ChatGPT … getting close to passing the Turing Test

[Andrew TinsleyParticipant @andrewtinsley63637]

I have questioned Chat GP recently concerning one of the topics that I was researching (technical problems with some theoretical physics).

It was quite obvious that the answers were complete rubbish, so much for artificial intelligence. Without the necessary information trawled from the internet and other sources, it was unable to string together any meaningful response.

Andrew.

 

[Michael GilliganParticipant @michaelgilligan61133]

A serious question for you, Andrew:

Do you think that the technical content of the answers has any relevance to the Turing Test ?

MichaelG.

[Kiwi BlokeParticipant @kiwibloke62605]

Early in this thread I posted rather sceptical and cautionary comments about AI, based on my very limited experience of its use. I stand by my previous comments, but further experience has led me to modify my view somewhat.

The output from ChatGPT (and other AIs) is very sensitive to the user’s input. You can, for example, control its style of ‘conversation’, get it to ‘understand’ your level of expertise in the subject of discussion, restrict it to certain texts or types of material to examine, and so restrict its tendency to speculate. It’s also, obviously, dependent upon being trained appropriately, thus ‘cutting edge’ or esoteric subjects may provoke hallucination in its attempts to respond (some can be hilarious). So, it’s a GIGO machine. Perhaps this partially explains the apparently divergent views of some of the users here.

I have also been playing with NotebookLM. This allows texts to be uploaded to the ‘cloud’, and then the AI can be asked to process or to answer questions on the uploaded texts exclusively. It may then recommend additional external sources, which the user can choose to allow, individually. I’ve used this to plough through a couple of Acts of Parliament and various Regulations (attempting to read which rapidly causes one to lose the will to live), in order to try to answer some questions on some forms I have to complete. I had had no idea how to approach the non-factual questions, despite being reasonably intelligent, and having reached a high level of academic achievement, but the AI found all the answers and suggested a suitable form of words to use. I had been stumped because I had got stuck in a linguistically pedantic mode, which didn’t suit the poor wording of the legislation or the forms, but AI interpreted both in ways I hadn’t been able to ‘see’. It would have been impracticable – and impossible – to have found a human who would have done this for me, let alone for free, in seconds.

My experience to date is that the AIs with which I’ve been playing are way beyond the Turing test (for what it’s worth) and have a super-human level and breadth of knowledge and an ability to simulate understanding that is, in some cases at least, greater than mine (well, perhaps that’s no big deal…). They are very useful tools, but should be treated as such, as helpers, not substitute thinkers.

AI is, however, very dangerous, for all the reasons already mentioned in this thread. Thank goodness I’m long retired – my job isn’t under threat! It seems to me that the most important and urgent AI-related task is to work out ways of controlling it in a way that doesn’t make it useless. If this is not addressed, human intellectual activity will be radically altered in unpredictable ways – we won’t need to learn anything or to think for ourselves – and the result will be unfortunate in the extreme.

 

[Kiwi Bloke]

On Kiwi Bloke Said:

… are way beyond the Turing test (for what it’s worth) …

 

At the risk of unnecessarily labouring my point … this topic was intended to specifically relate to the Turing Test, and I might suggest that participants read the Wikipedia page, which summarises that quite nicely

https://en.wikipedia.org/wiki/Turing_test

MichaelG.

.

The results would not depend on the machine’s ability to answer questions correctly, only on how closely its answers resembled those of a human.

 

[alecsParticipant @alecs]

If you read the comments sections on Facebook, social media, news sites etc, it’s obvious that it doesn’t take much intelligence to sound like a human. A computer with as much power as a digital watch could probably do it.

Getting answers right is another thing, as Turing said. But sounding human is barely any test of intelligence.

Seriously, anyone with an IQ of 80 can hold a competent conversation. But you wouldn’t want them designing bridges or flying aircraft. Or even advising you on the best way to machine a tricky casting.

[Andrew TinsleyParticipant @andrewtinsley63637]

Hello Michael,

My comments on AI were not intended to say if it passed the Turing test or not. It was intended to make the point that one should not take the output of AI as the gospel truth. If the information is not in the data banks that AI uses, then the results tend towards garbage.

Andrew.

[Michael GilliganParticipant @michaelgilligan61133]

Thanks for clarifying that, Andrew

Topic-drift is inevitable when humans are participating 🙂

MichaelG.

[blowlampParticipant @blowlamp]

I’ve seen that the cost of employing AI has reached the point where it can be more than the cost of employing humans. Couple that with the massive energy usage AI requires and it doesn’t appear to be plain sailing to me.

Martin.

[John HaineParticipant @johnhaine32865]

Any mathematical proposition by Alan Turing is almost certainly going to be absolutely correct.  But the “Turing Test” was entirely non-rigorous, there is no way that an AI “passing” it is necessarily truly intelligent.  Watch Hannah Fry’s documentaries on AI recently screened – she discusses the case of a young man who “conversed” with an AI chatbot which advised and encouraged him about how to go about assassinating the queen.  He was convinced by it but someone more mentally stable would have a different view.

[Author]

Posts