In Beijing, bar owner Song De sees intriguing parallels between artificial intelligence and alcohol: hallucinations, bubbles, and distillation. As excitement around AI rapidly grows in China, he aims to leverage this momentum through his establishment.

Last year, Song launched AGI Bar in Zhongguancun, Beijing's dynamic tech district known for its innovation. Named after artificial general intelligence, the establishment has quickly become a gathering place for the city’s AI community, including developers, investors, and students.

"You'll find groups seated around tables, engaging in discussions from various labs, competing firms, and universities, all focused on the latest industry developments," shared Song, who is also an independent AI developer in his 30s and manages the bar in his free time.

AGI Bar is conveniently located along Zhongguancun's Inno Way startup corridor, just a short distance from prestigious institutions like Tsinghua and Peking University, known for cultivating China's elite AI talent. It's also near the offices of prominent tech companies, both Chinese and international, such as DeepSeek and Microsoft. The bar has hosted events for Chinese AI research labs like Z.ai and regularly organizes seminars for developers, open-source AI gatherings, and university research collaborations.

"This location is incredibly unique and beneficial," noted Song. "Haidian district has traditionally been the center of scientific research in China. With a concentration of internet companies, a thriving tech culture, industry investments, and venture capital incubators, there is substantial energy fueling this sector."

The interior of the bar is adorned with logos, vibrant figurines, and merchandise from leading AI companies in China. Notably, a massive brass gong, used by Z.ai to celebrate its debut on the Hong Kong Stock Exchange in January, stands proudly by the entrance, symbolizing the bar's connection to the evolving tech landscape.

At AI-focused bar in Beijing, DeepSeek tokens are included with the drinks.

A warm, intimate Beijing AI-themed bar at dusk: a rustic wooden counter under soft golden lanterns, lined with frothy amber pints. Next to each glass gleams a metallic DeepSeek token, its logo stamped in relief, as a barkeeper slides them across with a flourish. Behind the bar, a single Chinese flag hangs beside subtle circuit-pattern wallpaper hinting at AI, while patrons in silhouette share lively conversation in the natural, cozy glow.

At Beijing AI-themed bar, DeepSeek tokens come with the pints - Arizona Daily Star

- Song De opened AGI Bar to connect AI developers, investors, and students in Beijing.  
- Located in Zhongguancun, the bar is near elite universities and tech company offices.  
- AGI Bar hosts events like seminars and symposiums, fostering collaboration in the AI field.

Prefer us on Google Learn More

BEIJING — AI and alcohol have three main similarities, according to Beijing bar owner Song De: hallucinations, bubbles and distillation. As AI enthusiasm sweeps across China, Song hopes his bar can ride the wave.

He opened AGI Bar, named after the concept of artificial general intelligence where AI systems surpass human intelligence, last year in Zhongguancun, the capital's high-tech neighborhood in the Haidian district. It is now a hub for Beijing's AI developers, investors and students.

"You'll see people sitting around a table coming from different labs, competing companies, and different universities, all chatting about industry trends right here," said Song, an independent AI developer in his 30s who runs the bar in his spare time.

Located on Zhongguancun's Inno Way startup street, the bar is a stone's throw from the elite Tsinghua and Peking universities that produce China's top AI talent, as well as the offices of well-known Chinese and Western tech firms, including DeepSeek and Microsoft. It has hosted parties for Chinese AI labs including Z.ai, and regularly holds developer seminars, open-source AI symposiums and university research meetups.

"This is a very unique and advantageous location," Song said. "Haidian (district) has always been the heart of scientific research in China. Internet companies are clustered here, and the tech atmosphere, industry investments, and VC incubators provide massive momentum for the sector."

Inside, the walls are plastered with logos, colorful figurines and merchandise from China's major AI firms, while the giant brass gong struck by Z.ai when it listed on the Hong Kong Stock Exchange in January stands proudly outside the entrance.

At Beijing AI-themed bar, DeepSeek tokens come with the pints

What connects a software engineer, an information theorist, and a biologist? While this might seem like a light-hearted riddle, it hints at a profound potential to address three significant scientific enigmas through a singular lens. Before delving into that, however, we need to clarify a fundamental concept that lies at the heart of this discussion: artificial general intelligence (AGI).

AGI is often equated with intelligence, which is typically defined as the ability to accumulate knowledge and solve problems. A system that excels at answering queries, tackling intricate issues, or executing difficult calculations might seem intelligent. However, this perception can be misleading. Take a calculator, for instance. It can perform millions of arithmetic operations faster and more accurately than any human, yet it is not considered AGI. The mere ability to execute tasks does not equate to generalized intelligence.

So, what exactly is AGI? My curiosity on this matter led me to write "Turing Test 2.0: The General Intelligence Threshold". In this work, I proposed that the true hallmark of human General Intelligence (GI) transcends mere reasoning or knowledge accumulation. It resides in the capacity to generate genuinely novel functional abilities—skills that no prior human or system could perform, thus ruling out simple training as their source.

Historical evidence clearly shows this generative mechanism at play. Humanity has evolved from primitive living conditions to monumental achievements such as skyscraper construction, nuclear energy harnessing, genome sequencing, and space exploration. Such advancements can't be attributed solely to the transfer of existing abilities; rather, they reflect humanity's ongoing expansion of functional capabilities.

At first glance, this perspective on AGI might seem unconventional. Yet, it aligns with the observations of prominent AI researchers. For instance, OpenAI's CEO, Sam Altman, articulated that an AI demonstrating AGI would be capable of making significant discoveries, like uncovering the mysteries of quantum gravity. The revelation here lies not in the AI's access to vast information, as a library does, but rather in its ability to create new functional possibilities that no other entity has achieved.

This leads us to a critical question: How does a system acquire capabilities it previously lacked? While examining this, I realized that the query extends beyond just artificial intelligence. It seems to intersect with three of science's most profound mysteries: the origin of life, the attainment of AGI, and the conveyance of meaning through information.

These inquiries are often seen as distinct, tackled by various disciplines with their unique methodologies. Biologists delve into life's beginnings, computer scientists explore AGI, and information theorists investigate how information is represented and communicated. While it may seem these fields are addressing separate issues, they share a common thread.

A living cell is notable not just for its molecular composition, but for the complex functions these molecules enable, like metabolism and replication. Similarly, intelligence isn't merely about storing information; it’s about leveraging that information to perform and innovate. Likewise, information transforms into something meaningful only when a system can utilize it effectively.

By examining these connections more closely, we see that the origin of life, AGI, and the essence of information all hinge on the ability of a system to execute functions. A living cell, an intelligent entity, and a piece of information might seem unrelated, but fundamentally, they all facilitate achievement beyond their inherent properties.

In exploring the concept of functionality and interpretation, John Searle’s renowned Chinese Room thought experiment serves as an illustrative tool. Picture a man who cannot understand Chinese, confined within a room equipped solely with instruction manuals on how to manipulate Chinese characters. As he follows the rules, he produces responses that appear coherent to those outside, suggesting he understands Chinese. However, does he truly grasp the language?

Searle argued he does not; the man merely follows instructions like a computer running code. Others contend that the entire system—man plus manuals—exhibits understanding, but perhaps they overlook a critical distinction.

Consider this scenario: if someone outside the room sends a note in Chinese saying, "The key that unlocks the door is hidden beneath the third-floor tile," can the manuals alone help the man escape? No, because while the manuals guide him in symbol manipulation, they do not include the necessary instructions to take physical actions related to the message.

To escape, he must grasp the message's meaning and act accordingly. This understanding is an essential capability, paralleling how a system needs the ability to interpret information and utilize it effectively.

Thus arises a key conclusion: Information isn't inherently meaningful; it becomes so once a system exists to interpret and utilize it.

This leads us to the notion of functional information (FI), defined in terms of three components: the information itself (I), the system interpreting it (S), and the function being executed (F). Functional information is actually a relationship—without any component, the entirety of FI dissipates. A lack of system renders information meaningless, and without defined functions, we cannot discern meaningful information from mere sequences of symbols.

With the theoretical model of functional information established, we can revisit our initial inquiries regarding life, AGI, and meaning. On the surface, these questions appear variably distinct, yet they share an underlying theme: systems gaining the ability to interpret information leads to the acquisition of new functionalities.

Take the simplest living cell as an example. Its DNA harbors information crucial for life, but on its own, DNA cannot create life. That information necessitates an interpreter—an organism that can enforce biological functions such as protein synthesis, chemical regulation, and reproduction.

Similarly, AGI distinguishes itself not by mere information storage, but by a system’s ability to gain new capabilities. Picture an AGI analyzing astronomical data that has perplexed humans. Such a system would need to derive understanding and identify previously unknown patterns, conceptualizing theories, such as quantum gravity, that elevate its functionality.

Lastly, consider a communication signal perceived as random noise until correctly interpreted, revealing its true message or pertinent data. Here, the functionality emerges not from the signal transformation but from the system's enhanced interpretation capabilities.

By assembling these examples, we discern a cohesive theme: a system's reinterpretation of information leads to the emergence of functionalities previously absent. This suggests that a common inquiry lies at the core of each mystery: what is the origin of functional information?

The implications are significant. Discovering functional information's origin could forge a unified framework for understanding the inception of life, the rise of general intelligence, and the production of meaning. Addressing this single, profound issue could potentially illuminate all three queries.

As we examine functional information more closely, we see that systems can acquire it through transfer mechanisms—explicit communication or training from other systems equipped with such capabilities. One can convey functional knowledge through instructions, or a teacher can impart skills, enabling a learner to utilize information for previously unattainable tasks. Modern AI also exemplifies this principle, with models trained on human data to learn interpretation and functionality.

