AGI, Are we there yet?

Artificial General Intelligence (AGI) is a hypothetical class of intelligent agents that has the ability to learn and perform any intellectual task that humans or animals are capable of. AGI, on the other hand, has been characterized as an autonomous system that excels at the most economically beneficial tasks. The development of AGI is the main objective of some artificial intelligence research and businesses like OpenAI, DeepMind, Meta, and Anthropic.

The evolution of AI

Since John McCarthy first used the phrase in the 1950s, the study of artificial intelligence (AI) has had a lengthy history. Since then, artificial intelligence has seen numerous stages of development, beginning with symbolic AI and expert systems in the 1960s and 1970s, moving on to neural networks and machine learning in the 1980s and 1990s, and finally to deep learning and natural language processing in the 2000s and 2010s.

However, as AI researchers attempted to address increasingly challenging and realistic issues, they ran into numerous obstacles and constraints. Among the difficulties are:

• Brittleness: Artificial intelligence systems frequently failed when confronted with unforeseen circumstances or inputs that differed somewhat from their training data.
• AI systems lacked the fundamental knowledge and comprehension of the world that humans take for granted.
• Knowledge Representation: AI systems struggled to represent and reason with complicated, unpredictable, and dynamic data.
• Scalability: To handle vast volumes of data, AI systems require a lot of processing resources and time.

To address these difficulties, AI research evolved into several subfields and paradigms, including probabilistic reasoning, expert systems, neural networks, fuzzy logic, evolutionary algorithms, computer vision, natural language processing, robotics, etc. For certain areas and activities, each discipline concentrated on creating specialized methods and applications.

As a result, AI research became dispersed and isolated. Different subfields and paradigms hardly interacted with one another or worked together. Additionally, research on general-purpose AI systems that could combine many capabilities and domains saw a loss of interest and funding.

However, most of these AI systems are regarded as weak AI or narrow AI, which means they can only tackle a single issue and do not possess general cognitive abilities. For example, a chess-playing program can beat human grandmasters, but cannot recognize faces or understand natural language. A face recognition system can identify people in photos but cannot play chess or write a poem.

The rise of LLMs

Many of us are now aware of large language models (LLMs). These are neural networks that undergo training on vast amounts of text data from diverse sources like books, articles, websites, social media posts, and more. LLMs exhibit the ability to produce coherent and fluent text on various subjects when given input or prompts. Additionally, they can handle other natural language tasks such as answering questions, summarizing texts, and translating languages.

Some examples of LLMs are GPT-3 by OpenAI, BERT by Google, and T5 by Google/DeepMind. These models have hundreds of billions or even trillions of parameters and require enormous computational resources to train and run. They have been shown to achieve impressive results on various natural language benchmarks and tasks, sometimes surpassing human performance.

Despite the advances in LLMs and other AI systems, AGI remains elusive. There is no consensus on how to define or measure AGI, let alone how to create it. There is also no agreement on whether modern deep learning systems are an early yet incomplete form of AGI or if new approaches are required.

?There are many research projects and companies that are pursuing the goal of AGI in various ways. Here are some examples of the organizations that are researching and investing in AGI:

