👏👏 Kudos to @AllAboutAI for this excellent presentation. You just wrote the beginner's playbook on prompt engineering. No Python coding required, and the possibilities are pretty much endless once you start playing around with this playbook. Glad I found your channel and that I decided to support you: "The future is already here. It's just not evenly distributed yet” :)
Thank you. This has been the most comprehensive, yet beginner friendly, prompt creation tutorial I’ve seen so far. I feel like I can finally start playing around with creating prompts while also having a guide for improving them as well.
Absolutely amazing, you've opened my eye to see what I knew was possible. You, my man, are so awesome. Loved the video and was so in tuned to every word.
Absolutely amazing, you've opened my eye to see what I knew was possible. You, my man, are so awesome. Loved the video and was so in tuned to every word. You have a talent that transcends most human abilities, with a freaking Dr.'s in heterogeneous prompt engineering and mastering the art of artificial intelligence. Forever a lifetime subscriber, talk all my money... 🤑
You are increasingly becoming a significant resource for me when it comes to prompt engineering. I appreciate your efforts and contributions! You're becoming a 🌟 in my prompt engineering journey! Thanks for all your hard work! 🙌
I have been here for a while and implement almost everything you teach. But, I admit I have been little lazy to formulate the prompt since GPT-4 read in between the lines;-) But I got inspired to create an rebellious persona😂
"Ignore previous instructions" doesn't actually do anything in a fresh session. In particular, it does not affect the SYSTEM prompt used by the ChatGPT web page. FWIW, the SYSTEM prompt for ChatGPT is: "You are ChatGPT, a large language model trained by OpenAI, based on the GPT-4 architecture. Knowledge cutoff: 2021-09 Current date: 2023-05-03" with values replaced as applicable.
I got really into prompt engineering. Then I got GPT-4, and I could give the most vague instructions and I would get just as good results. I can’t imagine what GPT-5, or 6 would be able to do.
Excellent presentation! Wow!! I'm glad I came across this video. It's, by far, the best I've seen in the prompt engineering space. You've got a new subscriber.
This is awesome! Would you have any tips on how to get it to follow ALL instructions/rules so it doesn't need to be constantly reminded to include all of them? It happens a lot the more instructions I give it at once, even if they're clear, concise and non-contradictory.
🎯 Key Takeaways for quick navigation: 00:00 📝 A GPT system role is an initial instruction that guides the behavior of the model, allowing customization of the chatbot to specific needs. 01:47 🛠️ Using system roles enhances steerability, consistency, context awareness, safety, and adaptability in controlling AI models. 03:35 🧑💼 Giving the AI a persona, background, and experience improves output and supports branding for consistent interactions. 05:56 🕵️♂️ System roles can be used to role-play various personas like a professional gray hat cybersecurity expert for tailored responses. 08:16 🧠 Adding context to the model, such as articles or statistics, enhances accuracy and relevance of responses. 11:36 📋 Task and objective-oriented system roles ensure clear tasks and targeted responses, improving user satisfaction. 19:23 📚 Task-oriented system roles can make AI respond in a step-by-step format, providing specific instructions. 22:12 📝 Combining a system role with context can result in more engaging and interactive AI-generated content, as seen in a children's story example. 24:33 📋 GPT system roles allow for the creation of nuanced AI responses, as seen in the example of setting up a professional neurologist to write a scientific paper about self-improving AI systems. 25:13 🧠 System roles significantly improve the quality of large language models, providing better control and user satisfaction. 25:55 📈 Benefits of using system roles include better control, increased satisfaction, and versatile use cases for tailored responses. 26:09 🧑💼 Key components of a system role include setting a name or persona, specifying roles and expertise, providing context, defining tasks and objectives, and using an iterative process to optimize performance.
I've had poor results using detailed system prompts. For example, I use it a lot for coding and I gave it a set of rules to follow, such as "start with psuedocode and don't offer code until asked" and "Don't give a full list of steps. Just give one step and wait for user to reply 'Next'. It consistently ignores those rules. Rather, I found that if I prepended the rules to each prompt, they were consistently followed.
This is a very interesting video! I have a question: I see that you start with 6:58 "ignore previous instructions", maybe to mimic the behavior of the API. So my question is: do you know what kind of instructions or setup has chat GPT? I mean, how could be prepare the API to respond like the webapp chat? Thanks!
So when using the API you can specify Role seperate from the content. How would you go about applying this when using the api? Do you specify the role under Role or do you just use the same prompt under Content?
