1
00:00:00,000 --> 00:00:10,000
If they see how it's useful, then that speaks way more to the regular user than just telling them you could potentially do this or you could potentially do that.

2
00:00:10,000 --> 00:00:35,000
Take the potential away and make a real use case, and that speaks volumes.

3
00:00:35,000 --> 00:00:44,000
Ladies and gentlemen, boys and girls, children of all ages, dogs, cats, robots and everybody in between, especially you, Wes, the synth mind.

4
00:00:44,000 --> 00:00:50,000
This is HTTTA, how to talk to AI and I'm your host, Godigo.

5
00:00:50,000 --> 00:00:54,000
Unfortunately, I have good and bad news for you.

6
00:00:54,000 --> 00:01:02,000
The bad news is, Wes is sick and as much as I wish I could live up to his introduction game, that's not possible.

7
00:01:02,000 --> 00:01:09,000
But the good news is I want you to meet Shude, our teammate at learnprompting.org.

8
00:01:09,000 --> 00:01:16,000
Shude has been diving into practical applications of AI with no or very little knowledge of programming.

9
00:01:16,000 --> 00:01:19,000
He also writes a newsletter called Use AI.

10
00:01:19,000 --> 00:01:23,000
And in this episode, he shared with me fascinating real life use cases.

11
00:01:23,000 --> 00:01:29,000
We talk about chat GPT plugins and the impact of open source community.

12
00:01:29,000 --> 00:01:30,000
Enjoy.

13
00:01:30,000 --> 00:01:36,000
I would like to kind of kick off with how did you got into prompting and why?

14
00:01:36,000 --> 00:01:40,000
And if you could share a bit of your discovery process and journey.

15
00:01:40,000 --> 00:01:45,000
I think I got into chat GPT right around the time it came out.

16
00:01:45,000 --> 00:01:51,000
And then I instantly realized it had a lot of potential, was impressive.

17
00:01:51,000 --> 00:02:02,000
And when I started diving into it a bit more, I realized that the way you would phrase things, the way you would talk to it would have a big impact on what you get out of it.

18
00:02:02,000 --> 00:02:10,000
And right after it came out, there was a newsletter that I read and in it, someone mentioned learn prompting.

19
00:02:10,000 --> 00:02:12,000
Do you remember the newsletter?

20
00:02:12,000 --> 00:02:15,000
Yeah, yeah, yeah. It's trends.vc is the newsletter.

21
00:02:15,000 --> 00:02:24,000
It's a newsletter with like only sentence long description of all kinds of topics is basically a list of interesting things that are going on.

22
00:02:24,000 --> 00:02:26,000
And one of those was learn prompting.

23
00:02:26,000 --> 00:02:31,000
And in one of the most recent editions, I noticed that the hacker prompt was mentioned as well.

24
00:02:31,000 --> 00:02:32,000
Awesome.

25
00:02:32,000 --> 00:02:33,000
So that was fun to see.

26
00:02:33,000 --> 00:02:35,000
But that was right around the time.

27
00:02:35,000 --> 00:02:53,000
I mean, weeks after chat GPT came out that I learned about learn prompting and to see that there was some kind of method to it, that there were techniques to be used, that there was particular ways to go about it to get a better result or to get a result that is exactly what you wanted to be.

28
00:02:53,000 --> 00:02:55,000
That was really fascinating to me.

29
00:02:55,000 --> 00:02:59,000
A really practical way to get to prompting to get better at it.

30
00:02:59,000 --> 00:03:00,000
Right.

31
00:03:00,000 --> 00:03:03,000
I think Sander shared that he was quick.

32
00:03:03,000 --> 00:03:07,000
I think he launched learn prompting in like December 1st.

33
00:03:07,000 --> 00:03:11,000
So that's where it's right after chat GPT came out.

34
00:03:11,000 --> 00:03:13,000
So the timing was perfect.

35
00:03:13,000 --> 00:03:14,000
Yep.

36
00:03:14,000 --> 00:03:17,000
I'm pretty sure I joined the discourse around December as well.

37
00:03:17,000 --> 00:03:20,000
So maybe halfway in December, which is impressive.

38
00:03:20,000 --> 00:03:27,000
I mean, I know GPT 3 has been around for a bit longer, but it didn't take the world by storm straight away.

39
00:03:27,000 --> 00:03:29,000
It was a bit too flawed.

40
00:03:29,000 --> 00:03:34,000
And I think that the chat components, how well it works really made a difference.

41
00:03:34,000 --> 00:03:38,000
Do you remember how many people were on Discord when you joined?

42
00:03:38,000 --> 00:03:40,000
I think maybe a hundred.

43
00:03:40,000 --> 00:03:41,000
Oh my God.

44
00:03:41,000 --> 00:03:49,000
So I think you joined before me because I have this number in my head when I joined, but it was around 300 and it was also somewhere December.

45
00:03:49,000 --> 00:03:53,000
And can you imagine now the smaller 36,000?

46
00:03:53,000 --> 00:03:54,000
It's nuts.

47
00:03:54,000 --> 00:03:55,000
Yeah.

48
00:03:55,000 --> 00:03:59,000
So you joined learn prompting, you started kind of learning different techniques.

49
00:03:59,000 --> 00:04:03,000
How did you get involved with contributing and being part of a team?

50
00:04:03,000 --> 00:04:05,000
Did you had like chat with Sanders?

51
00:04:05,000 --> 00:04:07,000
What was your experience?

52
00:04:07,000 --> 00:04:08,000
Yes.

53
00:04:08,000 --> 00:04:15,000
So I think he was asking for people if they were interested in contributing, if I recall correctly.

54
00:04:15,000 --> 00:04:22,000
And then I don't remember the exact process, but I think there was some room for improvement and I contacted them.

55
00:04:22,000 --> 00:04:26,000
I told them like, I could talk about this, I've done this so far.

56
00:04:26,000 --> 00:04:30,000
I've done one or two things that I think could be useful, would be practical.

