OpenAI GPT-4 Function Calling: Unlimited Potential

[@princemars6746]

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

[@zorawarsingh11]

…John Connor rolling over in his grave

[@kevinmaillet8017]

Sometimes I think they’re watching me when I’m in there with ChatGPT. For over six months I’ve been using JSON schemas to get out structured data. It makes the reliability of the response significantly better.

[@thequestion3953]

Aaron Jack had this same response when GPT-2 was released and I completely glossed over it. This was around 2 years ago. He was getting extremely excited. Function calling is most likely going to be a big deal in the hands of devs that know what to do with it. kzbin.info/www/bejne/ooLRfaqLm9yar6M

[@finnbuhse4775]

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.

[@imtryingmybestman]

wholesome ass comment man congrats on your future success

[@simply.nonchalant6006]

Same here man! :) It's crazy to see others on the internet on the same journey. sentdex is the 🐐

[@sciencelab4225]

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 😮

[@getgonged]

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!!

[@danielpaull309]

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.

[@marilynlucas5128]

😊Good job! You should start a KZbin channel. Your ideas are dope!

[@danielpaull309]

@@marilynlucas5128 I've thought about it... but I find the joy in making the stuff and I know how much work goes into running and managing a channel so for now just sticking to the dev side of things :)

[@marilynlucas5128]

@@danielpaull309 just press record!

[@ko-Daegu]

that's basically LangChain tools with agents free and open source

[@jeffreycordova9082]

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.

[@BlueyMcPhluey]

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

[@StoutProper]

Ask gpt, it will come up with a range of suggestions.

[@BlueyMcPhluey]

@@StoutProper good idea! Its results made me realise the most interesting potential applications to me come from speech to text and then feeding that result to the new GPT API to perform actions.

[@jas4768]

Ask GPT to do your thinking for you

[@GoldnetGaming]

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!

[@MarkFobert]

​@ThomasTomiczek you can provide a function that let's gpt ask for a function

[@StoutProper]

Couldn’t you get an LLM to a) suggest useful functions that currently don’t exist and b) write them for you?

[@StoutProper]

@@MarkFobertin fact write a function that generates functions, like a prompt generating prompt

[@skylark8828]

Yes great idea, I want it to write me detailed specs that can be turned into actual code

[@StoutProper]

@@skylark8828 gpt can do that. And if you give them to gpt engineer it’ll code the whole thing for you

[@alexdacat]

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

[@iamxenobyte]

need music for that game? :)

[@alexdacat]

@@iamxenobyte Sure!

[@ness3963]

Do you know how you would 'force' it to call a function? I'm having trouble with it in my own project

[@alexdacat]

@@ness3963 At this point I have not tried integrating this into that project. But I will let you know when I get to it!

[@jzam5426]

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!!!

[@redship7532]

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

[@memespdf]

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

[@avidrucker]

Love it!

[@sentdex]

Hah, this is a great idea!

[@wg2]

WRITE THAT DOWN WRITE THAT DOWN, in all seriousness GPT should be able to get automatically get context from well-documented functions

[@illeatmyhat]

It's just an OpenAPI operation spec

[@FranAbenza]

Mojo?

[@grafzhl]

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

[@FlyingScot911]

Ha ha yes. As I watched the video I kept asking myself. Where are the tilde he keeps on about...?

[@unconv]

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)

[@dawid_dahl]

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.

[@dirty_haute]

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.

[@bobsalita3417]

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

[@software_development]

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

[@trentleslie2888]

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.

[@tonydenion3557]

AutoGPT on steroids ? :D

[@avidrucker]

I'd love to see a side by side comparison across a range of tasks

[@StoutProper]

You still change temperature and top_p settings at the model level, right? Or can you do it at the chat/query/function level?

[@jeffreycooper6163]

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.

[@jaredhead9503]

There's a join button that pays him $4.99/month

[@Shaunmcdonogh-shaunsurfing]

“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

[@clivebird5729]

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

[@TeamUpWithAI]

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

[@alejandroraulleivailabaca5749]

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".

[@garbagechannel6514]

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

[@christianwooldridge406]

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

[@noahsplayground2564]

Yeah I’m doing the same thing with a fine tuned model, using a system prompt over and over in my training data to ensure that based on the user content, the assistant response is this nice structured format…so I guess this function calling replaces that?

[@noahsplayground2564]

I’m also using the fine tuned model because I want the output to use specific names that are found in a separate dataset so I can essentially have that connection between what GPT is responding with and what I meta data I have in my dataset

[@appaloosa-m]

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...

