Free YouTube Transcribe

Video transcript

Advanced Prompt Engineering for Cyber Security | Lawrence Home

BSides Sydney · 5,680 words · 26 min read

Want to search this transcript, jump the video from any line, or download it as TXT, SRT, or VTT?

Open in the transcript tool

Full transcript

0:05morning everyone welcome to bside Sydney

0:0820124 uh we're going to start off with

0:10our first um speaker here Lawrence to

0:13talk about gener of AI and some um

0:16engineering and cyber security for AI uh

0:18just a quick thing about the rooms and

0:21everything in this space there's a

0:22couple of um green rooms back here

0:25they're only for speakers so don't go

0:27into those rooms there plenty of um uh

0:30talks in this level the next level up um

0:32you can walk around most of those rooms

0:34a couple of places that are off limits

0:35they'll be well um shown and um call it

0:38on every floor Etc that's it have fun

0:41enjoy and I'll uh pass it over to

0:43Lawrence and um and let him go through

0:46this talk

0:54enjoy thank you good morning everyone

0:57thank you for being here before we dive

1:00in I want to address something important

1:02about the scope of today's

1:04presentation by the way if I am not

1:07speaking speaking clear enough for you

1:09please let me know don't be

1:13shy you might have read the description

1:15that that promises comprehensive

1:18coverage of advanced prom engineering

1:20for cyber security

1:23operations let me explain to

1:26you trying to do that in 40 minutes is

1:29going to do you a this

1:31service and why I have chosen the

1:33different

1:34approach first of all let us consider

1:37the the famous saying teach a person a

1:40uh give a person a fish you feed feed

1:43them a

1:44day teach them how to fish

1:48you give them fish a

1:53lifetime so the principle of this

1:56particular principle is very appable to

2:00Cy Pro Engineering in cyber

2:03security so if I try to talk about give

2:08in 40 minutes give you specific problems

2:10in every cyber security

2:13situations essentially giving you a fish

2:16it's going to cause some problems number

2:19one AI evolves quickly what I talk about

2:22today may be outdated very soon number

2:28two you it might not be applicable to

2:31your particular context or situation and

2:35number three you might not learn

2:37anything about how to create your own

2:39prompts for your own

2:42situations

2:44so I want to instead of giving you

2:49a catalog of prompts I want to focus on

2:52building un fundamental understanding

2:56giving you the skills to develop your

2:58own prompts so let us be let us begin

3:01let's start off with a beautiful woman's

3:08picture Okay I want to be on the same

3:10same make sure everybody's on the same

3:12page prom engineering is about asking AI

3:15chat bot like chat GPT question or

3:18requests and hopefully you get a good

3:20answer you're looking

3:22for so one day I receed SMS from a woman

3:25who looks at her she asked to be my

3:27friend so I have my ads so I asked CH

3:31GPT for help and this is a

3:35question so I got this asms from

3:40T by the way does anybody here not know

3:43whom T is okay I assume everybody knows

3:46who she is so I asked her I asked GPT

3:51she wants to be my friend friend I have

3:53never met her before never talked to her

3:55before is she for real or am I have I

3:58found the one and thankfully C GPT Saved

4:00Me From Myself it says most likely a

4:03scam so this is very good use of C

4:08GPT Oh by by the way you you guys who

4:11are late you miss the goodlook woman

4:14part so sorry man next time show out on

4:16time

4:17man so why use chat GPT I'll give you a

4:20very good

4:21reason you know everybody wants to be

4:24want to look smarter than they are so

4:26there are three ways you can do that

4:28three ways that that I know of number

4:31one you speak English with an English

4:33accent anybody who can speak English

4:35with an English

4:37accent if you are in America or you are

4:40what you going to be popular in

4:41America they think the English accent is

4:44Posh uh you ask you learn to ask better

4:47questions and number three of course you

4:49learn to use chpt and when I say chbt

4:52I'm talking about chpt and his

4:55competitors so these are the main

4:57competitors and recently I found this

5:00website you can actually ask one

5:01question and get a different answer from

5:04different models so you have model a

5:06Model B and you have and you can compare

5:08which one is is a better model for

5:10answering your

5:12question so some questions I will try to

5:14answer

5:16today

5:18limitations

5:20challenges why it is

5:23important why clear and precise

5:25instructions are important for PR

5:28engineering and arguments for Learning