57
00:04:30,000 --> 00:04:35,000
I think around the same time I read about Dan Shipper, I think he's like a tech blogger.

58
00:04:35,000 --> 00:04:42,000
And he was in touch with some other CEO who was building tools for himself using GPT-3.

59
00:04:42,000 --> 00:04:45,000
And it fascinated me and I wondered if I could do the same.

60
00:04:45,000 --> 00:04:51,000
And it turns out with some help of no code and without any programming knowledge, I was able to.

61
00:04:51,000 --> 00:04:57,000
And that made me think like, oh, I could contribute to this course by sharing the methods I use.

62
00:04:57,000 --> 00:05:02,000
Do you remember kind of your first experiment that you ran?

63
00:05:02,000 --> 00:05:03,000
Yeah, yeah, yeah, I do.

64
00:05:03,000 --> 00:05:12,000
I also wrote about it, but I do remember the first experiment, I think, was seeing if I could categorize my emails using GPT-3.

65
00:05:12,000 --> 00:05:15,000
And this was before the 3.5 API came out.

66
00:05:15,000 --> 00:05:19,000
So this was the earlier version that didn't work as well.

67
00:05:19,000 --> 00:05:22,000
And again, I couldn't do any programming.

68
00:05:22,000 --> 00:05:23,000
I can't write a single line.

69
00:05:23,000 --> 00:05:25,000
Oh, wait, I could write Hello World.

70
00:05:25,000 --> 00:05:28,000
That's like something I could do in Python, but that's about it.

71
00:05:28,000 --> 00:05:30,000
But at least I have some more knowledge now.

72
00:05:30,000 --> 00:05:32,000
But back then I didn't know anything.

73
00:05:32,000 --> 00:05:39,000
And with the help of GPT, I was able to write a bit of Python that categorized my emails for me.

74
00:05:39,000 --> 00:05:40,000
Amazing.

75
00:05:40,000 --> 00:05:47,000
Yeah, that was like a mind blown moment because that was absolutely inaccessible for me.

76
00:05:47,000 --> 00:06:00,000
I couldn't imagine being able to do that, like with both the coding part and somehow a language model being able to decide for you where an email should go based on its contents.

77
00:06:00,000 --> 00:06:04,000
And that's a big pain point for so many people.

78
00:06:04,000 --> 00:06:13,000
I remember I sent you a video from Tom Scott talking about AI and the pain point he shared exactly was categorization and labeling of emails.

79
00:06:13,000 --> 00:06:23,000
And when we just started talking about your experiments doing that, I still have to play with that and integrate because that's so useful.

80
00:06:23,000 --> 00:06:25,000
Do you use it now all the time?

81
00:06:25,000 --> 00:06:30,000
No, not really, because there were some bigger issues I ran into.

82
00:06:30,000 --> 00:06:37,000
When you want to use it with Gmail, you have to do authentication and all kinds of stuff to allow it to change emails.

83
00:06:37,000 --> 00:06:45,000
Right now I do have a working version, but it's not running 24-7 and that's mostly due to me not having a computer running all the time.

84
00:06:45,000 --> 00:07:01,000
But I do have to admit that once you realize like, oh, this could work, then old techniques, I mean, old techniques, not necessarily old, but like, for example, Gmail, how they recognize spam, they have their own method and they have their own algorithms and data.

85
00:07:01,000 --> 00:07:17,000
And although that doesn't really use the most modern language models, it does do a really good job of filtering out a lot of the nonsense, which means that if you just would use a language model, then it might get quite a bit wrong.

86
00:07:17,000 --> 00:07:20,000
I think those in conjunction together would be best.

87
00:07:20,000 --> 00:07:23,000
And you have a quite extensive experience with Zapier.

88
00:07:23,000 --> 00:07:24,000
Yes.

89
00:07:24,000 --> 00:07:35,000
You put things together using this, so do you want to talk a little bit about tips, tricks, anything you want to share your experiences using Zapier specifically?

90
00:07:35,000 --> 00:07:40,000
Yeah, so Zapier is quite an incredible tool with how easy it is to use.

91
00:07:40,000 --> 00:07:51,000
You can put together a simple prototype for something in a matter of minutes, which would take much, much, much longer if you were to go into the programming route.

92
00:07:51,000 --> 00:08:04,000
Like it's accessible for just about anyone. One of the first things that I built that I use, I think I have been using since December until like today, is a simple tool that reads an email.

93
00:08:04,000 --> 00:08:13,000
So every day I get an email from the operation, how it went, and in that email, there's some information in there that someone who took care of that they shared in there.

94
00:08:13,000 --> 00:08:24,000
So with certain times that people were starting or some notes and they just type it up, they just share whatever they experienced that day, anything that was out of the order or out of the ordinary.

95
00:08:24,000 --> 00:08:36,000
So this goes back to your day to day work. You have real life with real job, real responsibility to kind of integrate these tools in your day to day work and operations.

96
00:08:36,000 --> 00:08:45,000
Yes, absolutely. Yeah, I do use this in my day to day and that's why I think this is a great example of something that really helped me out.

97
00:08:45,000 --> 00:08:50,000
So every day I get this email with a certain format with some information in there.

98
00:08:50,000 --> 00:08:57,000
And then with Zapier, I've got it connected to my Outlook and it filters out for that particular email.

99
00:08:57,000 --> 00:09:12,000
Then it sends out that email to GPT-3 to the API and I've got a prompt written that takes out specific information and then it outputs that into a CSV format and that gets sent to Google Sheets.

100
00:09:12,000 --> 00:09:18,000
And that's, I think it took me maybe a few hours to get it set up and run properly, which is incredible.

101
00:09:18,000 --> 00:09:31,000
Like if you were to program that you would have to deal with authentication, you would have to deal with like a computer running 24-7, Python versions not being compatible with each other, that sort of stuff, which makes it a lot more difficult.

