AI Synergy: MetaGPT's Multi-Expert Collaboration Approach
Explore how MetaGPT is changing the game by enabling AI-powered multi-expert collaboration. Learn about the challenges of hallucination, MetaGPT's interaction workflow, and its human-expertise based approach. Discover the potential of MetaGPT for revolutionizing AI-powered work coordination.
The recent few months in AI focused on one single element:
How to automate tasks at scale using the power of LLMs?
But not any tasks, those that were supposed to be made by humans while existing automations tools took take of the daunting manual work they were used to do.
LLMs and ai foundational models got the ability to do creative tasks and give outputs very close to what expert humans in a specific industry could do, and were even getting close to AI surpassing the capabilities of humans.
However, a lot of chalenges remained:
- How to create tools that can automate tasks based on necessary skills?
- How could we have GPT like infrastructure that can interact between each other with a global task in ‘mind’?
- How can we make sure these models don’t hallucinate during their interaction?
These 3 questions alone led to billion dollars of investment in different startups, but the first move in that regard came from an existing player of social networks:
Meta launched MetaGPT.
MetaGPT’s mission is to infuse human workflow type of interaction in an LLM based multi ‘experts’ collaboration. But that won’t be easy:
MetaGPT challenge: The Hallucination Problem in Multi-Agent Systems
The goal of a tool such as MetaGPT is to give an output independently from the input given, you decide what the mission is and let collaboration of the experts come up with a solution/proposition.
But LLMs have the tendency to hallucinate, they can generate false information just to get the work done. See it as a confident output even if it’s the veracity of the information is not present in the trained data.
Well, when you interact with chatGPT and this occurs it’s in a close environment and with a single output:
- You ask a question.
- GPT answers you with a confident false answer.
- You see it and re-ask your question.
But what happens when using a tool such as MetaGPT in a professional setting?
- You give a mission.
- The agent interact between them.
- One agent gives a false output.
- The other base their interactions on this false output.
So the hallucination challenge gets amplified in the multi-experts interactions, MetaGPT therefore could be very dangerous when no backup verification plan is put in place.
Especially as the tool comes from Meta, some companies might believe it has ‘authority status’ and blindly believe in it’s ability to achieve complex tasks.
Is MetaGPT a real revolution?
MetaGPT is the tool with the most up to date state-of-the-art machine learning techniques for the text optimisation outputs.
As it is trained on a vast amount of data, MetaGPT achieves unprecedented levels of language understanding and generation which makes it easy to use by people from all type of backgrounds.
The architecture of MetaGPT, based on transformers, captures long-range dependencies and intricate language structures.
This allows it to generate text with a level of nuance and sophistication that was previously unattainable, contributing to more accurate and contextually relevant responses.
The approach of setting experts/agents with different capabilities each interacting with each other based on:
- The skills given
- The data as inputs
- The number of interaction
Is giving a new approach to conversational chat-based multi-agent systems. While chatGPT is mono discussion and needs your intervention at each step of the process, MetaGPT is getting a step further.
MetaGPT's interaction workflow model
So far what AI is missing is the ability to replicate the foundational structure of how we work and interact with each other as humans:
- Project based structure.
- Skill based allocation of work.
- Brainstorming and debate sessions.
- Divergence and convergence based creativity.
Having your different chat discussions on chatGPT is good, but each are evolving between closed doors are based on the same model without any type of real prompt engineering optimisation.
What MetaGPT allows you to do is to assign divers roles to the ‘agents’ based on real-world persona types.
You can then give them a mission and let them interact with each other to give you the output you wish. That is the only approach that can efficiently tackle complex problems and provide a concrete tool to help human coordinate their work and interact with AI.
On top of that the number of interactions is unlimited, which means you can re-start working on a project from scratch without impacting the energy level of your employees while keeping a certain level of context (previous datas).
MetaGPT is creating the base fundamental structure for coherent system of interaction with AI.
The Human-expertise based approach of MetaGPT
Any output given by a human is based on:
- Skills
- Experience
- Mood at a specific point in time
- Environment in which the interaction happens
- Corpus of data given as an input for the specific mission
MetaGPT incorporating human knowledge approach provides user of the tool a way to build unlimited ‘coworkers’.
This is the first time an internal barrier is build between the output of the AI and the quality of it, the expertise given to the agents is a way to validate internally the outputs and reducing the errors in the interactions.
If you haven’t tried yet Argil, what are you waiting for? Try it here