5:31and why some people struggle with prom

5:35Engineering in this a summary I want you

5:38to take away

5:39from AI is not an option it's the future

5:44you don't have a

5:45choice the usefulness of cck GPT is only

5:49up to your

5:52imagination remember garbage in garbage

5:54out and of course do not ignore

5:58Google and the difference between a

6:00basic prom and advanced prom is an

6:03academic paper because every single

6:06Advanced prom I have come across all

6:08started with an academic paper

6:12and they use magic words in their paper

6:16to prove or to try to prove that you by

6:19using the magic words you can get better

6:21output from from prom

6:25engineering

6:27contents so what we will not cover today

6:29what is prom engineering we covered that

6:32second Point earlier what makes a great

6:35basic prom prompting framework it's not

6:37your fault or is it your fault some

6:41examples so what we want to cover today

6:43we're not going going to cover all

6:45these Concepts that have something to do

6:48with Advanced prom engineering I only

6:51have 40

6:52minutes uh and so let's cover the next

6:56one prom prom framework what makes a

6:57basic prom good basic prom I just want

7:00to be sure we're on the same page so let

7:03me go this through this quickly so give

7:06you the Persona maybe you can ask it to

7:08be a doctor and ask ask it to explain to

7:11you this medication be specific with

7:14your request provide context about why

7:17you're asking the question provide

7:20specific constraints like how long is

7:21the output make it conversational like

7:24talking to a

7:26friend and have a clear and achievable

7:29goal

7:30before you ask a question do you want

7:33a what kind of format you want for

7:37example and this zero shot prom is a

7:40basic prom basically because zero shot

7:44prom is zero example you don't give AI

7:47an example to help you answer the

7:50question you are relying on the training

7:53of the AI to give you the example or to

7:57give you the answer

7:59so here we don't have a example next

8:02slide we're going to show a one shot

8:04prompt or one example

8:08prompt actually we skip that for now

8:10okay prompting framework there are a lot

8:13of them this is one example clear so

8:16these are the if you follow this one

8:18your prom should have all these

8:20ingredients next

8:24one you can just Google and see what

8:27they are if they are more or less the

8:30same

8:34ingredients it is not your fault that

8:37you cannot get the perfect prompt with

8:39AI check

8:40GPT why because it's very sensitive to

8:45the

8:46wording it's not what you say is how you

8:49ask how you say it so changing your

8:52phrase instruction

8:54can significantly improve performance

8:57the most effective changes are in these

9:00two

9:01features

9:03morphology look at the changes just by

9:06changing from anal analyze to analyzing

9:09or analysis could affect your output and

9:13also

9:17lexicon and here if you add Extra Spaces

9:20capitalization using the limiters I will

9:23explain what the limiter is later it's

9:26for now just know that basically it's a

9:28boundary that you set between

9:31your question and your data is something

9:35that make it clear to the AI which one

9:38which part is your question and which

9:40part is your

9:44data so there are a lot of assumptions

9:46that have been challenged According to

9:48some academic papers assumptions that we

9:51have about

9:53AI lower perplexity is no guarantee that

9:56you'll get better output perplexity is a

9:59measure of the AI model in is predicting

10:03in predicting a sample of tax in in

10:06coming up with an output in theory the

10:09lower the perplexity the better the

10:11output but that's not always guaranteed

10:13so this is one assumption we had

10:16about AI that uh that's being

10:19challenged okay using simple words is no

10:23guarantee that you get a better output

10:25sometimes if you use the rare words that

10:27that is in the training or of the AI you

10:30get better

10:32outputs

10:34and also the Lang of your output of your

10:38question the frequency of your keywords

10:40how clear your instruction is is no

10:42guarantee you'll get better

10:45output limitations and constraint there

10:47are more of them here so last year I ran

10:50this early last year when CH GPD just

10:52came out and this Pro that CHT is not

10:56particularly good at logical thinking or

10:58mathematics

11:00I think it's still the same case I

11:01haven't tried but this is one

11:04fundamental weakness about AI cheat

11:08GPT something to keep in mind when you

11:10do your

11:13prompting if you put your keywords in at

11:16the beginning or at the end it's better

11:17than putting it in the middle because

11:19the chat GPT might skip