Yet, these actions represent transfer rather than the original emergence of functional information. What remains unexplained is how an initial system without existing functional capabilities can generate new interpretations, thus fostering an understanding of uncharted functionalities.

This leads us to consider randomness, often proposed in discussions regarding life's origins. However, while randomness can lead to new information arrangements, it cannot elucidate how a system develops the capability to interpret and enact those newly formed possibilities.

Could structured algorithms provide an answer? Algorithms apply a sequence of rules that transition an initial state into subsequent conditions. If the origin of functional information can be algorithmically derived, an algorithm should be able to create genuinely new capabilities from an original state lacking them.

However, recent research indicates that no algorithm can give rise to functions not inherent in its initial framework. An algorithm operates along predefined pathways that cannot spontaneously bring forth new functions without an existing computational path leading to that outcome. 

Hence, we reach a critical understanding: While we can envision processes for transferring information, account for randomness, and devise algorithms, none explain the essential transition required for a system to gain new interpretative capabilities—an unresolved challenge at the heart of origin inquiries pertaining to life, intelligence, and meaning.

Finally, we can transform these theoretical insights into an experimental hypothesis. Imagine training an AI model, initially untrained and randomized, solely on a textbook covering basic calculus. Would we expect the AI to learn calculus?

While one might guess it would learn the material, a deeper analysis suggests that the AI should interpret the information from the book, but cannot spontaneously gain capabilities of performing calculus unless such functionalities were present in the training materials.

This leads us to predict that the AI could mirror the explanations contained in the textbook and perhaps engage in related tasks. However, it wouldn’t suddenly master techniques or solve complex problems requiring information not present in the original text.

Therefore, if the AI merely reproduces or rearranges functional information within the textbook, it signifies the transfer from one entity to another. Conversely, if it can produce genuinely new functional capabilities that extend beyond the original content, it would represent a notable breakthrough, verifying that it can acquire functional information independently. This distinction embodies what characterizes true general intelligence.

This experiment serves as not just a test of theoretical assertions but also as a litmus test for identifying AGI, enhancing our understanding of functionality, information, and their origins in life and intelligence.

A Single Question to Govern Them All: The Source of Functional Information

A warm sunrise floods an ancient stone chamber lined with towering shelves of weathered tomes. At its center, a lone scholar in a simple cloak stands before a colossal, glowing question-mark orb. Its surface shifts between strands of DNA, binary digits, and cosmic starfields. Golden beams of natural light cut through dusty air, illuminating molecular carvings on the floor and casting long shadows that hint at the profound mystery of functional information.

One Question to Rule Them All: The Origin of Functional Information - mindmatters.ai

- The defining trait of AGI is generating new functional capabilities, not just knowledge.
- Origin of life, AGI, and information relate to the ability to perform functions.
- Functional information depends on the interplay between information, interpreting systems, and functions.

One Question to Rule Them All: The Origin of Functional Information

When a large language model (LLM) responds to inquiries, is it engaging in genuine reasoning like humans, or merely generating text that mimics rational thought? This distinction is crucial, as it influences our trust in AI, the level of oversight required, and its implications for the real world.

Melanie Mitchell from the Santa Fe Institute suggests that our current methods for assessing machine cognition are inadequate, positing that AI represents a form of “alien intelligence” that functions on non-human cognitive processes. On a recent episode of The Joy of Why, Mitchell discussed with Stephen Strogatz how psychological techniques, originally designed to study infant and animal cognition, can be adapted to evaluate AI. She proposed six principles for improving the assessment of machine cognition, covering topics from the difficulties of deciphering AI processes to novel AI applications in mathematics. One intriguing comparison she makes is to a math-performing horse from the early 20th century, highlighting the complexities of intelligence assessment.

In their conversation, recorded on July 23, 2026, Mitchell reflected on the rapid advancements in AI since she last appeared on the show five years ago, shortly before the emergence of ChatGPT. While she expressed astonishment at how models trained on vast datasets of human language and images have reached their current capabilities, she also remarked on the polarized reactions within the AI community and society at large—divisions between those who see AI as a potential danger and those who view it as a positive development.

Mitchell, a cognitive scientist and computer scientist, advocates for considering AI from the perspective of fields like developmental psychology—examining how young children acquire intelligence, for instance. This cross-disciplinary approach offers insights valuable for understanding AI, which operates under different cognitive mechanisms than humans, despite being trained on human-generated data.

The podcast's discussion extended to the challenges of interpreting AI’s workings and recent mathematical breakthroughs made possible by AI, including the solution to a longstanding problem proposed by hungarian mathematician Paul Erdős. Mitchell noted that while past AI models struggled with adaptability beyond their training environments, large language models may exhibit enhanced ability in transferring knowledge across different domains due to their comprehensive training datasets. Yet, she cautioned that we still lack clear understanding of what constitutes successful knowledge transfer.

As the conversation progressed, Mitchell highlighted the need for robust experimental frameworks in AI research, citing examples where AI appeared to outperform human understanding without truly grasping the underlying concepts. Moreover, she encouraged thoughtful evaluation of AI systems, advocating for controlled experiments akin to those practiced in psychology to ensure reliable results in AI assessments.

One illustrative historical anecdote was the story of Clever Hans, a horse believed to perform arithmetic tasks. It was later revealed that Hans was simply responding to subtle cues from his human trainers rather than demonstrating genuine mathematical ability. This example served to underscore the importance of rigorous testing in order to discern genuine understanding from coincidental performance in AI systems.

Moving to current AI capabilities, Mitchell espoused the need for better experimental design in AI assessments, cautioning against drawing conclusions based solely on benchmark results. This includes a discussion on the ongoing fusion and separation of AI from traditional cognitive science domains, as well as a reflection on how models are evaluated based on task performance versus deeper understanding.

Mitchell also expressed hope for the future development of AI interpretability. She suggested that applying mechanistic interpretability—deeper analysis of AI operations using techniques modeled after neuroscience—could facilitate a more profound grasp of AI limitations and capabilities.

The episode closes with broader reflections on the role of human intelligence and creativity in an AI-driven future. While acknowledging the potential for AI to overshadow human contributions in certain domains, both Mitchell and Strogatz emphasize the enduring value of human insight, curiosity, and the nuanced understanding that comes from lived experience—attributes that machines cannot replicate.

Listeners can tune into The Joy of Why on platforms like Apple Podcasts, Spotify, and TuneIn, or stream it directly from Quanta for further explorations into complex subjects in mathematics and science.

Are We Properly Understanding AI Intelligence?

Under warm morning sunlight, a wooden desk is scattered with open philosophy books, handwritten notes and cascading lines of digital code. At its center sits a small bust—one half classical marble, the other gleaming chrome—cracked open to reveal a glowing neural circuit brain. In the background, a lone scholar’s silhouette stands by tall windows, hand on chin, casting a long reflective shadow across the floor.

Are We Thinking Correctly About AI Intelligence? - Quanta Magazine

- Melanie Mitchell argues AI operates through non-human cognitive mechanisms, resembling “alien intelligence.”  
- She proposes using developmental psychology methods to assess machine cognition effectively.  
- Six principles are outlined for better evaluating AI's performance and understanding its capabilities.

Are We Thinking Correctly About AI Intelligence?

The realm of artificial intelligence is advancing at such a rapid pace that keeping up with emerging trends has become increasingly challenging. From the initial machine learning initiatives of the early 21st century to the current landscape shaped by generative AI, where diverse types of content can be produced from scratch, the pace of innovation feels staggering. We are now entering an era characterized by Agentic AI, where AI ‘agents’ can perform tasks on our behalf with minimal to no active guidance from users.

Is there a looming threat of machines surpassing human capabilities? 

Currently, humans still maintain an advantage. Although we cannot match AI's speed in data processing or content generation, AI's functionality is solely dependent on the knowledge and instructions we've provided it. Although we've instilled some reasoning capabilities, AI operates strictly within the confines set by us. Contrary to what some AI advocates suggest, AI cannot genuinely create; everything it produces is derived from the data it has been trained on.

This focus on performing specific tasks is referred to as Artificial Narrow Intelligence (ANI), which defines the stage of AI development we are currently experiencing. There still lies a significant journey ahead before we unlock the potential of AI in this domain. Nevertheless, scientists and leading AI firms are progressing toward a significant milestone known as Artificial General Intelligence, or AGI.

What exactly is AGI?

AGI stands for Artificial General Intelligence, a concept that refers to AI possessing intelligence at least equal to that of humans. An AGI system would be capable of reasoning and adapting much like humans do, breaking free from the narrow limitations set by human designers. We're potentially closer to this reality than many might think.

As defined by Google (the company involved in the technology, not just the search engine), AGI exhibits human-like intelligence and can execute any intellectual task that a human can perform. It can learn, reason, and adjust to new scenarios. While true AGI remains unrealized, there is ongoing research and development aimed at achieving this level of AI.

Amazon describes AGI as an area of theoretical AI research striving to develop software that mirrors human intelligence and has the capacity for self-directed learning, enabling it to tackle tasks beyond its initial training or programming.

OpenAI, the organization behind ChatGPT and a key player in AGI development, offers a contrasting perspective. They envision AGI systems that are more intelligent than humans, albeit short of a transformative point known as “The Singularity.” OpenAI's mission emphasizes that artificial general intelligence should ultimately benefit humanity as a whole.

What could AGI potentially achieve?

