Hi.
I'm going to try and do a proper live demo.
So to set the scene here, I'm going to go through a demo where someone gets an inbound
sales email.
In this case, an email for someone that would like to try Mindstone.
So Anna from Mindstone is answering Sam based on an initial email that comes in.
So actually, I'm going to start, I need to do one thing first.
You will understand why in a second.
Okay.
So here, the very first thing that I'm going to show is the, I'm going to say, imagine
you're the world's best salesperson and you received the following email.
How would you reply?
This is probably what most people have done with chat to BT before.
So here, it is basically taking a very simple request and it is answering something about
XYZ corp, XYZ corp which is the company from Sam that's reaching out and Mindstone which
it is entirely making up.
At this point, chat to BT has no idea of what Mindstone really is.
It is taking what people think a learning and development platform might want and then
taking it.
So just to give you a bit of context on the email here, Sam reached out.
I hope this email finds you well.
I'm Sam, learning and development manager at XYZ corp, mid-sized tech company.
I'm intrigued by what Mindstone offers but would like to know more and how it can specifically
cater to my company.
So you go through this.
As I said, this is what most people have gotten to.
And the response you are getting is okay.
It is not great.
And I imagine most of the people in this room, actually what I should do is I should start
with chat to BT 3.5 on this because I imagine most people in this room don't have the commercial
version, don't pay the $20 a month so you are still stuck with 3.5 which is significantly
worse to what you can see here.
And most people stop at that.
They say, well, I hear all this stuff about AI but it is producing pretty crap results
for me so why would I go further?
It is overhyped putting it on the pile with web 3 and NFTs and blockchains.
It is not quite there yet.
But there is much more behind this.
So here it is made up an answer.
There are a few things to understand when you are interacting with chat to BT that will
dramatically improve the way that is able to answer.
One is you have to think about these models more as a person than as traditional technology.
And I say that this is not me saying that this is anywhere near an actual person.
We haven't achieved actual general intelligence yet.
If you were talking to a person, imagine you had the best head of sales standing next
to you and we have one of them right in the audience at the moment.
What context would you give them in order for them to be able to help you?
If I were to walk up to a great sales person and tell them, hey, this is an email I received,
how would you reply?
They can't do anything.
They have no idea what my business is about.
They have no idea what the unique selling proposition is, what the customer is about.
None of that.
And so if I had a person next to me, I would give them all that context.
Hey, we are building an upskilling business.
It is a SaaS platform.
It identifies these skills in any article podcast video that you come through.
It measures how your teams are upskilling and then gives you an ROI from that.
All of that stuff is extremely important.
But somehow, when people throw this into chat GPT, they expect it to just know all this
stuff, which it just doesn't.
Now, the second thing to know, I'm just going to need a little bit of drink.
Second thing to know is, and this is a little bit more technical, but that the as chat,
GPT or large language models in general, as they answer, they always predict kind of
the next token, basically the next letter slash word.
And so their answer is part of their reasoning.
So as it answers, it constantly takes into account what it has said before, and then
it continues the next phrase.
Now why this is important is that if you get chat GPT to outline the reasoning, to outline
what are the important parts of a question or of an answer before actually giving you
that final answer, the answer dramatically improves.
Now I'm going to put both of those things in practice, but I'm just going to first read
out what came back here so we can properly contrast.
So here, chat GPT created an answer that says, yeah, we can integrate with existing training
programs, it's tailored learning pathways, we don't do learning pathways in that way.
It is interactive and collaborative, there's analytics and reporting, and we have a bunch
of success stories.
We don't have a bunch of success stories, we only launched about a week ago.
So all of this is made up, but what is interesting is it does understand what are the core challenges
in the upskilling and reskilling market, and it's able to formulate some of this.
Now I am going to switch on now what is called custom instructions.
So again, for those that don't know what custom instructions are in chat GPT, custom instructions
basically allow you to give a set of context to chat GPT, which it is going to take into
account every single time it answers a question for your account.