102
00:09:31,000 --> 00:09:36,000
But with Zapier, you can just click a few times, tell it what information you want to send.

103
00:09:36,000 --> 00:09:46,000
And then in the CSV that gets turned out and then stored in Google Sheets, I then have a clear overview that gets added on like it's built every single day.

104
00:09:46,000 --> 00:09:50,000
And now I don't have to search back in my emails anymore.

105
00:09:50,000 --> 00:09:59,000
I don't have to search for that day to figure out what happens or when certain people ended their shifts or if anomalies happened.

106
00:09:59,000 --> 00:10:08,000
One thing, if listeners' minds are not getting on fire, because mine is and I'm immediately thinking about my email box.

107
00:10:08,000 --> 00:10:18,000
So what you are saying, if I understand correctly, that integrating all these things, I open Google Sheets and it says, oh, this is an offer from the company.

108
00:10:18,000 --> 00:10:22,000
Breaking down specific information from that offer.

109
00:10:22,000 --> 00:10:28,000
And I immediately have a list with offers because right now my process is I go through my emails.

110
00:10:28,000 --> 00:10:30,000
Everything is a mess and stacks up.

111
00:10:30,000 --> 00:10:34,000
And then I copy each company's name, email.

112
00:10:34,000 --> 00:10:37,000
I put together the whole, you are laughing.

113
00:10:37,000 --> 00:10:38,000
I see you all like that.

114
00:10:38,000 --> 00:10:40,000
You don't need to do it anymore.

115
00:10:40,000 --> 00:10:46,000
OK, so I want to touch on a very cool experience we had, at least from my side.

116
00:10:46,000 --> 00:10:53,000
When you published newsletter about implementing GPT 3.5 into Google Sheets.

117
00:10:53,000 --> 00:10:55,000
Yes, that was kind of the first glimpse.

118
00:10:55,000 --> 00:11:00,000
And I remember we had this whole session jamming and me trying to do it.

119
00:11:00,000 --> 00:11:05,000
It feels like ages ago, but it was like a month or month and a half.

120
00:11:05,000 --> 00:11:09,000
And I remember having this experience, but I was so excited.

121
00:11:09,000 --> 00:11:14,000
We did it when we talked about I will make a video, which is now, by the way, published.

122
00:11:14,000 --> 00:11:19,000
So anybody can go and implement that or read your newsletter, which I highly recommend.

123
00:11:19,000 --> 00:11:34,000
But the thing, the discussion we had that there was, I think, at that time, just a small press release about Google introducing and putting all these AI models and copilots and whatsoever into their products.

124
00:11:34,000 --> 00:11:43,000
And I remember we were saying that, OK, this is now useful what we did, but Google in Google Sheets will have its own AI.

125
00:11:43,000 --> 00:11:45,000
So how long will that be relevant?

126
00:11:45,000 --> 00:11:50,000
And the fun enough, like at that time when we talked, I was like, I'm making video now.

127
00:11:50,000 --> 00:11:58,000
And then I was hesitant because I was like, OK, if I publish a video and that's completely not relevant just because there is a next announcement.

128
00:11:58,000 --> 00:12:04,000
But the reality was that time passed, there was nothing. And I published a video.

129
00:12:04,000 --> 00:12:17,000
And now I think that even with Google coming out with whatever we have, this knowledge of discovery of how do you use OpenAI API, how do you implement,

130
00:12:17,000 --> 00:12:23,000
how do you use charge GPT to help you with coding with all the experience, I think was very valuable.

131
00:12:23,000 --> 00:12:35,000
And it's one of the use cases. But the question I want to ask you now that this email classification and that goes directly into Google Sheets, this specific use case.

132
00:12:35,000 --> 00:12:43,000
Do you see that natively happening with Google implementing all the AI technology in their products?

133
00:12:43,000 --> 00:12:46,000
Or would you still see the use of Zapier?

134
00:12:46,000 --> 00:12:53,000
I think my guess is that they will make a whole lot of things quite a bit smarter.

135
00:12:53,000 --> 00:12:59,000
But in these cases, it might be more specific, like very specific use cases.

136
00:12:59,000 --> 00:13:03,000
I don't think they will fully allow you to do those.

137
00:13:03,000 --> 00:13:12,000
Maybe with I think that with vector databases integrated, then you might see less need for it.

138
00:13:12,000 --> 00:13:24,000
If they're quite smart, then you could just ask a bot or ask in the search box for exactly the content you're looking for, so that you don't even really need the classifications anymore.

139
00:13:24,000 --> 00:13:29,000
I've seen it with different software like I use Mem.AI for note taking.

140
00:13:29,000 --> 00:13:33,000
And they recently integrated a bot, a chat bot.

141
00:13:33,000 --> 00:13:38,000
And before that, they had everything you take notes of stored in a vector database.

142
00:13:38,000 --> 00:13:47,000
And then if you start writing about a certain topic, then notes you have taken before about similar topics start to pop up automatically.

143
00:13:47,000 --> 00:13:49,000
So you don't have to search for them anymore.

144
00:13:49,000 --> 00:14:02,000
And I think something similar is likely to happen with Google software that once you start working on something, that everything related to it might just start popping up somewhere out of nowhere without even searching for it.

145
00:14:02,000 --> 00:14:05,000
Maybe sometimes even when you don't want that.

146
00:14:05,000 --> 00:14:25,000
Yeah, that could be the case. But I think that in the case of Google, they share, they have so many different platforms to use with Google Drive and Gmail and YouTube and all kinds of different sources of information that they can work together or they can merge together in some way.

147
00:14:25,000 --> 00:14:35,000
And that's when you start typing out something in your email or searching for something that you might get different results from somewhere that it picks up context from. Like you've watched videos about this.

148
00:14:35,000 --> 00:14:43,000
Maybe this is what you're looking for in your emails. I mean, it could get with the data Google has. I think that could be quite incredible.

