In today's rapidly evolving business landscape, where technological advancements are reshaping industries, progressive leaders recognize the imperative of staying ahead of the curve. Embracing the transformative power of AI is not just a strategic choice but a necessity.
Traditionally, there has been a tendency to allocate technical responsibilities within established organizational structures. However, the emergence of the AI wave presents a potential shift by introducing the possibility of AI Agents automating various departments. Visionary leaders recognize that navigating effective leadership in this era requires a synergy between business insight and technical proficiency.
This brief yet pivotal exploration of a technical framework offers an opportunity for visionary leaders to equip themselves with the knowledge needed to navigate the impending shift, ensuring not just survival, but a thriving future for their organizations.
The article guides executives through comprehending the evolution from GPTs' text generation to automating entire departments using advanced agents, heading toward achieving Artificial General Intelligence (AGI) or AI 3.0, within the upcoming year.
Exploring the progression of Gen AI business applications such as AutoGen from Microsoft or Assistants from Open AI in the context of a Cognitive Architecture framework, the article elucidates how cognitive elements like reflection and memory are swiftly evolving, enhancing Generative AI Agents' capabilities and leading to an emerging AGI paradigm that we call AI 3.0.
Armed with this insight, leaders can adeptly survey the market and strategically initiate immediate functionality integration while devising a timeline for gradually incorporating agent capabilities over the next few months and a year. This enables them not only to implement AI functionalities but also to plan for organizational transformation in tandem with these advancements.
Recent Developments that Necessitates a New Strategic Framework
The period from 2020, marked by GPTs, to 2023 has witnessed remarkable advancements in AI.
Developments have been so swift that it has become customary even for AI researchers to consider papers published just six months prior as outdated.
While the prevailing belief suggested that achieving artificial general intelligence (AGI) is distant or implausible, the rapid advancement of AI paints a different picture, with companies like Open AI founded for the purpose of realizing AGI, within a year, as some of the recent news about Open AI has started to reveal.
With AI agents approaching human-like capabilities, the urgency for businesses to integrate "AI-Agents" at their core becomes increasingly apparent. Neglecting this integration could lead to swift customer turnover and rapid obsolescence in the face of competition.
Pioneering companies navigating rapid technological changes must grapple with tough decisions. As AI progresses, it's poised to assume significant enterprise functions, necessitating reductions in procurement, sales, administration, and leadership layers. However, this shift also offers the chance to reallocate the workforce towards roles that add direct value to the products, emphasizing engineering, research, strategy, and areas reliant on human capital.
To navigate the dynamic landscape of Gen AI heading towards AGI, it's essential to underscore the significance of comprehending Cognitive Architectures. The rapid pace of developments in the AI realm may seem overwhelming even for those well-versed in the field.
For this reason, we propose that business leaders adopt an explorer's mindset, embracing advances systematically without succumbing to the sense of being inundated. Let's initiate our exploration by gaining insights into the origins of the current Gen AI revolution before delving into the swift advancements of Generative AI and its practical applications in the business domain.
Attention: Foundational Concept of Artificial General Intelligence
Gen AI's progress taps into a crucial link between natural and artificial intelligence, usually missed in AI talks. But if we look closely, they both start from attention—the focus that all active things, living or not, have in common.
In nature, animals like frogs pay close attention to important things around them. When a frog spots a fly, its brain area for recognizing moving things gets super active, showing how much attention it's giving the fly. This helps the frog notice details about how the fly moves or looks, making it easier to catch.
Attention works the same way for us, guiding what we notice and how we react. This idea of attention being central to consciousness isn't just for animals; it's a big deal in understanding how artificial intelligenembeddingsce systems, like those in AI, work too. The close connections between AI based Cognitive Architectures and Cognitive Neuroscience is highlighted in this article Humanizing AI: A Journey into Artificial Intelligence Through Cognitive Neuroscience.
2017: The groundbreaking AI paper: "Attention is all you need"
In the realm of AI, attention emerged as a pivotal element, analogous to the role it plays in human agents. Neural networks incorporated attention to learn how and where to focus on observed stimuli, refining their predictive capabilities about the world.
