In this tech insight, we look at whether AI models are nearing true general intelligence and what the arguments around this subject are and its relevance to society, innovation, and the future of technology development.

What Is AGI?

Artificial General Intelligence (AGI) is the theoretical development of AI systems capable of performing any intellectual task a human can do, i.e. reasoning, learning, problem-solving, and adapting across diverse and unfamiliar contexts without specific prior training. This is important because AGI could revolutionise industries and address complex global challenges by replicating human-like intelligence. Therefore, it remains one of the most ambitious goals in technology.

However, while significant strides have been made in AI, experts are divided on whether we are nearing AGI or are still far from reaching this milestone.

Why Is AGI Different To What We Have Now?

AGI is fundamentally different because whereas current AI systems are limited to specific tasks like language translation, image recognition, or gameplay because they rely on predefined training to do this. AGI would mean AI systems would be able to reason, learn, and adapt to entirely new and diverse situations, i.e. learn new things for themselves outside of their training without being specifically trained for them, mimicking human-like flexibility and problem-solving abilities.

François Chollet, a prominent AI researcher, has defined AGI as AI that can generalise knowledge efficiently to solve problems it has not encountered before. This distinction has made AGI the “holy grail” of AI research, promising transformative advancements but also posing significant ethical and societal challenges.

The pursuit of AGI has, therefore, garnered widespread attention due to its potential to revolutionise industries, from healthcare to space exploration, while also sparking concerns about control and alignment with human values. However, it seems that whether recent advancements in AI bring us closer to this goal remains contentious.

Recent Debate on the Subject

Much of the recent debate on AGI revolves around the capabilities and limitations of large language models (LLMs) like OpenAI’s GPT series. These systems, powered by deep learning, have demonstrated impressive results in natural language processing, creative writing, and problem-solving. However, critics argue that these models still fall short of what could be considered true general intelligence.

The aforementioned AI researcher François Chollet, a vocal critic of the reliance on LLMs in AGI research, makes the point that such models are fundamentally limited because they rely on memorisation rather than true reasoning. For example, in recent posts on X, he noted that “LLMs struggle with generalisation,” explaining that these models really just excel at pattern recognition within their training data but falter when faced with truly novel tasks. Chollet’s concerns highlight a broader issue, i.e. the benchmarks being used to measure AI’s progress.

The ARC Benchmark

To address this, back in 2019, Chollet developed the ARC (Abstract and Reasoning Corpus) benchmark, as a test for AGI. ARC evaluates an AI’s ability to solve novel problems by requiring the system to generate solutions to puzzles it has never encountered before. This means that unlike benchmarks that can be gamed by training on similar datasets, ARC may be more likely to measure genuine general intelligence. However, despite substantial progress, it seems that no system has, so far, come close to achieving the benchmark’s human-level threshold of 85 per cent, with the best performance in 2024 reaching 55.5 per cent.

Offering The ARC Prize To Spur Innovation

With the hope of providing an incentive to speed things along, earlier this year, Chollet and Zapier co-founder Mike Knoop launched the ARC Prize, offering $1 million to anyone who could develop an open-source AI capable of solving the ARC benchmark. The competition attracted over 1,400 teams and 17,789 submissions, with significant advancements reported. While no team claimed the grand prize, the effort spurred innovation and shifted the focus towards developing AGI beyond traditional deep learning models.

The ARC Prize highlighted promising approaches, including deep learning-guided program synthesis, which combines machine learning with logical reasoning, and test-time training, which adapts models dynamically to new tasks. Despite this progress, Chollet and Knoop acknowledged shortcomings in ARC’s design and announced plans for an updated benchmark, ARC-AGI-2, to be released alongside the 2025 competition.

Arguments for and Against Imminent AGI

Proponents of AGI’s imminent arrival point to recent breakthroughs in AI research as evidence of accelerating progress. For example, both OpenAI’s GPT-4 and DeepMind’s (Google’s) AlphaCode demonstrate significant advancements in language understanding and problem-solving. OpenAI has even suggested that AGI might already exist if defined as “better than most humans at most tasks.” However, such claims remain contentious and hinge on how AGI is defined.

