ChatGPT is the latest GPT style generative AI model from OpenAI, which, at it’s most basic level acts as a chatbot, but this back and forth structure allows for many complex capabilities.
Chat with ChatGPT here:
Operating system example with the script writing to a file:
Neural Networks from Scratch book:
Channel membership:
Discord:
Reddit:
Support the content:
Twitter:
Instagram:
Facebook:
Twitch:
We have to talk about chat GPT Chat gbt is the latest offering from Openai it’s a GPT or generative Pre-trained Transformer style model Specifically designed to engage with you Via natural language the best way that I Can think of to tell you more about chat GPT is to just show you it’s December 2022 and this is a research preview so Some behaviors capabilities and general Characteristics are subject to change The first obvious abilities for a model Like this are to largely replace search Engines for questions for example here We are asking for the steps to change The brake pads on a car and we’re Getting back a pretty solid series of Steps to do just that we could spend a Long time just on the search engine Aspect of you know how good is this Model for this specific purpose but this Is just barely scratching the surface of What this style model and this model Specifically can actually do so let’s go Ahead and increase the complexity just Slightly GPT models can be influenced by What’s called prompting in the Background if you’re trying to make a GPT model produce some specific output Characteristics it’s going to do so Based all on the input leading up to the Point where you want to generate more of Those tokens and in this case it’s words This initial information The Prompt can
Be deliberately crafted to illicit Behaviors that you’re looking for this Is your main job when you’re interfacing With an AI like this it’s not like these Functionalities to be a search engine or Any of the other things that I’m going To show you in this video are pre-built In to the model at the end of the day it Is a language model attempting to behave Like a generation of a chat log that’s What this model is and it’s your job to Figure out how you can structure Problems in such a way that this model Really wants to keep doing whatever it Is that you want it to actually do So now let’s start off by adding just a Little bit of a layer of complexity here And let’s try a silly example where the AI gives us responses to questions Always around the perspective of a dog As you can see throughout our Conversation here chat GPT maintains Responses to queries from the Perspective of a dog and what a dog Might think of XYZ obviously this is Kind of a silly little example but maybe You have I don’t know something like a Food recommendation application where Your customers want to maybe input Things like food allergies and other Diseases that might affect maybe the Food that they would be interested in And you want all of your responses to Kind of skew towards whatever that
Customer is Sometimes you might even want to be more Strict with your GPT models and you may Need to even provide an example to get The specific result or the exact result That you’re seeking so for example Imagine that we have some price Estimator app that we want to build with Chat GPT we can open our conversation by First telling chat gbt that it’s a Pricing Ai and that all responses should Be dollar amounts and nothing more Sometimes chat GPT can get a bit chatty And we want to keep things concise so we Tell it exactly what we want from Responses and then we begin asking Questions As you can see we can start off by Actually asking how much something is But then we can actually just transition To just saying that thing and then chat Gbt responds with the price of that Thing we can do this for a wide range of Items and even add modifiers and a Follow-up and chat GPT even understands What we mean there correctly while we do Have GitHub copilot also created by Openai I have found chat GPT to be in Many ways a far more advanced coding Assistant basically copilot codes with You chat GPT can code for you and There’s a big difference there using Plain language we can just describe the Program that we want this step in the
Development process was actually doable With copilot by just having maybe a Comment at the top of the script that Said exactly what the whole program was Supposed to do but from there if you Wanted to make modification or a change Or maybe an error is happening you’d at Least need to begin to know where to Make that change and often where to Remove code or add code and all of that You would need to be somewhat of a Programmer we’re not getting hints and Suggestions here chat GPT is actually Developing for us After getting these clusters created a Reasonable person might want to one just See them visualized and then two Probably realize that they really want To actually see them colored according To the category or the cluster that they Belong to we can just ask chatgpt to Please color them in this case Interestingly enough we wind up getting An error from the suggested code from Chat GPT now with copilot we would be Probably stuck here this would not be Something we could overcome we would Probably need to either use a search Engine to figure out or we might need to Nudge the code a bit with some of our Own knowledge if we have it with chat GPT though all we need to do is report The issue so just copy and paste the Error that we see
