[kwlug-disc] ChatGTP + Bing: we're entering a new paradigm and it's mind-blowing

[Ronald Barnes]

Khalid Baheyeldin wrote on 2023-04-06 12:53:

> A few weeks back, Europol was warning on the use of LLMs in crime
> 
> https://www.reuters.com/technology/europol-sounds-alarm-about-criminal-use-chatgpt-sees-grim-outlook-2023-03-27

The only noteworthy aspect of any technology (or any tool at all) being 
used for crime is how long it takes for the first instance.

Perhaps also the reach - with email / crypto criminals have global 
reach, but that is old news now.




> Then yesterday, the Washington Post had an article that is very
> concerning, even scary, to say the least.
> 
> ChatGPT claimed that a certain law professor sexually harassed a
> student on a class trip to Alaska, and cites a Washington Post
> article. There was no such trip, and no such article.

Yes, this is an issue that must be dealt with.

One thing that Bing integration did that I liked was to footnote much of
the chatbot's output. This will need further work.



> And Bing/GPT-4 drew from the professor's defence against the false
> ChatGPT claims to answer a similar question: so now we have an echo
> chamber of bots! Like Social Media Bubbles, but worse ...

Nothing new here. XKCD on "Citogenesis": https://xkcd.com/978/

(XKCD is up to 2760 now, so that's an *old* comic.)


Also, people need to not believe everything they're told, especially
when the "teller" explicitly advises them not to.



> ChatGPT also claims that a mayor in Australia served jail time for
> bribery.
> 
> And it led a journalist to contact a certain university professor
> citing an article she published, with dates. No such article exists
> ...

Yeah, that's a problem. Six months into a beta release, with OpenAI
strongly warning people to not trust the output as it is *not* a search
engine, I'm not setting my hair on fire yet.


It took Google & YouTube how many *decades* to start flagging
flat-earth, anti-vaxx, and other BS in their search results?


Accuracy and "internet" still have issues (that I agree need to be dealt
with), I just don't see LLMs as being unique in that perspective.



On the topic that began this thread though, "Turing test blown away", I
stand by that: 70+ years after its proposal, the challenge has been
exceeded to a degree that's shocking and revolutionary.

Your link backs this up (Europol's quote above).




It was only a few years ago that Google was proud that their assistant
could answer questions like this by following a conversation's subject
between questions:

"What time is it in Hong Kong?"
"X o'clock."
"What's the weather like?"
"Weather for Hong Kong is..." (*not* $current_location's weather)


We've gone from "can we make something that passes for human?" to "this
is virtually indistinguishable from humans, oh my god, what are we gonna
do?" in a *very* short period of time.

That is remarkable and a huge technical achievement.




> This is likely to ruin lives ...

That's probably true, but so have many aspects of the internet.  This is 
unfortunate but common with new inventions.




> And notably, a Princeton University computer science professor is
> quoted saying what Doug Moen said all along: these LLMs are Bullshit
> Generators ...

Wait a minute... From your own link above:

 > "ChatGPT's ability to draft highly realistic text makes it a useful
 > tool for phishing purposes," Europol said.

So, is it a "bullshit generator" or a useful (and quite remarkable) tool?



 From ChatGPT itself:

> ChatGPT Mar 23 Version. Free Research Preview. ChatGPT may produce 
> inaccurate information about people, places, or facts

It's a *chat* bot, not a search engine. It is wildly successful at that. 
It's also still in beta ("Preview").


rb