Critics argue that we are still far from achieving AGI. For example, Chollet’s critique of LLMs highlights a fundamental limitation, i.e. the inability of current models to reason abstractly or adapt to entirely new domains without extensive retraining. Also, the reliance on massive datasets and compute power raises questions about scalability and efficiency.

Further complicating the picture is the lack of a real consensus on what constitutes AGI. While some view it as a system capable of surpassing human performance across all intellectual domains, others (like the UK government) emphasise the importance of alignment with ethical standards and societal goals. For example, in a recent white paper, the UK’s Department for Science, Innovation and Technology stressed the need for robust governance frameworks to ensure AI development aligns with public interest.

Alternatives and Future Directions

For researchers sceptical of AGI’s feasibility, alternative approaches to advancing AI include focusing on narrow intelligence or developing hybrid systems that combine specialised AI tools. It’s thought that these systems could achieve many of AGI’s goals, such as enhanced productivity and decision-making, without the risks associated with creating a fully autonomous general intelligence.

In the meantime, initiatives like the ARC Prize continue to push the boundaries of what is possible. As Mike Knoop (co-founder of Zapier and the ARC prize) observed in a recent blog post, the competition has catalysed a “vibe shift” in the AI community, encouraging exploration of new paradigms and techniques. These efforts suggest that while AGI may remain elusive, the journey toward it is driving significant innovation across AI research.

The Broader Implications

The pursuit of AGI and the thought of creating something that thinks for itself has, of course, raised profound ethical, societal, and philosophical questions. As AI systems grow more capable, concerns about their alignment with human values and potential misuse have come to the forefront. With this in mind, regulatory efforts have already begun, e.g. those being developed by the UK government, aiming to balance innovation with safety. For example, the UK has proposed creating an AI ‘sandbox’ to test new systems in controlled environments, ensuring they meet ethical and technical standards before deployment.

What Does This Mean For Your Business?

From a business perspective, the current state of AI—powerful but far from true AGI—presents both opportunities and threats.

Opportunities

1. Enhanced Tools for Specific Tasks: Current AI excels in narrow applications, giving businesses access to highly specialised tools that can improve efficiency and reduce costs without waiting for AGI to materialize.
2. New Markets in Innovation: With benchmarks like ARC exposing AI’s limitations, there’s room for startups and R&D-heavy businesses to innovate and fill these gaps, potentially leading to lucrative intellectual property.
3. Incremental Value Creation: The gradual path to AGI allows businesses to benefit from ongoing advancements in narrow AI, staying competitive and future-ready without betting the farm on AGI’s arrival.
4. Leadership Through Thought Clarity: Companies that articulate clear AGI strategies, even amidst the lack of consensus, can establish themselves as thought leaders and attract investment.

Threats

1. Hype-Driven Overinvestment: Ambiguity around AGI’s definition can lead to wasted resources chasing vague goals or overestimating timelines for true innovation.
2. Dependence on Narrow AI: Relying heavily on current systems with limited reasoning capacity may create vulnerabilities, especially if competitors leap ahead with paradigm-shifting breakthroughs.
3. Regulatory and Ethical Complexity: AGI aspirations attract scrutiny. Businesses must navigate a murky landscape of emerging regulations, ethical debates, and public perception.
4. Talent Wars: The race for top AI talent is fierce, and unclear definitions of AGI may exacerbate competition, driving up costs for hiring and retention.

Bottom Line: Businesses should focus on exploiting narrow AI’s proven value while investing selectively in AGI research. Clear-eyed strategies that balance ambition with practicality will outpace rivals lost in the hype cycle.

Amid these debates, the ethical and societal implications of pursuing AGI demand equal, if not greater, attention. Governments, particularly in the UK, are already taking steps to establish governance frameworks that aim to harness AI’s potential responsibly. Balancing the push for innovation with safeguards against misuse will be critical in shaping the future of AGI research.

For now, the path to AGI remains uncertain. However, the efforts of initiatives like the ARC Prize suggest that the journey is as valuable as the destination, driving forward new ideas and collaborative research.