The foundational neural networks for these AI agents were denoted as transformers. Transformers are the underlying technology powering Large Language Models (LLMs) and the field of Generative AI.
Transformers use a mathematical formula to implement attention in Neural Networks. Given a word sequence, like "how are you," the goal is to predict the next likely word, such as "doing." Transformers convert the sequence into embeddings, a powerful format.
Each word, like "how," considers nearby words (e.g., "are" and "you") to predict the next word, adjusting its value based on an attention score. These scores create embeddings, enabling accurate predictions for the next word in the sequence.
From 2018: GPT - Generative Pre-trained Transformer
As the transformer analyzed an increasing number of sentences, its ability to attend improved progressively, reaching a point where extraordinary models like ChatGPT emerged. These LLM models, captivating the world's imagination, possess the ability to generate predictions or take actions based on observed input with an uncanny resemblance to human thought processes.
Think of GPT models like how we develop instinctual habits to perceive and act in the world, without thinking too much. GPT learns to do something similar, but in the digital world of words, pictures, and sound.
In a whirlwind of advancements, OpenAI's ChatGPT, powered by GPT-3.5, broke records with 100 million users in just two months, marking a milestone in text generation. Meta has disrupted the trend of AI companies keeping their models proprietary by openly releasing the code for Llama 2, its advanced large language model, potentially posing a challenge to ChatGPT's dominance. Entities such as Hugging Face surfaced to democratize large language models by making them open source, effectively becoming the GitHub equivalent for LLMs and significantly hastening the pace of open-source development.
Meanwhile, specialized LLMs like Google's Med-PaLM targeted specific domains like clinical knowledge. Salesforce introduces Einstein GPT for customer relationship management, while Anthropic's Claude introduces ethics into AI with constitutional AI.
Businesses are now leveraging ChatGPT for various tasks including generating summaries, crafting SEO-friendly keywords for specific topics, brainstorming new ideas, composing customer service emails, elucidating complex concepts, building responsive chatbots using ChatGPT's API, and assisting with web development and coding, among other applications.
Prominent contenders, including Google Cloud Platform (GCP), Amazon Web Services (AWS), and Microsoft Azure, are fiercely competing by introducing their respective cloud services for generative AI, as this technology requires vast amounts of data and computing power. These providers create an entire Gen AI ecosystem comprising of text, code and image foundational models, both commercial and open source model catalogs, vector databases, ML Pipelines, fine tuning capabilities, low code development and code completion.
Furthermore, companies like Microsoft excel in delivering Gen AI capabilities alongside comprehensive offerings, including cloud computing, productivity tools, and ERP software, providing a holistic and convenient solution for enterprises.
2020: Domain Expertise through Retrieval Augmented Generation (RAG)
As these transformers underwent training on an extensive general corpus, they acquired broad knowledge of the world. The next state of the evolution of transformers was the question of how to adapt them to specific domain. To enhance their adaptability to specific contexts, one could either retrain/finetune them or supplement them with domain-specific knowledge sources.
Retraining AI models is like learning a new skill; the model weights adapt to the new domain. Retrieval augmented generation, akin to using a manual, lets AI access specific data without changing their overall learned weights. This method combines general skills with specific domain knowledge, improving tasks like question answering and dialogue systems with more contextually fitting responses.
RAG, a versatile tool, brings several business benefits. First, it improves websites by adding responsive chat features for easier navigation and user assistance. Second, it simplifies data queries, making complex information more understandable for teams to make informed decisions. Additionally, it speeds up responding to RFPs/RFIs, making the contract securing process more efficient. Lastly, it provides personalized recommendations across different areas by interpreting preferences, offering tailored suggestions linked to the right platforms.
Several companies provide solutions for implementing RAG in businesses. Azure Machine Learning enables RAG integration. OpenAI's ChatGPT Retrieval Plugin enhances responses using document databases. HuggingFace offers a RAG model transformer. IBM Watsonx.ai uses RAG for accurate outputs. Meta AI combines retrieval and generation in one framework.