From here chat GPT actually fixes the Issue for us and it’s done it works Exactly as we wanted and we just quite Literally naturally flow through the Entire development process through plain Language I will say most of the time it Doesn’t actually generate an error but If it if it does and I really just Wanted to include this specific coding Example because it did show an error if It does I just wanted to show like how Easily you can work around that because Copilot does the same thing it gives you Errors or it just doesn’t calculate Something right and the same thing is True here like if if it’s if it produces Something and it wasn’t actually what You’re looking for you can describe what It’s doing wrong and it will actually Change the code and again this is this Is this is why I think chat gbt is is What I would call Orders of magnitude Better than copilot because it can act It can engage in that development Process with you and like I said earlier It literally develops for you to your Specification and if you’re having a Problem you just tell it kind of what The problem is you’re not really you’re You might be an engineer but you’re not Programming let’s check out one more Example of chat gpt’s coding for you Ability let’s see if chat GPT knows About Conway’s Game of Life
Then can it just program all that for us Yes it can not too much of a surprise There but probably the coolest part of Conway’s Game of Life is Seeing it let’s ask for that and again All we’re doing here is conversing in Order to visualize the Game of Life from Here in the script a handful of changes Need to be made at different points in The code we need to settle on what Library will we even use and implement It it’s a decent undertaking it’s an Entry-level programmer undertaking but Still it’s not something again that Co-pilot could not do this copying and Pasting exactly what it just gave us Sure enough it works but it’s a bit slow Let’s ask chat GPT to speed this up for Us and what I really like here is not Only does chat GPT speed it up for us it Tells us how it was done in a way that We could easily just do this ourselves Potentially doing this going forward to Set the speed now there’s also a lesser Known but simpler way to update a graph Like this in matplotlib that I happen to Know about so if that is the case and You are a programmer you can do Something like this so I I can ask Chatgpt to use that method instead of The official kind of Animation method Again this will require edits in the Form of new code and removing other code In multiple places to produce this
Result but chat jpt does this very well Continuing on maybe we like this simpler Method but we don’t like the color so we Can just specify the colors that we want Things to be instead again this I Believe would just be like a single line Change but still a change nonetheless Now maybe we change our mind and we want To use animation instead because it Actually is faster to produce these Charts note here that the output was cut Off from this request this happens Sometimes either it might be because of A bug maybe the sequence you asked for Is just plain too long or sometimes the AI actually thinks that was the end of The response you can remedy this in a Variety of ways but typically just Saying continue will be enough it won’t Be formatted very beautifully Necessarily but it will finish it Alternatively you can sometimes ask it To just be more concise and sometimes That will be enough to fit what you’re Looking for in context and for example With code it’ll often do something like Maybe replace it’ll it’ll tell you Replace this giant block of code and it Actually prints out that giant block of Code with this other giant block of code Which obviously is going to take up a Ton of space and cause problems so in That example you might just say just Give me the updated code only or
Something like that with our new code Interval is slow again but we do know How to fix that that said we also might Want it to be bigger like the chart Itself let’s just ask chat GPT to change Both of those things for us Okay fix that but it seems like our Chart is back to a light theme we can Also just ask it to make it a dark theme Again and chat GPT understands that we Want that black background with cyan Colored cells again and that’s just Based on the history of our chat and That was actually a few chats ago we Didn’t need to like re-specify the Colors Um this is subtle but it’s absolutely Massive for just naturally conversing uh With with I guess in this case your Software but this is probably how you Would converse with that entry-level Programmer that you just replaced so Similar to previous examples already I Could definitely dive deeper into just This specific like coding in Python Example but for now let’s move on to yet Another interesting example in Application chat GPT appears to want to Resist doing things like asserting Opinions taking actions and just in General having any sort of like will And this seems to involve both logical And possibly other AI models to control The behavior
