ChatGPT Writes a Chatbot AI 

Who else is missing the wonderful and nostalgic sentdex tutorials in IDLE 🙃

Many congrats and best wishes for the new family members! As someone who has been following your channel since you had these open air GPUs in background, the new home server is looking really cool in the background!

Congratulations, double the family brain count can't be bad

You’re low key one of my favorite RU-vid creators in this topic. Good luck with the multitasking 💫💫💫

Really nice, I did something like this but kept it to python interaction and was pretty blown away when the code just worked. Thanks for the in-depth exploration and report.

There is an property called "stream" into the model, which is this character animation in ChatGPT. That means "No" the full text is not already know after enter the prompt. Stream does exactly this, respond token by token.

Congrats and great demo. Clicking the waffle top left to bring back all the auto-bookmarked (crazy how it auto names our conversation too) history (effectively prompt engineering) of a previous conversation will make it even quicker to conceive the next version of your bot :-) Perhaps we will see just those 'priming' prompts (i.e. the stuff we actively tell it after, like you say, whatever prompting it's been given before they present the bot interface to us) made available like software in the future for particular use cases (of course this is already there in a small way with the drop down of example presets in the OpenAI playground).

Congrats on the family!!! Wishing you the best!!

Congratulations on the twins man. Always learned a lot from you. Amazing channel.

I love that you show pics of your twins! Congratulations!! You are a blessed man!

We miss your videos bro ! Congrats for the little guys !

Seeing your video since you're were using python IDE 2.7 till now. You've come a long way 👏

First and foremost, congratulations on the babies!!! I think the last couple of minutes of your video are very valuable and people should make a note, if not already noticed, that ChatGPT can be used to build simple applications, simple websites and that without no big knowledge from the requestor. the best part is, if you run into an error, it will provide you a solution. And this is very very helpful. You learn better like this because you get also explanation. As example, I am in no way a programmer. Just hobby stuff and nothing big. And yet, thanks to ChatGPT we wrote together one script, then when I decided to create something similar, the errors which I was getting were very similar and i was dealing alone with them, not needed the Bot. Hence, it has quite the positives for your learning curve as well.

This video is edited beautifully! Great info, ppl like you make RU-vid priceless

Congratulations to you. We also had twins six months ago. It took me two weeks before I had time to make coffee again. And you manage to make RU-vid tutorials and videos!!!! I don't know how that's possible :) Respect! BTW: great content :)

I LOVE this video! Thanks man! And good luck with 2 kids at once! 1 baby is mega hard for at least the first 3 months, more like 6. I like your geeky pythonic credits. You really look like (my memory of) a guy my team used to collaborate with.

congrats on the twins! mine are a little over 1.5 years old and I can confirm its a lot of work! It gets better every half year, though, hang in there ;) Kind regards from a long time fan!

Congratulations on the twins! Wish you all the best

Congrats with the kinds :) thanks for the videos

Congratulations on the newborns! This is awesome, I would love to be able to run my own instance of chat GPT that I could really train myself to help me get things done! I believe in order to really do that we will need to have our own models in some way shape or form!

Many congrats and best wishes for the new family members

You should integrate the Whisper speech transcription model and a TTS-model so you can actually _chat_ to it! Also, congrats on the twins!

I don't think it's slowed-down, remember, the generations are done autoregressively and the models are already loaded on the GPUs, which means that you can either wait for the complete output (which can take some time even if every token costs like 3ms to generate), or you can print the output one token at a time and spare the humans some impatience.

17:26 ElonGPT: great, now pay 8$ to unban

Nice to see a ChatGPT video with a lot of specific, insightful and original info unlike the deluge of generic make $10,000 in a day with ChatGPT variant.

Congratulations on your twins! Also, funnily enough, I attempted to do this when chat GPT first became available, I unfortunately got stuck behind many errors, and my project was no where as in depth as yours.