2022: Emergence of Agents
As people used transformers more for complex tasks, they found that careful prompts led to better answers. Amazon's multimodal-CoT model goes beyond GPT-3.5 by using "chain-of-thought" reasoning, similar to breaking down a math problem into steps for a better solution, like human thinking.
This shift in AI lets machines think deeply, highlighting the importance of reasoning for better results. But relying only on deep thinking isn't always enough—it's like how human reflection can lead to occasional mistakes.
To improve, it was observed that AI needs to interact with the real world, gather data, and learn from actions taken. This philosophy, called praxis, complements deep thinking with action and helps AI understand reality better. But how did this evolution happen?
The chain of thoughts approach, realized through frameworks such as Langchain, transformed AI by blending chat functions with diverse data sources. This enabled AI to actively engage with data while maintaining coherence, advancing beyond reflective thinking. Such developments highlight the fusion of reflection and real-world interaction, setting the stage for theoretical frameworks like Reflexion and ReAct.
AI agents emerged to stand out from previous LLMs in three primary abilities: they merged reasoning, action, and reflection in a closed loop formulation.
Several companies started to offer AI Agents over a short period of time. AI agents from ThinkChain streamline deal evaluation, portfolio management, and financial modeling in real time, offering updated insights and swift decision-making support. GoCharlie announced the first of it's kind multimodal content creation agent for marketing. Aomni emerged as an AI agent specifically designed for business intelligence.
The Next Leap: Collaborative Conversational Agents
In the autumn of 2023, the landscape of Large Language Models (LLMs) experienced significant excitement with the emergence of multi-agents like XAgent, AutoGen, MetaGPT, and BabyAGI, among others. A notable discovery was made, indicating that collaborative efforts among multiple agents prove to be a more powerful paradigm for tackling complex tasks.
AutoGen from Microsoft, a platform with a specific focus on developing conversational AI applications using multiple agents, stands out in this landscape. It offers various features tailored for building conversational AI applications, including support for multi-agent conversations and effective context management. AutoGen adopts a graph-based approach, allowing components to be interconnected in diverse ways to create intricate conversational flows.
Automating Entire Departments using Advanced Agents
Having tracked the evolution of LLMs from their inception to the era of multi-agents, let's now explore their potential to revolutionize enterprises by automating entire departments.
Consider for example automating the procurement department. Imagine an AI-driven journey in procurement where smart agents start by predicting demand shifts using historical and market insights, aligning this with internal data for precise forecasts. Collaborating with stakeholders, they swiftly craft schematics for new parts, ensuring alignment with business objectives.
These agents then dive into the world of supplier selection, evaluating data for trustworthy partners and devising adaptable sourcing strategies across geographies. As they engage suppliers, streamline RFP cycles, and mediate stakeholder discussions, they're also recommending the best-fit suppliers based on communication analyses.
Seamlessly, they assist in creating purchase requests, drafting contracts, and placing orders while optimizing ERP interactions for time efficiency. Continuously, they monitor risks, suggest improvements, and enhance collaboration, ensuring a resilient and evolving procurement ecosystem.
These AI agents, mastering procurement, hint at a broader future—where they seamlessly predict sales trends, optimize HR functions like talent acquisition, and refine operations through streamlined workflows. Their potential spans across departments, unifying data insights, automating tasks, and revolutionizing the enterprise landscape for enhanced efficiency and innovation.
Emergence of a Cognitive Architecture with System 2
Looking ahead, teamwork among agents will enhance learning strategies for handling complex tasks. This shift involves moving beyond simple weight adjustments to incorporating long-term rules and abstractions.
These conceptual frameworks will capture high level patterns in the world, crucial for agents during reasoning. This reasoning process, based on a tree-of-thoughts search, will separate into system 2 (cognition) from the lower-level system 1, which primarily perceives and acts.
