AI’s Greatest Hits: Seeing the Path to AGI
I’m Ylli Bajraktari, CEO of the Special Competitive Studies Project. In this week’s edition of our newsletter, P.J. Maykish and Joel Nelson of SCSP discuss how a Deepmind definition of artificial general intelligence (AGI) promotes a historical perspective of where we are and where the world is progressing on a path to AGI. You will enjoy this newsletter if you just want to step back and visualize how much AI has changed.
Over two hundred years separate the creation of the steam engine from the National Ignition Facility (NIF). And that invention, which created a star for one billionth of a second, would have been inconceivable to those who first harnessed steam, for the machine age was marked by a steady march from the steam engine to the cotton gin, the telegraph to the telephone.
The pace of change was steady and incremental but still shocking. When direct current (DC) was stored in a battery by Alessandro Volta in 1800, it would have been impossible to imagine Nicola Tesla’s invention of alternating current (AC) motors in the 1880s leading to continental energy grids. The number of general purpose technologies, those which can transform our societies, are increasing in number and rapidly being woven into most functions of society, combining and mutating. Just as the intersection of the machine and electrical ages led to the Industrial Revolution, so too will the meeting of computers with AGI inaugurate a new age. To grasp the transformative potential of AI as a new general-purpose technology, it is instructive to consider the stages the world has gone through in just the last few years. Unrecognizable iterations of inventions no longer take a century to be made—they’re happening yearly.
Using Deepmind’s position paper on AI development that leads to AGI and beyond, it is easier to visualize the progression of our human future with AGI. The below narrative expands on the 10 levels initially outlined in Deepmind’s incredibly useful framing table.
Five Narrow AI Levels: From Calculators to Superhuman Mastery in AlphaZero
Narrow Level 0: The Foundation (1967)
In the beginning, there were calculators. These precise tools laid the groundwork for computational thinking and an early form of human-machine teaming that could be found in homes and classrooms across the globe. Calculators were building blocks that would eventually support more complex intelligent systems.
Narrow Level 1: Emerging Intelligence (1968-1970)
A decade ago, “GOFAI” (Good Old-Fashioned Artificial Intelligence) systems like “SHRDLU” took a step toward human language understanding. This was a computer program that could follow instructions and answer questions within a limited virtual world known as "blocks world." These apps were designed to follow specific rules and could perform only simple tasks in controlled environments. It was like a "recipe" for intelligence, where the system could only follow pre-defined steps to generate something that felt like more than number crunching. While these systems were limited in their capabilities, they helped researchers understand the challenges of developing the next levels of AI.
Narrow Level 2: Competence Arrives (2011-2014)
Smart speakers like Watson, Siri and Alexa marked a turning point. These AI systems could now perform specific tasks as well as a human could. They could: answer questions, filter content, and perform some specific tasks at a level similar to, or in excess of, human capabilities. These systems are used in a variety of applications, from filtering out harmful online content to helping us with everyday tasks. Building on IBM Deep Blue’s defeat of chess master Gary Gasparove in 1997, IBM’s Watson even won $1 million in February 2011 by beating two former (human) Jeopardy champions. State-of-the-art LLMs also touch this category for specific tasks like short essay writing and coding. The development of these systems was a step towards creating AI that can seamlessly integrate into our lives.
Narrow Level 3: Expertise Unleashed
Systems like Grammarly and DALL-E 2 demonstrated new skills in narrow domains. These AIs didn't just assist; they often started to transform human fields including communication, art, and design through unprecedented precision and creativity or, compositionality. From the beginning of Grammarly in 2009 to OpenAI’s release of DALLE-2 in 2022, “compositionality” of AI at this level of AI continues to unfold. These narrow expert systems also led to competitive advantages for humans who adopted them. At level 3, it is possible or even probable for individuals and organizations to fall behind competitors if they do not leverage this new form of human-machine teaming.
Narrow Level 4: Virtuoso Performance (2016)
Starting with Deep Blue (1997) and crescendoing with AlphaGo (2016), AI virtuoso performance was a change that caught the attention of nations - AI systems that didn't just compete with human champions, but developed novel strategies in complex domains including chess and Go. These machines could not just replicate human thinking, but potentially change human approaches. After the Go professional player Lee Sedol lost 4-1 to AlphaGO in 2016, he stated that a certain move made by the machine (move 37) would change the way humans play the game. These systems have not only defeated human champions but have also demonstrated new strategies and ways of thinking about these games. The development of virtuoso narrow AI systems is pushing the boundaries of what’s possible in areas like game-playing and strategic decision-making.
