in Chinese writing, but the underlying precept is going to be the same).There is a French masculine version, chien and a feminine version, chienne for dog, but lets simplify things and go with just using for the sake of discussion word chien (thanks for playing along).If you dont know French, and if I showed you the word chien, Id bet that you wouldnt know what the word means.This makes sense that you wouldnt know. For example, the word dog has the letters d, o, and g, but none of those letters exist in the word chien. The French word for dog doesnt seem to resemble the English word for dog. You are unable to readily figure out that they are essentially both the same words in terms of what they signify.Dog And Chien Have Roughly The Same FactorsSuppose we went ahead and did a scan on the Internet to find the word chien and identify other words that seem statistically related to that word.What would we find?The odds are that you would see that chien is associated with the words fluffy, four-legged, tail-wagging, and the like.And what could you therefore say about the word dog versus the word chien?Well, both of those words are associated with roughly the same set of other words. Since they are nearly associated overwhelmingly with the same set of other words, we could reasonably conclude that both those words probably have the same assorted meaning. They are two peas in a pod.The crux is that the word dog and the word chien can be treated as the same, not because you and I in our heads know them to refer to the same thing, but because they both point to other associated words that are approximately the same set of other words.LLMs Pickup Other Languages When Data TrainingThe deal is this.When doing the initial data training of generative AI and LLMs, the widespread scan of the Internet is usually aimed primarily at English words (kind of, thats true of English-oriented LLMs for which English-speaking AI developers tend to build). During my talks about AI, attendees are often shocked to learn that while the data training is taking place, there are bits and pieces of other languages getting scanned too.This is more incidental than purposeful. You can see why. The scanning is moving from website to website, and sometimes there might be content in something other than English, maybe just a page here or there. The chances are pretty high that the scanning is going to eventually touch on a wide array of languages other than English, such as French, German, Chinese, etc. Not at a full clip, just on a random wanton basis.What does the AI do with those said-to-be foreign words?If it was you or me, and we were trying to read all kinds of websites, the moment you hit upon a page that had something other than English, you might be tempted to set aside the verbiage. You might be thinking that since your principal pursuit is English, discard anything that isnt English.The usual approach with AI is that the AI developers just let whatever language is encountered be encompassed by scanning and pattern-matching. No need to try and kick it out. Just toss it into the pile and keep churning.This produces an exciting and quite intriguing outcome, keep reading.Bringing The Dog Back Into The PictureImagine that an Internet scan is taking place, and the word dog is encountered. Later, the words fluffy, four-legged, tail-wagging and others are found and determined to be statistically related to the word dog.The same might happen with the word chien.Then, the AI mathematically and computationally construes that dog and chien appear to be referencing the same thing. It is almost as though the AI crafts an internal English-French dictionary associating English words with French words.The downside is that since that wasnt the main goal, and since the volume and variety of French words encountered might be relatively slim, this English-French dictionary is not necessarily going to be complete. Gaps might readily exist.Various AI research studies have shown that English-focused LLMs often end up being able to readily switch to using other languages that have been previously scanned during data training, see my analysis at the link here. The phenomenon is an unintended consequence and not particularly planned for. Also, the switching is not necessarily going to be fluent in the other language and might be flawed or incomplete.You can likely envision the surprise by AI developers that their LLM suddenly was able to spout a different language, such as French or Chinese. Their first thought was heck, how did that happen? Researchers eventually found that the smattering of any other language that was encountered can lead to the AI devising a multi-lingual capacity, of sorts, in a somewhat mechanical way.Mystery Part 1 Is ExplainedReturning to the mystery at hand, how is it that o1 can suddenly switch to Chinese, French, German, or whatever other language beyond English?The answer is straightforward, namely, the AI picked up an informal smattering of those languages during the initial data training.Boom, drop the mic.Whoa, you might be saying, hold your horses. It isnt just that o1 displays something in a language other than English, it