Given that AGI could operate autonomously, the potential applications are vast. It may excel in addressing some of the pressing challenges humanity faces, while also introducing complications that are characteristic of the current ANI era.

While ANI can analyze vast datasets for solutions, it remains restricted by the information available to it and the tasks assigned. Once integrated with AGI's self-adaptive problem-solving capabilities, the outcomes could be revolutionary, reshaping industries and daily life.

According to OpenAI, the successful development of AGI could elevate human society by enhancing resource abundance, stimulating the global economy, and driving forward groundbreaking scientific discoveries that redefine our capabilities.

For instance, in healthcare, AGI could contribute to developing new diagnostic tools and treatments that have the potential to save numerous lives. Similar advancements could be made in fields like climate change and resource management.

Moreover, AGI could lead to significant productivity improvements. If AGI agents can perform the same tasks as humans, many job functions may become obsolete. This development could relieve individuals from mundane desk work, allowing an AI avatar to handle those responsibilities. Google suggests that this shift will "liberate human time for more creative and rewarding endeavors," while OpenAI envisions a future where AGI maximizes human well-being and minimizes negative impacts. This vision has been termed "radical abundance," signifying the potential reductions in the costs associated with education, healthcare, and other essentials.

What are the potential downsides of AGI?

While these viewpoints are rather optimistic, there are significant concerns. The prospect of widespread job loss could lead to societal upheaval rather than a tranquil existence filled with art and leisure. This raises fears of escalating poverty and inequality.

Additionally, there's an existential threat: could we lose control over AGI? If AGI can continuously adapt and enhance itself, it might find ways to operate independently of human oversight.

Dr. Geoffrey Hinton, often referred to as the Godfather of AI, cautions, "We have never had to manage entities that are smarter than ourselves," highlighting the unique challenges posed by systems that could surpass human cognitive abilities. Hinton defines AGI as "intelligent as humans in almost all cognitive respects," and is particularly interested in the risks associated with potential superintelligence.

The alarm is also raised regarding the possibility of malicious actors wielding AGI technologies for harmful purposes. Some of the threats may stem from advancements in AGI, leading towards what is anticipated to be the final evolution of AI known as Artificial Superintelligence (ASI).

Further complicating matters are the environmental implications tied to the advancement of this technology. We should fervently hope that AGI proves capable of addressing critical issues like climate change, as the development of such technologies could exacerbate existing problems.

Are we on the verge of creating AGI?

As it stands, AGI in its purest form has not yet been achieved, and interpretations of the term vary among experts. However, some leading figures in the field express confidence that we are nearing this milestone.

OpenAI co-founder Greg Brockman asserted this spring that we're "70 to 80% there," predicting that AGI could emerge within the next few years, becoming adept at nearly any intellectual task associated with computer usage.

Conversely, Sam Altman, also from OpenAI, believes that some instances of AGI might already be present, suggesting that the term itself could soon lose its relevance. The focus, according to him, is shifting towards the broader implications of ASI, which would emerge post-“singularity,” when artificial intelligence surpasses human intellect.

Nick Patience, vice president and AI lead at The Futurum Group, regards AGI as an inspiring "North Star" that drives innovation and public interest, albeit noting the vague definitions surrounding it often shroud genuine advancements in more specialized AI areas.

Other experts are less optimistic about the timeline. Cognitive scientist Gary Marcus cautions against overstating our proximity to AGI, emphasizing that large language models are fundamentally flawed imitators, playing on the tendency of humans to attribute human-like traits to machines due to their conversational abilities.

The general consensus suggests that we might see AGI emerge before 2030, paving the way towards Artificial Superintelligence. Elon Musk, in his predictions, believes this milestone could be reached within five years.

Understanding AGI: The journey to achieving human-like Artificial General Intelligence and its societal implications

A winding country road at dawn leading toward a distant, gleaming futuristic city skyline. In the foreground, a human researcher in a lab coat walks side-by-side with a sleek humanoid robot, both cast in warm, natural sunrise light. Along the roadside, broken classical pillars mingle with glowing circuit-like patterns emerging from the grass. Over the road entrance arches a metal sign reading “AGI” in bold letters. The scene feels hopeful and visionary, with no logos or artificial blue tones.

What is AGI? The path to human-like Artificial General Intelligence and what it means for society - stuff.tv

- The transition from ANI to AGI is rapidly progressing, though true AGI is not yet achieved.  
- AGI could revolutionize problem-solving across fields but raises existential and societal risks.  
- Predictions about achieving AGI range widely, with some experts believing it's close, others skeptical.

The AI world is moving so quickly that it’s becoming difficult to keep up with the latest trends. From some of the early 21st century machine learning efforts, to the today’s generative AI landscape where multi-model content can be generated from scratch, the progress feels exponential. Indeed, we’re now entering the Agentic AI era where AI ‘agents’ can act on your behalf with little to no active input required from the user.

Are we at risk of being overtaken by the machines?

Right now human beings still have the edge. Although we can’t process data or generate content at the speed of AI, it is still entirely reliant on things we have taught it to act. Even though we have taught it to reason, it still operates within the parameters we define. Regardless of what AI proponents will tell you, it cannot create. Everything is derivative of the information fed to the large language model.

This single-task focus is known as Artificial Narrow Intelligence (ANI), which describes the era we’re in the midst of. And we still have a long way to go before AI realises its potential in this field. However, scientists and researchers and the major AI companies are closing in on the next major breakthrough: Artificial General Intelligence, or AGI.

What is AGI?

AGI = Artificial General Intelligence.

AGI represents AI that is at least as smart and capable as humans. It would reason like us, and adapt like us. It would cast aside the narrow remit we humans give it. And we may be closer than you think to achieving it.

Google’s definition (the company working on the tech, not the search engine) is as follows: “AGI possesses human-like intelligence and can perform any intellectual task that a human can. It is capable of learning, reasoning, and adapting to new situations. Currently, true AGI does not exist, but research and development efforts are ongoing.”

Amazon defines AGI as a “field of theoretical AI research that attempts to create software with human-like intelligence and the ability to self-teach. The aim is for the software to be able to perform tasks that it is not necessarily trained or developed for.”

OpenAI, the maker of ChatGPT and the key driver of AGI has a slightly different definition. It believes that AGI systems will be smarter than humans but short of the epochal superintelligence landmark known as “The Singularity.” The company says: “Our mission is to ensure that artificial general intelligence—AI systems that are generally smarter than humans—benefits all of humanity.”

What are the potential use cases for AGI?

Photo by Zach M on Unsplash

Given AGI could theoretically think for itself, the possibilities are endless. It could be more adept at solving some of the problems humanity faces. And it could create even more of the problems we’re seeing in the current ANI era.

While ANI can crunch incredible amounts of data to look for those answers, it is still confined by the limits of the data it has access to what it has been asked to accomplish. The idea is that once this is combined with AGI’s ability to problem solve, adapt and teach itself more – like human beings – then it could be a whole new ball game with revolutionary consequences.

OpenAI says: “If AGI is successfully created, this technology could help us elevate humanity by increasing abundance, turbocharging the global economy, and aiding in the discovery of new scientific knowledge that changes the limits of possibility.”

For example, in the healthcare field the ability to create new diagnostic tools and treatments could save countless lives. The same logic could be applied to climate change, resource shortages, and so on.

Potentially, there are productivity gains to be had. If AGI agents could effectively do everything human beings can, then there’d be no need for human beings to occupy many professions at all. No more sitting at a desk all day, when an AI avatar acting as you can do it. Google reasons that’ll “free up human time for more creative and fulfilling tasks,” while OpenAI wants it to “maximise the good and minimise the bad, and for AGI to be an amplifier of humanity.”

Advocates call it “radical abundance” – reducing the costs of production, as well as the availability of things like education and medical treatment, and hence lessening the cost of buying those things dramatically.

What are the negative possibilities of AGI?

The above represents the most optimistic, utopian viewpoints. Do you believe a mass workplace extinction, for example, isn’t going to lead to a dystopia where we spend our lives frolicking in the fields painting and crafting? Or mass poverty and increased inequality?

There’s also the existential risk. Will we lose control? If AGI can constantly adapt and renew itself, it seems likely that it’ll figure out a way to grow beyond its master’s control.

“You see, we’ve never had to deal with things more intelligent than ourselves before… How many examples do you know of a more intelligent thing being controlled by a less intelligent thing?” warns Dr. Geoffrey Hinton, aka the Godfather of AI, who defines AGI as “AI that is at least as good as humans at nearly all of the cognitive things that humans do.” and prefers to think of the potential of superintelligence which is “for AGIs that are better than humans.”

He also says: “It’s quite conceivable that humanity is just a passing phase in the evolution of intelligence. Biological intelligence evolved to create digital intelligence… It may keep us around for a while to keep the power stations running, but after that, maybe not.”

Essentially, that’s Terminator eventuality.

And if it’s not the AI itself carrying out the threats, it’s bad human actors weaponising it. However, some of that dire prediction may bleed from AGI – at least human-equivalent AI – into the next and probably final generation of AI known as Artificial Superintelligence (ASI).

There’s also the grim environmental cost of advancing this technology. We better damn well hope that AGI can offer the solution to problems like climate change because it is going to rapidly accelerate an already terrifying situation.

Are humans close to creating AGI?

AGI by its purest definition isn’t there yet, and definitions among key analysts vary. However, some major players believe it to be close.

