🔴 This Agentic AI Workflow Will Take Over 🤯 Algorithm + Papers Explained

what if I told you that you've been using AI the wrong way and that there is a much better optimized way to get the best out of even a simple generative model researchers in artificial intelligence believe that the pass to AGI or artificial general intelligence is by leveraging the agentic AI workflow but what exactly is an agentic AI workflow and how can we use it in our daily task and business decision making to find out we will go through several important Concepts in very recent AI developments using examples from prominent research studies that I will share with you today most probably you've been interacting with AI as a zero shot use of large language models and chat Bots for example asking in llm or chatbot to generate a text story code or image for you and sometimes getting disappointing results but there's a much better way of interacting with an llm or anyi tool and that's to make an agentic workflow to give an initial simplified understanding of it an agentic system ask the same model to refine its outputs in an iterative process using a series of feedback to get to the final answer or results on a task this is pretty much how he humans complete a task we just don't sit there and complete a task from start to finish instead we make a series of revisions on multiple iterations of our works by redrafting proofreading fixing the errors asking our colleagues for critical iCal comments and so on to complete a project this is illustrated by this research study selfrefined iterative refinement with self feedback by medon atau in 2023 as they put it the main idea is to generate an initial output using an llm then the same llm provides feedback for its output and uses it to Define itself iteratively self- refine does not require any supervised training data additional training or reinforcement learning and instead uses a single llm as the generator refiner and feedback provider they tested this new method across seven different tasks including text and code generation and found that the selfrefined versions out perform their base generative models such as GPT 4 they also noticed that the outputs keep improving with more iterations this ability of an AI system or language model to correct and refine itself is also refer refer to as reflection you can see an instance of it in this new framework reflection language agents with verbal reinforcement learning by shinol in 2023 the Au showed that using linguistic feedback and asking the model to self- refine itself is more effective than the traditional reinforcement learning method for example by updating weights as you see the new reflection method outperformed previous refinement methods on reasoning and decision making as well as code generation you might have already experimented with this before for example by asking chab or any local llm to give you an alternative or better output or to correct a faulty code but did you know that now algorithms can do this for you too let's take a look at one of these algorithms the reflection method of agentic AI uses verbal reinforcement to help agents learn from prior failings reflection converts binary or scalar feedback from the environment into verbal feedback in the form of a textual summary which is then added as additional context for the llm agent in the next episode an evaluation signal which helps the system to find the cause of an error or problem further amplifies the feedback in the previous step these two feedback blocks remain in the long-term memory and then pass to the actor which is built on a large language model this itative process continues until the desired output is received from the agent I highly recommend you read this paper to see how this interesting evaluator component works but that's not all to agentic AI systems a main component in the agentic AI workflow is the ability of these generative models and large language models to use tools just like humans to perform a task if that's new to you keep watching this paper even proposes a gentic systems that can access a variety of tools via API calls the users have also made a tool called gorilla which is based on llama and can effectively write API calls with accurate arguments look at this example where gorilla outperforms both gp4 and CLA in an API call they successfully tested gorilla on a document retriever task and even built a comprehensive API data set that all llms can use as part of completing a user task is that not so cool the third important Concept in an agentic AI workflow that I wanted to share with you is planning in the words of Andro in it is giving the llm the chance to work more slowly and plan the completion of a task in steps for example by explaining the reasoning behind each result in clear steps this concept is explained in this paper as Chain of Thought prompting or providing a model with a series of intermediate reasoning steps that they claim can help llms tackle complex reasoning tasks they show how prompt engineering with a Chain of Thought outperforms standard prompting across various language models and test data sets feel free to pause the video and read these examples of the Chain of Thought that are implemented on a famous reasoning test GSM AK I've already made a video about this test data set I'll put a link to it in the description box as well this concept of a as a controller of other a models has already been implemented in hogging GPT as described in this paper what the authors have done is essentially use chpt as a controller to receive the task prom from a user and then divide it into subtasks the controller then uses any relevant model available in the hogging pH repositories to perform the subtask and retain results and even combine the results of multiple AI models for very complex task that involve multiple mod it such as text speech and vision an example of it is to generate an image of a girl reading a book with hair pose being the same as another image provided to the agent and finally to describe the new image in your voice this really fascinating framework follows four stages of task planning task selection task execution and response generation then you can also use to complete your projects I will leave the full reference of this work in the Des description box so you can also check out these examples such as visual question answering and object detection as if this advancements aren't cool enough we have now a very exciting framework for multi-agent collaboration as well just as multiple AI models or llms can collaborate on one task to make an agentic system we can get multiple agentic systems to collaborate with each other too think crew AI or autogen what's even more more interesting about systems such as autogen is the collaboration between humans and multiple agentic systems this is so important because in some cases human input and judgment become necessary to direct the course of action by llms and their agentic designs this step is usually done by customizable and conversible agents the cherry on top is that many of these systems and algorithms are already open source which means you can use them right off the bat to build your own applications the cles of autogen for example showcased six applications and use cases such as math problem solving retrieval augmented chat Alf chat for example based on the idea of Alf World in this application as an AI detector and Navigator multi-agent coding Dynamic group chat and conversational chess but these are not just potential use cases the chat Dev application has already been built on the same idea that is fully described in this paper communicative agents for software development Chad Dev is a virtual company that employs multiple agents with different roles in the process of software development such as designing coding testing and documenting agents each of these agents could use multiple llms or generative models that mimic what human Engineers do and Converse in a real V company for example self-reflection instructions and testing and so on the author's document that how this multi-agent workflow significantly reduces the costs of software development even when using API calls but they emphasize that because these systems are in their initial development stages there are some biases and risks involved in the process depending on the choice of llms and their training data if you like this video consider subscribing and let me know if you have a plan or idea to use an agentic a workflow for your projects

Channel: Analytics Camp

Views: 10,276

Keywords: ai, llm, agents, agenticai, chatdev, autogen, crewai, reflexion, llm agents, ai agent, hugginggpt, huggingface, api, gsm8k, multi agents, multimodal agents, gpt4, chatgpt, ai tasks

Id: lA3Tju4VUho

Length: 9min 29sec (569 seconds)

Published: Tue Apr 23 2024