is also that it suddenly does this seemingly out of the blue.Whats up with that?I hear you.We need to resolve that second part of the mystery.When Something Points To Something UsefulGo with me on a handy thought experiment.Free your mind. Throughout all the instances of the English word dog, suppose that at no point did we encounter the word whisper while scanning the Internet. Those two words never came up in any connected way. Meanwhile, imagine that the French word chien at times was statistically found to connect with the word whisper. Please dont argue the point, just go with the flow. Be cool.Heres the clever part.When AI is computationally trying to solve a problem or answer a question, the internal structure is typically being searched to find a suitable response.Pretend I typed this question into generative AI.My entered prompt: Can a dog whisper?The AI searches throughout the internal structure.There arent any patterns on the word dog and the word whisper. Sad face.But remember that we have the word chien exists in there too, plus we had found that chien has an association with the word whisper. Thats good news, due to the AI associating dog and chien as essentially the same words, and luckily the word chien is associated with the word whisper.Stated overtly, you might remember those days of algebra where they kept saying if A is to B, and if B is to C, then you can reasonably conclude that A is to C. Remember those rules of life? Nifty. Here, in essence, dog is to chien, while chien is to whisper, and thus we can say that dog is also to whisper. Logic prevails.The AI is going to be able to answer the question, doing so by accessing the French words that perchance were picked up during the initial data scanning.Internally, suppose the AI has this sentence that it composes: Oui, un chien peut chuchoter. That is generally French for saying that yes, a dog can whisper.An answer was generated, scoring a victory for generative AI, but we need to do a little bit more sleuthing.Final Twist That Deals With Displaying ResultsWould you be likely to normally see a French sentence as a displayed response when using an English-focused LLM?No. Not if you are using an English-language-based LLM that is set to show principally English responses, and if you havent explicitly told the AI to start displaying in French (or whatever language). The AI might have the French sentence internally stored and then convert the French sentence over into English to display the English version to you.Thats our final twist here.Remember that the report by users is that the language switcheroo only seems to happen when the chain of thought is underway. The chances are that language switching isnt necessarily active for the chain-of-thought derivations. It is activated for the final response, but not the intervening lines of so-called reasoning.This also explains why the AI suddenly switches back out of the other language and continues forward in English thereafter.The basis for doing so is that English in this case is the predominant form of the words that were patterned on. The switch to French was merely to deal with the whisper resolution in this instance. Once that happened, and if the prompt or question had other parts to it, the AI would simply resume with the English language for the rest of the way.Boom, drop the mic (for real this time).The Logical Explanation Is Satisfying In recap, most generative AI and LLMs tend to pick up words of other languages beyond English during the initial data training and scanning of the Internet. Those words enter the massive statistical stew.They are considered fair game for use by the AI.If those non-English words are going to be helpful during generating a response to a user prompt, so be it. As they say, use any port in a storm. The AI is programmed to seek a response to a user inquiry and spanning across languages is easy-peasy. It might also be flawed, depending on how much of any other respective languages were involved in the data training.A significant clue of the o1 mystery is that the reported instances are infrequent and only seem to arise in the chain-of-thought. This can be linked to the notion that while the AI is composing a response, there isnt a need to convert from a non-English language to English. Those are just intermediary steps that are merely grist for the mill. The AI doesnt have any computational need to convert them from one language to another.Once a final result is ready, only then would a language conversion be warranted.That is then one reasonably sensible and altogether logical reason for resolving the mystery. Of course, I had mentioned at the get-go that there are other logical possibilities too. I just wanted to share an explanation that seems to cover the proper bases. Now then, some might be tempted to reject the logic-based route entirely and argue for something more sinister or incredible, perhaps ghosts hiding inside o1 or the AI is starting to take on a life of its own. Imagine all the wild possibilities.Lets end with a final thought expressed by the great Albert Einstein: Logic will get you from A to B. Imagination will take you everywhere.
