You commented TRICK. Here are the pages behind the reel: how your phone keyboard guesses your next word, why ChatGPT is the same guess on repeat, and what that means for how you use it.
1. Your keyboard is a language model
Since iOS 17, iPhone autocorrect runs on "a transformer language model, a state-of-the-art on-device machine learning language model for word prediction." (Apple Newsroom, iOS 17)
At WWDC23 Apple said the iPhone can "run this model every time you tap a key." (WWDC23 keynote) The words above your keyboard are the ones you'd "probably type next, based on your past conversations, writing style, and even websites you visit in Safari." (Apple support: predictive text)
Android works the same way: Google's Gboard uses language models "trained to predict the next word given input text." (Google Research)
2. ChatGPT is that guess on repeat
GPT stands for Generative Pre-trained Transformer, the same family of model. (Google ML glossary) OpenAI says its models "first learn through pretraining, a process of predicting the next word in huge amounts of text." (OpenAI: why language models hallucinate)
The difference is what happens after the guess. Your keyboard shows you one and stops. ChatGPT takes its own guess, adds it to the sentence and guesses again. Google's glossary calls this auto-regressive: "auto-regressive language models predict the next token based on the previously predicted tokens." A token is a word, part of a word or punctuation. (OpenAI: tokens)
And it's a lot bigger. The GPT-3 paper describes "an autoregressive language model with 175 billion parameters." (Brown et al., 2020) OpenAI hasn't published the size of the models behind ChatGPT today.
3. Why it can sound sure and be wrong
OpenAI defines a hallucination as "instances where a model confidently generates an answer that isn't true," and says plainly: "ChatGPT also hallucinates." Part of the reason: "arbitrary low-frequency facts, like a pet's birthday, cannot be predicted from patterns alone." (OpenAI)
4. How to use it, knowing it's a guesser
Give it the facts. Everything earlier in the chat becomes context for the next guess. (Google ML glossary) Paste the email, the doc or the numbers in, and ask it to work from those.
Don't trust it on one-off facts. Names, dates, prices and quotes are exactly the "arbitrary facts" patterns can't predict. Check them.
Ask for the source. If it can't point to where a fact came from, treat it as a guess.
Try it now: open Notes, type "I'll be there in" and keep tapping the middle suggestion. That's ChatGPT with the brain switched off.
Mike
PS: I build AI tools for small businesses at OptiMAX. If you want one working from your own facts instead of guessing, reach out at optimax-ai.com.