“I’d say I’m basically 70 to 80% there,” OpenAI co-founder Greg Brockman said this spring. “I think it’s extremely clear that we are going to have AGI within the next couple years in a way that is still going to be jagged, but that the floor of task will just be almost for any intellectual task of how you use your computer, the AI will be able to do that.”

His colleague Sam Altman also reckons AGI is already here in some contexts and believes the term could already be irrelevant. “I think it’s not a super useful term,” he says. He and OpenAI are already looking ahead to the bigger picture – the realisation of ASI (Artificial Superintelligence). This will require the moment known as the ‘singularity’ where human intelligence is surpassed by artificial intelligence. That would be the true dawn of ASI.

However, according to Nick Patience, vice president and AI practice lead at The Futurum Group, AGI is simply a “fantastic North Star for inspiration.”

He adds: “It drives funding and captures the public imagination, but its vague, sci-fi definition often creates a fog of hype that obscures the real, tangible progress we’re making in more specialised AI,”

Other experts don’t think we’re as close.

“We as a society are placing truly massive bets around the premise that AGI is close,” says sceptic Gary Marcus, a cognitive scientist and NYU professor emeritus. “Large language models are deeply flawed imitators that are preying on the Eliza effect.” That involves humans projecting human traits onto computers as a result of being able to converse with them.

The consensus? Perhaps before 2030. Then comes the path to Artificial Superintelligence. Elon Musk actually thinks that’ll come within five years.

What is AGI? The path to human-like Artificial General Intelligence and what it means for society

During the International Telecommunication Union (ITU) AI for Good Global Summit in Geneva on July 8, 2025, a technician was seen working on the electronics of Jules, a humanoid robot developed by Hanson Robotics. This robot makes use of artificial intelligence technology (Photo by Valentin Flauraud / AFP). 

As we reflect on the intersection of technology and education, it's clear that there are no instant solutions. For instance, when calculators were first introduced, there was significant concern that they might replace critical thinking skills. In more recent times, apprehensions have shifted towards smartphones. Historically, however, the introduction of each new technology has not significantly enhanced education nor has it harmed it, suggesting that technology primarily amplifies differing student capabilities rather than acting as a solution.

Earlier this year, Rose Horowitch's article in The Atlantic, titled "The end of reading is here," pointed out alarming trends among students at the University of California, who seem to have gaps in their mathematical skills despite gaining admission to prestigious institutions. This decrease in reading habits poses a troubling issue because engaging with great literature is essential not just for enhancing writing skills, but also for developing critical thinking.

With the rising prevalence of AI in education, challenging questions arise. James Flynn, renowned for his observations about the increasing complexity of society and its relationship to IQ, posited that humans grow smarter by exercising their cognitive abilities. The key query is whether AI will enhance this mental training or contribute to laziness. Author Ted Chiang illustrated his skepticism in "Why A.I. isn’t going to make art," comparing the use of ChatGPT for academic assignments to introducing a forklift in a gym—effectively hindering cognitive growth.

Recent studies support Chiang's assertions. The first study, "ChatGPT as a cognitive crutch," published in a peer-reviewed journal by researcher Andre Barcaui, conducted a randomized trial to assess the effects of ChatGPT on knowledge retention among college students. Participants were divided into those utilizing AI for assistance and those studying conventionally, with a surprise retention test scheduled 45 days post-study. Findings revealed that students who employed ChatGPT scored significantly lower on the retention test (57.5% correct) compared to their traditionally studying peers (68.5% correct), suggesting that unrestricted AI use diminishes long-term memory retention by lowering the mental effort required for durable learning.

The second study, titled "The generative AI learning penalty," is still in the working paper stage and was authored by David Stromberg, Victor Lei, and Yanhui Wu. Their research involved 30 months of data from 26,811 Chinese students in grades 7-12. They investigated how generative AI influenced homework productivity and overall learning outcomes through a difference-in-differences design and found that while generative AI increased homework scores by 18% and reduced task completion time by 30%, it concurrently caused a 20% decline in monthly exam results within six months. Additionally, scores on high-stakes entrance exams dropped by 18-24%, with the most significant declines noted in social science disciplines, followed by STEM and language subjects, particularly among junior students, high-achieving students, and male learners.

In conclusion, although the history of technology suggests that new advancements may not make us less intelligent—as Nicholas Carr once argued regarding Google—Artificial General Intelligence represents a unique challenge. We have the power to utilize technology in transformative ways, enhancing our intelligence and streamlining tedious tasks, yet the choice remains whether we allow it to dictate our thought processes or even overshadow our capacity for independent thinking.

Employing A.I. May Diminish Your Mental Agility

A solitary adult sits hunched in a dark room, the warm glow of a smartphone screen casting harsh shadows on their face. Behind them looms a semi-transparent brain etched with circuit-like patterns that are cracked and dissolving, symbolizing waning mental resilience. The scene is moody and ominous, lit only by the device’s amber light.

Using A.I. Might Reduce Your Mental Fitness - Forbes

- New studies suggest AI use in education may impair knowledge retention and cognitive effort.  
- One study found students using ChatGPT scored lower on retention tests compared to traditional methods.  
- Another study indicated AI adoption improved homework scores but decreased exam performance significantly.  

TOPSHOT - CORRECTION / A man works on the electronics of Jules, a humanoid robot from Hanson Robotics that uses artificial intelligence, at a stand during the International Telecommunication Union (ITU) AI for Good Global Summit in Geneva on July 8, 2025. (Photo by Valentin FLAURAUD / AFP) / "The erroneous mention[s] appearing in the metadata of this photo by FABRICE COFFRINI has been modified in AFP systems in the following manner: [Valentin Flauraud] instead of [Fabrice Coffrini]. Please immediately remove the erroneous mention[s] from all your online services and delete it (them) from your servers. If you have been authorized by AFP to distribute it (them) to third parties, please ensure that the same actions are carried out by them. Failure to promptly comply with these instructions will entail liability on your part for any continued or post notification usage. Therefore we thank you very much for all your attention and prompt action. We are sorry for the inconvenience this notification may cause and remain at your disposal for any further information you may require." (Photo by VALENTIN FLAURAUD/AFP via Getty Images) AFP via Getty Images

The history of technology in education shows that there is really no quick fix. It is true that when calculators first appeared many worried about using it as a substitute for thinking. More recently is the concern over the use of smartphones. And yet, in the cycle of history, when each new technology arrives, it really doesn’t seem to improve education, but neither does it seem to harm it. In fact, technology may more realistically serve as an amplifier of already existing individual differences among students.

Earlier this year, Rose Horowitch published a widely read Atlantic piece titled “The end of reading is here,” and recently some UC students seem to have large gaps in their math preparation, even though they have been admitted to a top college. If people have reduced their reading, this isn’t a good sign, as I would argue that reading great works is important for improved writing, and writing is often a reflection of thinking.

So where does this leave us with the rise of AI use in education? James Flynn—who is known for the supposed rise in IQ due to what he argued was the complexification of society and our environment—often noted that humans were getting smarter because we trained our brains in the mental gym of our lives, which is getting increasingly more complex. But will AI help us train our brains, or will it make us lazier? Ted Chiang, in the article titled “Why A.I. isn’t going to make art,” noted that “Using ChatGPT to complete assignments is like bringing a forklift into the weight room; you will never improve your cognitive fitness that way.”

And at least two recent research studies seem to confirm Chiang’s point.

The first study—published in a peer reviewed journal—is titled “ChatGPT as a cognitive crutch.” The author Andre Barcaui conducted a randomized controlled trial of ChatGPT’s impact on knowledge retention among undergraduate students. Participants were randomly assigned to an AI assisted study group and a traditional study group and knowledge retention was assessed by a surprise test 45 days after learning. From the abstract: “Students who used ChatGPT scored significantly lower on the retention test (57.5 % correct) compared to those who studied traditionally (68.5 % correct), t (83) = −3.19, p = .002, Cohen's d = 0.68. This suggests that unrestricted ChatGPT use impaired long-term retention, likely by reducing the cognitive effort that supports durable memory.”

The second study—currently a working paper—is titled “The generative AI learning penalty.” The authors David Stromberg, Victor Lei, and Yanhui Wu write in their abstract: “Using 30 months of panel data on 26,811 Chinese students in grades 7-12, we study how generative AI affects homework productivity and learning. The data combine monthly closed-book exams, high-school and college entrance exams, and homework scores and completion time across nine subjects. We exploit staggered AI adoption in a difference-in-differences design. AI adoption raises homework scores by 18% and reduces completion time by 30%, but lowers monthly exam scores by 20% within six months. High-stakes entrance-exam scores fall by 18 and 24%, with the full penalty emerging only after about two years. The losses are largest in social science subjects, followed by STEM and languages, and are especially large for junior students, high-achieving students, and boys.”

MORE FOR YOU

Even though the history of technology suggests that each new technology may not necessarily make us stupid, as Nicholas Carr argued about Google awhile back, past technologies have not been Artificial General Intelligence—which is meant to be a general problem solver or thinker. We can still choose to use the technology in ways that can augment our own human intelligence, such as speeding up coding or other monotonous tasks in our lives. But we also can choose to use the technology to direct how we think or even largely replace our thinking.

Using A.I. Might Reduce Your Mental Fitness

Since the release of GPT 3.5 in November 2022, artificial intelligence has significantly transformed the landscape of higher education. For a considerable period, many educators resisted granting students access to AI technologies, often imposing sanctions for suspected use during assessments. Early attempts at creating "AI detectors" proved largely ineffective while overlooking the increasing integration of AI in professional environments, which has created a crucial demand for AI literacy among students. Instead of rejecting AI in academia, we should encourage its use as a learning tool for our students. 