Narrow Level 5: Superhuman Breakthrough (2017-2018)
AlphaZero and AlphaFold went further, solving challenges once considered exclusively human. AlphaZero was a revolutionary AI program that mastered the games of chess, shogi, and Go without any human input besides the rules. It achieved this through self-play and reinforcement learning, playing millions of games against itself, learning from its mistakes, and constantly improving its strategies. This marked a new level of AI because it demonstrated the unprecedented ability to surpass human expertise and develop novel strategies in complex games, all without relying on human knowledge or data. AlphaFold was an AI system that predicted the 3D structure of proteins from their amino acid sequence. Determining protein structure was previously a time-consuming and expensive process, often requiring years of experimental work. Then AlphaFold arrived. Its ability to accurately predict protein structures revolutionized drug discovery and disease understanding by providing scientists with crucial insights into the building blocks of life. The DeepMind paper defines “Superhuman” as “outperforming 100% of humans. For instance, we posit that AlphaFold… is a Level 5 Narrow AI… since it performs a single task (predicting a protein’s 3D structure from an amino acid sequence) above the level of the world’s top scientists.” AlphaFold and AlphaZero are examples of AI outstripping human ability.
General AI Levels: Here Today and Tomorrow
AGI Level 0: Human-Augmented Intelligence (2005)
An early stage of AGI was reached with Amazon Mechanical Turk (MTurk) and other programs where humans and machines began to work together seamlessly. MTurk is a large crowdsourcing platform through which businesses can hire people to complete small tasks that computers struggle with, including identifying objects in images, transcribing audio, and writing product descriptions. Essentially, MTurk acts as a digital marketplace where human intelligence is a commodity, allowing companies to tap into human capabilities to solve problems that AI can’t yet handle. While not true AGI, it demonstrates the potential of integrating a human and AI to achieve a range of tasks beyond the capabilities of either alone.
AGI Level 1: Sparks! (2023)
Large language models like OpenAI’s ChatGPT began generating accurate human-like text, showing potential for more generalized intelligence. When GPT-4 was released in March 2023, it exhibited impressive abilities in various domains, including coding, writing, and problem-solving, surpassing previous AI models in its general intelligence. Researchers were astonished by its capacity to learn, adapt to new situations, and even exhibit creativity. Its capabilities led to the publication of “Sparks of AGI,” a paper detailing experiments conducted with GPT-4 that revealed its capabilities and sparked debate about whether it demonstrated early signs of AGI. The paper argued that GPT-4 showed “sparks” of AGI, exhibiting rudimentary forms of reasoning, planning, and understanding, thus representing the next leap toward future forms of general AI.
AGI Level 2: Better-Than-Human Science
Strides are continuously being made globally. Using DeepMind’s position paper definitions, OpenAI’s “o1” model performs both competently and expertly across a range of tasks. “OpenAI o1 ranks in the 89th percentile on competitive programming questions (Codeforces),” its authors write, and it “places among the top 500 students in the U.S. in a qualifier for the USA Math Olympiad (AIME), and exceeds human PhD-level accuracy on a benchmark of physics, biology, and chemistry problems (GPQA).” This is not just an incremental change in the level of AI. It is the beginning of human-machine teaming in the sciences. AI no longer only mimics human intelligence; it is now surpassing it. A new form of early AGI is arriving and it has the potential to redefine the boundaries of scientific discovery, technological innovation, and even artistic expression.
AGI Levels 3-5: Seeing the AGI Equivalent of the NIF and Global Grids
The next three levels of AGI portrayed by the Deepmind authors are not here yet. The velocity of AI's evolution defies linear projections, powered by the synergistic convergence of three megatrends. These are, first, the exponential growth in large language models; second, AI novel paradigms or algorithmic progress and; and finally, the proliferation of AI infrastructure: advanced compute/microelectronics, data, energy, networks. This trend is multiplying the raw material of very large artificial systems that can substantially improve civilization if developed properly. This confluence of trends will yield an AGI landscape so transformative that the human mind struggles to imagine it unless you trace this basic progression (from calculators to level 2 AGI) to imagine the next levels of AGI (“AGI levels 3-5”). For now it is as though we are riding steam trains, and being told of the jet engine future trying to imagine a globalized world of modern air travel. Imagining a positive AGI world is essential to helping that version of the world come into being. While the pieces are in flux we must chart a path to leverage AGI’s full potential to make the world a better place.
AI+Expo for National Competitiveness
On June 2-4, 2025, SCSP will host its second AI+ Expo at the Walter E. Washington Convention Center in DC. The AI+ Expo is the place to convene and build relationships around AI, technology, and U.S. and allied competitiveness.
Interested in joining us? We are seeking additional exhibitors and sponsors that are eager to share and discover new breakthroughs with us this summer. Find out more at expo.scsp.ai.