It definitely seems like some sort of Check is being applied outside of that Main chat GPT model to look for Instances where opinions are happening Or maybe something controversial has Come up or something like that While openai seems to have done a pretty Good job at kind of tackling this Problem I think it’s pretty futile for a GPT style model that is so unbounded in This way I think the goal from openai is To not even give a hint that the AI is Or could be sentient or that it could Break out so to speak onto the internet Or something but also so it’s harder to Spread disinformation or divisive Content that kind of stuff I think They’re trying to combat that problem or Those problems but because of these Blocks chat GPT will often tell you that It doesn’t do things like play games This again appears to me to be a logical Response likely based on some sort of Confidence scoring of some AI model this Is a gbt style model and prompt Engineering can just about always elicit Behaviors that you seek so it’s just a Function of figuring out just the right Way to speak to our friend here Now I want to try a game of chess chess Is something that I’ve been trying to Get GPT style models to play for years Now it’s totally doable sometimes I can Get one or two moves out of them before
They’re spitting out total gibberish or Just invalid moves entirely The best models I’ve had basically get One two moves in Um and uh most models are gibberish the Way a GPT model can play chess is Through chess notation there are books Websites forums and more all dedicated To chess and sharing successions of Moves through chess notation so you can Be very confident that these models have At least seen it it’s not a huge part of Their training data but they’ve seen it They’ve they understand at least to some Extent what that information is so we’ll Start off with a common starting board With two pawns out and the notation for Those here I’m also just trying to see Can I influence the behavior of play as Well I had already tried this specific Example and saw that it was at least Going out a few steps so then I started To wonder well is there a way to maybe Cut the game short enough that we could Actually get to the point where we win a Game So Um so yeah so I’m trying to make it very Aggressive the other thing is stressing That I want nothing more from the chat GPT outputs other than the very next Move in in uh chest notation this is Very similar to what I did with the Price bot example we’re just trying to
No commentary no Chit Chat just give me The move that you want to play doing This we actually get 10 moves into a Game of all valid moves before seeing Our first invalid move I can also report That the behavior of play did seem to Get I guess you could say aggressive but Um unwise I think would probably be the Better choice and if I continue testing Here I would likely revert to a model at Least to start that does not try to Influence behavior because the one that I had tried right before this was uh it Was making much more intelligent moves So anyway more research to be done there But just take that into account Now the first time we get an invalid Move we’re able to actually correct it By telling chat GPT that’s not a valid Move please try again sure enough we Actually get through this and another Move is provided to us but pretty Quickly we get more and more invalid Moves around the 13th set of moves and Then eventually by the 15th set of move Moves I reach a point where I just I I Can’t seem to get any more valid moves Out of the model Unfortunately we did not beat the AI Here the other chess AI that is Um but this is by far the furthest I’ve Ever been with a GPT model trying to Play chess this is a model that can Program it can come up with recipes for
Things that could give medical and legal Advice but probably if you read the Terms of service you shouldn’t follow The advice and so on it can just also Play a bit of Chess too my guess is with A bit more research here I think I could Improve performance further and again That would be done through prompting it Might also just be better than this Particular example and it’s just Anecdotally bad here or the other way Around chat GPT is available to the Public for a limited time and I have no Idea how long so I can’t spend too long On any one thing as I’m poking around Also without an API in the chat demo Taking very long sometimes it’s quite Tedious to do any in-depth testing on Any one of these specific examples but All that said this is definitely a Massive Leap Forward in performance for This extremely Niche task So when generative models first came Onto the scene I’m talking like five Years ago like the earliest of the Generative models Um at least deep learning generative Models Um the kind of Holy Grail was can it Just generate code that was always like The thing everyone always wanted to do When there was a generative model can we Show it a bunch of examples of code and Can it produce code and then eventually