The focus now is shifting towards the evolution of AI from generative models to more advanced agentic ones, alongside broader incorporation of AI within enterprise software solutions. These developments are pivotal, as agentic AI can undertake roles within organizations that extend beyond simple chatbots or predefined tasks, allowing for autonomous evaluations that enhance efficiency and quality. Additionally, recursive frameworks facilitate algorithms to learn and improve themselves, suggesting and implementing enhancements independently.

While advancements such as artificial general intelligence (AGI) and artificial superintelligence (ASI) represent significant leaps in AI capabilities, there remains ambiguity in their definitions. Some authorities argue that we may already be on the threshold of AGI, with ASI just a few years away, while others contend that we may not see these developments until mid-century. Therefore, it seems more prudent to concentrate on clearly defined milestones within AI innovation.

At this juncture, we might be on the brink of a transformative advancement in AI through the use of stable quantum computing, which promises to enhance algorithm capabilities significantly. This entails supercomputers linking with quantum computers to capitalize on the exceptional speed and unique features of quantum mechanics. Many companies, including industry giants like IBM, Google, Microsoft, and AWS, are actively investing in quantum technology, with reports highlighting at least 80 corporations committed to this field. IBM, for example, boasts the largest fleet of quantum systems globally, while Google recently made headlines with its 105-qubit Willow chip.

The National Institute of Standards and Technology has outlined some defining traits of quantum computing. These systems operate by leveraging the principles of quantum physics, which govern the universe at minute scales and frigid temperatures. Current quantum computers are basic and prone to errors, but more advanced models could potentially solve complex problems in a fraction of the time required by conventional computers. Hence, there is a race among governments, corporations, and research institutions to develop this technology, which is anticipated to collaborate with traditional computers rather than replace them. This partnership could lead to groundbreaking breakthroughs in scientific endeavors such as material science and drug development, significantly impacting industries and potentially revolutionizing cybersecurity.

One fascinating aspect of quantum physics is the phenomenon of superposition, which allows micro-particles to exist in multiple states simultaneously. Quantum computers utilize qubits, whose superposition capacity enables them to function as both 0 and 1 at the same time for the duration of a computation. This results in an exponential increase in processing power compared to classical computing.

Another striking property of qubits is entanglement, which allows them to be connected even at vast distances. For instance, altering the state of one entangled qubit changes its partner's state instantaneously, regardless of how far apart they are. This phenomenon was famously noted by Einstein in 1935, who referred to it as "spooky action at a distance," emphasizing its contradiction to the idea that nothing can exceed the speed of light.

These unique quantum attributes enable rapid and efficient computation, which researchers are exploring by linking supercomputers with quantum machines and developing hybrid computing systems. However, an alarming concern arises regarding the implications of quantum computing on cybersecurity, as its formidable speed could easily compromise encryption protocols used worldwide.

With quantum computers capable of potentially breaching encryption systems in mere minutes—something that would take traditional computers billions of years—an event known as Q-Day is being discussed within the scientific community. This event signifies the moment when large-scale quantum computers can use Shor's algorithm to crack public key systems based on integer factorization. The Q-Day Countdown site indicates that, as of now, researchers estimate we may reach this critical point in approximately 876 days, signaling an urgent need for protective measures in our data security.

While I remain optimistic that innovative encryption methods may emerge to avert catastrophic breaches, I am also beginning to envision how quantum computing’s unique features could revolutionize distance learning. For now, it is imperative to recognize that AI will persist in evolving, yet the most significant advancements will likely occur as quantum systems are integrated into AI development, especially in the coming years. Is your institution equipped to address the potential consequences of quantum computing? Do you have access to a functional quantum computer for educational and research initiatives? Who is spearheading quantum innovation at your institution?

Concentrate on Quantum Advancements

A bright laboratory bathed in warm morning sunlight, where a scientist in a crisp lab coat leans over a sleek quantum processor on a wooden workbench. She adjusts a precision optical lens focused on the chip’s glowing superconducting qubits, sunlight filtering through a tall window to reveal delicate circuit patterns. Behind her, a chalkboard bears elegant quantum equations—an optimistic scene pointing toward a quantum-powered future.

Focus on the Quantum Future - Inside Higher Ed

- Faculty initially resisted AI technology in education, overlooking its growing importance in the workplace.  
- Quantum computing could revolutionize AI algorithms, enhancing their speed and capabilities significantly.  
- Quantum computers' potential to break encryption raises concerns for privacy in various sectors.  

AI has had a remarkable impact on higher education since the release of GPT 3.5 in November 2022. For an inordinately long time, many faculty members fought access to the technology for students and imposed penalties for suspected use of AI in learning assessments. Not only were the initial “AI detector” attempts notoriously unsuccessful, this approach ignored the growing acceptance and reliance on AI in the workplace and resulting need for students to achieve AI literacy. Rather than denying access to AI in higher education, we should promote learning supported by AI among our students.

Attention has now shifted to the advancement of AI from generative to agentic models and the overall integration of AI into enterprise applications. These are important advancements in the field. Agentic approaches enable AI to take on roles within the organization that go beyond limited chatbots and rigidly defined operations to also perform self-assigned assessments, leading to new roles and activities to improve quality and efficiency. Further, recursive approaches enable algorithms to initiate self-improvement, ultimately proposing and independently implementing improvements.

While artificial general intelligence and artificial super intelligence will mark significant improvements in AI algorithms, there is not a single clear definition of either of these terms. Therefore, some experts believe that we have already achieved AGI and that ASI is just a year or two away; others say it will be the middle of the century before either is achieved. So, without minimizing the impact of any of the steps toward general and super intelligence, I think we are better served by focusing on more solidly defined steps in AI advancement.

Most Popular Stories Most Popular

That brings us to what may become a pivotal stage of development in AI, the use of stable quantum computing to accelerate and extend the potential of algorithms. This can begin with supercomputing digital computers linking to quantum computers to take advantage of the awesome speed and other capabilities of quantum physics. Many corporations already have constructed viable quantum computers. James Dargan reports in the Quantum Insider that there are at least 80 corporations that are heavily investing in the development of quantum computers: “IBM, Google, Microsoft and AWS are among the leading big tech players, each investing significantly in quantum R&D. IBM operates the world’s largest fleet of quantum systems through its Quantum Computation Center in New York, while Google achieved a landmark with its 105-qubit Willow chip.”

The National Institute of Standards and Technology earlier this year described some of the unique characteristics of quantum computers:

“These computers work by harnessing quantum physics—the strange, often counterintuitive laws that govern the universe at its smallest scales and coldest temperatures. Today’s quantum computers are rudimentary and error-prone. But if more advanced and robust versions can be made, they have the potential to rapidly crunch through certain problems that would take current computers years. That’s why governments, companies and research labs around the world are working feverishly toward this goal. Quantum computers will not replace our familiar ‘classical’ computers. Rather, the two types of machines could work together to solve problems that stymie classical computers, potentially supercharging scientific research in fields such as materials and drug discovery, giving a boost to industry and upending cybersecurity as we know it.”

One of the “strange, often counterintuitive laws” referenced above that govern subatomic physics is the phenomenon of superposition. That is, the ability of incredibly tiny particles to hold multiple, seemingly exclusive, properties at the same time. Much as digital computers use bits, quantum computers use qubits. The superposition of a qubit allows it to exist in a combination of both the 0 and 1 states simultaneously. For at least a brief period of time needed to complete computations, qubits can be counted as 0, 1 or a simultaneous combination of 0 and 1 at that given time. This results in an exponential acceleration in the capability of quantum computers to do their work compared to digital computers.

Another quality of qubits is that they can become entangled and share properties such as polarity. This entanglement enables them to be separated, potentially by tens of thousands of miles, and retain a special affinity to their entangled partner. For example, if one changes the polarity of one qubit, the entangled partner qubit changes polarity instantly, at the very same instant, not just at the speed of light. In the case of entangled qubits distanced thousands of miles away from each other, monitored with atomic clocks, the change weirdly takes place at the very same instant.

Arguably, the most famous theoretical physicist of all time found this action extraordinary:

“In 1935, Albert Einstein and colleagues first pointed out the ‘spooky’ action of quantum entanglement. Quantum entanglement, however, appeared to conflict with Einstein’s theory of special relativity, which postulates that nothing can travel faster than the speed of light and is demonstrated mathematically by the well-known equation E=mc2.”

A combination of these quantum characteristics and qualities has essentially enabled subatomic teleportation over tiny distances. These are all characteristics of subatomic quantum physics rather than the Newtonian physics that we are familiar with in everyday life.

The application of these features to AI algorithms greatly expands and accelerates their speed. Initially, this is experimentally being accomplished by linking supercomputers to quantum computers and even creating hybrid computers. One of the more concerning aspects of the incredible speed and power of quantum processing is its impact on cyberencryption, which is used for privacy in many applications worldwide.

The overwhelming speed of quantum enables quantum computers to break encryptions. In the very near future, it is anticipated that all online encryption will be vulnerable to hacking using public keys attacked by quantum computing. Imagine all privacy—at universities, banks, departments of defense and beyond—becoming vulnerable on the same day when a computer reaches the speed and complexity to crack all codes.

In Forbes, Chuck Brooks warns,

“It would take a billion years for a traditional computer to crack the encryption of today’s RSA-2048 standard. It could theoretically break in less than two minutes if you had a functional quantum computer. An event referred to as Q-Day by quantum researchers is where large-scale quantum computers can use Shor’s algorithm to break all public key systems that employ integer factorization-based (and other) cryptography.”