[@goodtothinkwith]

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

[@darke7777]

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

[@davidbeesley5076]

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.

[@willholland1697]

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

[@ajnart_]

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

[@albertgao7256]

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)?

[@GeorgePolzer]

good questions! Have you figure out the answers? Please share.

[@dr.benway1892]

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.

[@JChen7]

I'm feeling the same way. As a programmer, I admittedly get some comfort in the deterministic nature of programming in the classical sense. This new stuff feels like trying to sculpt desired behavior out of squishy clay to me. Will definitely take some time to wrap my head around it, hopefully before it takes my job (:

[@countofst.germain6417]

Good work you were quick with this one.

[@adandrea]

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

[@danielb.2873]

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

[@kelvinprimo]

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

[@MarkWernsdorfer]

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

[@TheHavyxon]

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. :)

[@SeanKula]

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

[@alimohammad2541]

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

[@karlosfy]

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 :)

[@davidbeesley5076]

100%. This is why I think the “auto” example didn’t work. The function is named backwards.

[@nootajay]

Bro scaled up from tones to dozens real quick

[@chrisbarber7291]

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.

[@robhess3878]

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

[@ILikeAI1]

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.

[@StoutProper]

What system prompt is that?

[@tiffanitaylor8373]

The one that goes at the top of your array/list of message objects. It's the only system prompt 😅 I know of anyway. Did you read the docs?

[@ThePacifistanarchist]

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

[@obliojoe]

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

[@techfren]

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

[@AbdullahOllivierreIT]

Any way you could also provide an AI summary of the video and provide the tool used for that ? I'm sure this must be a fairly easy task to do given the video is north of 20 minutes

[@techfren]

@@AbdullahOllivierreIT yeah definitely

[@Artorias920]

this is actually perfect for something Im working on :D

[@Jamman88888]

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

[@MuhammadUsman-ix6jo]

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

[@JT-sv9bi]

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.

[@ander300]

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

[@StephenRoseDuo]

This is indeed absolutely insane

[@mattkim96]

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

[@WandafoolAI]

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

[@homerdus9668]

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

[@WT83]

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.

[@YouBetterBeYou]

If I am getting this right, function calls allow us to instruct GPT to iterate on its own response without making multiple API calls, correct? So, some simple applications would be "format_response_html" where you can make GPT provide a html/markdown version of its response, correct? That means you can probably also make more elaborate iterations on the content like "formulate_1_leading_question" that takes the response and turns it into a leading question that nudges the user towards the desired outcome instead of directly telling the user, correct? So whereas with prompt engineering, fine-tuning and embeddings you give up-front instructions for the text generation--which fails to deliver the desired results consistently, we can now put it into a function as way of post-processing the response. Did I understand this correctly?

[@maxernststockburger9420]

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?

[@erickmarin6147]

Using a tower of GPUs running GPT4 to dim the lights is such a meme use

[@sentdex]

I hope you know that's hyperbole and the real cost to actually perform this task with GPT4 would be something like a penny, and a small fraction of a penny with GPT 3.5... because you wouldn't be actually using "stacks of gpus" no more than you use millions of dollars of infrastructure to check your email.

[@hadtomakeachannel]

@@sentdexyeah but you still need the stack of GPUs, if only for a very short time. And checking email *does* require millions of dollars worth of infrastructure!

[@torley]

"Extracting structured data, period." @ 7:00 EFFICIENT SUMMARY!

[@mytechnotalent]

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.

[@sentdex]

I may do something like this at some point, gotta finish NNFS videos first though. Karpathy did a sort of GPT from scratch series if memory serves during his transition from Tesla to OpenAI. Have you seen that one?

[@mytechnotalent]

@@sentdex yea its pretty amazing but moves quick. I like how you really take your time to break it down and make it digestible Harrison it is what really defines you.

[@jwadaow]

@@mytechnotalent Also it uses tiny-Shakespeare and tokenizes differently to Tiktokens. It is educational for certain aspects of the Google paper that spawned transformers. There is room to expand on it further.

[@parseeval]

I'm wondering if it would work even better when it is worded the other way around. So not "get a list of commands to run", which is like a description of the function in first person, from the functions perspective. That sounds odd to me and maybe this isn't what it is trained on. I wonder if "what can I, ChatGPT, do with it", like "Show shell commands to the user" would work better. Sadly I don't have GPT4 access to test what would work best.

[@ghosthermes]

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

[@illustratum42]

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

[@KP-sg9fm]

The ultimate programming language would just be in pure English so that makes sense

[@illustratum42]

@KP-sg9fm that's why I love python. It's pretty dang close...

