Offering AI Chatbot

[Quickbots]

I am a startup and wanted to introduce my company. It is called <link removed> and offering chatbots powered by chatgpt as a foundation to train on your business and answer questions your customers have.

 

[fisicx]

Welcome to UKBF @Quickbots

Not yet seen a chat bot in any flavour that was of any use. This weekend I had a technical question about a product and the chatbot couldn’t answer the question. It just spewed out a load of guff.

Just had a look at yours and it asks for my name and email so immediately got rejected.

 

[Quickbots]

Welcome to UKBF @Quickbots

Not yet seen a chat bot in any flavour that was of any use. This weekend I had a technical question about a product and the chatbot couldn’t answer the question. It just spewed out a load of guff.

Just had a look at yours and it asks for my name and email so immediately got rejected.

Thank you for your feedback. Our bot trains on your website data and we have a knowledge base section within the bot where you can add manual information which is not on your website that can also learn from. Not only this it also can be trained on pdf or docs.

In relation to our website it does have a name and email but is Optional so you will see ‘x’ button to close it if you do not want to give this information but it is there incase you want to capture leads. Thank you once again for checking it out. Happy to get one trained on your website if you like to test it out.

 

[fisicx]

Look at the answers I got from the chat I’ve just had. They aren’t very good and mostly irrelevant to the question.

 

[Mark T Jones]

Welcome @Quickbots

The difference between forums and real life is that on a forum people will jump on your idea and pick it apart, whereas in real life they will simply walk by if it doesn't jump out and shout at them.

I'm not your target customer - my absolute plan is to be the 'human voice' as things go more AI.

The 3 most important questions for you at this point are:

1. Who is your ideal target customer?
2. What clear, compelling, quantifiable benefit will AI deliver to them?
3. Who you, rather than one of the many others jumping on this bandwagon?

You don't need to answer on here - but time & effort spent on those 3 questions will probably be the best investment you can make in your business.

 

[japancool]

And how do you deal with data protection and security?

How do you stop an unauthorised customer services person from getting access to accounts or personal data?

 

[antropy]

Not yet seen a chat bot in any flavour that was of any use. This weekend I had a technical question about a product and the chatbot couldn’t answer the question. It just spewed out a load of guff.

That's the biggest problem with LLMs at the moment - they make stuff up, and so you can't rely on them being honest. If you can't trust the truthfulness of the answers then they're as good as useless.

However accuracy is increasing a lot and if these are trained on actual data then they might be good?

Current LLMs pretty much never say "I don't know" they make up lies instead as they don't really have a concept of truth.

Paul.

 

[japancool]

Current LLMs pretty much never say "I don't know" they make up lies instead as they don't really have a concept of truth.

So they're politicians?

 

[Mark T Jones]

That's the biggest problem with LLMs at the moment - they make stuff up, and so you can't rely on them being honest. If you can't trust the truthfulness of the answers then they're as good as useless.

However accuracy is increasing a lot and if these are trained on actual data then they might be good?

Current LLMs pretty much never say "I don't know" they make up lies instead as they don't really have a concept of truth.

Paul.

My experience has been that they are mostly factually correct but in a woolly, imprecise kind of way.

I've yet to see a chatbot offer a clear or definitive answer, and the 'human' narrative is oddly stilted and unnatural.

Like you say, it will evolve and there are undoubtedly some good applications for it. Meanwhile just watch the snake-oil & Emperor's New clothes.

 

[fisicx]

Current LLMs pretty much never say "I don't know" they make up lies instead as they don't really have a concept of truth.

Just had this response from a bot:

Sorry, I do not understand, please contact us via email or completing the online Contact Form.

 

[antropy]

So they're politicians?

Haha yes basically

Paul.

 

[antropy]

Just had this response from a bot:
"Sorry, I do not understand"

Probably because it was prompted specifically to say that in that situation.

Paul.

 

[Quickbots]

Look at the answers I got from the chat I’ve just had. They aren’t very good and mostly irrelevant to the question.

Thank you i had a look and now i have improved to give better response based on your 2 questions. So it has room to improve and be tweaked. Looking at overall list of questions it did respond well apart from 2-3 questions. Thank you for the feedback.

 

[Quickbots]

And how do you deal with data protection and security?

How do you stop an unauthorised customer services person from getting access to accounts or personal data?

The bot is only trained on what it is fed. So if trained on your website it is already public knowledge and if you want to add further documents again these will be for public knowledge. It cannot access databases at the moment. It is a great tool for training on business processes and general customer support questions such as product returns or policies. We are looking at how we can incorporate certain CMS so it can interrogate and ofcourse data protection will be forefront with relevant steps taken.

 

[ctrlbrk]

My experience has been that they are mostly factually correct but in a woolly, imprecise kind of way.