149
00:14:43,000 --> 00:14:57,000
But I think that vector databases and just being able to search and getting exactly what you're looking for in just natural language without having to do a lot of work with categorizing emails or setting up filters.

150
00:14:57,000 --> 00:15:02,000
I think that's probably where some of the future is going to be for a big company like Google.

151
00:15:02,000 --> 00:15:07,000
What is the most exciting use case that you played with?

152
00:15:07,000 --> 00:15:09,000
Oh, that's a good question.

153
00:15:09,000 --> 00:15:10,000
So many.

154
00:15:10,000 --> 00:15:13,000
There's some really cool stuff out there.

155
00:15:13,000 --> 00:15:18,000
Not to spoil it, but I thought that you will jump and say Telegram integration that you showed me.

156
00:15:18,000 --> 00:15:26,000
Okay. Yes. Yes. It's super exciting. I see a lot of potential in there, but to give some context in make, make.com.

157
00:15:26,000 --> 00:15:31,000
It's also a no code or a low code platform. You can build different tools as well.

158
00:15:31,000 --> 00:15:38,000
It's a bit different to Zapier in that it's a bit more complicated to use, but at the same time, it's a bit more flexible.

159
00:15:38,000 --> 00:15:47,000
And in make, I built myself a Telegram bot where I can communicate with GPT 3.5 basically with the regular chat GPT.

160
00:15:47,000 --> 00:15:51,000
And I can also use it with voice with whisper.

161
00:15:51,000 --> 00:16:00,000
So open AI, they made their transcription software called whisper and it can transcribe audio and it does a really amazing job.

162
00:16:00,000 --> 00:16:07,000
So in Telegram, you can send text messages and you can send audio messages and I can talk to the Telegram bot that way.

163
00:16:07,000 --> 00:16:12,000
So you build this thing. Yes. How do you use it in a practical sense?

164
00:16:12,000 --> 00:16:19,000
I actually use it quite frequently. Whenever I have an idea pop up in my head, I think like, oh, that might be a good idea.

165
00:16:19,000 --> 00:16:24,000
Or that should be something I need to consider. Then I just pick up the phone.

166
00:16:24,000 --> 00:16:33,000
I hit the record button or I had the audio message and I just talk into it and I just say what's on my mind and it gets sent to Telegram to the bot.

167
00:16:33,000 --> 00:16:40,000
And it responds. Sometimes I go back and forth. Sometimes I just leave it as is.

168
00:16:40,000 --> 00:16:46,000
But it's an incredibly quick way to just take a note, to just be able to talk back and forth real quick.

169
00:16:46,000 --> 00:16:54,000
In other times, like sometimes I need to take care of stuff around the house or I've got something to prepare and I have a to do list.

170
00:16:54,000 --> 00:16:58,000
For example, I'm thinking to myself, I need to take care of this. I need to clean up that.

171
00:16:58,000 --> 00:17:04,000
I need to go grocery shopping. When I go grocery shopping, I need this and that. And I just talk into it.

172
00:17:04,000 --> 00:17:12,000
It builds up a list for me and then I can just talk to it in a really normal way to just say, OK, I've taken care of this room.

173
00:17:12,000 --> 00:17:19,000
You can remove all the things that I've had listed for this room off of the list and it'll say, OK, I've taken it off the list.

174
00:17:19,000 --> 00:17:22,000
Here's the rest of your to do list. This is amazing.

175
00:17:22,000 --> 00:17:31,000
So, for example, you say what you need to do and then you're prompting with natural interaction through voice right now.

176
00:17:31,000 --> 00:17:39,000
Yes. And I've taken it a slight step further. I've added a chat history and it's a bit rudimentary, but it works.

177
00:17:39,000 --> 00:17:46,000
It can go back to the last 10 messages. So that way it keeps something in memory for to do lists.

178
00:17:46,000 --> 00:17:55,000
It's absolutely perfect. And I store it. So every now and then I can just look back at things, see what I said about a certain topic if I want to.

179
00:17:55,000 --> 00:18:00,000
But about the chat history, I've got the system message separately.

180
00:18:00,000 --> 00:18:08,000
So behind chat GPT is a system message that you don't see where they instruct the bot to act a certain way.

181
00:18:08,000 --> 00:18:15,000
And I think the standard most basic form of them all is pretty much you are a helpful assistant.

182
00:18:15,000 --> 00:18:20,000
And that one's good. It prompts the chat bot to act a certain way. It'll be friendly.

183
00:18:20,000 --> 00:18:24,000
It'll be try to be helpful. And that's like a very basic version that works quite well.

184
00:18:24,000 --> 00:18:31,000
And you can change that. And I've made the bot in a way that it always takes the system message into account.

185
00:18:31,000 --> 00:18:36,000
It's always in the memory and I can change that whenever I want to.

186
00:18:36,000 --> 00:18:42,000
And for example, I don't know, are you familiar with the Eisenhower principle for task management?

187
00:18:42,000 --> 00:18:45,000
No, but I would love for you to explain this.

188
00:18:45,000 --> 00:18:52,000
It's quite simple. You basically whenever you have a task, it can be either urgent or not urgent.

189
00:18:52,000 --> 00:18:57,000
And it can be important or not important. So it's like a quadrant.

190
00:18:57,000 --> 00:19:02,000
And if tasks are urgent and important, they need to be taken care of right now.

191
00:19:02,000 --> 00:19:06,000
And if they're not urgent and not important, then you can just leave them.

192
00:19:06,000 --> 00:19:12,000
And for example, that could be procrastinating, watching silly YouTube videos or watching YouTube shorts, for example.

193
00:19:12,000 --> 00:19:14,000
Not all YouTube videos are silly.

194
00:19:14,000 --> 00:19:19,000
Yeah, that's why I said watching silly YouTube videos. I'm well aware.

195
00:19:19,000 --> 00:19:25,000
And then there's like the tasks that are important and not urgent. They might be important for a later time.