[@AIonRails]

I'm sure it will soon be a reality. Can you imagine how many side effects could occur in large-scale software that is entirely produced by AI? I'm not a pessimist; I just want to imagine what new challenges software engineers will face

[@Massenko]

Awesome, thanks a lot for this demo ❤

[@USS_Daedalus]

i am curious that multi part request is possible that you for example send PDF and post messages that GPT should upload it in my Cloud storage.

[@TheMullela]

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.

[@liog4127]

Great Intro. TNX!

[@vinicius6385]

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

[@CosasCotidianas]

Is this the beginning of web 4.0? Buying products, searching for real time data, controlling stuff, anything!!

[@jhadesdev9576]

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.

[@shreygupta7877]

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

[@JazevoAudiosurf]

The default function call = auto doesn't work well for me. In the system message, I tell it to "ask the user questions if prompt is unclear" and then add an unclear prompt from the user. Provided a function "ask_user", description "Ask the user a question", it still chooses a non-function question every time, unless the user specifically tells it "ask me a question"

[@tomirkpl]

I don't understand. How will know your script what function has to give with the prompt from a user? Isn't it a negation of GPT possibilities that you must choose a function instead of GPT choosing a function?

[@kyber.octopus]

*read in dramatic voice* This... Changes... Everything...

[@sundarswaminathan3054]

So I get that you get the arguments regardless of whether the function exists... but for producing the final output that you showed here, you had to create those functions or the LLM is creating those functions on the fly based on those decriptions? That part was not clear so it's confusing

[@adamrozycki9822]

Great explanations, cheers! 👌

[@diego-castro]

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

[@GuidoSalimbeni]

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?

[@criostasis]

A function that allows it to create its own functions for new tasks you give it.

[@jainicz]

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?

[@sentdex]

my guess is almost all plugins were created up to this point using forms/preprompts and theyll all be soon re-written to use function calling. For an end-user (or middle user?) using some plugin...you probably wont notice any difference.

[@memespdf]

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?

[@ylazerson]

thanks! this is out of this world!

[@RipYaZa]

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?

[@iamacoder8331]

To be honest, I don't think its gonna change anything in programming landscape. We need to be sure that product for the client is working in deterministic way. We cannot rely on probabilty.

[@aaronmineen5511]

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

[@technolus5742]

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.

[@drdanoosh]

Was it intentional to have San Francisco referenced in the boston weather function? Was San Francisco a default or something like that

[@Bandude231]

This will be nuts

[@isMatvei]

Very cool shishtuff. 😮

[@mpfmax0]

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

[@therestartprince6418]

I have a question that might be important, so if no one answers this in a day or 2 I will ask using a different method. Is it possible, with this new feature, to make, similar to adding a "child", a sub or child function that answers to the main intended function? I have been understanding how important this option is but also how limited it might be if you want to create a more complex function that leverages GPT s ability to do things the human mind is not able to do. I want to use the extra function for an idea I have. The idea is for an algorithm that is predictive in nature that is based on a human skill I have developed personally. I have been trying to use gpt to help me translate and find comparative data that turns experience into "algorithmic code". But all attempts seem so disjointed and I am not getting data that is matching the whole of my complex request, since I have to break up my requests into pieces to get a descent answer. But there are so many pieces. keeping a consistent output, that doesn't need an infinite amount of tweeks, is not working for me. I was hoping that this feature would help with consolidating multiple requests into a master request that I could use to gather each section of data without losing context. Or is there another way to handle this? This algorithm could be extremely valuable if I could just get it translated.

[@deejai1220]

Guys, it's the same prompt engineering you could do back in March to call functions. It's a nice convenience, not a new paradigm

[@mattizzle81]

Yeah it still breaks. I tried it and it fails enough to not be useful for me. Wait until people use it and find out.

[@seanfuller3095]

Can this be used with "context" (i.e. top_k matches from a vector db similarity search)? LangChain is cool, but I have found that the moment I want to customize functionality in any way, things go downhill fast. So I have been generally been using normal ChatCompletions instead. So....could function calling be used like LangChain prompt templates?

[@i_accept_all_cookies]

If I understand correctly, function calling is a lot like using Langchain agent tools.

[@hamzaoner4623]

So isn't using LangChain in a project better because on the long term you can replace the LLM you use anytime if another one will more fit your needs

[@deepster5150]

How does this impact or alter usage of LangChain? Is this competing with the Agents capabilities? Pardon my ignorance but ramping up as fast as I can...

[@ness3963]

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