OMG! Congrats on the twins! Also, condolences, but mainly congrats!

Always interesting watching your take on these...

really inspiring stuff man. I was underestimating ChatGPT and just using as a lousy poet and therapist. Its wayyy more powerful as a rubber duck. Major.

Hey sentdex, congrats with your twin newborns! Thanks for keeping up with amazing content.

I made soooo much money with ChatGPT! I am in love with the tool at the point where i just decided to start a tutorial channel about GPT! This tool is fantastic!

You're GENIUS really! I even couldn't imagine this and I just only found some RUST codes with ChatGPT.

Twins!? How unusual batch size, congrats man!!!

This iterative process in 2:45 could be a vscode plugin that modifies the code and writes commits. We're not that far from creating an issue on github and receiving a well-tested PR, entirely AI generated. We're approaching the singularity guys!

Hello i have a question about something you said, if anyone know please tell me. I understand that ChatGPT is a self attention model, and that means that the output is created token by token. GIven that, what do you mean when you say that the delay of the answer of chatgpt is artificial? I mean, i understand that probably it is but isn't it possible that if the answer is too large the model is still computing the ending?

How did you do this? I often get: Too many request within an hour, timeout :/

After following for last 7 years. the output was great. Twins.:D

How can I create my own model from large texts?

Fantastic. Absolutely.

I am confused, can you clarify what you mean when you say that sequence already know and they intentionally slow down the text generation? Doesn't chatGPT run auto-regressively?

How do you even get into the ChatGPT? It seems the server is always full.

Congrats on your twins! Best wishes and happiness!

Can you teach the chatbot with information from a website, so it can call on that info to make replies in a chat questions?

😁😁😁Congratulations for the twins Harrison, all the best in the coming years!

OpenAI: We Present ChatGPT, Its like Stack Overflow But On Drugs! Me: MmmM, mAkE TreE PoeM.

I could be wrong but I think that RL done to the model has a huge impact on it. As far as I know they used human input-out pairs to create a neural net which acts as the reward function (hence the millions of dollars of funding needed) then optimse the LLM to the reward function while disallowing output distribution drift w/ the kl divergence. Basically ensuring that the model optimised on the reward function would still output coherent english. I just don't think it's all pre-prompting, I think most of the magic comes from the RL step.

Congrats on the twins!!

First congratulations and my best wishes for you and your beautiful family! and secondly thank you so much for teaching us with such a high level of quality and entertaining content! Have a great day sentdex

20:44 yeah it's been shown in many papers - the more you generalise, the better the model becomes at a narrow task

I've been able to get chatGPT to give a transcript of the conversation up to that point. I could not get it to reveal a pre-prompt though. Also, I think that chatGPT's reinforcement learning is key to it's performance. The way that they say they did it is through ranking four responses by quality. Then a discriminator learns how to rank the model output with that data and the model uses that discriminator's rankings in its loss function. Instead of tweaking prompts and modifying regexes to find and remove every edge case, this reinforcement learning process can learn to find edge cases unseen by humans.

Also 13:00 onward: yeah, coming from the linguist clique, I definitely am completely confounded by what we're seeing already. People have their hangups about incorrect replies, but by God, remotely subjective topics can be iteratively resolved quite a bit with chatgpt and I can't get enough of it. I really thought the novelty would wear off at some point, but let's face it, it's only going to get so much more intriguing. Game dev is going to be nuts - is what I would say if it wasn't going to impact every single nonsense trade in the whole wide world.

i cant wait to see this code lol I have been sitting around for awhile trying to get it to do this but i am not a coder so the learning curve is there. I have been teaching it Hydra visual synthesizer and getting it to write working scripts for visuals and more. I can get it to write html and python but idk what to do it with it lol I been at it all day so I just end up getting tired before I get through the loop of "debugging". I kinda gave up that it would actually give me a full working script. with my limited knowledge it may be hard. thanks for the video and hope haha ;)

For something like this. is it possible to have it remember things from prior sessions? like if you told it your favorite color is blue. is it possible for it to retain that and 3 weeks later ask if it remembers your favorite color?