11:24it and we came across this earlier

11:27adding Extra Spaces limit uh

11:30calization using the

11:32limits and this is the most important

11:35thing you want to

11:37remember when you ask a generic question

11:40cat GPT has a lot of options to choose

11:42from about how to answer your

11:45question in his from his training so if

11:48you ask a generic a general question C

11:50GPT has so many different options it's

11:52going to give you General answer so you

11:54want to focus check GPT on a specific

11:56domain that you interested in so if you

11:59are

11:59cyber security expert you want want to

12:01ask about incident respon then say you

12:05are an incident respon expert give it a

12:07Persona then you will focus on that

12:08particular domain in this training and

12:10then you will come back it it will most

12:12likely come back with

12:14a a better output better quality

12:19output so basically right now Che GPT is

12:23a black box I mean there are things we

12:26know about cat GPT and how they trans

12:28transform input the from in their the

12:30data their training data into and output

12:33there there are things we know about but

12:35there are still a lot of things we don't

12:36know about that's why if you you should

12:39you should understand what this one

12:41means some people have say check GPT is

12:44a troll and I can agree with

12:49that so now we're going to cover is it

12:51your fault that you cannot come near a

12:54come up with a near perfect

12:57prompt by the way if you if you think

12:59you're missing out downstairs their talk

13:01is on YouTube one of the speaker gave

13:03the same talk last a few months ago in

13:06Las

13:09Vegas uh if is it your fault that you

13:11cannot get to per near perfect

13:14prom I have a few theories why some

13:16people have uh are struggling with Form

13:21Engineering Theory number one remember

13:24we came across this

13:26earlier ask better questions

13:30the way we ask questions in life is very

13:32similar to the skills we use in prom

13:35engineering and this

13:38is pointed out by this paper I don't

13:41know whether it's peer review but it

13:42sounds right up my

13:44alley the art of creative inquiry the

13:47conclusion the skill you use in asking

13:49better question every day is the same

13:52skills you use in prom

13:56engineering

13:57now and another another theory I have

14:01about why people struggle with some

14:03people struggle with prom engineering is

14:05if you don't like writing if you dislike

14:08right thing you're going to have problem

14:10I think

14:13because

14:14you interacting with C GPT right now

14:17requires a lot of back and

14:19forth so there are some examples I want

14:21to go through let me see where we are

14:23right now oh 9:14 w wow

14:27good okay

14:29zero shot we came across this earlier

14:31zero example zero shot means zero

14:34example that means you are purely

14:35relying on the training of the training

14:38data of the AI to give you the

14:40output hopefully they will give you the

14:43one that the output that you

14:46want the next shot is going to show you

14:48the one one

14:50example prompt so on the left side the

14:54example we're talking about is a the

14:57answer is 11 so on the left side here

15:00it's actually called One Shot One Shot

15:03prompting and because it has one example

15:06so here you ask a question and then you

15:08give an example of how you how AI should

15:11arrive at the answer and then you ask a

15:13different question hopefully AI is smart

15:15enough to follow your first example and

15:17come up with an answer but unfortunately

15:20in this case it does it doesn't work so

15:23some people came up with the Chain of

15:25Thought

15:27prompting so in this example you

15:29actually show AI how you step by step

15:33come to your conclusion and then

15:36hopefully AI smart enough to follow your

15:38step by step and come up with a answer

15:41and in this case it

15:43works we are entering Advanced prom

15:47engineering area

15:49now so the different as I mentioned

15:52earlier a difference between a basic

15:53prom and an advanced prom is an academic

15:56paper because every single Advance prom

15:59I have come across so far started with a

16:02acade academic

16:04paper I'll talk about magic words

16:07later this is another Advanced PR

16:11I'm I'm very happy that I discover let

16:14me show you why this is a basic problem

16:16you ask CBT to come up with a python

16:20code based on his training data and the

16:23next line is the magic words in this

16:26particular strategy

16:29you ask AI to review their answer and

16:31come up with another answer find

16:33problems and then you ask AI

16:36again so this is from the paper they

16:40started they come they came out with

