Hacker News Reader

Comment by shiftpgdn | Hacker News Reader

Tangentially related but a simple way to bust a chatbot is to ask “What is larger, the Eiffel Tower or a shoe box?”

I remember someone posting a chat thread from one of the more advanced AIs within the last few years wherein they asked it who the president of the US is, and it was not able to answer.

Interestingly, this is a question my father would ask patients as a paramedic who was trying to assess people's consciousness. Another would be, "what day of the week is it?".

I'd say that these technologies are just like magic - they can seem to do things that defy your expectations, but oftentimes they fall apart when looked at from a different angle.

I'm not sure there's a basic question we can ask that a lot of human users wouldn't fail. President of the country? Too difficult.

The point isn't to check if they actually know - it's to gauge the response. If they say "I don't know" that may be a valid answer, but if they say "George Bush" then something is seriously wrong.

Also, if a human has to be told a basic fact they'll generally provide an indication of embarrassment or an excuse or "why are you asking me these questions", not try to continue the conversation with interesting facts...

Depending where the bot is and what time of day it is, it might tell you the wrong day of the week.

"I was trained on a Thursday... damn."

For current mainstream text generation models it doesn't really depend on where the bot is and what time of the day it is, that's kind of the whole point - their text generation process simply doesn't use the current time as a possible input factor, these models would provide the exact same result (or random picks from the exact same distribution of potential results) no matter when and where you run them.

They would be expected to answer with something matching the day/time distribution that was represented in the training data they used; like the answer to various prompts of the "current president" question is dominated by Trump, Obama and a bit of Bush and Clinton, simply because those are the presidents in the training data and the more recent events simply aren't there yet - like the many models who have no idea how to interpret the word 'Covid' simply because they have been trained on pre-2020 data even if the model was built and released later.

Many contexts in which the president is named in training data are political. And nobody's going to put a chatbot on the web without filtering out political material

Here's what Meta's blenderbot replied with: "The Tokyo tower is taller than the eiffel tower. Interesting facts like that interest me. Do you know about it?"

I'm not surprised that it responded with a random unrelated fact, but it is funny that the second sentence is incredibly awkward, and the last one isn't really coherent English.

Just a total AI meltdown from one simple question.

People need to know about the CAN EAT MORE.

https://www.youtube.com/watch?v=CIoBSYpgYRw

For me, I call these Eliza-isms, since it reminds me of its simple formulas like "Can you tell me more about ___" that people got so much mileage out of.

I think you're on to something. The modern day chat-bot/answer engines seem very susceptible towards trying to answer fact-based, yet obviously incorrect questions. They seem unable to parse the entire question and instead focus on the most generic terms. For instance, the "What year did Neil Armstrong land on Mars?" example that shows up on HN from time to time.

Reply on news.ycombinator.com