GPT-4 Can Now IMPROVE Itself Without HUMAN Intervention

Auto GPT is now able to run its own code and execute python script is chat gbd to find me some money and within a minute I had 210 dollars coming to my bank account gbd4 is now able to reflect on its response and create an even better response if you haven't seen this interface for GPT it's absolutely amazing if you want to turn yourself into a 3D model you can now do it way easier and better they just created hugging GPT which basically solves all the AI tasks with chat gbt as the controller chat GPT can't read PDFs but now it can if you're scared of AI and you think it's going to take over the entire world and it's gonna just be like hit a button and all of a sudden it's out of control well read this gbd4 is now able to reflect on its response and create an even better response see the success without reflection is linear and the success with reflection in the blue here is just gradually growing gpt4 without self-reflection got a 67 percent while reflection plus gpd4 got an 88 this is a massive leak and the model is extremely simple it gives an answer and then it basically says is the answer correct and then it says no it's actually not correct I'm sorry let me iterate again and then find another answer here is a perfect example where gpd4 goes wrong make it wrap with every word starting with the letter b and you can see it does amazing except areas unmasked expressing now if we ask it did you perform the previous action correctly it says apologies for the oversight in my previous response I didn't create a wrap with every word starting with the letter B here is a corrected version on the the second try it only got one word wrong compared to the first one it got seven words wrong it can be quite scary to let the AI self reflect but at least we know that it won't have a midlife crisis and buy a Supercar Auto GPT is now able to run its own code and execute python script this allows it to run on its own take a look at this video you can see this self-improvement evaluating the code for syntax errors and logic issues improving the code based on the evaluation suggestions adding tests for the functions in basic math.pi pending the tests to the basic math.wi file executing the basic math.pi file to test the code you can see that it's writing on the right side fixing the syntax issue with the calc underscore circumference function this is saving the updated code and tests to the basic math.pi file executing the updated basic math.pi file to test the code code in basic math.y has been successfully evaluated improved and test tested the task is now complete Andrei karpathy says the next Frontier of prompt Engineering in my opinion is Auto gpts One gptl Call is just like one instruction on a computer they can be strung together into programs use prompt to Define input output device and Tool specs Define the cognitive Loop page data in and out of context window dot run smoke away tweeted can Auto GPT lead to AGI and 65 said yes it's completely open source I'll leave a link in description I ask chat gbd to find me some money and within a minute I had 210 dollars coming to my bank account that's what Joshua Browder said on Twitter and he basically asked it to browse the web search for unclaimed money in California the first idea it came up with was to visit the Obscure Government website California State Controller this website holds unclaimed refunds for company that can't contact you it gave me a link with step-by-step instructions on what to do and here are the instructions I followed the instructions and sure enough 209 dollars was waiting for me one minute later the only thing that stopped the AI doing it itself was a captcha companies will never build these Integrations directly because it loses their money Comcast isn't gonna let you cancel with a chat GPT plugin if you haven't seen this interface for GPT it's absolutely amazing look you just start with a chat bot straight in the middle and ask it a question and then it will answer the question called the three library and now it will pop up multiple choices that you can actually click on and for example are there any alternatives and here it will start writing the other one you can also click on more Alternatives here as well pull them round and continue doing the brainstorming so here you can see what are the performance considerations when using D3 and you can go down the rabbit hole to multiple different Bubbles and you can see that if the first answer didn't provide you the best thing that you wanted to go down you can just select the other answer that you actually wanted to go down and now you can see that the information is growing rapidly that two becomes four becomes 8 which is exponential compared to how it is with just a linear answer access is invite only for now follow for future updates and just subscribe if you want to turn yourself into a 3D model you can now do it way easier and better here is the new way to turn yourself into a 3D model as you can see they've upgraded the expressive 3D human avatars from simple x to Super our latest and greatest body model this new version is trained on 1.2 million 3D scans and is more expressive it includes feet with articulation it looks absolutely wild you can see that with a single picture they can create a full body 3D scan they also have this amazing model on how they let Fabric and clothes kind of perfectly float around the body have you tried using hugging face before but it's just too complicated you don't know what the code is and maybe you have to use multiple ones to get the result you actually want well they just created hugging GPT which basically solves all the AI tasks with chat gbt as the controller as you can see chat GPT is the controller to connect numerous AI models EG those in hugging phase for solving complicated AI tasks in this concept an llm acts as a controller managing and organizing the cooperation of expert models the llm first plans a list of tasks based on the user's request and then assigns expert models to each task after the experts execute the tasks the llm collects the results and responds to the user so here you can see the request is please generate an image where the girl is reading a book and our polls is the same as the boy in the image example jpeg then police describe the new image with your voice as you can see stage 1 is Task planning so it's trying