Google AI or ChatGPT

[RogerS]

I've been using AI's for research more and more. For example, LOML and I were trying to remember the female cast member in a TV comedy series from years gone by. We knew Rowan Atkinson was in it but that was all.

My approach to this in the past would have been to go to IMDB.com, enter Rowan Atkinson and scroll down for ever and a day seeing if I could spot a programme that would fill the gap. Then again using IMDB, or maybe Wiki or just plain old Google, enter that programme's name and see which search result gave her name. Quite long-winded.

I have found that Google's AI is not to be trusted, makes things up if it doesn't know and when challenged will back down and have another go...ending up with often another wrong answer.

ChatGPT on the other hand is galaxies away far better. In this instance I simply asked it "What comedy series on BBC TV did Rowan Atkinson appear in"...long list but there it was 'Not the Nine O'clock News'. I asked who the main female cast member was. Pamela Atkinson.

But that was trivial for ChatGPT and I now use it as my online expert GP/consultant where it excels. I have yet to find it make a mistake. It remembers our previous conversations and will tie in facts it gleaned from those to see if they have a relevance to the latest health concern.

Out of curiosity I took Windows OP in his 'Nervous about the surgeon' thread substituting 'my sister' in place of 'I'. Take a look.

 

[the bear]

Pamela Stephenson surely?

 

[Windows]

I use Grok through X. They keep cutting back how much you can use it for free, but 20 questions a day is enough for me. It’s pretty useful, but it’s always a case of how much broken glass you want in your custard, isn’t it? I like my custard to be glass-free. Grok mostly is just custard, but then occasionally… glass. I try to keep my wits about me and spit it out instead of chewing it.

I’ve recently been using it for researching law and also helping with purchase decisions. I always ask for quotes and direct links to the relevant legislation so I can read it myself and not rely solely on summaries.

I agree that Google is poor and ChatGPT and Grok are both much better.

 

[Cabinetman]

Google’s reasonable as long as you use a bit of brain power too, trying to get a micro switch out from behind the clutch pedal on mr Mazda, Google was very handy telling me where it was what it looked like and that it was a bayonet filling and twist clockwise, struggled for a minute and tried anti clockwise, out it came.

 

[RogerS]

Pamela Stephenson surely?

Oops..senior moment !

 

[RogerS]

Google’s reasonable as long as you use a bit of brain power too, trying to get a micro switch out from behind the clutch pedal on mr Mazda, Google was very handy telling me where it was what it looked like and that it was a bayonet filling and twist clockwise, struggled for a minute and tried anti clockwise, out it came.

Hah! Told you Google’s AI was inaccurate

 

[Cabinetman]

Hah! Told you Google’s AI was inaccurate

More of a helpful nudge than a definitive roadmap?

 

[Dr.Al]

 

[Accipiter]

Google’s reasonable as long as you use a bit of brain power too, trying to get a micro switch out from behind the clutch pedal on mr Mazda, Google was very handy telling me where it was what it looked like and that it was a bayonet filling and twist clockwise, struggled for a minute and tried anti clockwise, out it came.

Did you put in the year/model number/reg plate number? Just in case of some changes for different year/model of...

 

[Accipiter]

Sincerely hope Roger doesn't mind and feel I'm breaking any confidence in responding to this as I used Google for a recent query I had for medication I'm prescribed - following a recent discussion/posting to the forum.

Roger and I have had a private discussion re health matters a a med we are both using and each shared the responses we got from the respective AI we used: I with Google and Rog with ChatGPT.

Both, I feel, covered significant areas with their respective responses to the questions asked. In one instance Google asked a question, following on from an answer given, regarding times (for meds) taken whereas Rog said it hadn't occurred to *him* to ask/mention such. *that's how I read/interpreted Roger's text - hopefully correctly

Google also came up with plans/approaches on how to discuss my concerns with my doctor on my different meds when I see him/them (don't always see *my* doc!) next.

I've yet to ask/put my questions to ChatGTP (Apologies Roger - I said I would but haven't yet got to do so ) so I can compare the two AI things side by side. I guess I've used Google AI because it's *there* when using Chrome on/for searches - I don't have to go looking for ChatGTP... which I'm going to have to do to ask the questions I put to Google

NB:
I'll just state that I don't take any AI response/replies as 100% correct but I *consider... think for myself*... on the presented information and then go from there. I'm realising that they (Google) respond to your questions on what you've written... which then could need further information from you to get the more *helpful* information

In my case, re the meds... I asked if omeprazole was a correct treatment for hiatus hernia. Reply was "Yes" but then other questions followed on such as: how long taking/strength of ...
From there I mentioned other meds taken - Google then saying about interference of between some... Which then led to other questions from Google to my replies to more Qs...

 

[Steve Maskery]

I have some batten lights in my workshop and barn. The first lot I brought over from the UK, three-and-a-half years ago. After just a few months one of them started to flicker. Then another and another. I contacted the vendor (via Amazon) and got a refund for the duff ones.

