ChatGPT Explained in 5 Minutes

ChatGPT has taken over Twitter and pretty much the whole internet, thanks to its power and the meme potential it provides. We all know how being able to generate memes is the best way to conquer the internet, and so it worked.Since you’ve seen numerous examples, you might already know that ChatGPT is an , that you can chat with. It is also called a chatbot, meaning you can interact with it conversationally, imitatting a one-on-one human discussion.What you might not know is what it is and how it works... Watch the video to learn more!

References

►Try it:
►OpenAI's blog post:
►What is GPT-3:
►What is Reinforcement Learning:
►Join our Discord community:
►Twitter:
►Support me on Patreon:

Video Transcript

0:00you've seen it everywhere Chad GPT has0:02taken on Twitter and pretty much the0:04whole internet thanks to its power and0:06the meme potential it provides we all0:08know being able to generate memes is the0:11best way to conquer the internet and so0:13it worked since you've seen numerous0:14examples you might already know that0:16chatgpt is an AI recently released to0:19the public by openai allowing you to0:21chat with it it's also called a chatbot0:24meaning you can interact with it0:25conversionally imitating a one-on-one0:28human discussion what you might not know0:30is what it is and how it works0:32chadjupiti is a model based on0:35reinforcement learning and the GPT0:37series of models from openai I will0:39refer you to a video about reinforcement0:41learning we recently published with my0:43friend Elias to learn more about the0:46subfield of AI but quickly reinforcement0:48learning is a way to train algorithms by0:51trial and error aiming for rewards just0:54like humans would do by learning with0:56positive feedback more specifically chat0:58GPT was built following three steps the1:02first was to take an already powerful1:04model and fine tune it with supervised1:06learning what does this mean it means1:08that they took a model specifically its1:11GPT 3.5 and improved and up-to-date1:14version of gpd3 which they trained once1:17more on conversation examples1:19specifically instead of being trained on1:21pretty much their whole internet as gpt31:24was this means they are trying to narrow1:26its potentials strictly to conversations1:28making it theoretically better at1:31conversing compared to gpt3 since a1:34specialist is almost always better than1:36a generalist at a specific task if you1:38are still not familiar with the GPT1:40series of models I would suggest1:42watching the short introduction video I1:44made covering gpt3 when it came out the1:47second step is to add our reinforcement1:49learning magic which will allow the1:51model to practice and get better as you1:53know practice makes perfect more1:55precisely in this step we will use the1:57model to chat with humans directly have2:00it provide multiple possible answers and2:03ask the human to rate the answers from2:05best to worst this data will then be2:07used to train another model called our2:10reward model learning to replicate our2:12human annotators this leads to our last2:15step where our new reward model will2:18give feedback to the chat GPT model's2:20answers as a reward function to help it2:22converge toward the best answers over2:24time this final step is to further train2:27our algorithm after the initial fine2:30tuning step we explained this is why it2:32is companies like open AI that release2:34those kinds of amazingly powerful models2:37it will be unfeasible for universities2:39or individuals as it requires way too2:42much Computing and time for training2:44still what they achieve is quite2:46remarkable and I believe they are worth2:48doing and worth sharing to Advent2:50science and voila after coupling the2:54already powerful and most recent GPT2:56based language model fine-tuning it to2:59conversations and finally using3:01reinforcement learning to make it3:03practice its conversation skills you3:05obtain chat GPD as you have seen before3:07the model is quite promising but also3:10sometimes a very dumb and doesn't seem3:12to have any logic whatsoever it is still3:15just an algorithm and far from being3:17either intelligent or conscious though3:20it will depend on how we Define both it3:22definitely has its limitations3:24nonetheless the outputs it gives are3:27often surprisingly interesting and3:29pertinent chatgpt is definitely a step3:31forward in conversational Ai and quite3:34promising especially working on the3:36prompt engineering side of the model to3:38leverage its true potential and limit3:41failure cases I hope you've enjoyed this3:43video and I'd love to see your3:45experiments please tag me on Twitter at3:47what's AI if you share them or join our3:50Discord Community where we created a3:52channel specifically for it I will see3:54you next week with another amazing AI3:57research4:00foreign4:04[Music]