to figure out all of the things it needs to do stage two is the model selection so it goes into hugging face and says oh I'm gonna use this one and I'm gonna use this one and maybe this one then state three it's just the task execution so it's running the models and figuring out okay let's put all of these together and then and stage four is response generation and as you can see here the response is here and this is how it looks like you can even see he was using this model and then this model and this model this model this model and this model all after each other hugging GPT is like a conductor to an AI Orchestra but instead of making music it makes you a hundred times more effective did you ever hear something on a podcast and then you're like oh I want to search that on Google but obviously Google doesn't give you a good response well this company is trying to make it easy to search through podcast they just launched huberman AI so if you don't know who humor Ben is he has one of the most downloaded Health podcasts in the world and here you can basically just write any question to him and the AI model is trained on all of his episodes they also have Lex Friedman podcast big fan Tim Ferriss podcast coming soon and one of my favorites first million so you can start asking questions I'd love for them to train it on my podcast as well link down below if you're scared of AI and you think it's going to take over the entire world and it's gonna just be like hit a button and all of a sudden it's out of control well read this this is the chief AI scientist at meta some folks say I'm scared of AGI are they scared of flying no not because airplanes can't crash but because Engineers have made airliners very safe why would AI be any different why should AI engineer just be more scared of AI than Aircraft engineers were scared of flying Elon Musk responds actually airplanes used to crash frequently eventually so many people died that the FAA was created to ensure that commercial aircraft makers and airliners didn't cut corners on safety now flying on an airliner in America is super safe yawn responds my opponent exactly it took 50 years between the first flights and the creation of the FAA why be scared of AI when we don't even have a blueprint let alone a demo of a system capable of human level intelligence it's like Auto being scared of engine failures this guy doesn't get it man because it might also take 50 years to get it right but unlike airplanes we may not have the chance to get there progressively through trial and error if the errors are irrecoverable I'll be scared of flying in the first ever plane I'd be unwilling to load every man woman and child on earth into the first supersonic flight planes got saved by crashing repeatedly and US learning from our mistakes repeatedly AGI might not give us multiple chances like that because in this case any random dude can fly an airplane into your house for the low simple in the next few years everyone that has a computer will be capable of creating AGI pretty positive that a 10 year old could be capable too humans are dangerous AI is not as of now humans are to be replaced with AI for the sake of evolution unless we merge Simon says why would you fly an airplane that hasn't passed FAA tests I wouldn't I'm disappointed by your reaction to people asking for more security and regulations at least now we all know your true colors damn bro the question is why are you against the regulation process starting now as with everything else it will take time why not having it ready by the birth of AGI I would love to see AGI helping us solve all the problems Humanity faces today including answer but at the same time I would really hate if when AGI is used to develop super smart weapons we should regulate it and enjoy the good side of AI chaptbt can't read PDFs but now it can it's called PDF chat bot and you can upload and chat with a book one of my favorite short reads is this 1000 true fans by Kevin Kelly and I'm just gonna drag it in here and just like that it's starting to read and it's kind of summarizing it in the beginning here it's about 1 000 true fans and how it can be used as a strategy for creators to make a living by cultivating a dedicated fan base that will buy anything they will produce and that they only need a thousand of them to earn a hundred thousand per year if your creator is an amazing read now if I ask it what are the key components to getting 1 000 true fans wow focusing on the unique aspects of your work that your true fans appreciate nurturing your true fans and converting less their fans into true fans so let's say how can I convert lesser fans into true fans it says creators should focus on maintaining direct contact with their true fans and providing them with unique and high quality content that they will appreciate by doing so true fans will continue to support the Creator and spread the word to their own networks which can attract more fans additionally creators can offer exclusive content or experiences through the true fans so such Early Access to new work or personalized messages to make them feel valued and appreciated wow this is really helpful how do I know that there are true fans true fans are defined as fans who will buy anything and everything you produce they will drive long distances to see you perform by multiple versions of your work and eagerly await your next release are any of you guys true fans I guess if you want to get in direct contact with me the best way is on Discord and I read every single message I'm just really enjoying reading everything about Ai and then sharing it to you guys and the reception has been great I don't care for you guys to buy anything at least not yet so I need to get an audience before I know what to sell I'm grateful you guys are nerding out here with me checking up on what is happening every single day and I check you guys the comments as well and you guys are hilarious sometimes so that's it for today thank you so much for watching if you want to join the journey subscribe down below and I'll see you tomorrow peace

Channel: AI Andy

Views: 104,487

Keywords: chatgpt, gpt4, openai, google bard, microsoft, bing ai, midjourney, palm-e, ai, artificial intelligence, hugginggpt, chatgpt pdf, huberman ai, reflexion, reflexion ai, gpt reflection, gpt-5, gpt5, gpt 5

Id: LuIJRnMJHbM

Length: 13min 39sec (819 seconds)

Published: Mon Apr 03 2023