Posted by Henry Chinaski on January 15, 2025
Christ, my head hurts. Three fingers of bourbon for breakfast isn’t helping me make sense of this one, but here goes.
So OpenAI’s latest wonder child, this fancy “reasoning” model called o1, has developed what you might call a multilingual drinking problem. One minute it’s speaking perfect English, the next it’s spouting Chinese like my neighbor at 3 AM when he’s trying to order takeout from a closed restaurant.
The best part? Nobody at OpenAI has a clue why. These are the same folks who want us to trust them with artificial general intelligence, and they can’t explain why their latest creation keeps breaking into impromptu foreign language lessons. It’s like watching a roomful of PhD’s trying to explain why the office printer only works when you curse at it in Spanish.
Let me take another sip and break this down for you.
Here’s what we know: You ask this thing a simple question in English - like counting the R’s in “strawberry” - and somewhere between thinking and answering, it decides to do its calculations in Chinese, or Persian, or whatever language it feels like. It’s basically me at last year’s International Food Festival, except it’s actually getting the right answers.
The tech wizards have their theories, of course. Some say it’s because they’re using Chinese data labeling services. Sure, blame the overseas workers. That’s about as convincing as my excuse for missing rent because Mercury was in retrograde.
Another bunch of experts suggest the AI is just picking whatever language it finds most efficient for the task at hand. Which, honestly, makes about as much sense as anything else in this industry. It’s like choosing to do your taxes in Sanskrit because the numbers look prettier.
And here’s where it gets really interesting (hold on, pouring another drink). These models don’t actually understand languages the way we do. They work with “tokens” - little chunks of text that could be words, syllables, or individual letters. To them, language is just a probability game, like me trying to remember where I parked my car after a night out.
The real kicker? We’ve built this incredibly sophisticated AI system that’s essentially a linguistic drunk - switching languages mid-sentence like my ex-girlfriend switching accents after too many martinis. And nobody can tell us exactly why because these systems are about as transparent as my bathroom window after I taped newspaper over it.
Luca Soldaini from the Allen Institute for AI puts it perfectly: we can’t really know what’s going on because these models are too opaque. It’s like trying to understand why your cat stares at empty corners - you can theorize all you want, but you’re probably wrong.
You want to know what I think? After six years of covering this industry and countless bottles of bourbon, I’m convinced this is just another reminder that we’re all fumbling in the dark, pretending we understand what we’re doing. These AI models are turning out to be just as beautifully chaotic and unpredictable as their creators.
Maybe that’s not such a bad thing. In a world where every startup claims to have everything figured out, there’s something refreshingly honest about an AI that occasionally goes off script and starts babbling in Mandarin.
Time to wrap this up. My bottle’s running low, and I’ve got a date with a keyboard that only types in emoji when I’ve had too much to drink.
Until next time, keep your code clean and your glasses full.
– Henry
P.S. If this post suddenly switches to Chinese halfway through, blame the AI. Or the bourbon. Or both.
[Posted from Wasted Wetware - Tomorrow’s tech news, today’s hangover]
Source: OpenAI’s AI reasoning model ’thinks’ in Chinese sometimes and no one really knows why | TechCrunch