So I have two or three things that I did here.
One is I gave it information on the business we're building.
So I'm saying here, mine sounds like dueling or for any kind of learning mixed with the
professional showcasing of LinkedIn.
Very simple, I'll give it a little bit more information about what the value proposition
is for an individual, what it is for a business, and then who I am directly.
Now on top of that, I tell it to first outline its thinking, never answer a question directly.
But one, first ask if I am not giving you enough context, ask me questions before you
give me the answer so that you're basically completing your own context.
And second, outline your thinking before you give me the final answer.
And this is that second bit I was talking about, because that forces chat GPT to start
outlining what are the core parameters to answer this question in a good way and it
improves the final result because it's taking into account the first bit of that answer
to improve the end.
So I'm going to switch this on.
And then I'm going to try this again.
So here you can see what it's doing now is it's outlining its thought process, right?
So understand the client's needs first.
Second, emphasize the relevance of Mind Stone.
It now knows what Mind Stone is about.
Integrate with existing programs.
So highlight unique features.
And then here you can see the difference of approach already.
So it is a slightly more prose-oriented feature response as it goes through.
It now specifically goes through, specifically for tech companies like yours, Mind Stone's
ability to track progress in technical and soft skills is invaluable.
It aids in identifying skill gaps and tailored learning paths to individual employees.
And then it is still making up success stories because, again, we are in the middle of actually
creating these.
So this is a slightly better answer already.
And I would say in many cases, depending on how good you are at formulating a sales
email, it probably passes for a decent one.
So long as you remove the hallucinations, you kind of want to make sure that that doesn't
stick.
Now what I'm going to do is I'm going to take a second email from Sam.
I'm going to say, now, imagine you wrote that email and Sam replies with the following.
Now would you answer?
So here you can see now, because I have my custom instructions, it will do the same thing.
It will first outline its chain of thought, what are the important criteria based on the
email that I got back from Sam and what are the points I need to address.
Now in the email from Sam, he says, we have some budget constraints.
This year we are putting relatively tight budget, but next year we have a little bit
more.
The CEO has some expectations, the CFO has some expectations, and the head of technical
requirements also has some questions on security constraints.
So all things that you would take into account as you are trying to address their sales email,
it goes through all of this, so addressing budget constraints, meeting CFOs, CFOs, scrutiny.
And again, it's important it does this in order to then start to write the reply.
And it goes one by one.
So it addresses the budget constraints, so MindSound offers scalable solutions that can
adapt to varying budget constraints, we have flexible pricing models, so on.
CFOs expectations, so we have documented success stories in data that demonstrate MindSound
positive impact, because the CFOs' expectations here were about ROI, so how can you actually
prove ROI?
And then IT here, this was about infrastructure and security, so it goes through all of this.
Now you can take this at face value.
And again, I would say this is a decent answer to start with, so long as you remove some
of the hallucinations and make it your own.
But what is more interesting is asking the next question.
Imagine you are the best salesperson in the world.
How would you rate this reply from zero equals to non-effective and unlikely to lead to a
sale to 10 equals to extremely effective and very likely to lead to a sale?
It's going to grade its own homework.
So here you can see it has rated itself 8 out of 10.
And it is starting to tell me why the answer it provided wasn't as great.
What is missing from this answer?
Now as I said, this is an actual example.
I used this myself, not enough, and this goes to one of the keys around AI at the moment
that I have to change my own muscle before I hit send.
I'm like, no, no.
Don't hit send.
Run this through chat.
I'm probably missing something and it's real rewiring in my own brain because I've been
doing this for 20 years.
I know how to write, I thought I knew how to write decent sales email.
Two weeks ago I did it, I got a 4 out of 10.
That's the reality of what my sales email was.
I went through this process a few times.
So you can see here it's outlined everything that is missing.
It's doing a decent job at addressing specific concerns, but not perfect yet.
The worst thing is evidence and case studies.
It mentions them but doesn't clarify them.
