Wow this is what I always wanted! It was so cumbersome to get this functionality before. Now it's supported and reliable! Your idea of providing functions which don't exist is awesome, I would have never thought about that 😮

Big love for you my bro. I remember when I was first introduced to Python in what you would call elementary school and I loved it. So I went home to watch your Python game tutorials (the tank one was my favorite, even tho I had no clue even what a quadratic was, somehow I got it to work lol). Completing those felt so rewarding and it really inspired me. Now I'm 19 and soon to be an apprentice at a software company which was my goal. Seeing this video from you after so long felt nostalgic and I'm so happy to see that you're still making videos. You make dreams come true you absolute legend.

His authentic happiness shows just how much of a bombshell function calling will be, and for them to nonchalantly release it is interesting.

Been hoping to see a video on this new feature. Everyone is talking about the larger 16k context but not much mention about these intelligent functions. This is a game changer. Great explanation & great job!!

21:18 - THIS!! I've been reading blogs and watching videos about function calls wanting to really understand what all the fuzz was about and no one, NO ONE has highlighted this extremely important point more than you did. THANK YOU!!!

This is EXACTLY what my business partner and I have needed for our project, can’t wait to implement this!

This is actually very cool. It's definitely tough to constrain the response to something that can be reliably integrated into a system. It seems like with this new feature we can now directly pass these constrained responses into an api. Super cool.

Looked at other videos then replayed yours. Finally clicked. Your stream of consciousness chatter and enthusiasm was tops.

awesome. you're always presenting sth extra nice in addition to only raw functionality.

Returning structured data using a function_call is awesome. Thank you!

Slick. Much better than the, "Give me the results in the following json format {'foo':'bar'}" that was required before. Im excited to try a set up 2 functions for input to gpt as well as output.

Just want to thank you again for this awesome video. I’m a dev as well and I’ll be holding a presentation at my company on Friday, echoing a lot of the amazing points I learned in this video. It really seems like people haven’t really understood how powerful this Function Calling feature really is. This video I feel is the only one where the true potential of this new tech is revealed, with how *any* kind of structured data can be generated using the almost magical smarts of these LLM systems.

Aww your genuine enthusiasm made this so much more engaging and entertaining :)

Now the trick is to implement this into an assistant, and if the current input cannot find a function, we ask to program, to ask gpt4 to create this function syntax with the description and anything + the actual code that would execute this function, an auto-writing/evolutive AI assistant

This is definitely a different way than I thought of using function calls, very interesting

Kudos man, what you did with this is awesome, very creative!

Thanks for covering this. It's funny, I feel like simultaneously this opens up so many doors, but also I have no idea what those doors could even be. I find it really hard to think of applications for this kind of thing

This is how I imagine the programming in star trek. If you've ever watched the show and someone writes a program or a subroutine, I imagine their just putting in basic functions and descriptions and the computer turns it into actual code. This is so cool and I am super hyped to see this evolve

This is so cool, I've been nerding out this last days rewriting the whole code of my chatbot to use function calls. Weird fact: it seems that functions are still in some part of the system prompt, and when you ask particular questions like "how can I get the weather?", gpt will respond "using the 'get_weather' function, to which you have to specify your location".

I actually made a system to do this back in April and it was a lot of work!! This is great. My system had a function with description, parameters etc and then vectorise the description, then when you could ask the bot to do a task and it would search the index for the closest match to the task and use ChatGPT to get the parameters from the original task and if it needed more requests it from the user. (I also had a loop to write a function if the function didn’t exist which for basic tasks worked about 80% of the time). It would then execute the function and return the result.

Good work you were quick with this one.

I'm using ChatGPT to make a choose your own adventure game, and this looks like a perfect way to get back JSON that is formatted correctly! I can get it correctly 70% of the time, but this looks perfect

“AI functions” (see Marvin and AutoGPT projects) have been a thing for a while. We’ve had some work arounds but it’s nice this exists now. Definitely a step forward

I gave it filesystem functions to write and read files and it can create multi-file applications from a single prompt now (with recursive prompting)

I think you could probably have hundreds of functions stored somewhere in SQL or a vector database and use a search function (any algorithm really) to get the top X relevant/necessary commands. With this new feature you could even let the LLM write an array of queries to use for searching relevant commands. After that you would then pass that into the functions array with the prompt like you showed. I know that would be an additional prompt, but in the long run, that would probably be cheaper than filling up the context size with all existing commands!

Your whole channel is full of cool stuff I'd like to do. I'm just commenting to let KZfaq's Algo know I want more recommendations from this channel.