16:41this RCI

16:43thing here's the

16:46paper oh by the way uh at the end I will

16:49give you a link to Google Docs and you

16:51can find all the references in Google do

16:56uh and you can find all this

16:59pictures in in the reference in the in

17:02the academic papers so here we have us

17:04one questions ask and then we down here

17:07we have four different Advanced prom

17:09engineering well three Advanced prom

17:12engineering technique and one standard

17:14basic prompt so this question Val earns

17:17$5,000 a month so all these four

17:21prompting techniques trying to come up

17:23with the answer and they are over here

17:25on the left side they all wrong and RCI

17:28which we came across

17:30earlier just by asking this second

17:33line which is I highlighted here it give

17:37you a better output down here so this is

17:40one of my favorite

17:41prompt prompting

17:44technique and also this is another one

17:46of my

17:48favorite okay here it's a bit

17:52overwhelming okay here you have two

17:54parts initial problem your initial

17:56problem and then you have instruction in

17:58the instruction you have two P two steps

18:00or in the

18:02instruction first one first part you are

18:05asking AI to come cat GPT to give you

18:08five

18:09examples in the example AI chat GPT will

18:12give you the question and the answer of

18:14each example and then the next step chip

18:18was by looking at these five examples it

18:21came up with it going to try to solve

18:23your initial problem which is this one

18:26here I'm going to show you what the out

18:28looks like so you have you have a better

18:31idea so here here here cat GPD came out

18:34with five example I only have two up

18:36here because too long so in the in each

18:39example there's there's question and

18:41there's answer question and then there's

18:43answer chpd tried to find in his

18:46training similar relevant an analogous

18:49problems that you have compared to uh

18:53anal analogies to your initial problem

18:55and down here C GPT will try to solve

18:57your initial problem based on is five

19:00example up there it's very

19:04useful oh I'm used by the way I'm using

19:07this to write a a movie script so next

19:11time you see me I'll be on I'll be

19:14accepting my

19:15oscar make sure you chewing

19:20in okay another Advanced prom

19:23engineering techniques very simple one

19:26can ask AI to relax just relax

19:30supposedly it will improve improve the

19:33output by the way chat GPT they don't

19:36have feelings they are not humans some

19:39people think they do but they are not

19:40humans they are just predicting the

19:42words that they that they should use to

19:44to reply to your

19:47request okay favorite compon components

19:50of magic words we came across magic an

19:52example of magic words earlier so these

19:55are some of the magic words I like to

19:57use

19:59each one of them each one of them has a

20:01academic academic paper behind them so

20:04this is a personality you are giving AI

20:06chat

20:07GPT and then you

20:10have this one I usually ask at the end

20:13of the of my

20:16prom this is another one I use uh the

20:20limiter we talked about earlier

20:22boundaries that we set to distinguish

20:24your question from your data so I use

20:27this one usually as a my

20:29delimiter to separate my question from

20:31my

20:34data this is the one we came out came

20:36through came came from uh we look at in

20:39RCI

20:42earlier and this one I also use a lot

20:45because I usually find very good

20:50improvements and this one this is

20:52another theory I have about why some

20:53people have

20:55problems with learning prom engineering

20:59if you are not used to challenging

21:01Authority and if you look at CH GPT as a

21:04Authority on everything you're going to

21:06feel not very comfortable challenging

21:08Authority so if you are for example born

21:12and bred in a country like communist

21:15China where the the authority is always

21:17correct then you're going to have I

21:19think you will have a lot of

21:21problems getting the hang of

21:25GPT and there are dozen or hundreds of

21:28different promoting techniques in this

21:30paper they came up they distill do 15

21:34that they say are relevant

21:37for programmers but I think they are

21:40also very useful for other domains not

21:43just programming because all this you

21:45can

21:46see in other domain as well so if you

21:49want to start with 15 techniques start

21:53with this 15

21:58okay so now we going to ask one example

22:01one one question of chat GPT and then we

22:03see how diff some different Advanced

22:07prompting techniques will give you we'll

22:09phrase we'll we'll phrase the

22:12question so here we're going to ask CH

