The Path to AGI – Book Launch with John Thompson

Executive Summary

This webinar explores the multifaceted landscape of Artificial General Intelligence (AGI) and its wide-ranging implications. John Thompson addresses the inherent challenges associated with AGI development, explores future prospects, and emphasises the ethical considerations vital for responsible advancement. The webinar extends to the impact of AI on the job market, the role of Generative AI in Data Management, and the complexities of Metadata integration. Furthermore, it highlights the critical intersection of AI and data analysis, advocating for the importance of critical thinking in harnessing AI’s capabilities.

John examines the balance between democratisation, commercial applications, and the rise of autonomous systems. The last part of the webinar considers the unique perspectives brought forth by autism and multidimensional thinking regarding innovation. In closing, John underlines AI’s potential in bridging diverse viewpoints to foster progress in various industries.

Webinar Details

Title: The Path to AGI – Book Launch with John Thompson
Date: 24/03/2025
Presenter: John Thompson
Meetup Group: DAMA SA User Group Meeting
Write-up Author: Howard Diesel

Artificial General Intelligence: A Practical Perspective

Howard Diesel opens the webinar and shares an anecdote about attending John Thompson’s sessions at Data Modelling Zone. Howard highlights the interactive nature and challenges faced, particularly regarding understanding concepts like vectors and embeddings.

John shares encouragement for engaging with the evolving field of Generative AI and reflects on his own personal experiences in AI over the last 30 years. He goes on to account for holding significant roles at IBM, Dell, and EY, where he led the development of a major secure Generative AI environment. John then emphasises the importance of asking questions and fostering a learning environment amidst the rapid advancements in AI technologies.

One of John’s motivations for writing his book came about as he experienced scepticism regarding the predictions of imminent Artificial General Intelligence (AGI), citing views from notable figures like Elon Musk and Ray Kurzweil. John adds that he feels he aligns instead with Rodney Brooks, who estimates AGI’s arrival to be 130 years away compared to John’s own projection of 125 years.

Challenges of Artificial General Intelligence

John makes the distinction between Artificial Intelligence (AI) and Artificial General Intelligence (AGI), with the latter defined as an AI system that possesses the capabilities of an above-average human, including memory, recognition, empathy, and creativity. He goes on to note that despite the advancements in AI over the past 70 years, such as those demonstrated by tools like Siri, there remains a significant gap to achieving AGI, primarily because current AI lacks memory and cannot understand or interact with information as humans do. Additionally, John shares that while AI excels in tasks like rote memorisation and solving basic mathematical problems, it struggles with nuanced creativity and long-term comprehension, indicating that reaching AGI may take much longer than anticipated.

Two critical aspects are concerning with regard to AI’s development: the distinction between individual AI systems and AI ecosystems and the implications of synthetic data. While synthetic data can help address biases and shortcomings in existing datasets, there are significant concerns about data poisoning and misinformation, which could undermine the integrity of AI models.

John notes that as we progress towards Artificial General Intelligence (AGI), the focus will shift from single models to integrating various types of AI, including causal, generative, and foundational components. He emphasises the importance of using ensemble approaches to harness the full potential of AI technologies while ensuring data governance.

Future of AI: Challenges and Prospects

Large Language Models (LLMs) lack short, mid, and long-term memory architectures, and thus, AI does not have the ability to develop human-like memory capabilities. Although numerous research papers explore augmenting AI with memory, John notes that practical applications have yet to emerge from academia into innovation. He goes on to share his visions of a future where composite AI—merging generative, foundational, and causal AI—will evolve over the next 20 to 30 years, enabling more complex problem-solving, including the pursuit of Artificial General Intelligence (AGI).

John expresses concern about the current hype surrounding Generative AI, which is creating confusion among businesses. The urgency from companies to adopt these technologies has led to a panic, as many feel they’re falling behind, despite the fact that only a few leaders have successfully integrated AI to enhance operational effectiveness and profitability. John suggests maintaining a realistic understanding of probabilistic models in technology and data professions. While various vendors offer impressive model capabilities, users should recognise that these systems do not produce consistent answers for the same prompts, which can be challenging for professionals accustomed to definitive responses. Additionally, an attendee points out the ongoing challenges posed by vendors who may oversell AI capabilities for profit, echoing past issues experienced with Business Intelligence systems where poor Data Quality led to misleading outcomes.

Ethical Considerations in AI Development

Ethical considerations are imperative in the development of Artificial Intelligence (AI) and Artificial General Intelligence (AGI). John shares that while AGI remains a distant prospect, ethical issues surrounding current AI models must be addressed, especially as they are widely accessible to the public.