The original system 1 loop of perception-action will still remain for its efficiency. The perception module will still observe the world's state, while the actor module would generate actions via a policy module. This process would operate reactively without utilizing the world model or cost considerations. as explained by Yann LeCun in his article A Path Towards Autonomous Machine Intelligence. This article Cognitive Architectures for Language Agents on cognitive architectures is also enlightening, showcasing the integration of concepts from Knowledge AI 1.0 with the accomplishments of extensive language models associated with Data AI 2.0. The outcome is Artificial General Intelligence that we categorize within the AI 3.0 paradigm.
These trends clearly show that the development of Generative AI will move beyond the current GPT architectures which are impressive and fast executors but slow and costly learners.
Thinking, Fast and Slow" by Daniel Kahneman is like a tour guide through your mind. It talks about two thinking modes: System 1, which is like autopilot, quick and automatic, and System 2, which is more like a deliberate, slow-thinking process.
In the next quarters we will see the emergence of a Cognitive Architectures where reflection and reasoning as System 2 emerge as a distinct entity, apart from the perception-action cycle. Through System 2, the agent will gain the ability to contemplate potential outcomes without physically testing them in reality. This circumvents the need for numerous risky trials to determine optimal actions.
Through system 2, the agent would envision various actions and anticipates their potential consequences before taking any concrete steps. Through ongoing cycles of self-reflection, the agent would begin to discern which strategies of reflection are beneficial and which are not.
AI 3.0: Cognitive Architecture for Artificial General Intelligence (AGI)
Through the process of reflecting on strategies, the meta-cognitive abilities of System 2 will acquire diverse higher-level concepts about the world, encapsulated within episodic, procedural, and semantic memories, through the process of self-learning.
During reasoning processes, it will draw upon and retrieve information from these memories based on the specific requirements at hand.
As System 2 that is based on explicit reasoning starts to produce better outcomes, it will in-turn train System 1. This process is analogous to how we train our habits by means of conscious deliberation.
Additionally, there will be a push for multi-modal integration, incorporating vision and connecting AI more deeply with the physical world. As AI delves into modeling human intentions, it will increasingly align with the our own psychological reality, paving the way towards achieving Artificial General Intelligence (AGI).
All in all, the developments outlined above will create AI 3.0 Cognitive Architectures that combines the System 1 capabilities of deep learning or Data AI with explicit knowledge representation and reasoning techniques typically associated with expert systems or Knowledge AI. Ideally, the integration should foster transparency, reliability, and trust in AI.
Conclusion: Business and Economic Impact of AI 3.0
Generative AI, like ChatGPT, is getting a lot of attention, but it's crucial to look at its history to grasp its fast changes and use its potential. The move towards collaborative agents, shown in frameworks like Reflexion and React, emphasizes reflective thinking. As these agents work together, future improvements focus on explicitly learning higher level concepts in an explainable manner, showing a shift towards Artificial General Intelligence (AGI).
Companies dealing with quick tech changes face tough choices. As AI gets better, it can handle many business tasks, which means jobs might be significantly reduced in areas that are categorized by accountants as indirect costs such as procurement, sales, management, and mid-level leadership. But this change also gives a chance to move workers into roles that adds direct value to products, like engineering, research and business strategy.
Looking at things this way, a new kind of AI 3.0, the third generation, combines reasoning and learning, knowledge and data. This is seen as vital for tackling big economic challenges in areas like manufacturing and agriculture, where planning and monitoring needs to be combined in a closed loop fashion for increased efficiency. The trigger and relevance of AI 3.0 for sectors in the real economy are elaborated further in the article Driving Growth: The Interplay of Macroeconomics and AI 3.0.
The integration of AI into these processes will bridge the gap between innovation and production, bringing them closer to enterprises and the supply chain. This has the potential to revolutionize productivity in critical sectors of the real economy, including manufacturing and agriculture, potentially leading to abundance and widespread prosperity. Needless to say, a crucial prerequisite for this transformative progress is the democratization of AI development.
To conclude, the purpose of this article was to inspire a brain inspired Cognitive Architecture framework that helps leaders guide their organizations toward using AI more, bringing in a future where smart AI use sparks innovation and adds value through human work.
Krishna Sridhar
CEO, Sparsa AI
Comments