22:15chity to generate some python code so

22:19here on the left side you have some BS

22:21one two three four but eight eight

22:24different prompting Advanced prompting

22:25techniques over here are the examples

22:28we're going to focus on these two in the

22:30middle zero shot Chain of Thought and

22:32RCI which we talked about earlier so the

22:35same question as you can

22:37see is phrased differently depending on

22:40the technique that you use so with zero

22:42short Chain of

22:43Thought the magic words is less things

22:46step by step so just by asking a simple

22:49question less things step by step you're

22:53going to get

22:54supposedly get a better output and then

22:58here with talk about this earlier so I'm

23:01going to skip it but do you have

23:03question about this Slide by the way I

23:05know it's a bit

23:06overwhelming you can find in this

23:09paper if you need

23:11proof when you are tired of watching TV

23:15yes let me introduce you some some

23:18weekend weekend

23:23reading so different domain different

23:26papers I got to make sure

23:29I'm same

23:33distance yeah in this paper is in the in

23:36the references list I'm going to give

23:38you at the end of this

23:40talk sorry you want to take a picture

23:45ready let me know later on if you want

23:48to take picture of any slide I'm happy

23:50to go back here has another

23:52one it's going to keep you out of

23:54trouble for a while

23:59yeah and this one as

24:04well this is a very good paper you want

24:07to start with the paper start with this

24:09one my one of my favorite

24:13papers a bunch of people like at least

24:1612 12 authors in that

24:18paper thought process of a prom engineer

24:21if you want something even a bit more

24:23advanced yeah this one here you will

24:25give you the the an

24:31example this this is from the same paper

24:33I talked about earlier one of my

24:35favorite papers is in the appendix

24:38section so one book to read okay because

24:42I wasn't planning on having so much free

24:44time right now

24:46after after all the slides so I I I I

24:50didn't I want to introduce you this book

24:53for cyber Security Professionals this

24:55book is written about CH GPT prompting

24:58for cyber Security

25:01Professionals I could

25:04have copy and pasted all the prompts but

25:07I think it's better for you to read this

25:10is the

25:16book it covers a lot of

25:21domains it's in the reference list here

25:24I will show you I'll give you later so

25:27here here are the chapters

25:34and this is an example prompt from the

25:39book so here he's using a personality

25:42you are professional cyber threat

25:45analyst and this is a very long promt

25:49here the second book to read it's

25:52written by a very handsome man

25:58I have bad news and good news for you

26:00the bad news is I'm quite happy with the

26:03content it's version 1.1 the bad news is

26:07the formatting is really sucks so if you

26:10want to read better use a computer

26:11rather than a a handphone and the good

26:14news is today Saturday 7 P.M onwards you

26:20have 24 hours to download for free from

26:22Amazon yes if you have bad things to say

26:25about this book email me if you have if

26:28you have good things to say about this

26:29book give me a five

26:32star I need to

26:35impress

26:37okay summary we came across this earlier

26:40so as as mentioned earlier AI is not an

26:42option if you well it's the future you

26:46don't have a choice the usefulness is

26:49limited to your

26:51imagination remember this

26:53one and do not ignore

26:56Google and the de and of course the

26:59difference between a basic prompt and an

27:00advanced prompt is always an academic

27:02paper so if you want to uh to be better

27:05at prompting read an academic

27:07paper let's see here thank you oh here's

27:09the

27:18references any questions

27:29hi uh thanks for the talk uh it seems

27:31like a lot of the prompt engineering

27:33techniques rely or or at least that I

27:36saw rely on uh you knowing the correct

27:39answer you know what I mean like if

27:41you're refining the answer and you're

27:42asking it to check what's wrong say

27:44you're generating secure python code how

27:47do how do you know if you yourself don't

27:49know whether it's secure when to stop

27:51asking it to find problems and how do

27:52you know it's not hallucinating problems

27:54in its previous answer and actually

27:56giving you less secure code

27:58yeah it's a good question that's why I

28:01use more than one AI chatbot to cross

28:04check each other even then it's no

28:06guarantee they they know what they're

28:08talking about so it's always have good

28:10to have a human friend at hand who can

28:12guide you along the way as

28:14well I don't know if that answer your