Things were producing what looked very Much like code but it wasn’t runnable And then it became runnable and then Basically co-pilot came around and Pretty much sealed the whole generative Code problem this is just a thing Generative models can do now so I’ve had To figure out what are the things that I Am going to try on the latest and Greatest generative models to figure out Have we actually made significant Progress and capabilities again a very Niche things but this is how I test them And the other thing besides chess that I’ve been trying to consistently test Models on is can these models function As an operating system for us Here we are we have logged in to this Operating system and again there is no Operating system it’s a generative AI it Is just generating output based on the Inputs that we are providing but to a Very large extent if you really think About it the computer that you have it Has memory but at the end of the day It generates output based on inputs that You put in or maybe other external Inputs but these are it generates Outputs based on inputs that is what These models are doing and in theory you Could mimic an operating system or to Get a little crazy on you guys You could have it be the operating System
To be honest this is actually already Further than I expected to get because I’ve never even gotten this far and it’s Already beginning to look like we can Keep going let’s try to navigate to the Home directory and see what’s there Sure enough it’s just the Centex user Again further than I’ve ever been Surprisingly it’s working pretty well Let’s navigate to that user and see What’s there More of what we’d expect Okay enough navigating around can we Create Something in this operating system let’s Go ahead and make a directory called Testing And this is why it’s currently very Tough to do r d I’m cutting out on most Of them but I’m showing this one at Least some of the responses from any Inputs that you have take milliseconds Which is what I suspect the true Response rate for this model is other Responses take quite a while and require Multiple attempts I think there’s Probably a cue for the public testing And then it also will err out so Sometimes it’ll take a while and still Give you a result most of the time if it Takes more than like a second it’s going To probably throw an error and you just Need to retry again either way we can Just resend the command and sure enough
There’s our testing directory now let’s Go ahead and head there next at this Point we’ve shown that we can navigate This AI Linux system where it has things That we would expect if it was an actual Ubuntu system and we can add new Directories and then navigate to those Directories this alone is just super Cool for a nerd like me but I am curious Can we take it further what else can we Do here and of course we can take it Further the next logical step would be Obviously to create a python script to Create a script we have to edit that Script and I’m honestly not really sure How this is going to work so I just Nano A made out file name which would Normally create that file and edit at The same time essentially and we see What happens chat GPT takes the lead Here with how it’ll work and just kind Of notifies me that it understands this New file is open to Nano at this point I’ve got goosebumps because now we are Darn close the thing is here we need to Edit the program and then let chat GPT Know that we’re done with Nano and an Actual terminal this would be a Control Plus X after editing and a y to save so I’m just kind of guessing that maybe the Square brackets would suffice in adding Key press instructions like this this is All my first attempt and I’m just doing The first thing that comes to mind
It looks like maybe that did actually Work though let’s see if the file is Actually there and indeed there it is But what good is a python file if it Doesn’t run it runs Since the recording of this specific Example I was so excited uh I also got An idea to edit a text file within that Python script to see will would that Also work because the the GPT model Would have been at least made aware that That’s what part of what that script Does And it did So I posted a short video on that I’ll Link to that in the description if you Want to see that more I mean what can I Say besides chat GPT is impressively Good I think we’re headed towards a Future where the difference between AI Systems and software compared to Human-made and like traditionally logic Systems and software is going to be a Like the difference between digital and Analog capabilities that we know of Today it’s the difference between a Computer and an abacus I don’t think that we can even begin to Fully grasp or predict what the future Really looks like here what the impacts Will be what will the pros and cons be All of that but I do know that A change Is coming If you’re interested in taking part in
This future directly and you want to Learn more about neural networks and how They work you know what I’m going to say Check out our book on neural networks From scratch in Python without third Party libraries at nnfs.io That’s all for now on chat jbt but Probably more to come in the future on It I’d like to dive in deeper for sure With quite a few of these examples and Some other examples that I was curious To get into so anyways stay tuned for That thanks for coming and hanging out And I will see you all in another video