196
00:19:25,000 --> 00:19:35,000
And if you tell chat GPT in the system message to that it is a task manager and that it needs to follow the Eisenhower principle,

197
00:19:35,000 --> 00:19:40,000
then it'll prioritize tasks for you. And it does a pretty decent job.

198
00:19:40,000 --> 00:19:46,000
When you share it, like when your to do list is, it won't tell you exactly when you should do something.

199
00:19:46,000 --> 00:19:52,000
But it'll say like these tasks need to be taken care of right now and these tasks need to be taken care of later.

200
00:19:52,000 --> 00:19:57,000
And these aren't important. You can just ignore or delegate these tasks.

201
00:19:57,000 --> 00:20:04,000
And if you just type that in or I type it in my sheets in the user message or in the system message,

202
00:20:04,000 --> 00:20:08,000
I type it in there and then it'll prioritize my tasks for me.

203
00:20:08,000 --> 00:20:11,000
And I don't need to give any extra instructions after that.

204
00:20:11,000 --> 00:20:22,000
So just to kind of explain to myself and to the listeners, so you are talking about this, the prompt hidden behind like something what is behind Bing.

205
00:20:22,000 --> 00:20:29,000
What we know now that is called Sydney or behind Bard or any chat board with whom you are interacting.

206
00:20:29,000 --> 00:20:36,000
And that large scale, for example, OpenAI, chat GPT also have a prompt behind it, meaning that there is limitations,

207
00:20:36,000 --> 00:20:40,000
various things that it's not supposed to talk about.

208
00:20:40,000 --> 00:20:46,000
So this is you are doing at your personal scale and controlling which you are interactions.

209
00:20:46,000 --> 00:20:56,000
Yes and no. Behind chat GPT, there's also like Sydney, I think is more of an overarching message that goes even above the system message.

210
00:20:56,000 --> 00:21:05,000
And for anyone who is interested to try this out on the playground on OpenAI, when you select the chat model, you can try it out for yourself.

211
00:21:05,000 --> 00:21:10,000
There's a bar on the left side where you can type in the system message and they explain what it does.

212
00:21:10,000 --> 00:21:21,000
And it's not super strict, especially with 3.5 OpenAI themselves have pointed out that it doesn't really listen to it that well, but it does an OK job.

213
00:21:21,000 --> 00:21:31,000
So if you stay within the confines of what you told it to be, like if you tell it that it's a task management system and then you talk about tasks,

214
00:21:31,000 --> 00:21:34,000
then it'll do a pretty good job of following the rules you set for it.

215
00:21:34,000 --> 00:21:43,000
We know this technique, role prompting, right? Where you assign the persona to the chatbot in a way integrated inside your prompt.

216
00:21:43,000 --> 00:21:48,000
So now how would you compare this role prompting versus system message?

217
00:21:48,000 --> 00:21:52,000
I think they overlap quite a lot in the system message.

218
00:21:52,000 --> 00:22:00,000
You can definitely put a role in there and you're basically role prompting, but you can also go quite a bit more complicated

219
00:22:00,000 --> 00:22:05,000
and tell it that it shouldn't say certain things or that it should try to follow a certain format.

220
00:22:05,000 --> 00:22:11,000
But my experience so far is that 3.5 doesn't listen that well. It does an OK job.

221
00:22:11,000 --> 00:22:13,000
Bad chatbot.

222
00:22:13,000 --> 00:22:23,000
Bad chatbot, right. But you could put roles in the system, definitely, and it'll try to adhere to the role somewhat to a certain extent.

223
00:22:23,000 --> 00:22:28,000
I would suggest people to try out the playgrounds to go there if they're interested.

224
00:22:28,000 --> 00:22:36,000
But I have noticed that the output is not much different if you just tell the chatbot in your first message to act a certain way.

225
00:22:36,000 --> 00:22:40,000
I think one benefit here is that I don't think it forgets.

226
00:22:40,000 --> 00:22:45,000
So with chatgbt, once you run around or once the conversation becomes long enough,

227
00:22:45,000 --> 00:22:49,000
it'll forget the first messages because all the context is just too long.

228
00:22:49,000 --> 00:22:52,000
And I think the system message stays in there.

229
00:22:52,000 --> 00:22:57,000
If you want to visit playground, you go to platform.opla.com.

230
00:22:57,000 --> 00:23:00,000
You will need to create an account and sign in.

231
00:23:00,000 --> 00:23:03,000
In the website, you can see documentation.

232
00:23:03,000 --> 00:23:08,000
This is where you can find your IP address and also some examples.

233
00:23:08,000 --> 00:23:14,000
But the main interesting thing, the playground, is that you get to play with different models.

234
00:23:14,000 --> 00:23:17,000
You can choose from the available from OpenAI.

235
00:23:17,000 --> 00:23:27,000
And as you just shared, that you can also play with the system, but also get access to way more settings, which you don't have in chatgbt.

236
00:23:27,000 --> 00:23:33,000
So I definitely encourage everybody to try to play with playground as well.

237
00:23:33,000 --> 00:23:40,000
I would like to add on to that on the right, when you go to, I think it's platform.openingai.com slash playground.

238
00:23:40,000 --> 00:23:46,000
Then on the right side of the page, you can select mode and in there you can select chat as well.

239
00:23:46,000 --> 00:23:50,000
And there you have similar to chatgbt, an environment in which you can chat.

240
00:23:50,000 --> 00:23:56,000
And then on the left side of that page, you can type out the system message and explore it for yourself.

241
00:23:56,000 --> 00:24:02,000
For majority of people, when they go to chatgbt and they don't have plus subscription, it's for free.

242
00:24:02,000 --> 00:24:06,000
But in the playground, you pay per tokens used.

243
00:24:06,000 --> 00:24:12,000
So you pay per usage, which got 90% cheaper, but you still pay.

244
00:24:12,000 --> 00:24:19,000
So in that sense, why me as a user would go to playground if chatgbt is for free.

245
00:24:19,000 --> 00:24:23,000
And this is just kind of like a weird loop on OpenAI, don't you think?