28:16question but you can never trust

28:20AI maybe you should just ask your human

28:22friend to help you from the start the

28:25problem is your human friend may not be

28:27so so free to entertain your question so

28:29maybe you want to get the basic out of

28:31the way then if maybe you it will help

28:34you come up with a better question then

28:36you're happy so that your human friend

28:38will be happy to

28:43answer any

28:45question I can't I can't I can't I can't

28:48believe every oh but uh before before

28:50you before I answer a question I want to

28:52thank the organizers for allowing me to

28:56be one of the speakers today I'm very

28:58deeply honored to be one of your

29:00speakers today is to be able to speak to

29:04at least 100 I hope a 100 people Prof

29:07highly qualified

29:10professionals is one is was on my bucket

29:12list so that one is

29:15T so and they I believe they put up

29:19together a very excellent event today

29:22and so I'm very very gr grateful to be

29:25part of this yes sir

29:28yeah thanks for the talk man it was

29:30really good I love your accent man oh

29:32cheers

29:34buddy um so look just for Contex this is

29:37actually my first bides event so like

29:39I'm I'm I'm new to a lot of this stuff

29:41um but I myself have been using chat gbt

29:44in like so many areas of my life it's

29:47kind of caught me off guard how often I

29:49use it these days um but look I've

29:53worked help Des for about seven or eight

29:55years now I'm currently been trained as

29:56a CIS admin um chat gbt in in in my

30:01experience in the it side of things at

30:04the moment is

30:05still like it's a bit of a Pandora's Box

30:08in a lot of ways and I think look you

30:10probably understand that better than

30:12anyone um I suppose my question is for

30:15you how far off do you think we are to

30:20adopting chat gbt and AI at large into

30:25just the it space as like a mandatory

30:31option very good

30:33question that question is better

30:35answered by the one the two ladies

30:37downstairs because they're talking about

30:38hacking AI right

30:40now uh but from what I understand

30:43there's a lot of problem

30:45with llm large

30:48language

30:50model so a lot of problems you can see

30:54on on the news about company

30:59having problems in in in in installing

31:02making use of

31:04AI to be honest with you I'm not that

31:06qualified to tell you when we will see

31:08that happening hopefully within my

31:14lifetime but I'm happy to pass on your

31:16question to the two lovely ladies

31:18downstairs yeah that'll be good man

31:20cheers

31:32thanks for that um do you have a view on

31:36which of those prompting techniques you

31:38think gives the best productivity

31:52gains it really depends on the context I

31:55think that's why you want to read as

31:58many papers as possible to to

32:00familiarize yourself oh yes of course

32:03read my

32:04book because my book actually is a

32:07collection of many different examples

32:09from different sources some of them you

32:11have to pay so you save you some money

32:14by going to my book and actually I

32:17categorize the prompts according to

32:19functions like

32:21extracting words or summarizing or

32:25problem solving stuff like that you know

32:30of so really basically you want it

32:34really depends on context and so you try

32:35different techniques and the ones that I

32:38the the favorite ones that I show on the

32:39slide are the ones I use most often yeah

32:43I guess they are my favorites and they

32:46the one that work for my context yeah I

32:49hope I answer your question

33:04what AI

33:06or chat GPT would you recommend for

33:10running on pram like off the

33:14internet I hope somebody here knows this

33:17more than me because to be honest with

33:19you I haven't tried any one of them this

33:21is on my on my bucket

33:23list uh

33:35I don't want to

33:46invest but some of the papers they

33:49actually talk about the the the machines

33:52that they use so maybe from there you

33:54can get some

33:55Inspirations also of course remember

33:58this

34:00website go

34:04back why is this not

34:08working

34:10stuck

34:13okay

34:19the

34:22remember remember this one yeah you can

34:25test the same question on different

34:27seven models according to the website

34:30and see which one is better uh just to

34:32answer the previous question if you're

34:34interested in running like a model like

34:37an AI locally like chat GPT um you can

34:41look up oama or you can look up uh

34:44hugging face um which is like GitHub for

34:47AI models um and then there's a program

34:50called LM Studio which lets you test

34:53them and and have a look but o Lama or

34:55hugging face um on Google and that'll