6:12 warranted a like from me. It's nice to see how enthusiastic you are about this

In the past I made a home assistant that required this type of functionality. The way I achieved it was through fine tuning a model which was kinda pain because I had to make a bunch of training data but it did work very well. Excited to see if this works better then my approach

Awesome, thanks a lot for this demo ❤

Great video. To tell gpt to produce content as well as a function call you could try adding a function argument named something like "comment" or "description" which gpt would populate with a string. For get_commands, I feel like the problem you've noticed there is that you've written it from YOUR perspective not from gpts. You the user are calling a getter... but for gpt it's not calling a getter because its writing the commands. I think it'd be more effective to call it "put_commands" or "call_commands" as its more descriptive to gpt

Thanks for sharing this... I only had a minimal sense if what this feature provided. This is really cool!

Dude - you continue to provide the best content for a technical audience. Can I give you money or something? It's been like 10 years of just binging your content and getting inspiration and training for so many great things.

This is HUGE! Pretty much what's needed to embed AI anywhere.

It should be possible to write a Python library that automatically extracts GPT function specs from functions with docstrings. Perhaps a `@openai` decorator would be cool

Great explanations, cheers! 👌

This is a tilde ~, what you mean is a backtick ` 😁 Cool explaination!

So stoked you covered this. Getting creative output in a structured format is huge due to its contradictory goals. I assume you can still change temperature? I didn't see anything about that, guess I'll just have to try.

I'm excited to have this implemented into a humanoid robot. Function call it to do chores. It would be really cool if it can write its own custom functions based off what you asked

I was puzzled… and then my eyes got wide when I realized what this really was. Deterministic structure enables it to be used as a piece in a larger thing… you could put a thousand of these together, building a “program” full of these “functions”… utterly brilliant move by OpenAI

this is actually perfect for something Im working on :D

very interesting, thank you!

thanks! this is out of this world!

Great Intro. TNX!

This is amazing, I am using this to parse data 🙂 and its working very well

Interesting. I was already doing this by including a "menu" of functions in the initial prompt and then 1st ask which of these functions would be applicable given the actual prompt context. Then with that answer, I'd run the function and create a new prompt with the initial prompt + the result of that specific function. This is now a more official, structured way of doing it. Given the limited amount of functions you can pass it probably still makes sense to work with a tree structure of sets of functions which you pass in a second prompt, after you've asked which set is most applicable first.

whoa this is awesome, especially that it works with gpt 3.5

Thanks for sharing, just might be simpler to show this example as a chart

great video, thanks for explaining this! Project Idea - Automated Trading/Stock Advisor bot: Scrape a financial news page, send article to GPT to determine if the stock should be sold or bought, then output the stock symbol & trade (buy, sell, nothing) with functions. Maybe GPT could check your positions & portfolio first before making the final decision...

I was trying to make it consistent forever, this is way better and the capability is huge, I wonder how far it can goes.

This is indeed absolutely insane

Thanks for covering this topic! How are temperatures handled in the varied_personality functions? Could it be possible to include a temperature in the functions? Or to let GPT think about 'the right temperature' for generating the response? How do you see this? Will this be the 'prompt engineering part' people need to finetune themselves? Or will GPT do that for us?

This feels like a new programming paradigm. Had to order an API key just to test this. It's a bit frustrating to know there's loads of possibilities with this but you have to figure out those possibilities. New territories to be found.

To populate "content" for example for some confirmation or other message - it could work simply to put it to the function call output. For example: "message" : "{"type":"string", "description":"Reply with confirmation that the function get_current_weather was called."} Should work - I used it in the old "form" way of using chatgpt to get custom confirmation replies or questions for required fields. In the end, all this function calling is then achievable also in gpt-3.5 with high success rate with a good prompt, but this is really nice structure. :) Also very good practice would be to use function calling for intent recognition and then pass the input to identified function to save tokens for the parameter blocks for individual functions. :)

This will be nuts

I think your function should be called “set_commands” or “run_commands” instead of “get_commands” as the commands as provided as a parameter. “get_commands” sounds like a getter function so you would expect to be returned the commands instead of providing them. I think this might be why the “auto” version of get_commands didn’t work.