Developers and technologists often engage in structured thinking, allowing them to effectively program AI, but the general population may lack this framework, leading to potential misuse. Each AI model operates under a system prompt containing ethical guidelines, and organisations can implement additional constraints to ensure that the models adhere to their ethical principles. Therefore, if users are careful in their interactions with AI, and if ethical considerations are integrated into their prompts, the risk of ethical breaches can be mitigated.

The Impact of Artificial Intelligence on Job Market

John shares the fear of his students at the University of Michigan, who fear the impact of Artificial General Intelligence (AGI) on their future employment. However, he believes that humans will still be essential in the workforce, as AI excels at simple, mechanistic tasks but will not replace programmers and developers for many decades to come. Additionally, a recent experience of generating 4 million lines of AI-produced code at EY did highlight AI’s capabilities, but the majority of that code was not viable for production. John notes that this indicates that AI is not poised to take over the world. This is presently where John feels confident that there will continue to be meaningful roles for people in the job market for the foreseeable future.

Future of Artificial General Intelligence

There seems to be a general agreement that achieving Artificial General Intelligence (AGI) could take 20 to 30 years. Ethical questions arise regarding the tendency to anthropomorphise AI models. John shares this concern, emphasising that AI lacks true reasoning or original thought and functions instead through inference and recursive processing. John reiterates that current models lack the capacity for original thought and deception, arguing that any potential for misleading information can be managed through specific prompts that restrict their behaviour. He cautions against anthropomorphising these models, noting that even pets exhibit greater reasoning abilities. Lastly, John reassures that since models cannot truly design or mislead on their own, there is little to worry about regarding their misuse.

The Impact of AI Models

The potential bidirectional influence of AI models on human behaviour and societal capability raises important concerns. For instance, a study involving teachers and students using ChatGPT for essay writing led the students to reconsider their reliance on the tool, highlighting fears of diminished learning and cognitive skills.

As technology increasingly takes over tasks like calculations and information retrieval, there is a risk that people may lose essential skills through lack of practice. The reliance on quick searches, such as Googling information, often results in superficial understanding, overlooking nuances and context. This situation prompts critical discussions about the quality of information we integrate into these AI systems and the broader implications for mental capacity and knowledge retention in society, especially when a significant percentage of information might be deemed inadequate or unreliable.

In the evolving landscape of Artificial Intelligence, particularly with Large Language Models (LLMs) from companies like Google, it’s essential to recognise that these models are not the sole solution for all problems. While LLMs provide extensive knowledge drawn from the internet, they can also perpetuate inaccuracies; therefore, they should not be relied upon solely for tasks requiring high accuracy, such as specific file retrieval.

The future will likely see a proliferation of both large and small domain-specific language models, enhancing our ability to deliver accurate and transparent answers. Organisations can build and control their own models within this ecosystem to ensure ethical standards and improve accuracy, creating a balanced approach without being solely dependent on proprietary LLMs. However, challenges remain regarding the consistency of terminology and definitions when integrating information across different models, emphasising the need for careful management within each organisation’s context.

Generative AI and Data Management

The rise of Generative AI has significantly heightened the importance of data accuracy, as demonstrated in an early project where outdated documentation led to the client’s dissatisfaction with the AI model’s performance. This highlights the necessity for organisations to audit their data thoroughly, particularly when integrating AI systems that draw from various information sources.

Thought leaders like Andrew Ng emphasise the need to establish solid Data Management practices before advancing AI capabilities. Fortunately, organizations can leverage AI itself to assist in auditing their data, identifying inaccuracies, and navigating the inherent ambiguity of evolving terminology and document updates. Effective Data Management is crucial to successfully harnessing AI’s potential.

The concept of “Artificial General Explainability” (AGE) addresses the need for clarity in high-stakes decision-making contexts where ethical dilemmas arise. It emphasises that decisions influenced by Artificial Intelligence cannot be separated from embedded values, as those values shape the interpretation of data.

As technology advances, particularly in Generative AI and neural networks, research is focusing on understanding how these models activate and function, which is crucial for enhancing explainability, transparency, and interpretability. Over the next 10 to 20 years, significant developments are expected in these areas, potentially providing deeper insights into AI decision-making processes and the values that underpin them.

Understanding the Challenges of Metadata in Data Management and AI

The challenges associated with its storage and processing highlight the importance of Metadata in managing data effectively, particularly as it relates to artificial intelligence. Increased Metadata fields are necessary for better control and understanding of data quality, despite the potential costs involved.

John highlights that the advantages of the costs associated with Metadata storage are limited when weighed against the potential risks of ethical violations and mishandling of data, which is particularly pertinent when talking to the business.

John points out that an absence of a uniform set of Metadata fields and lack of diverse viewpoints—like those of employees, customers, and regulators— can cause problems when creating Metadata frameworks. However, this approach to Metadata can be intricate and may lead to confusion among stakeholders if they do not receive adequate guidance. Additionally, the business’s data faces a challenge with regards to how it changes and decays at different rates. While “fresh data” is valuable for gaining insights, older data—though labelled as “stale”—can also provide significant historical context and utility, depending on its application.