I ordered more of the same in France, different brand but same product, a pack of 5. I installed two, the other three stayed in the box for a couple of months. When yet another one went I came to instal one of the three new ones. Nothing. The other two were the same.

So I decided to take a look inside. There is a little circuit board in one end, and it was obvious that it had got hot. But I figured that if I could get a new board, the light would work again. So I asked Copilot.

He wanted a picture of the board. Tick. He could identify the make and model Tick. The good news is that replacements are widely available in France. Great, where? "Just before I find one for you I need this info". OK, done. "Oh just a bit more". OK, where can I find one?

This went on and on and on. "You need to search for this XYZ specification". Please give me a link!

After about an hour of this nonsense, he admitted that there is no supply in France at all, thy are made in China (NS Sherlock) and I would have to find one in China. I'd already checked Ali Express and the like. He couldn't find one either. What a complete waste of time. If he were real I'd punch him in the diodes.

BTW, these lights are supposed to have a 5 year Warranty. Some haven't lasted 5 months.

I am not impressed. Neither by the lights nor Copilot.
S

 

[Woodbloke]

Pamela Stephenson surely?

Who married Billy Connolly - Rob

 

[RogerS]

Who married Billy Connolly - Rob

Mrs Connolly ?

 

[Woodbloke]

Mrs Connolly ?

The very same Rog - Rob

 

[Spectric]

Can AI be more inteligent than it's creator, ie humans. It can logically sort data very fast and so appears to be some super brain but look at a real time operating system which appears to multitask, ie does many things at once but in reality it is not. It actually does just one thing at a time but so fast you perceive it as multitasking so AI is fast but not sure it is intelligent, in reality it is just pattern matching.

 

[SamQ aka Ah! Q!]

so AI is fast but not sure it is intelligent, in reality it is just pattern matching.

Hear, hear!!

 

[spb]

I have found that Google's AI is not to be trusted, makes things up if it doesn't know and when challenged will back down and have another go...ending up with often another wrong answer.

All current LLMs have a tendency to do this to varying degrees, and it takes careful prompting to avoid it as much as possible. In some of them the 'system prompt' - the instructions that are given to it before you interact, that tell it how to respond, how to refer to itself, and so on - is better than others, but hallucinations like this are very much just a feature of the current models.

Deep down, they're text generators. They generate text that, based on the vast volume of text in its training set, seems like a plausible response to the prompt you give it. That means if you give it a prompt that even subtly suggests you want a certain type of answer, it'll try hard to give you what it thinks you want regardless of what actual information it has access to. This is why they flip-flop so much when you point out their mistakes as well; what they really want to do is agree with the most recent thing you said, and if you said its first response was wrong then it'll come up with a completely different one for you.

Also, an integral step in the response generation process is a random number generator, which means that sooner or later it will just completely hallucinate something. Even when it doesn't, it doesn't know how to evaluate the quality of different sources in its training set. Short version: never blindly trust any information an LLM gives you. They're good at summarising, but if you want anything reliable then ask it for sources and check them.

 

[RogerS]

All current LLMs have a tendency to do this to varying degrees, and it takes careful prompting to avoid it as much as possible. In some of them the 'system prompt' - the instructions that are given to it before you interact, that tell it how to respond, how to refer to itself, and so on - is better than others, but hallucinations like this are very much just a feature of the current models.

Deep down, they're text generators. They generate text that, based on the vast volume of text in its training set, seems like a plausible response to the prompt you give it. That means if you give it a prompt that even subtly suggests you want a certain type of answer, it'll try hard to give you what it thinks you want regardless of what actual information it has access to. This is why they flip-flop so much when you point out their mistakes as well; what they really want to do is agree with the most recent thing you said, and if you said its first response was wrong then it'll come up with a completely different one for you.

Also, an integral step in the response generation process is a random number generator, which means that sooner or later it will just completely hallucinate something. Even when it doesn't, it doesn't know how to evaluate the quality of different sources in its training set. Short version: never blindly trust any information an LLM gives you. They're good at summarising, but if you want anything reliable then ask it for sources and check them.

Based on my experiences with ChatGPT, I have not seen any of the issues that you raise TBH.

 

[Blackswanwood]

Can AI be more inteligent than it's creator, ie humans. It can logically sort data very fast and so appears to be some super brain but look at a real time operating system which appears to multitask, ie does many things at once but in reality it is not. It actually does just one thing at a time but so fast you perceive it as multitasking so AI is fast but not sure it is intelligent, in reality it is just pattern matching.

That is a flawed comparison. RTOS are more likely to be designed to be linear and not progress until the previous task has been completed. Even then tasks may be performed simultaneously. AI can and does multitask - the two most common examples are multi-task learning and multi-agent processing. Pattern matching may be one of the tasks AI can perform more effectively than humans but that’s just the tip of a large iceberg.

There are a lot of misconceptions about AI which imo are amplified by the choice of the word intelligence when it was named.