And then room for improvement.
While the response is thorough, it could include more direct references to how MindSome has
specifically helped similar companies.
Then you have a little bit of extra here.
The response highlights MindSome's flexible pricing and customizable features.
But it's 8 out of 10.
So now, very simple.
You can say, now give me a version that would rate 10 out of 10.
Fairly simple.
It seems simple once you do it.
But it's now going to address its own mistakes by creating a better version of said sales
email.
Now, in my case, as I went through this a few weeks ago, I had written an email that
was mostly about 2024.
And I had started addressing that because that was the bigger thing.
And I had totally forgotten.
There was actually also a conversation about the end of 23.
And as I put it into chat, you kind of forgot to address the immediate opportunity and partnership
here, which you really should at least be mentioning.
And I used that, send the email, and it actually converted afterwards.
It's literally just because I didn't pay enough attention that that would not have converted
had I not mentioned it in the first email.
So here you go.
You have a, well, there you go.
Yeah.
So you have the actual improved response, which is now taking into account everything
that itself thought could be improved into a new email that you can take from here.
So it's saying, did it say here?
Yeah, it's not saying that it did a 10 out of 10.
I wonder if it, yeah, OK, it did.
Here, so to elevate to a 10 out of 10, this is the response that comes through.
So there you go.
You can use this rating system for many other things within chatTBT.
But take away these three things.
One, for anything you ask chatTBT, give it the context that you would give a normal
person when you're having a conversation with them.
Two, ask it to outline its reasoning before giving you the final answer.
And three, ask it to rate itself and then improve upon itself.
And you will see your answers go from useless to pretty mind blowing sometimes, at least
in my case.
So hopefully this showed you how chatTBT can be useful on a day-to-day basis if you
have any questions.
Afterwards, come in.
Actually, I can take a few questions now if you have, we'll have a few minutes.
Sanford.
Can you now ask it to rate this result?
Sure.
This is actually a very good exercise.
You can do this multiple times.
It will probably come back with an eight or a nine.
It's not interesting.
It's actually rating it 10 out of 10 at this point.
I've had it in the past, to be very clear, I've had it in the past tell me an eight
or a nine because it finds additional improvements and you can iterate on it.
So it's a really good point and I should actually make that point so that you can continue to
refine your own answer.
Sorry, yes, right, front.
Why did I create a new chat for the second time I did the demo?
So the first time, I didn't have my custom instructions switched on.
So all of the context about the company, all of the context about how it should answer,
weren't taken into account in the first one and in order for chat to be taken into account
my custom instructions, I needed to create another chat because otherwise that one chat
was not going to have the context.
Sorry, can you repeat that question?
Yeah, no, yes, if you switch on your custom instructions, let's stay within the same chat,
the custom instructions are going to be ignored.
Yeah, sorry.
You can put that into your own instructions.
So actually one of the things I didn't run through is that part of my custom instructions
specifies the style in which I wanted to respond.
So I'm pretty succinct.
I try and be a little bit to the point.
I like it to use bullet points.
You will have seen some of this, right?
It's using some of that very specifically in the way that it's answering.
So you can fine tune that and you can ask it to reiterate.
At this point actually, and I'll do a demo on this one day, maybe next time, you can
feed it a lot of your own context.
It has enough context now that you can feed it your own content and then it will replicate
the same tone of voice in the same way of phrasing so that it reads like you.
Sorry, yes, sorry.
Yeah, you can, absolutely.
I can, absolutely, but I would say that we probably want to move on to the next talk.
Do we have Daniel in the audience now?
There we go.
Amazing.
We have another great talk right after.
I'm going to be around for the next few hours.
You can definitely go and get me.
All of this is filmed.
I will upload this to the AI mind space afterwards so you can definitely rerun it, but I would
suggest you try it yourself and hopefully you understand now that there's a lot more
behind church at BT that just the first answer you tried.
So thank you very much.
Thank you.
Thank you.
Thank you.