No, they're not Mark.

ChatGPT and Google Bard regularly provide completely wrong answers on any topic.

I asked Bard to give me the summary of a court case and it completely made it up. Nothing to do with the case I gave it whatsoever.

Same thing with ChatGPT.

Sometimes they just make stuff up.

 

[Quickbots]

No, they're not Mark.

ChatGPT and Google Bard regularly provide completely wrong answers on any topic.

I asked Bard to give me the summary of a court case and it completely made it up. Nothing to do with the case I gave it whatsoever.

Same thing with ChatGPT.

Sometimes they just make stuff up.

It all comes down to your prompt. You have to tell it in baby language therefore need to give it enough information as possible so it can respond more precisely. Again different types of language models using various algorithms. Perfecting prompts will improve your answers.

 

[ctrlbrk]

It all comes down to your prompt. You have to tell it in baby language therefore need to give it enough information as possible so it can respond more precisely. Again different types of language models using various algorithms. Perfecting prompts will improve your answers.

I don't accept that premise.

If I ask Google Bard "provide a bullet point summary for the court case at https://example.com" and it comes back with a summary that looks like it's from a court case, I don't expect it to be made up with facts and names that are nowhere near the case provided.

This is not a scenario where a better prompt will provide a better answer. It's a case where a clear prompt provides inaccurate results.

 

[japancool]

It all comes down to your prompt. You have to tell it in baby language therefore need to give it enough information as possible so it can respond more precisely. Again different types of language models using various algorithms. Perfecting prompts will improve your answers.

There are tools that already do it. And yes, OpenAI powered.

 

[Nick@Daydot]

Thank you for your feedback. Our bot trains on your website data and we have a knowledge base section within the bot where you can add manual information which is not on your website that can also learn from. Not only this it also can be trained on pdf or docs.

I think that's a large part of the problem.

For a start, many many websites are not well written and are written from the point of view of the website owner rather than what the customer wants to know. The same often goes for help files or supporting documentation. Usually it takes a degree of customer feedback, attention to detail and openness to feedback to get internal documentation working for customers.

So, when someone resorts to customer services, including a chatbot, it's often because they've looked at the site or read the docs, and still can't get the answer. So whilst some people are "lazy" and a facile answer will do, many will need a person to really understand what it is they are asking and how to pitch the answer. A lot of humans aren't very good at this also.

 

[fisicx]

I had a question today about the use of Dobânda Anuală Efectivă (DAE) in calculating repayments. DAE is referred to on the website but only as an option in the settings. What it doesn't do is explain how DAE is calculated. Which means the chatbot wouldn't be able to answer the question. But if a human was available the visitor would get the answer they wanted and be more likely to convert. Which means I really want your chatbot to ask me for help rather than just post guff.

 

[Quickbots]

I had a question today about the use of Dobânda Anuală Efectivă (DAE) in calculating repayments. DAE is referred to on the website but only as an option in the settings. What it doesn't do is explain how DAE is calculated. Which means the chatbot wouldn't be able to answer the question. But if a human was available the visitor would get the answer they wanted and be more likely to convert. Which means I really want your chatbot to ask me for help rather than just post guff.

Yes if you had a definitions file with the methodology/ calculation or you can add it in the knowledgebase and it would pull this through for the customer querying. You can also tweak the output by querying the bot and then implementing the answer in the backend.

 

[fisicx]

I can probably do all sorts of things. But 99% of the people signing up for a service such as yours won't have a clue what a definitions file is or how to use it.

I could now add some content about DAE on the site but the likelihood is I'll never get asked the question again. This is the first time I've has a query about DAE in 8 years!

Ideally the bot should see the mention of DAE and then ask me (the site owner) for more information or clarification.

 

[Newchodge]

The big problem is that, generally, if I ask a question it is because I don't know the answer. If I receive an answer that looks OK, I can only be sure that it IS right by checking everything it says. Which seems pointless. In my own area, law, there have been some cases in the US where courtroom advocates have given the court lists of cases to cite as precedents. When checked by their opponents the cases did not exist - the advocates had used AI to do their research. The judges involved were not amused. I am not sure what sanction they received but it will not have been comfortable.

 

[johndon68]

I am not sure what sanction they received but it will not have been comfortable.

It appears they got away rather lightly, the two lawyers involved were given a joint fine of $5000.

 

[Newchodge]

It appears they got away rather lightly, the two lawyers involved were given a joint fine of $5000.

Being 'spoken to' by a judge is never an easy situation. It happeend to me once 'in chambers' so the only witness was the prosecution lawyer. But it was not comfortable! I wonder what their bosses said - I think the firm they worked for was also fined, and the judge made a point of mentioning the paucity of technological research materials at their firm. I looked it up after my previous post.