Great. OpenAI staff DOESN'T KNOW why ChatGPT, trained on English, uses other languages

[highplainsdem]

Just as they don't know why it "hallucinates" (makes things up).

They hope someone can explain their own creation to them.

Business Insider article on ChatGPT generating a fake news story on Chinese traffic policy that went viral and caused problems: https://www.businessinsider.com/chatgpt-made-fake-chinese-press-release-shocking-residents-china-ban-2023-2

Which was bad enough. But this was the part of the story that really caught my eye:

Jan Lieke, OpenAI's Alignment Team lead, expressed his confusion in a tweet on why ChatGPT — which he said is "exclusively trained on English" — is able to follow instructions that are written in other languages.

"We still don't know why," Lieke said. "I wish someone would figure this out."

And if that doesn't give you enough confidence that that these people working on AI know what they're doing, here's a little AI humor from the same guy, a week later:

Link to tweet

If AI ever goes rogue, just remember to make yourself really tall.

It will be intimidated and leave you alone.

[Tetrachloride]

1. Show the code. Data traces.

[edisdead]

6. Why?

[Tetrachloride]

8. AI is not magic or self aware.

As a former coder, the idea that AI is out of control is ridiculous as if to blame the software and not the coders or management. sorry for run on sentence

[edisdead]

11. Agreed but

Why should they have to show their code?

I am a developer/software engineer for 25 years. I see no reason to force them to show their code.

[Tetrachloride]

12. i speak rhetorically for when "someone" blames software.

[Silent3]

15. This kind of code is self-generated and damn near incomprehensible

Even if you try to trace what's happening inside the tangled, opaque code AIs generate (which could require millions or billions of steps from input to output), you probably won't be able to make sense out of what's actually happening at each decision point.

[Happy Hoosier]

18. I think it gets more complicated....

When machine learning is involved.

I'm not a software engineer, but I'm currently working with a machine learning application that can "teach" itself to recognize radio signals it's not porgrammed to understand, and it can determine the message formats and starts to to interpret the contents of the message. It's becuase the software is able to generate it's own routines to accomplish data tasks. It's a bit craapy to watch it do its thing. Even the guys who wrote it sometimes will pop out with a "well, THAT was unexpected." The creepiest bit is when it started trying to reconfigure the software configurable radio itself, without prompting. It was attempting to configure the radio to receive UHF voice comms frequencies. Weird!

[SheltieLover]

2. Wth don't people use their own brains?

Why the need to talk to a computer?

Lazy, imo.

Thx for sharing!

[ret5hd]

17. Sounds like something a...

computer frustrated with humanity would say.

[2naSalit]

3. So...

Is it like a pandemic level virus that once released can't be contained?

[sanatanadharma]

4. Google Translate

It is not perfect but certainly explains the phenomena.

[genxlib]

5. The John Oliver show did their main story on AI this week

Really interesting and of course funny.

The short story is that self learning technology depends a great deal on the data is has to work with. Even then, it is difficult to know what it will learn. Sometimes it gets downright weird.

Definitely worth a look.

[edisdead]

7. I mean not much different than this place.

[sdfernando]

9. I shutter to think what happens when someone asks ChatGPT

about SkyNet.

[durablend]

14. It got stuck thinking of a response

Yeah, we're fucked.

[yardwork]

10. Looks like it taught itself other languages.

Languages are based on logic, after all.

[keep_left]

13. It reminds me of the chat bots that Facebook supposedly "pulled the plug" on...

...back in 2017. The "plug pulling" was incorrectly reported, but the story took on a life of its own. Still, the bots (fairly dumb ones, actually) started to make up a pseudo-language, or perhaps a form of robot shop-talk.

https://www.theregister.com/2017/08/01/facebook_chatbots_did_not_invent_new_language/

An excerpt ("Alice" and "Bob" are the bots):

Bob: i can i i everything else . . . . . . . . . . . . . .
Alice: balls have zero to me to me to me to me to me to me to me to me to
Bob: you i everything else . . . . . . . . . . . . . .
Alice: balls have a ball to me to me to me to me to me to me to me
Bob: i i can i i i everything else . . . . . . . . . . . . . .
Alice: balls have a ball to me to me to me to me to me to me to me
Bob: i . . . . . . . . . . . . . . . . . . .
Alice: balls have zero to me to me to me to me to me to me to me to me to

Skynet this ain't.

[haele]

16. I suspect the clean-up and quality checks missed some cut and paste code.

That's usually the base cause, they were so much in a hurry to show off their working product, they forget to clean up the random trash code hiding behind slash marks.

Haele