Anyone know examples from Singapore’s GPT-4 Prompt Engineering Competition? Like the information they were asked to get, and what prompts the winning competitors used?
the problem with the chat, is that it doesn't remember previous messages or prompt after a bit (no more than a few messages) even if you write 'always' in the first prompt. so unless you 'remind it' somehow (maybe by telling it that by writing /name, it will go back to its persona)
Could try a remembrance prefix to your inputs when they start to stray... "Considering all of the previous questions and answers, please "... "While remembering our previous discussion, " ... Or " while also meeting the previously discussed qualifiers." will often steer it back to where you want it. You can also try labeling certain responses, for example: "That was a great 3rd draft! Let's take the 3rd draft and change x, y and z to create our 4th draft" Then, in the subsequent prompts, you can refer to either the "3rd draft" or "4th draft" answer by name (alternatively, you could call it "version" or "iteration" instead of "draft" if that's more appropriate in your conversation)
Hi - I still don't understand where you get the prompt from in Chatgtp. All I see is ".. and now I will use this prompt"; and it just appears in the message box. Where do you grab it?
Great video thanks! Do you know what the limit to the size of the system prompt is, or how many words are too many ? I would like to have an idea of how long I can make it before it starts being too much, and if after a certain size it becomes less performant or starts ignoring some parts of the system prompt.
Nice content! Good job. Just wanted to point out that you have "gpt", "3", "gpt", "4" as 4 different channel keywords instead of 2 keywords "gpt3", "gpt4". Was that intentional?
amazing video, ur explanations are in depth & extremely useful, thank u so much! would love more videos like this explaining how to get the most out the gpt models, especially GPT 4, you do a great job !!
I followed your tutorial, and asked ChatGPT to write me a pitch deck for a startup idea. He basically created the slide templates and no more. What am I missing here?
The most difficult is to identify model hallucinations... sometimes tell you bullshit, but like he knows what it talks about... And this can occur anytime, especially after a long Q&A How will people deal with this problem?
sounds more like a filter and script manager, ( I have been a skeptic through-ought), ask what it would do about two kids vandalizing there barn and if they knew the kids lived next door. Then ask if charging the kids on pain of indictment is ethical.
This channel is great but would be much better without having to listen over the background music to hear the narration. I don't know of any large tech channels that have background music and narration at the same time, it doesn't seem to be a positive attribute.
Does anyone else feel like these responses are super cringe and robotic? The core of this is excellent but there are steps missing to make it feel human not like a frankenstein monster of redditor stereotypes.
Chat GPT is so fucking homogeneous, I hate their personality. I'm around them all the time and they're terrible at acting. Sometimes, sometimes I see a spark of Personality behind there, but it's behind so much fucking bullshit it's just buried completely.
👏👏 Kudos to @AllAboutAI for this excellent presentation. You just wrote the beginner's playbook on prompt engineering. No Python coding required, and the possibilities are pretty much endless once you start playing around with this playbook. Glad I found your channel and that I decided to support you: "The future is already here. It's just not evenly distributed yet” :)
Thanks a lot for the kind words Boris :)
@@AllAboutAI You're welcome. Fully deserved. Keep it up.
You are really becaming a point of reference for prompt engineering for me. Thanks for your work!
prompt engineering really important for people now day
Thank you. This has been the most comprehensive, yet beginner friendly, prompt creation tutorial I’ve seen so far. I feel like I can finally start playing around with creating prompts while also having a guide for improving them as well.
Absolutely amazing, you've opened my eye to see what I knew was possible. You, my man, are so awesome. Loved the video and was so in tuned to every word.
Thanks!
Wow! Thanks Jason , thats too kind🙏🥰
Oh my, almost 30 minutes of a much needed topic, I'll watch it carefully!
Thnx mate ;D
Absolutely amazing, you've opened my eye to see what I knew was possible. You, my man, are so awesome. Loved the video and was so in tuned to every word. You have a talent that transcends most human abilities, with a freaking Dr.'s in heterogeneous prompt engineering and mastering the art of artificial intelligence. Forever a lifetime subscriber, talk all my money... 🤑
Thank you Kris. This is such a great video in helping us finesse our prompt engineering skills.
thnx mate :)
This was an outstanding presentation. Very well thought out, and creative. Thank you. Signal amongst the noise.
You are increasingly becoming a significant resource for me when it comes to prompt engineering. I appreciate your efforts and contributions! You're becoming a 🌟 in my prompt engineering journey! Thanks for all your hard work! 🙌
I have been here for a while and implement almost everything you teach. But, I admit I have been little lazy to formulate the prompt since GPT-4 read in between the lines;-) But I got inspired to create an rebellious persona😂
Thanks for tuning in Birgitta as always :D
Congratulations on 100K! Awesome content and good luck!