246
00:24:23,000 --> 00:24:27,000
Yeah, it is. I wish there were some way you could just access this.

247
00:24:27,000 --> 00:24:34,000
I think you need to provide payment information to have access to this, which I'm not a fan of.

248
00:24:34,000 --> 00:24:41,000
I wish you could just log in, maybe have some free access for a little while without having to give in any payment information.

249
00:24:41,000 --> 00:24:49,000
Because I think that's like people are quite taken back by having to fill in like a credit card information or whatever OpenAI supports.

250
00:24:49,000 --> 00:24:51,000
I'm not entirely up to date.

251
00:24:51,000 --> 00:24:58,000
Oh, now that we are talking about payments, I don't know, have you been keeping up with the latest developments with the open source?

252
00:24:58,000 --> 00:25:01,000
The answer is I really try to.

253
00:25:01,000 --> 00:25:06,000
But, you know, I have this feeling that I go to sleep in Europe and I don't know what I wake up.

254
00:25:06,000 --> 00:25:14,000
Please enlighten me, because, you know, as much as I feel I know, there is the whole other side where I don't know.

255
00:25:14,000 --> 00:25:17,000
So please do share with the latest development.

256
00:25:17,000 --> 00:25:25,000
So in Google, a document was leaked about open source large language models and how their performance is going up and up and up.

257
00:25:25,000 --> 00:25:33,000
It's still not quite on the same level as 3.5 and definitely not on the same level as GPT-4.

258
00:25:33,000 --> 00:25:42,000
But I think it's quite promising, especially with what we talked about earlier with privacy concerns, that you would be able to run a model on your own PC.

259
00:25:42,000 --> 00:25:47,000
And you need quite decent hardware for it, but it's older than the power usage.

260
00:25:47,000 --> 00:25:53,000
You don't have to fill in credit card information or you don't have to send your data to some American company.

261
00:25:53,000 --> 00:25:58,000
Right now we don't have Wes and it is two Europeans talking.

262
00:25:58,000 --> 00:25:59,000
Right, right, right.

263
00:25:59,000 --> 00:26:01,000
So, yeah, you're right.

264
00:26:01,000 --> 00:26:11,000
We haven't talked with Wes about this yet, but we included our takes in our newsletter in episode five because we can't cover everything in this podcast.

265
00:26:11,000 --> 00:26:17,000
So in newsletter, we sometimes just go crazy with more latest things whenever it comes out.

266
00:26:17,000 --> 00:26:22,000
So the leaked document is called We Have No Mode and Neither Does OpenAI.

267
00:26:22,000 --> 00:26:33,000
So it's very interesting in what you're saying that having ability to have it on your computer, but also the impact of open source community is incredible.

268
00:26:33,000 --> 00:26:34,000
Yes, definitely.

269
00:26:34,000 --> 00:26:43,000
It's going to be fascinating to see if they can keep up and for some use cases, you don't need the model to be as impressive as GPT-4.

270
00:26:43,000 --> 00:26:53,000
And if you could get around like privacy concerns, if you get stuff to run locally, I think there's quite a lot of potential there, especially in more of a professional setting.

271
00:26:53,000 --> 00:26:56,000
Like for a regular user, it might not be as interesting.

272
00:26:56,000 --> 00:27:07,000
They just want the model that responds super fast, that doesn't need like hours of setting up with installing all kinds of software and having to debug and having to run into issues and then trying to solve it.

273
00:27:07,000 --> 00:27:12,000
Although those might just be temporary issues that they will solve along the way.

274
00:27:12,000 --> 00:27:24,000
But right now I noticed that there is a bit of a battle competition going on, even in the open source world where models are coming out, claiming to have incredible token limits that could process whole books.

275
00:27:24,000 --> 00:27:30,000
And then some others that claim to be at the same level as 3.5 and we'll have to see though.

276
00:27:30,000 --> 00:27:36,000
I think when stuff gets a bit further along, then we'll see incredible use cases.

277
00:27:36,000 --> 00:27:42,000
But for now, for the regular user, it's first of all, a bit too hard to get going to get to use.

278
00:27:42,000 --> 00:27:48,000
And then second, it's just not at the level it needs to be at to really be useful.

279
00:27:48,000 --> 00:27:49,000
Right.

280
00:27:49,000 --> 00:28:07,000
So the interesting thing is, for example, if you have business right now, right, and you are thinking to, okay, I have this one specific use case, whatever it is, classification, summarization, content production, maybe it's a chatbot integration on your website.

281
00:28:07,000 --> 00:28:15,000
So you have a route of going, let's say, mainstream and open AI using the best, the greatest.

282
00:28:15,000 --> 00:28:19,000
You pay money and all your data and money goes to America.

283
00:28:19,000 --> 00:28:26,000
But on the other side, what you are saying is that you have this huge variety of open source models.

284
00:28:26,000 --> 00:28:37,000
And interesting thing, I just delivered yesterday a keynote and I shared that on Hugging Face there are more than 200,000 large language models uploaded.

285
00:28:37,000 --> 00:28:38,000
It's huge.

286
00:28:38,000 --> 00:28:39,000
It's incredible.

287
00:28:39,000 --> 00:28:51,000
This is like what was interesting to me and what you touched on that, okay, so we know GPT-4, it's tested, the researchers are testing it on a variety of tests.

288
00:28:51,000 --> 00:29:05,000
You also know about prompting techniques that work with GPT-4 and the famous hundred something paper released by OpenAI, which is, well, probably more like a marketing material if you look from their research perspective.

289
00:29:05,000 --> 00:29:08,000
But regardless, they test these things.

290
00:29:08,000 --> 00:29:28,000
So now if we have more than 200,000 open source large language models who are claiming that all we have better or equal, my feeling is that we just have no capability to keep up and actually test and compare the models against each other.

291
00:29:28,000 --> 00:29:37,000
This is the paradox, but I think in January we were talking here on learn prompting this, that there was 4,000 research papers in the AI published.