6:10 "do cool shhhh...stuff." nice save haha. Good video, i wonder if a lack of functions like this is why the snapchat AI used to gaslight people? It would tell them it didn’t know their location for privacy reasons, then in the same sentence tell them where the closest macdonalds was, or the weather in their area. Also, any chance on a new gantheftauto video? Its been 2 years, and its my all time favourite of yours :p

Very cool shishtuff. 😮

this is awesome. text-davinci was the only model I could get good performance out of (without access to gpt4) so this is really exciting since gpt3.5-turbo is so much cheapter. There also might be some token saving compared to typical zero shot learning (in some cases)

Damn this will be so useful. I'm no programmer, I've actually decided to work with chat gpt to help me create some apps that I wished existed and kinda leaning along the way. Bit I've been in IT for years so I get how things should operate but coding syntaxes is something else entirely. My idea while using chatgpt with its access to web search and noteables via plugins, was to create a series of routines and shorthand. So instead of spending all my time working and reworking my prompts I could create a series of routines and subroutines that would tell it exactly how I wanted it to behave. I use various tagging techniques and had it form an index to refer to. So for a bit I could say something in a prompt like: [rt:1] and based on the documents in Noteables it would go find that routine and then follow it. The problem is these tags mean less and less as the conversation goes on because from my understanding the weight of the initial context means less and less. Still fun tho

Awesome video! Just wondering about the "get_commands" example, I don't know if I understood correctly but I think the function should be "execute_commands", and the argument is a list of commands to be executed, so GPT gives you the commands to be execute, right? Anyway, great content :)

You can loop the GPT call as well so that GPT can call multiple functions one by one if it wants, before responding to the user.

Bro scaled up from tones to dozens real quick

I wanted to make a table in chat gpt and have it populate it with items that make sense. I wanted to do some manipulation of the table so I told gpt to pretend it was database software and the table was a table was in the database. I could make up functions on the fly and they all worked as long as I told it what the function was supposed to do. This feels like an official version of that.

Hi... i really wnjoyed this video even thou i sidnt understand much, but ayilk sone and really enjoyed OP's enthusiasm... So , hmmm, kind of awkward but here i am , 44 years old, travelled all my life around the world, didnt get any college defeee, just finiahed highschool at 18 and have been travelling since, qnd loving life .. but rn , after 25 on the road z ive decided to go back live closer to my family, in belgiumz and im gling throuhh what i thinkis my middle life xrisis 😅, and im fascinated with all the AI tools andntechnology, and kind of want to sice deep into it, and hope i could turn my life around and actually fet a good ceeative job in the future, uaing these tools, or programming them / engeneering prompts etc... Im just not aure where to start really , neither where k want to get to with it.. i really loce art aongraohic design/ video/ music is really attractive to me (i have 29!years of autodidact experience in electronic music production) but maybe programing will be better as far as fiture job ? I can see myself atudy that too ... Okz the queation for whoever is atill readuing : where should i start ? What advice woukd you give me? .. i feel likenwith ai techs and tools i ahouldnt lose time ayatting from zero, that things are sifferent nownthan when you all started , maybe? ... I hear outhon, i hear JSON, i hear open source, qnd many of these termsz but honestly I don't even have a great laptop at this point, thats why i didnt even started mingling with open source (except for aome that are usable withoit API , or i donno ...hmm . WEBUI? ..aee, i know terms but im not sure what they are 😅 Thats it. Please, any info or advice to guide me ? Id appreciate it dearly

For the second function call about the dehumidifier, would you not get a similar answer just from using that info in a prompt? And you are saying, with the call function, that this is basically the inner workings of how chat GPT reads prompts? So it makes sense to accurately write your prompts as if you are writing a call function? Except that inputting the call function itself is more accurate? Im wondering if this can be used to make GPT less forgetful in tree of thought conversations through multiple prompts, or not lose track of context as easily

Wow that’s impressive

00:00:01 - Exhibition of OpenAI Function Calling Capability 00:02:34 - Programming - Creating Functions for Weather Data Extraction from GPT-4 00:05:05 - Describing and setting parameters for a 'get current weather' function. 00:07:44 - User describes calling function with JSON object & options for running function. 00:10:06 - Converting JSON object to dict for term GPT with GPT 3.5 00:12:41 - Stream discusses extracting structured user data & terminal commands for a GPU-tensorflow install. 00:15:12 - GPT4: Describing a function and its parameters for a structured response 00:17:39 - Injects varied personality responses into AI functions to generate personalized responses to user queries. 00:20:14 - Introducing sassy and sarcastic response structure from GPT4 models created with timestampgenius.com

Nice! Gonna try that as soon as possible. The functions I define will be added to the tokens that i use? So in theory, the number of functions I pass in is limited by the context window or?

Great Video

Cool shtuff!

I liked very much! and I only managed to learn like you teaching! Could you demonstrate how it would be to create a food recipe, adding title, ingredients, preparation method? I think it would be amazing example

I wonder if they use microsoft's guidance library for the structured data, it feels like it (along with appropriate prompting). I was looking at using Guidance to create more accurate JSON responses from LLMs, this will do it for me though.