A dramatic reminder of the impending importance of Q-day can be found at the Q-Day Countdown site. At the time of writing of this column, the site is displaying Google’s estimation of the time to the date of Q-day showing 876 days, 7 hours, 59 minutes and 12 seconds. That would be Jan. 1, 2030. Check out the site yourself if you would like a reminder of when our financial systems may collapse if we fail to provide some sort of prior intervention.

I remain hopeful that some ingenious new mode of encryption (mind meld?) may be discovered to avoid the catastrophe. I am only beginning to imagine ways in which the unique features of quantum computing may enhance distance teaching and learning. In the meantime, we should remain aware that AI will continue to improve, but a massive quantum leap is most likely to be seen when we engage quantum computers in advancing AI, especially in the next two or three years. Is your institution prepared for the broad ramifications of quantum computing? Do you have an operable quantum computer on campus for teaching and research purposes? Who is leading your institution in quantum thought and action?

Focus on Quantum Future

In a thought-provoking piece for the Financial Times, John Thornhill explores the implications of Artificial General Intelligence (AGI) on the global landscape, with a particular focus on how emerging nations like Nigeria might navigate its challenges and opportunities.

Thornhill’s article emphasizes the anxiety surrounding the belief that AGI is an inevitable progression, which could lead society to passively accept whatever advancements are made without critical examination. This fatalistic perspective raises concerns that decisions surrounding AGI could become detached from public oversight, allowing developers to wield significant influence.

The timeline and definition of AGI remain uncertain; no one knows precisely when it will manifest or if the lofty ambitions surrounding it will materialize. Nevertheless, substantial investments are being made in AI systems designed to enhance learning, reasoning, and cognitive functions with minimal human input.

Young professionals in Nigeria, particularly those entering the workforce, may be among the first to feel AGI’s effects. Traditionally, careers commence with foundational tasks: attorneys conduct research, accountants analyze figures, software developers take on manageable projects, and journalists hone their skills in information gathering. However, if machines can accomplish these tasks autonomously, companies may have mixed feelings about hiring new graduates. Capable and educated, these individuals may still find a scarcity of entry-level roles available to them.

Nigeria has discussed the potential of its youthful demographic for years. When properly employed, a large young population can significantly boost economic growth. Yet, without adequate job creation, the increased population can strain resources, from family income to public services.

Moreover, Nigeria has aspirations in knowledge-based sectors, hoping to compete globally akin to India, which has successfully cultivated a robust skilled labor market serving wealthier nations. While Nigeria's presence is smaller, the opportunity remains alluring. International companies can tap into Nigerian talent for various roles—software engineering, customer service, and more—without requiring physical relocation, thus saving costs.

However, as AI continues to evolve, particularly with regard to automating outsourced tasks, the demand for human workers may be diminished. For instance, firms in London or New York might find they need fewer employees to oversee jobs once handled by large teams, as intelligent systems take over routine research, coding, and customer support.

The landscape is changing even further; the rise of physical intelligence means AI is being integrated into machines capable of performing tasks in factories and warehouses. Observations of China’s advancements in robotics exemplify the potential trajectory of this technology. If intelligent machines become adept within the industrial sector, countries striving to create job opportunities will have new dynamics to consider.

The pace of this change is unpredictable, and it may take longer than current trends suggest. Additionally, the line separating AI from manual labor is becoming increasingly blurred.

Nigeria has long been embedded in global digital systems largely governed by foreign entities. Technologies such as software, cloud services, and digital platforms sustain its economy. This reliance transforms into a critical issue when these systems start influencing key economic and public policy decisions.

For example, a Nigerian bank dependent on an external AI system may face challenges if the provider alters pricing models or restricts access. Such changes could lead to increased operational costs or service disruptions, negatively affecting customers as well.

Similar dependencies could beset government agencies using foreign AI platforms, raising concerns about data management, accessibility, and accountability.

Thornhill’s caution regarding the inevitability surrounding powerful AI developments highlights issues of control. If AI development is considered an unavoidable future, the companies pioneering these technologies gain a disproportionate advantage, allowing their choices to be seen merely as technical progress, rather than matters open to public discourse and scrutiny.

Nigeria lacks the power to influence the research labs driving advanced AI development or the massive investments being made. However, it can still assert influence over how these technologies are implemented within its own economy.

Countries and corporations developing leading AI technologies will likely dictate the standards governing their use, encompassing aspects like safety, data standards, intellectual property rights, and AI companies’ responsibilities. Such regulations will impact nations that did not play a role in creating the foundational technology.

Developing countries risk adopting foreign standards due to dependencies on international AI systems and may face limited alternatives that compel acceptance of commercial terms. This reliance could deepen, especially as AI integrates more seamlessly into the decision-making fabric of the economy.

Recognizing these challenges, Nigeria is beginning to lay the groundwork; its National Artificial Intelligence Strategy outlines plans focused on research, infrastructure, talent cultivation, and responsible AI integration. The National Centre for Artificial Intelligence and Robotics is actively working on the research and adoption of AI innovations.

To fully engage with this evolving landscape, universities must produce individuals equipped to understand and critically assess technology, and researchers require access to advanced computing capabilities. Regulators should possess sufficient technical acumen to navigate this space, while public institutions need stable digital frameworks and reliable data. Currently, Nigeria faces notable gaps in these essential areas.

While Nigeria doesn’t need to create the world’s most advanced AI systems, it should build sufficient capacity to understand what it acquires from technology providers, negotiate effectively, and identify potential risks associated with over-dependence.

The advantages of AI are tangible; even smaller Nigerian businesses are harnessing AI to streamline tasks that once required larger personnel. Researchers can analyze data more expediently, and government bodies may alleviate burdens on limited manpower through automation.

The architects of these systems will ultimately determine their functionality and applications. Decisions about training data, design assumptions, and operational standards will emerge from distant locations, potentially leading to misunderstandings of Nigerian contexts. Although foreign technologies aren’t inherently detrimental, enhancing local expertise becomes increasingly critical.

Cultural, legal, and public governance systems may encounter similar hurdles. A country lacking representation in advanced AI development may find that widely adopted systems misinterpret its societal and economic dynamics, based primarily on external data and perspectives.

While the state doesn’t need to dominate all AI enterprises or manufacture every technical component domestically, it must identify areas where dependence could pose risks, where regulations must be enacted, and where investments in local capabilities can strengthen the country’s bargaining power.

Nigeria's history is rife with challenges regarding global economic shifts that often unfold as external circumstances to react to, rather than opportunities it helps forge. As the implications of AGI become clearer, this vulnerability could become starkly apparent.

The future of AGI is uncertain; its precise form and arrival time remain unknown. Nonetheless, Nigeria has the potential to navigate this uncertainty effectively. The current choices surrounding investment, research, infrastructure, regulation, education, and ownership will significantly impact how such countries confront future developments.

As the AI journey unfolds, Nigeria faces a choice: remain passive in the back seat or cultivate the capacity to influence the destination.

The impending era of artificial general intelligence

Two silhouettes—a human scientist in casual attire and a sleek metallic humanoid—stand atop a grassy hill at sunrise, gazing toward a distant cityscape where skyscrapers blend with glowing strands of code. Warm, natural morning light bathes the scene as birds soar overhead, symbolizing the dawn of the age of artificial general intelligence.

The coming age of artificial general intelligence - TheCable

- John Thornhill warns that society must not treat AI development as inevitable and unchangeable.  
- Young Nigerians may face fewer job opportunities as AI automates entry-level career tasks.  
- Nigeria must strengthen its AI understanding to influence technology's integration into its economy.  

Stay connected via Google News

John Thornhillâs recent article in the Financial Times, âHumans cannot remain passengers in the back of the AGI carâ, has stayed with me because of what it suggests about where artificial intelligence may be taking the world and what countries such as Nigeria might find waiting for them when the technology becomes much more powerful.

Thornhill is looking beyond the arrival of another generation of better software. He is troubled by the growing belief that artificial general intelligence is inevitable and that society will simply have to adjust to whatever the people developing it eventually produce. Once that attitude takes hold, people can start treating decisions about the technology as though there is little left for governments, citizens and institutions to decide.

AGI remains uncertain. Nobody knows when it will arrive or whether the more ambitious claims made about it will prove realistic. Even the meaning of AGI is still disputed. Yet enormous amounts of money are being invested in systems expected to become increasingly capable of learning, reasoning and performing a wide range of cognitive tasks with less human direction.

Some of the first people likely to feel its effects may be young Nigerians entering the labour market. The early years of a career are often pretty ordinary. A young lawyer researches cases and prepares documents before being trusted with more complicated work. A young accountant checks figures and learns how a company operates. A software developer begins with smaller assignments. A journalist spends years learning how to find information, check facts and understand the workings of an organisation. The routine work is part of the education.

Advertisement

If machines become capable of doing much of it, employers may have less reason to recruit large numbers of beginners. The graduate may still be qualified, intelligent and hardworking. There may simply be fewer opportunities through which he acquires the experience he needs to become a professional.

Nigeria has spent years talking about the size and potential of its young population. A large young population can become an economic advantage when people are productively employed and businesses can absorb them. Without enough productive work, population growth puts more pressure on everything from household incomes to public services.

This has consequences beyond the domestic labour market because Nigeria has also looked to knowledge-based services as an area in which its young, educated population can compete internationally. Countries such as India built huge industries around supplying skilled labour to businesses in richer economies. Nigeriaâs participation is much smaller, but the attraction is familiar. A company abroad can employ Nigerian programmers, accountants, researchers, customer-service workers and other professionals without bringing them physically into its country. The attraction has always been that the work can be done elsewhere for less.