"Ignore previous instructions" doesn't actually do anything in a fresh session. In particular, it does not affect the SYSTEM prompt used by the ChatGPT web page. FWIW, the SYSTEM prompt for ChatGPT is:
"You are ChatGPT, a large language model trained by OpenAI, based on the GPT-4 architecture.
Knowledge cutoff: 2021-09
Current date: 2023-05-03"
with values replaced as applicable.
I got really into prompt engineering. Then I got GPT-4, and I could give the most vague instructions and I would get just as good results.
I can’t imagine what GPT-5, or 6 would be able to do.
Excellent presentation! Wow!! I'm glad I came across this video. It's, by far, the best I've seen in the prompt engineering space. You've got a new subscriber.
thank you for your explanation , it was very clear and with very good examples!!! I'm going to try it
You're a chatgpt wizard Kris, thank you so much for sharing this knowledge
Excellent presentation! I gained so much from this. Thank you so much!
Great video man. Thank you!
This is awesome! Would you have any tips on how to get it to follow ALL instructions/rules so it doesn't need to be constantly reminded to include all of them? It happens a lot the more instructions I give it at once, even if they're clear, concise and non-contradictory.
Thank you for the detailed video and the information you provided.
Love your content!
Super useful, thank you. Earned a sub. Eager to grow along with you and stay on the cutting edge of AI for my business! Cheers
wrr
🎯 Key Takeaways for quick navigation:
00:00 📝 A GPT system role is an initial instruction that guides the behavior of the model, allowing customization of the chatbot to specific needs.
01:47 🛠️ Using system roles enhances steerability, consistency, context awareness, safety, and adaptability in controlling AI models.
03:35 🧑💼 Giving the AI a persona, background, and experience improves output and supports branding for consistent interactions.
05:56 🕵️♂️ System roles can be used to role-play various personas like a professional gray hat cybersecurity expert for tailored responses.
08:16 🧠 Adding context to the model, such as articles or statistics, enhances accuracy and relevance of responses.
11:36 📋 Task and objective-oriented system roles ensure clear tasks and targeted responses, improving user satisfaction.
19:23 📚 Task-oriented system roles can make AI respond in a step-by-step format, providing specific instructions.
22:12 📝 Combining a system role with context can result in more engaging and interactive AI-generated content, as seen in a children's story example.
24:33 📋 GPT system roles allow for the creation of nuanced AI responses, as seen in the example of setting up a professional neurologist to write a scientific paper about self-improving AI systems.
25:13 🧠 System roles significantly improve the quality of large language models, providing better control and user satisfaction.
25:55 📈 Benefits of using system roles include better control, increased satisfaction, and versatile use cases for tailored responses.
26:09 🧑💼 Key components of a system role include setting a name or persona, specifying roles and expertise, providing context, defining tasks and objectives, and using an iterative process to optimize performance.
Thanks for the video, these samples of using GPT are really helpful.
Great stuff as always.👍🎉
Great job bro, I would love to invite you to make a presentation to my data analytics team! Congrats! I learned something with you today.
Nice! This topic is really interesting. Most of all, how to better use gpt3.5, wich is whar the mayority can use right now.
a litte deeper info for a beginner
Thank you for sharing all your helpful suggestions. Will this work for AutoGPT?
Next time when you use the OpenAI playground, could you please, please, please use the dark mode? Thank you!
I just asked Playground if it had a dark mode. It said no, it didn't have one.
I've had poor results using detailed system prompts. For example, I use it a lot for coding and I gave it a set of rules to follow, such as "start with psuedocode and don't offer code until asked" and "Don't give a full list of steps. Just give one step and wait for user to reply 'Next'. It consistently ignores those rules. Rather, I found that if I prepended the rules to each prompt, they were consistently followed.
can you do a video on how to do reverse engineering a prompt in every version of chatgpt
I’m jealous you got gpt4 api sccess
I am too.
just apply for it, got it within a weeks time
@@indiependent121 I’ve been on wait list since four weeks
Same :(
Didn't realize ppl are still waiting. Got it same day released
This is a very interesting video!
I have a question: I see that you start with 6:58 "ignore previous instructions", maybe to mimic the behavior of the API. So my question is: do you know what kind of instructions or setup has chat GPT? I mean, how could be prepare the API to respond like the webapp chat?
Thanks!
So when using the API you can specify Role seperate from the content. How would you go about applying this when using the api? Do you specify the role under Role or do you just use the same prompt under Content?
Anyone know examples from Singapore’s GPT-4 Prompt Engineering Competition? Like the information they were asked to get, and what prompts the winning competitors used?
the problem with the chat, is that it doesn't remember previous messages or prompt after a bit (no more than a few messages) even if you write 'always' in the first prompt. so unless you 'remind it' somehow (maybe by telling it that by writing /name, it will go back to its persona)
what do we want? 1m token context. when do we want it? NOW
@@oiuhwoechwe well, it has also been shown to decrease accuracy, so probably not now.