Thanks for sharing. Do you know if there is a way to format empty response. One issue of Gen AI is to say I don't know ;-) But with some patience and proper prompts you can get to the point where you have an empty response. I noticed that it was then pretty hard to get a consistent formatting of the response and that quite often GPT 3 and 4 where returning a string that was not matching the requested format.

You have a much better time using auto if you use the system prompt to tell it that it prefers to use commands over dialogue whenever possible.

Part 10 of Neural Net from Scratch, about analytical derivatives??? Please bring the series back!

Interesting. So it seems like this is more or less just the ChatGPT plugins approach, but exposed in their developer API and with a model fine-tuned to handle it better. I say "just" but it's pretty cool! ...I kind of wonder if this could be used with a vector database like Pinecone to implement the Generative Agents paper's memory stream architecture more easily?

I suppose there is a hacky way to still get the content. Just define the function call to contain a content parameter of type string and instruct gpt to populate it with the content you desire.

Did this not already exist before even ChatGPT in applications like Google Dialogflow, where there were intents and entities?

and can an argument of the function by an array, something like a list of 5 suggestions for a blog post title? I couldn't find the syntax to define an input argument of the function as a list. Thank you for sharing, awesome content.

Holy shit this is amazing! 😮

4:27 "note this is plural"... Me: "note this is an array" 😂 awesome video

The functions data passed to GPT looks like a JSON-ified version of OpenAPI 3.0 spec. Is there any information about if this is intended? Maybe OpenAI used public OpenAPI specs for training?

Couldn't you make a call that makes it number the messages as a reference? You could then make ChatGPT act as basically every profession in the world and use it for educational purposes.For example you could build a library with cases on how it should reply to a student/client in different scenarios and then build a session summary to continue next time and so on.... So many possibilities now!

My example variation of ChatGPT as an assistant was also a sassy and sarcastic assistant. I wondered why we would do that to ourselves haha

Hi everyone! And thanks to sentdex for the video. I have a beginner's question. Is there a way to save the model and its functions in OpenAI without having to set it up for every new prompt or instance I want to run?

Wait until fine-tuning for gpt4, it's going to be amazing

amazing!

Wow. This is perfect for me. I always understood how to describe it, but hard coding just isn’t my thing. With this I just can describe what ever I would like to have AND get a result.

Harrison this is crazy. I really wish you would create a GPT from scratch series. I know you had a 2 years ago GPT series but what do you think? I am also helping new students with your book. They are buying the ebook.

Using this functionality to do cool shhi.... stuff. 😅. Nice Catch

amazing.

Thanks for the video, I am a bit confused about why use this function, for example, in that get_weather() function, all the GPT does it to extract the information to a parameter that we can feed into a function, so why not we just write a prompt to get the location as a json directly? is it because you do not know what the user would ask? so you can provide a list of functions that you wrote, and from the output, we can determine, whether the user invokes one of our functions or not, if yes, then we use the GPT output to do a further processing? but this is actually complex, after getting the result from the your own function, how do you know how to construct the proper answer to the user? what about the user triggers 3 functions, and what should happen next? how could you provide an outcome to the user in this case (more than 1 function has been invoked)?

0613 refers to the date this version was launched. In this case, 13 June.

Thanks for your video! I wonder how is this different from, say, a weather plugin? Doesn't a plugin do the same thing by calling some "functions" that were written by human developers?

I have been using the chat completion api to get gpt-3.5-turbo to output JSON strings with the aspect, sentiment and text segment of product reviews, which I then use to build dataframes and make aspect based sentiment analysis plots and stats. I just send it 5 examples of user/assistant answer/response as a prompt. I wonder if using this would make the outputs more reliable/accurate? I want to understand what is the best way to conduct aspect based sentiment analysis using GPT models with openai api

Great video. How can force to run all the function call when there are more than one? I cannot find documentation . You used ('name': 'the function name} ... but how to ask to force run 2 functions?

4:37 wait for the engineer to make a marvel reference challenge

This would be really great if you could force it to use any function, not just one. Part of why this is useful is the fact that its able to determine which function to call. If I have to figure that out ahead of time that almost defeats the point

Re 9:50 I'd imagine (but haven't yet tried) that if you want to also get a message, you can simply add another "parameter" to the function called "message" or "content" and describe it as such. You could even make it required

One very funny thing I noticed, as a result (I think) of the training they did for this, it seems GPT is now CONVINCED that it has access to a Python interpreter. For example, through the ChatGPT website (or through the completion API, I found either is affected), try asking "Are you able to run Python code?". It may respond correctly, but if you regen a couple times it will tell you at some point that it can. I noticed this bug as I was messing around trying to get function calls to work, and ChatGPT repeatedly tried to call a function called "python" with python code, even though I had no such function or anything related defined. This affects GPT 3.5, not sure about GPT 4.