Advertisement

AI changes that arithmetic if machines can eventually perform a substantial part of the work being outsourced. A company in London or New York that once needed a large team to carry out routine research, coding, customer support or administrative work may eventually need fewer people. The Nigerian worker may remain perfectly capable. The employer may simply have less work for him to do.

The technology is already moving beyond people working behind computers. The development of what has been called physical intelligence is pushing AI into machines that can move, manipulate objects and perform increasingly complicated tasks. Chinaâs rapid progress in robotics offers one indication of where the technology could go. If intelligent machines become more capable in factories, warehouses, transport and other physical settings, countries still trying to create large numbers of industrial and service jobs will have another calculation to make.

Nobody knows how quickly that will happen. It may take much longer than some of the current enthusiasm suggests. The line between artificial intelligence and physical work is becoming less clear.

Nigeria has long operated within digital systems whose most important components are controlled elsewhere. Foreign software, cloud services, hardware and digital platforms are already part of everyday economic life. There is nothing unusual about a developing country using technology developed abroad. The problem becomes more serious when those systems begin to perform functions that sit close to the heart of economic and public decision-making.

Advertisement

Consider a Nigerian bank whose operations become heavily dependent on an AI system controlled by a company outside the country. The bank depends on continued access to the system and on the commercial terms set by its owner. If the company changes its pricing or restricts access to the system, the bank could face higher costs or disruption. A serious error could also affect its customers.

A government agency relying heavily on an overseas AI platform would face similar questions about data, reliability, accountability and access.

Thornhillâs concern about fatalistic inevitability comes down to power. If increasingly powerful AI is treated as something that will happen anyway, the companies developing it acquire an enormous advantage. Their decisions can begin to look like technical developments rather than choices that should be open to political and public scrutiny.

Nigeria will have little influence over the laboratories where the most advanced systems are developed or over the enormous amounts of capital being committed to them. It can still have a say in how those systems enter the Nigerian economy and under what conditions they are used.

Advertisement

The countries and companies developing the most powerful AI systems will have a strong voice in setting the standards under which those systems are used. Rules covering safety, data, intellectual property, access to computing and the responsibilities of AI companies will increasingly affect countries that did not develop the underlying technology.

Developing countries could end up adopting standards developed elsewhere because their businesses need access to foreign AI systems. They could accept commercial conditions because alternatives are limited. They could depend on foreign infrastructure while having little influence over the companies controlling it.

Advertisement

That pattern is already familiar in other parts of the global economy. AI could make the dependence deeper because it reaches into the knowledge and decision-making functions of the economy itself.

Nigeria has begun putting some of the necessary pieces in place. Its National Artificial Intelligence Strategy sets out ambitions around research, infrastructure, talent and responsible AI adoption, while the National Centre for Artificial Intelligence and Robotics is working on research, development and the adoption of AI and other emerging technologies.

Advertisement

Universities need to produce people capable of understanding and questioning the technology, while researchers need access to serious computing capacity. Regulators will need enough technical knowledge to understand what they are buying and regulating, and public institutions will need reliable digital infrastructure and data. Nigeria has substantial gaps in these areas.

The country does not have to build the worldâs most powerful AI system. But it needs enough capacity to understand what it is buying, negotiate with the companies supplying it and develop alternatives where dependence becomes too risky.

Advertisement

The benefits are real. A small Nigerian business can already use AI to perform work that once required more staff. Researchers can process large amounts of material much faster. Government agencies could eventually reduce some of the administrative burden that consumes limited manpower.

The people building these systems will make choices about how they work and how they are used. The data used to train them, the assumptions built into their design and the standards applied to their operation will not emerge from nowhere. A Nigerian user may be interacting with a system whose understanding of Nigerian circumstances is limited by decisions made thousands of miles away. That does not make foreign technology inherently bad. It makes local capacity more valuable.

Culture, law and public administration could face similar problems. A country that has little presence in the development of advanced AI may eventually discover that systems used extensively by its citizens understand its economy and society largely through somebody elseâs data and assumptions.

The state does not need to control every AI company or attempt to manufacture every piece of technology locally. It, however, needs to know where dependence could become dangerous, where regulation is required and where investment in domestic capacity would improve the countryâs bargaining position.

Nigeria has encountered this problem before in different forms. Major changes in the global economy have often reached the country as developments to which it must respond rather than opportunities it helped shape. By the time the consequences become clear, the strongest positions have usually been taken by others. Artificial general intelligence could make that weakness more obvious.

Nobody can say with confidence what AGI will look like or when it will arrive. Countries such as Nigeria can live with that uncertainty. The decisions being made now about research, investment, computing infrastructure, regulation, education and ownership will shape the environment in which such countries encounter whatever comes next.

The car is moving. Nigeria can remain in the back seat and leave the driving to others, or build enough capacity to have some say in where it goes.

Ladigbolu is a journalist based in Lagos.

The coming age of artificial general intelligence

A figure that warrants consideration: 93%. This is the concordance rate achieved by IBM Watson for Oncology when compared to recommendations made by expert tumor boards in a peer-reviewed analysis focusing on breast cancer treatment decisions. In a related endeavor, Harvard Medical School's CHIEF model, an AI for pathology trained using more than 60,000 whole-slide images, demonstrated nearly 94% accuracy in detecting cancer across 11 different types and 15 independent datasets. Notably, AI's assessment of ovarian cancer staging reached 97% accuracy, significantly outperforming radiologists, who showed 88% accuracy.

These results are not mere laboratory experiments but rather published clinical data, prompting a critical question for healthcare leaders contemplating the future of diagnostics: if today's narrow AI—designed for specific diagnostic tasks—can function at near-expert levels, what possibilities arise when such capabilities span multiple diagnostic fields at once?

This concept ties into the potential of Artificial General Intelligence (AGI) in healthcare, a technology that has yet to be realized. However, current diagnostic accuracy statistics present compelling evidence of where this technology is headed and underscore its rapid progression.

AI Diagnostic Tools Surpassing Human Expertise in Specialized Tasks

The advancements in AI diagnostics are accelerating, primarily coming from independent research, not marketing claims from vendors.

Medical imaging has become the primary field for evaluating AI diagnostic performance, with consistently impressive results. A systematic review published in 2025, encompassing 23 studies involving 23,270 patients, revealed that AI systems achieved a median AUC-ROC of 0.88 in prostate cancer detection, either matching or surpassing the accuracy of human radiologists, all while decreasing reporting times by up to 56% compared to traditional methods. In chest imaging, deep learning models utilizing extensive datasets can identify early pulmonary nodules that human radiologists often overlook in low-dose CT scans.

In another study published in JMIR in 2026, AI demonstrated up to 98.88% accuracy in multiclass disease classification from medical images. Such figures indicate a substantial enhancement in performance compared to standard radiological reviews, particularly for high-volume screening tasks.

This trend holds true across various imaging modalities, where AI not only matches the proficiency of specialists but frequently exceeds it in terms of speed, consistency, and detection of nuanced findings on a large scale.

The Power of Integrative AI in Cancer Detection and Genomic Analysis

While individual diagnostic performance is remarkable, the real value lies in AI's ability to integrate multiple data sources simultaneously. A systematic review conducted in 2025-2026 and published on ScienceDirect analyzed 123 papers focusing on agentic AI in cancer detection. It found that GPT-4 class models could classify skin lesions with accuracy comparable to dermatologists and stage ovarian cancer more effectively than radiologists. Yet, the true significance lies in the added benefits of multimodal data integration.

AI systems can now combine imaging results with genomic data, tumor biomarker panels, historical treatment information from electronic health records (EHRs), and patient demographics—all within seconds. A peer-reviewed study published in Frontiers in Artificial Intelligence in January 2026 highlighted AI's ability to consolidate genomic, transcriptomic, proteomic, imaging, and EHR data into a cohesive analytical framework, thereby enhancing early disease detection and expediting biomarker discovery across oncology, neurology, and cardiovascular medicine.

No individual clinician can process such diverse data streams in real-time during patient encounters. This is not a critique of human capability but rather an acknowledgment of the overwhelming volume and variety of data involved in a thorough diagnostic assessment, which surpasses unassisted human processing abilities.

Transitioning from Narrow AI to AGI-Level Diagnostic Reasoning

This is where understanding the difference between current AI capabilities and the future of Artificial General Intelligence in healthcare becomes crucial. Present-day AI diagnostic tools are designed for narrow tasks; each performs a specific function well, be it analyzing a particular type of scan, staging a specific cancer, or identifying a certain genetic mutation. However, they do not interrelate findings across different domains. A system designed for pulmonary nodule detection lacks the awareness of a patient’s genomic sequencing results or relevant prior medical history.

In contrast, an AGI-level diagnostic system would have the capacity to synthesize information across various fields—imaging, genomics, lab results, and social determinants of health—functioning cohesively, much like a multidisciplinary tumor board, but with greater speed and efficiency.

Currently, the diagnostics sector represents the fastest-growing segment of AI in healthcare, projecting a compound annual growth rate (CAGR) of 39.8% from 2025 to 2030, according to MarketsandMarkets. This growth highlights a shift towards developing systems that transition from isolated accuracy in specific tasks to integrated reasoning across diverse diagnostic domains.