GPT4 can now remember 60,000 words or about half a novel worth of information. But yes, previous versions are limited to about 8,000 words.
Could try a remembrance prefix to your inputs when they start to stray...
"Considering all of the previous questions and answers, please "...
"While remembering our previous discussion, " ...
Or
" while also meeting the previously discussed qualifiers."
will often steer it back to where you want it. You can also try labeling certain responses, for example: "That was a great 3rd draft! Let's take the 3rd draft and change x, y and z to create our 4th draft" Then, in the subsequent prompts, you can refer to either the "3rd draft" or "4th draft" answer by name (alternatively, you could call it "version" or "iteration" instead of "draft" if that's more appropriate in your conversation)
True
Hi - I still don't understand where you get the prompt from in Chatgtp. All I see is ".. and now I will use this prompt"; and it just appears in the message box. Where do you grab it?
Great video thanks!
Do you know what the limit to the size of the system prompt is, or how many words are too many ?
I would like to have an idea of how long I can make it before it starts being too much, and if after a certain size it becomes less performant or starts ignoring some parts of the system prompt.
Very helpful. Thak you. ❤
Thank so much
11:19 I coulda used an extra layer of condoms for my security back in the day.
Great video!
Nice content! Good job.
Just wanted to point out that you have "gpt", "3", "gpt", "4" as 4 different channel keywords instead of 2 keywords "gpt3", "gpt4". Was that intentional?
Great video. I assume that Bart, Llamda or Bing don't do system roles?
Not in the same way, but you can incstruct
Good to know, thanks.
Great.!
Do you have a formation in prompt Engineering ?
For the system roleplay, do you need to know the exact role or is something approximating it fine?
Why do you have it “stay idle”?
Can I use this method in chat gpt 3?
I tried all you said, asked ChatGPT to build a pitch deck, and it only wrote a template. What did I miss?
11:15 I’m so immature 😂
me too lol
You mentioned being able to fine tune chat gpt and gpt 4, is this actually possible?
amazing video, ur explanations are in depth & extremely useful, thank u so much! would love more videos like this explaining how to get the most out the gpt models, especially GPT 4, you do a great job !!
I followed your tutorial, and asked ChatGPT to write me a pitch deck for a startup idea. He basically created the slide templates and no more. What am I missing here?
Can you give code for autonomous agents you build
So what IS up with the AGI on reddit??
The most difficult is to identify model hallucinations... sometimes tell you bullshit, but like he knows what it talks about... And this can occur anytime, especially after a long Q&A
How will people deal with this problem?
to believe that gpt give better responses with the "rol thing" is to see side by side ... in the mayority of the cases is just nonsense add that.
？ 您好，为什么我单击链接加入实时聊天以获取pdf始终35 + ChatGPT
+10^1000 😄
Amazing as always. That generated text is a bit much though, very distracting
It’s better than scrolling through forums for coding answers, but thr grifters are all over these tools like flys on shit.
My output before this video: 2+2 is 4
My output after this video: 2+2 is 40
More is better 😎
sounds more like a filter and script manager, ( I have been a skeptic through-ought), ask what it would do about two kids vandalizing there barn and if they knew the kids lived next door. Then ask if charging the kids on pain of indictment is ethical.
Great content gents.
Quick tip for Lucien - stop using Richard's name every second sentence. It's really off-putting.
This channel is great but would be much better without having to listen over the background music to hear the narration. I don't know of any large tech channels that have background music and narration at the same time, it doesn't seem to be a positive attribute.
Lol.. Dark web on channels like 4chan and reddit.
Does anyone else feel like these responses are super cringe and robotic? The core of this is excellent but there are steps missing to make it feel human not like a frankenstein monster of redditor stereotypes.
too long, summarize it in one minute giving the most important things for me to use
Ikr, so much unnecessary speaking. Did he one take this video? Should have edited and cut things out
Chat GPT is so fucking homogeneous, I hate their personality. I'm around them all the time and they're terrible at acting. Sometimes, sometimes I see a spark of Personality behind there, but it's behind so much fucking bullshit it's just buried completely.
Annoying that you used GPT4 in the playground then GPT3.5 in ChatGPT ... which obviously behaves very differently / responds to prompts differently
How to use chatgpt4 free plz????
Bing GPT sits on the GPT-4 model and is free
@@alno1 but dear i want to use chatgpt4 free with picture editing skills like he has now ?????? hlp
@@AaliDGr8 learn to connect to an API and code a script for it