AI, Data Management, and Industry Applications

It is important to provide quality data in training AI models. John emphasises that feedback from introductory AI classes often highlights the dominance of biased data, such as a model that inaccurately claims the world is flat. John has 28 years of experience in traditional AI and notes that various tech vendors, such as IBM, offer industry-specific libraries for training Large Language Models; however, the effectiveness of these tools remains untested, as far as John is aware.

AI, Data Analysis, and the Role of Critical Thinking

Industry-standard Metadata and libraries require vetting as these resources may not align with individual perspectives and needs. John notes a specific case where a project involved building a Domain Language Model for tax, requiring 7 billion words for training. By gathering tax codes from 37 countries and additional legal documents, they successfully created a model with enough data for effective training, illustrating the challenges and strategies in developing specialised language models.

John moves on to emphasise the importance of utilising a well-structured and regulated corpus of data when creating Large Language Models, highlighting that standard definitions and terminology contribute to clarity. An attendee expresses optimism about the practical applications of AI, contrasting it with the exaggerated claims often made by vendors. John notes that there are exciting opportunities in data and analytics, suggesting that the field is only beginning to evolve positively.

Democratisation vs Commercial Applications vs Autonomous Systems

The potential for commercial applications to create exclusive “AI ecosystems” highlights the tension around the democratisation of AI. It raises concerns about whether Artificial General Intelligence (AGI) will be universally accessible or dominated by powerful entities, reducing individuals to mere data points. John then reflects on the resources required to develop advanced AI, noting that while AI may appear open to all, significant computational power is still a barrier for many. Ultimately, he reiterates his belief that AGI is still far from realisation—potentially 120 years away—due to inadequate governance and societal frameworks, but it is expected that when it does emerge, governments will adopt it as a tool for their interests.

The development of Artificial General Intelligence (AGI) is expected to be a race among governments and companies, the idea being that early achievers will gain a significant advantage, yet they won’t be alone in the pursuit. The landscape is dynamic, with rapid advancements occurring continuously, suggesting that the journey towards AGI will be marked by pivotal developments and prerequisites, such as the democratisation of technology and shifts in investment returns for big tech firms.

Despite some perceiving AGI as remote and unworthy of immediate interest, John argues that AI has historically been an exciting field, and working towards AGI is a valuable pursuit. He emphasises the importance of continued contributions to the field while viewing the potential challenges and long timelines as motivating rather than discouraging.

Autism and Multidimensional Thinking on Innovation and Progress

An attendee highlights the contrast between jocks and nerds portrayed in “Back to the Future,” noting that while jocks excelled in mechanical skills, nerds brought unique thinking through technology and creativity. They go on to emphasise that many individuals on the autism spectrum have contributed significantly to analytics and innovation, challenging the notion of a singular “normal” standard of intelligence. Additionally, another attendee shares his own personal experiences with teaching and counselling children with spectrum disorders, illustrating that, with the right support, these children can excel academically, often outperforming their peers. John adds to the discussion by sharing his advocacy for recognising the diverse capabilities across the spectrum and acknowledges that advancements in fields like AI may benefit from these unique perspectives.

A personal experience is shared by an attendee who identifies as slightly autistic. They share on their experience of multidimensional thinking and express a passion for helping others perceive complex ideas beyond traditional models. John shares his understanding of the attendees’ perspective of Buckminster Fuller’s architectural designs, such as buckyballs, which represent a blend of various perspectives and data sources. Noting that their most successful projects integrated 10 to 20 diverse data sources, highlighting the challenge of conveying unique insights in a business context without a translator familiar with the nuances of the organisation.

Another attendee adds their own reflection on a robotics and AI competition sponsored by Dell in the UK, which included both elite schools and schools catering to students with learning difficulties. Notably, teams from schools with students facing challenges performed better, highlighting the benefits of multidimensional thinking and perseverance. The Attendee notes that a representative from Dell emphasised that the success of these students illustrates how resilience and the ability to adapt are crucial skills, often fostered in environments where challenges are present.

Role of AI in Bridging the Gap between Different Perspectives

The webinar wraps up with a discussion on neurodivergence and the potential use of AI to help individuals with conditions like ADHD and ASD better fit into societal norms. An attendee explores their own personal experiences of being diagnosed and emphasises the importance of viewing neurodivergence not as a limitation but as an opportunity to showcase different perspectives and innovative thinking. They then suggest that AI could serve as a translator, enabling those with unique viewpoints to communicate their reality through a more conventional lens, thereby facilitating understanding. John then closes off by highlighting the value of deep thinking and detail-oriented problem-solving.