• ?OpenAI: A research organization that aims to create artificial intelligence that can benefit humanity without being constrained by profit motives or human biases. It is also developing an AGI alignment framework that ensures that AI systems align with human values and goals.
• DeepMind: A subsidiary of Google that focuses on creating artificial intelligence that can learn from its own experience and achieve general intelligence across a wide range of domains. It is best known for creating AlphaGo, AlphaZero, and AlphaFold, which have achieved superhuman performance in games and protein folding respectively.
• Anthropic: A research company co-founded by Demis Hassabis (the founder of DeepMind) and his wife Daphne Koller (a Stanford professor) that aims to create artificial intelligence that is understandable, aligned, and scalable. It is also developing tools and methods for interpreting and steering large-scale neural networks.
• Vicarious: A research company co-founded by Dileep George (a former colleague of Jeff Hawkins at Numenta) and Scott Phoenix (a former entrepreneur) that aims to create artificial intelligence that is inspired by the computational principles of the human brain. It is also developing a cognitive engine that can perform various tasks across vision, language, and reasoning. Vicarious was acquired by Intrinsic, a robotics software and AI company at Alphabet.
• Numenta: A research company co-founded by Jeff Hawkins (the inventor of the Palm Pilot) and Donna Dubinsky (a former CEO of Palm) that aims to create artificial intelligence that is based on the theory of hierarchical temporal memory (HTM), which models the structure and function of the neocortex. It is also developing applications of HTM for anomaly detection, natural language processing, and robotics.
• GoodAI: A research company founded by Marek Rosa (a video game developer) that aims to create artificial intelligence that can learn, communicate, and generalize across domains. It is also developing a framework for gradual learning that allows AI systems to acquire skills and knowledge incrementally.
• Neuralink: A neurotechnology company founded by Elon Musk and others that aims to create brain-computer interfaces that can connect humans and artificial intelligence. It is also developing a device called the Link that can implant electrodes into the brain and record or stimulate neural activity.
• SingularityNET: A decentralized platform founded by Ben Goertzel (a pioneer of AGI research) and others that aims to create a network of artificial intelligence agents that can cooperate and exchange services. It is also developing a protocol for AI interoperability and governance.
• Hanson Robotics: A robotics company founded by David Hanson (a former Disney Imagineer) that aims to create human-like robots that can express emotions, communicate, and learn. It is also developing a platform called Sophia, which is a social robot that has been granted citizenship by Saudi Arabia and has been featured in various media outlets.

The future of LLMs and AGI

LLMs are still far from achieving AGI. They are still narrow AI systems that are specialized for natural language processing tasks. They do not have the same level of cognitive abilities and flexibility as humans. They do not have self-awareness, emotions, or goals. They have many potential applications in various domains, such as education, entertainment, business, health, etc. They can also be combined with other AI systems, such as computer vision, speech recognition, knowledge graphs, etc., to create more intelligent and interactive systems.

However, LLMs also have limitations and challenges, such as data quality, scalability, interpretability, robustness, ethics, etc. They require careful design, evaluation, and regulation to ensure they are safe, reliable, and beneficial for humans.

AGI is still a distant goal, but it is a fascinating and important one. It has many implications for science, technology, society, philosophy, etc. It could be the greatest invention of humanity or the last one. It could be a friend or a foe. It could be a dream or a nightmare.

The future of LLMs and AGI depends on how we address the current limitations and challenges of LLMs, as well as how we leverage their potential and benefits. Some possible directions and scenarios include:

• Developing more robust and reliable LLMs that can handle noisy, incomplete, or adversarial data, and generate outputs that are accurate, consistent, and trustworthy.
• Developing more interpretable and explainable LLMs that can provide insights into their internal workings and reasoning processes and justify their output and behaviors.
• Developing more adaptable and transferable LLMs that can learn from new domains or tasks with minimal supervision or guidance and generalize their knowledge and skills across different contexts.
• Developing more ethical and responsible LLMs that can align with human values and norms, and avoid generating outputs that are harmful, offensive, or misleading.
• Developing more interactive and collaborative LLMs that can communicate with humans and other agents in natural language and cooperate or compete with them in various scenarios.
• Developing more creative and innovative LLMs that can generate novel and original outputs that are useful, interesting, or entertaining.
• Developing more autonomous and self-aware LLMs that can set their own goals and preferences and monitor and improve their performance and behavior.

Conclusion

Are we there yet? The answer is no. We are not close to creating AGI or even understanding what it is or how to achieve it. But progress is being made in AI research and development, especially in LLMs and other deep learning systems.

Large language models are a recent breakthrough in artificial intelligence research. They are powerful systems that can generate and understand natural language for various tasks. They are also a potential stepping stone toward artificial general intelligence.

The future of large language models and artificial general intelligence depends on how we overcome the current challenges and limitations of large language models, as well as how we harness their potential and benefits.

I'm part of?Tata Consultancy Services' Architecture, Innovation, and Transformation (AIT) group led by? Prakash Mehrotra . Our mission is to collaborate with customers and support their business transformation goals. Using technologies AIT has explored various use cases that can be a starting point for your transformation journey. To learn more and engage with us, please contact your dedicated delivery or sales contacts at Tata Consultancy Services, who will connect you with our team.