34:57give you the answers oh this one

35:05yes good day thanks for the talk U just

35:08a question um is there an inherent bias

35:11against AI especially for cyber security

35:14that we expect AI to be right all the

35:17time and we sort of say that it's bad if

35:20it's not and it shouldn't be trusted but

35:22in reality most cyber people aren't

35:25right most of the time in in in all

35:27things so we have a higher expectation

35:30for AI to be write secure code yet in

35:33reality very few people write secure

35:35code uh but we is is is there any

35:39herab I think we do I mean we expect too

35:43much from AI there's a lot of hype about

35:45AI maybe that's one of the problems yeah

35:49so we have to be realistic so the more

35:51you play with a the more you find that

35:53it's not so intelligent I mean CPT

36:05what's your thought with chpt but o one

36:08the new one that is thinking because

36:11also with the I tried like a basic

36:15prompt and the reply was very good

36:18because he's thinking instead of let me

36:21doing like a a better prompt you're

36:25talking about

36:28no no not the paid one I I guess because

36:30I'm paying probably I have access is the

36:3401 basically when when I ask when I do a

36:38prompt it takes maybe 10 15 40 seconds

36:42to reply so he got all the steps you can

36:46see all the steps and the reply is

36:49incredible comparing with the 40 what's

36:53your thought about that compared to the

36:55older version yeah so you're happy with

36:57the new free version no I'm using 01

37:01you're using the o one o1 it's a free

37:03one right ER I don't think so you're

37:07you're paying for an old one yeah I

37:09guess because it's a preview but H AI is

37:14thinking once you ask a prompt when you

37:17digit a prompt then doesn't reply

37:20straight away wait wait you can see what

37:24what he AI is doing and then you're

37:27going to give you you a a very detailed

37:29reply okay are we talking about chpt the

37:32brand yeah chpt open AI chpt oh open AI

37:36yeah I have in the phone the browser one

37:38or the the in the app in the phone in

37:42the browser yeah the one in the browser

37:45not the one that you run through the API

37:48no no no in the browser okay so you're

37:50saying you using you're paying for chbt

37:52and you're happy with the answer hey

37:54it's just $20 is not uh oh okay so but

37:58the new one is thinking thinking so with

38:01a basic prompt yeah basic I mean yeah

38:04you have a very good reply anyway yes

38:07correct so you're saying why should I

38:09bother learning no no no no no what what

38:11what's your thought because AI is going

38:12quick yes yes I mean we have a lot of

38:15paper but if it's from let's for me my

38:18opinion like from

38:202023 for me sounds very old yeah Am I

38:25Wrong Here's my thought because prom

38:28engineering as I mentioned earlier help

38:30you develop your questioning skill so in

38:33real life when you don't have cat GPT

38:35next to you you have to learn to ask

38:37better question tra using chat GPT is a

38:40very good way to train yourself to make

38:42you smart look smarter in the conference

38:45table around the conference table so

38:47that's why whether we have whether no

38:50matter how good chat GPD gets to become

38:52in answering your question I believe the

38:54training that you get in learning how to

38:56ask better questions

38:58help you in real life when you don't

39:00have internet

39:02connection I don't know I hope I hope

39:04they answer your question

39:11yeah can try Forex I from Google and um

39:16they have the option that you can change

39:18the parameter to uh turning your model

39:22so I I tried that one already and um yes

39:27it's thinking so it's not just reflect

39:30um immediately and uh take more two or

39:34three second or sometime 10 second based

39:37on your complexity of your questions

39:39yeah

40:03I think um when people talk about the AI

40:05thinking I think it what open AI are

40:09doing is basically just automating the

40:11recursive process so it's starting with

40:14an input and an output and then feeding

40:16it back into itself looping it a few

40:18times and so that's where I think like

40:20prompt engineering is a skill it's

40:22something that the product people open

40:25Ai and Google they don't want end users

40:26to be good prompt engine is they want it

40:28to automatically work and so they're

40:30building this Research into the layers

40:32where it just does it automatically for

40:35you thank

40:37you any more

40:40question I hope I I I haven't wasted

40:42your

40:45time40 okay thanks for coming

This transcript was generated from the captions YouTube publishes for this video. Get the transcript of any YouTube video atfreeyoutubetranscribe.com: free, unlimited, no sign-up.