Can’t wait for video from ChatGPT, how it re-created itself with the help of human!

Congratulations!

Congratulations 👏

Reminds me of Neo talking to the Oracle in Reloaded (I think)... "programs creating programs"

Hi sentdex, I happen to have access to 8 A100 with 40gb ram each at work. I work at a speech recognition startup. So far I used ChatGPT to help me with kubernetes. If I find some time and the gpus are empty I could try to run the large model.

Fascinating!

Congratulation! for one year you will have some time for videos, after that you are done , full time busy, I got triplets I can watch around 3 min videos at once.

Simply awesome!

Congratulations and good luck with the twins 🙂

I gave this a try myself a little while ago and found that using a json formatted chat log works pretty well as preprompt, would look something like this: [{"author":"John","sent_at":"unix timestamp here, doesnt add much but does let the ai know what time it is","content":"Hello!"},{"author":"Bot","sent_at":"current time","content":"

19:44 unfortunately I don't have time for this, but the compressing idea for when you run out of tokens sounds like something ChatGPT can do

Congrats on the twins! I have 2 on the way myself. Good luck to us both.

Hey I need a small help , I completed Course on Data Science but not getting good job and low knowledge , that reason few years on i moved doing other jobs , It's any possible chance of get a job in Data Science.

congratulations for your baby twins! thay are so cute!

how to use?

Congrats on the twins. I think you do know how lucky and blessed you are. Its my dream to have twins.

Sentdex your ml Playlist is awesome i have learnt alot.

Thanks!

Congratulations! Also where did you get that thumbnail of your video at???

This video is a gem!

congratulations for you baby!!!

ChatGPT just aprove a thoughts, that "googling skill" - is most important thing in your life. You can have ultimate powerfull AI, but if You can't ask "right questions" - AI will not help you.

brilliant!

Can you do a tutorial on making a chatbot using ontology, AI, nlp?

How fast would it work with a 3090 do you think? I had like 16 3090s and 30 3080tis when I was mining last year. Sold many of them though.

wow Pinkman really made a switch

You should make another video titled "ChatGPT Writes a Chatbot AI which I used to create another ChatBot"

One thing that AI needs is compute power. Whenever crypto crashes, there's a whole bunch of compute farms sitting around not doing much.... uh oh

You might be able to side holster the kiddos and be the John Wick of baby hammocks and have it at and angle so you have some sway motion and access to pockets~ Congratulations Congratulations!!!!!

always ask ChatGPT to ask you questions to clarify your request to enhance your experience with the program

hoe to get html code

If the response chat GPT will give is already known when the response gets printed then why does chat GPT sometimes print a word and then delete it just to replace the word with a new word? It would only do that if it is updating something live based on every character printed.

I know there is an Prompt for MidJourney in the transcript somewhere

Congratulations!!!

Hey btw the ChatGPT Pro member model is insanely fast. The entire script in a few seconds

why did u not continue then...

Thank you.

The reinforcement learning was also used to limit it

Have you asked ChatGPT about it's instructions? It told me it's pre-prompt, also it explained interestingly how it uses the attention mechanism on the history which is just given with the latest prompt

As someone that actually knows nothing about programming I've been doing similar things. Managed to get my own chatbot that was working on a terminal not on a web page. After watching this I realise I probably could have just asked it to make it into a web page

17:25 he's just like me frfr

Happy you seem happy with the twins i hope everyone can be happy

congrats! :)

Congratulations

Congratulations!!!!!!!

Harrison congrats to you with the babies!!!!!!!!!!

I love this thing. As a programmer, it makes me slightly concerned for my future as a professional programmer, but I shan't sweat it. I use it a lot to start my projects, which it's great at. It also starts my stories for my novel and gives me advice on how I should do that too.