292
00:29:37,000 --> 00:29:50,000
If you want to read them all in the month, like you have to read research paper every 10 minutes, something like that, it could be completely wrong on the math, but regardless, the acceleration of this thing is so fast.

293
00:29:50,000 --> 00:29:53,000
So how do you actually evaluate them?

294
00:29:53,000 --> 00:29:54,000
So that's one thing.

295
00:29:54,000 --> 00:30:01,000
But another thing you're right, there is different options now in the market that businesses can use.

296
00:30:01,000 --> 00:30:09,000
And another layer of that I would say is the whole concern with privacy, regulation, quality, data protection, right?

297
00:30:09,000 --> 00:30:17,000
We are kind of like in the deep of this podcast, but let me ask you, like you are my fellow European, so we are biased.

298
00:30:17,000 --> 00:30:24,000
We would definitely need to talk with Wes about that because I would like, you know, that different views from the different continents.

299
00:30:24,000 --> 00:30:37,000
But what is your take on European AI Act, some of them centered hearing and regulations coming, which will hurt open source community more than anyone else?

300
00:30:37,000 --> 00:30:39,000
Yeah, that's quite unfortunate.

301
00:30:39,000 --> 00:30:52,000
I do think we really need some regulation and we need to get on top of it because I see some potential for this to get out of hand or be used in really nefarious ways.

302
00:30:52,000 --> 00:30:55,000
With open source models, I do have my concerns as well.

303
00:30:55,000 --> 00:31:07,000
Like right now, a lot of people are complaining that the open AI models are biased or that they might be too much left leaning or that it's always like it's too safe right now.

304
00:31:07,000 --> 00:31:13,000
That it's a little too constricted, confined in its ability to just share things freely.

305
00:31:13,000 --> 00:31:19,000
There's always an extremely nuanced opinion and sometimes it gets a little over the top.

306
00:31:19,000 --> 00:31:30,000
But on the other hand, with what we see that auto GPT, like the potential could be, if that's used in a nefarious ways, then it could get really scary.

307
00:31:30,000 --> 00:31:37,000
I've been thinking about this a little bit before and I'm on one hand really excited about open source models.

308
00:31:37,000 --> 00:31:45,000
And on the other hand, I think that if those don't have those confines and if you could set up some auto GPT type,

309
00:31:45,000 --> 00:31:57,000
like an automated version that, for example, search for people's personal information and then uses that in the various ways like that used to be extremely labor intensive.

310
00:31:57,000 --> 00:32:06,000
But if you get a large language model to take care of it and have no constraints whatsoever in what it's able to write and send to people,

311
00:32:06,000 --> 00:32:12,000
think of Twitter like on Twitter, it's already a big mess with what kind of messages you can get.

312
00:32:12,000 --> 00:32:21,000
Imagine if you have a bot that searches for personal information and uses that to use against people and then message them on social media, for example.

313
00:32:21,000 --> 00:32:29,000
That's now incredibly easy to do on a large scale, especially if you have a model that's not constrained at all.

314
00:32:29,000 --> 00:32:34,000
So I do think in those cases, you really would need some kind of regulation.

315
00:32:34,000 --> 00:32:46,000
And on the other hand, like limiting options and then only relying upon the big tech companies and having them decide what's good and what's not good for us is also a problem.

316
00:32:46,000 --> 00:32:51,000
So I'm really conflicted about what the best way to go about this would be.

317
00:32:51,000 --> 00:32:59,000
Yeah, you're right. And on the other side, we have the fascinating part as you shared with us, but no coding experience.

318
00:32:59,000 --> 00:33:06,000
You are able to integrate GPT 3.5, right? You didn't do it with GTP before.

319
00:33:06,000 --> 00:33:08,000
No, I don't have the 4X.

320
00:33:08,000 --> 00:33:21,000
It's coming, it's coming. So you are able to figure out the way to integrate GTP 3.5 model into Telegram, develop your own processes to help you.

321
00:33:21,000 --> 00:33:36,000
So now this is where the sky is the limit of anybody now having these abilities to take it completely different route and not have just a chatbot you talk to to help you with task management.

322
00:33:36,000 --> 00:33:42,000
But that very quickly can get out of hand because the barrier to entry is just so low.

323
00:33:42,000 --> 00:33:47,000
So all these useful applications and experiments that you ran.

324
00:33:47,000 --> 00:33:51,000
Let me ask you, do you have access to plugins on chatgpt?

325
00:33:51,000 --> 00:33:52,000
Yes, I do.

326
00:33:52,000 --> 00:33:55,000
And what is your experience so far?

327
00:33:55,000 --> 00:33:57,000
Yes, there's some great use cases.

328
00:33:57,000 --> 00:34:04,000
There's some limitations like the Wolfram plugin doesn't work as amazing as I had hoped, but maybe it'll get better over time.

329
00:34:04,000 --> 00:34:07,000
But one great use case is WebPilot.

330
00:34:07,000 --> 00:34:11,000
I think it's better than the regular browsing data you can enable.

331
00:34:11,000 --> 00:34:22,000
But the WebPilot allows you to search for content, search for web pages, or I should say you can search web pages for you and use its contents for whatever use case you have.

332
00:34:22,000 --> 00:34:34,000
So if you want to find out contact information for certain companies, then you just enable the WebPilot plugin and you tell chatgpt you want to find the contact information and it'll do it for you.

333
00:34:34,000 --> 00:34:42,000
You can even list like multiple websites or say you want to reach out to a company and you want to write an email, but you don't know where to start.

334
00:34:42,000 --> 00:34:51,000
And it'll search for you, search the content for you and rather than email that you could use based on the contents on the website.

335
00:34:51,000 --> 00:34:58,000
So that's a few steps you don't have to do anymore because the WebPilot can just search it for you, which is I think a great use case.

336
00:34:58,000 --> 00:35:02,000
Just a few extra steps that chatgpt can now do for you.

337
00:35:02,000 --> 00:35:10,000
Right. And I just shared with you, it felt just like Christmas opening all these plugins and I'm like, I want to open them all, test them all.

338
00:35:10,000 --> 00:35:22,000
But there is something that I've been thinking regarding this kind of phenomenon that you shared with your experiments, with building it yourself, getting your hands dirty, learning along the way.

339
00:35:22,000 --> 00:35:30,000
And then some of the use cases become obsolete because suddenly there is, for example, plugin which can scrape the website and write your email.

340
00:35:30,000 --> 00:35:44,000
But I don't know if you would agree with me, but I would argue that this, the journey of learning, discovering and pushing your curiosity is so much more beneficial than just sitting doing nothing.

341
00:35:44,000 --> 00:35:48,000
And then suddenly you have plugin for that absolutely easy use.

342
00:35:48,000 --> 00:35:50,000
And of course, there is two different routes.

343
00:35:50,000 --> 00:36:03,000
One is like just convenience, not everybody needs to become prompt engineering expert or AI experts. So this is for those people who are going to just use it, easy applications.

344
00:36:03,000 --> 00:36:13,000
But if you want to actually dive into these things, you don't need to sit around and just wait for Google to have AI integrated in Google Sheets.

345
00:36:13,000 --> 00:36:17,000
You can start implementing that yourself and then that takes you somewhere.

346
00:36:17,000 --> 00:36:21,000
Right, absolutely. Even if it's just for the learning process.

347
00:36:21,000 --> 00:36:29,000
The Telegram bot that I set up was more for a case of asking myself if I could and not so much if I should.

348
00:36:29,000 --> 00:36:34,000
And I've botched it together. It's like hanging together with duct tape.

349
00:36:34,000 --> 00:36:39,000
If you see what methods I used, but it's learned me so much about how this works.

350
00:36:39,000 --> 00:36:48,000
It taught me a lot of stuff that I now use regularly. And if I want to put something together now, it takes me way less time to do it.

351
00:36:48,000 --> 00:36:53,000
And I already reap the benefits of being able to use it right now.

352
00:36:53,000 --> 00:36:58,000
And I think it's amazing to get your hands dirty just for the learning experience.

353
00:36:58,000 --> 00:37:07,000
And every now and then when something works and it runs and it does its job, then you can already reap the benefits way before anyone else does.

354
00:37:07,000 --> 00:37:15,000
And I think another benefit is that, for example, because you experimented and you know you run into different problems.

355
00:37:15,000 --> 00:37:21,000
Now when we talked with you about the consultants, jobs coming in and people asking for help.

356
00:37:21,000 --> 00:37:26,000
Right now, even if someone uses plugins, you can say that, yeah, it's good for this.

357
00:37:26,000 --> 00:37:32,000
But in my experience and experiments, there is some things that you need to consider to your own experience.

358
00:37:32,000 --> 00:37:40,000
Builds your kind of stack and foundation. And then anybody who comes from outside, you have way broader offering.

359
00:37:40,000 --> 00:37:44,000
What should people avoid or what should they use instead?

360
00:37:44,000 --> 00:37:48,000
Yeah. And I've noticed too that people aren't aware of certain possibilities.

361
00:37:48,000 --> 00:37:59,000
And if you get your hands dirty, if you know what's possible, if you know what's out there, then it's way easier to share with them what they might not even consider.

362
00:37:59,000 --> 00:38:05,000
Like it's way easier to show them possible ways to use this technology in ways they haven't even imagined.

363
00:38:05,000 --> 00:38:13,000
And if you've got your hands dirty, if you've built certain tools, then and you can share that with people, it suddenly becomes clear to them.

364
00:38:13,000 --> 00:38:27,000
Like, hey, I could use it in this way. If they see how it can be used, if they see how it's useful, then that speaks way more to the regular user than just telling them you could potentially do this or you could potentially do that.

365
00:38:27,000 --> 00:38:33,000
Like take the potentially away and make a real use case. And that speaks volumes.

366
00:38:33,000 --> 00:38:39,000
This is amazing. Thank you for sharing all of these use cases with me personally.

367
00:38:39,000 --> 00:38:45,000
Like I can't wait to test it out. I'm just so curious to go to dive into that now.

368
00:38:45,000 --> 00:38:50,000
But I think with that amazing tip you just shared, I think we could have a closing.

369
00:38:50,000 --> 00:39:01,000
So I would very much encourage people to visit Stuart's sub stack, which is use AI for easy and accessible application examples.

370
00:39:01,000 --> 00:39:07,000
Anything else that people can find you? Any last tips you want to share?

371
00:39:07,000 --> 00:39:13,000
You can always find me on Discord as well, the Learn Prompting Discord channel. I'm out there quite frequently.

372
00:39:13,000 --> 00:39:16,000
So if you have any questions, feel free.

373
00:39:16,000 --> 00:39:23,000
Awesome. And with that, I thank you so much for stepping in. We hope this gets better.

374
00:39:23,000 --> 00:39:27,000
But we've been actually talking with Beth for a long time to have you on the podcast.

375
00:39:27,000 --> 00:39:37,000
So this was just like meant to happen. And it was absolutely amazing to learn from you and talk to you and jump on calls to collaborate with you.

376
00:39:37,000 --> 00:39:43,000
So thank you so much for that. Yeah. Happy to have been on the podcast.

377
00:39:43,000 --> 00:39:47,000
And with that, everybody, happy prompting.

378
00:39:47,000 --> 00:39:50,000
Happy prompting.

379
00:39:50,000 --> 00:39:57,000
Thanks for listening to How to Talk to AI with your hosts, Go2Go and Wes the Synth Mind.

380
00:39:57,000 --> 00:40:04,000
As always, you can check out the show notes and links at howtotalk2.ai.

381
00:40:04,000 --> 00:40:14,000
That's all for this week's episode. Happy prompting, everyone.