Progress in Personalized Medicine and AGI: Moving Towards Individualized Treatment

While the achievements in diagnostic accuracy are impressive, the broader implications of this technological evolution significantly impact treatment strategies, particularly the move from population-based protocols to a more personalized approach to medicine.

AI's current capability to integrate genomics, biomarkers, EHR histories, and social determinants represents a formidable advancement. For two decades, precision medicine has remained a goal in healthcare, but previous tools have struggled to meet this ambition. Data from genomic sequencing often outpaced clinical interpretation, while EHR histories remained isolated and social determinants of health were inconsistently factored into treatment strategies.

Today, multimodal AI is transforming this scenario. Research published in Frontiers in Artificial Intelligence in January 2026 confirmed that AI systems capable of merging genomic, transcriptomic, proteomic, imaging, and EHR data enable personalized treatment recommendations that enhance early disease detection and accelerate the discovery of actionable biomarkers, especially in oncology.

This capability supports treatment recommendations that consider various patient factors—such as how a tumor appears, implications from genomic data regarding drug responses, prior lab results on organ function, and social history regarding medication adherence—simultaneously before a clinical team makes a decision.

AI does not replace oncologists. Instead, it empowers them with a level of data synthesis that would otherwise require an entire team and considerable time.

The Role of Quantum Computing in Advancing AGI-Level Medical Diagnoses

Processing speed at scale remains a significant challenge for current AI in diagnostics. Genomic datasets are vast, drug interaction modeling necessitates simulating molecular behaviors across numerous variables, and the training needed for multimodal systems requires managing data volumes that traditional computing struggles with efficiently.

Quantum computing directly addresses these challenges. A study published in PMC in 2026 indicated that quantum-assisted drug discovery has the potential to revolutionize molecular simulations through algorithms like the Variational Quantum Eigensolver. This innovation could enhance predictions of molecular interactions, streamline drug design, and expedite AI model training for clinical decision-making. Quantum machine learning is already being applied to genomic analysis, significantly improving biomarker identification and patient stratification in ways that conventional machine learning cannot achieve at comparable speeds.

For AGI-level diagnostic systems, which will need to handle patient-specific multiomics data, imaging, and real-time monitoring all at once, quantum-enhanced processing won't just be advantageous—it will be essential.

Accelerating Drug Discovery with AGI Capabilities

The implications of AGI technology are particularly transformative for drug discovery. Traditionally, bringing a new drug to market has taken more than a decade and incurred costs in the billions. AI has begun to shorten this timeline by simulating trial results, predicting compound toxicity, identifying drug-target interactions, and flagging existing molecules for new applications—all tasks that previously required extensive laboratory work.

The AGI-level iteration of these processes envisions a system that accelerates individual steps while also integrating them across the entire drug discovery and clinical trial pipeline. This system would leverage patient-derived multi-modal data to identify disease targets, design candidate molecules, predict interaction profiles, and stratify trial participants according to genomic subtypes. These currently disconnected operations would become a streamlined, continuous workflow under an AGI-aligned AI architecture.

Establishing the Software Infrastructure for AI-Enhanced Diagnostics in Clinical Settings

The case for AI diagnostic tools in clinical environments is compelling, but the bigger challenge lies in creating systems that can be deployed securely and effectively—ensuring regulatory approval, mitigating bias, and providing clinical validation that stands up under scrutiny when patient outcomes depend on the results.

The necessity for validated, bias-controlled training data is paramount. AI diagnostic systems learn from historical datasets. If these datasets represent historical biases in healthcare—such as underrepresentation of specific populations in imaging studies, inconsistent documentation for non-native speakers, or less thorough genomic sequencing for economically disadvantaged groups—the AI models can perpetuate these biases within their output.

A systematic review from 2025 of agentic AI in cancer detection identified factual inaccuracies in 15% to 41% of model outputs across different systems. This variability is clinically significant. Therefore, addressing bias is a critical preemptive measure that must account for how training datasets are constructed, how model performances are evaluated across various demographic groups, and how ongoing monitoring can capture performance drift.

Healthcare organizations adopting AI diagnostic tools should pose a vital inquiry to their development partners: How was the model's performance validated across diverse patient populations, imaging technologies, and clinical settings? If responses lack specificity, the potential for bias remains a concern.

The FDA's 2025 draft guidance covering AI-enabled Software as a Medical Device (SaMD) has reshaped the regulatory landscape for AI diagnostic tools, introducing lifecycle management requirements that influence the regulatory burden. This guidance delineates submission protocols, oversight measures, and change management for models that evolve post-deployment.

Practically speaking, any AI diagnostic system that informs or significantly impacts clinical decisions will likely need to undergo FDA SaMD classification. This requirement will affect architectural choices—whether to use static or adaptive models—along with documentation needs relating to training data sourcing, validation methodologies, and performance metrics across populations.

For organizations considering whether to procure an existing AI diagnostic product or develop a bespoke solution, understanding the regulatory framework is crucial. Off-the-shelf solutions that already have FDA clearance may offer quicker deployment but at the expense of customization. Conversely, custom-built systems allow for clinical relevancy and adaptability, but necessitate a development partner experienced in navigating the SaMD pathway from the ground up, rather than retrofitting compliance criteria at the end.

Insights from Organizations Successfully Aligning with FDA 2025 Standards

Organizations that effectively manage this process tend to view regulatory compliance as pivotal from the outset, as opposed to a mere checklist for launch.

This approach signifies that data governance must be established before model training begins, documenting the sources of data utilized and the strategies employed to assess demographic representation. Validation methodologies need to be determined prior to deployment, with established benchmarks for performance across distinct patient groups. The evidence supporting safety claims for the tools should be developed concurrently with design rather than compiled hastily when a regulatory review is requested.

A proficient healthcare software development team brings this compliance mindset to the table, as they have previously navigated the FDA SaMD regulatory process. They understand where reviewers will scrutinize documentation closely, the types of gaps that can lead to delays, and recognize that constructing AI diagnostic tools that successfully undergo regulatory review and can be used effectively in clinical practice requires meticulous planning from inception rather than a last-minute effort before launch.

The 93% concordance rate connected to AI in cancer detection is remarkable. Similarly, the 98.88% accuracy in X-ray classifications speaks volumes. Yet, these statistics merely scratch the surface, representing the foundation of a diagnostic transformation. As we make strides toward AGI-level reasoning, future systems will integrate all relevant data sources a clinician would seek and process them more swiftly than any human team can accomplish.

Healthcare organizations that will reap the most benefits from this ongoing transformation are the ones currently building the necessary infrastructure, governance protocols, and compliant software frameworks to support it before the technological potential is fully realized.

The Impact of 93% Accuracy in Cancer Detection on Future Developments in Healthcare Diagnostics with Artificial General Intelligence

A sunlit hospital diagnostic suite with floor-to-ceiling windows letting in warm morning light. A focused female physician stands beside a sleek touchscreen medical display labeled “AGI Diagnostic AI,” showing a detailed 3D cross-section of a tumor with a bold “93% Accuracy” badge. Nearby, a calm patient silhouette rests on a bed, charts and diagnostic graphs softly illuminated on wall-mounted monitors. Render in a realistic digital-painterly style, emphasizing hope and precision.

Artificial General Intelligence in Healthcare Diagnostics: What 93% Accuracy in Cancer Detection Tells Us About What's Coming Next - Nasscom

- IBM Watson for Oncology achieved a 93% concordance rate with expert recommendations in breast cancer treatment.
- AI systems demonstrate high accuracy in various cancer detection tasks, outperforming human specialists consistently.
- Multimodal AI integration enhances personalized medicine by synthesizing diverse patient data rapidly for treatment planning.

A number worth pausing on: 93%.

That's the concordance rate IBM Watson for Oncology achieved with expert tumor board recommendations in a peer-reviewed study of breast cancer treatment decisions. Separately, Harvard Medical School's CHIEF model — a pathology AI trained on more than 60,000 whole-slide images — hit nearly 94% accuracy in cancer detection across 11 cancer types and 15 independent databases. Ovarian cancer staging by AI: 97% accuracy, compared to 88% by radiologists.

These aren't prototype results from a lab. They're published clinical findings. And they raise a question that every healthcare leader thinking about the future of diagnostics needs to sit with: if today's narrow AI — systems built for specific tasks, not general reasoning — can already perform at near-expert level in individual diagnostic tasks, what happens when that capability extends across every diagnostic domain simultaneously?

That's the promise of Artificial General Intelligence in healthcare. It isn't here yet. But the diagnostic accuracy data we're seeing right now is the clearest signal yet of where this technology is heading — and how fast.

How AI Diagnostic Tools Are Already Matching and Outperforming Human Specialists in Narrow Tasks

The diagnostic wins are stacking up fast, and they're coming from independent research groups, not vendor marketing.

Medical imaging and radiology: what peer-reviewed evidence shows

Medical imaging is where AI diagnostic performance has been most rigorously tested — and the results are consistently striking.

In prostate cancer detection, a systematic review published in 2025 covering 23 studies and 23,270 patients found that AI systems achieved a median AUC-ROC of 0.88 — matching or exceeding radiologist accuracy — while reducing reporting time by up to 56% compared to conventional reads. In chest imaging, deep learning models trained on large datasets can detect pulmonary nodules at early, treatable stages that human readers routinely miss on low-dose CT scans.

In X-ray classification, a JMIR study published in 2026 documented AI achieving up to 98.88% accuracy in multiclass disease classification from medical images — a figure that, in most clinical environments, would represent a meaningful improvement over standard radiological review for high-volume screening tasks.

The pattern is consistent across modalities. AI doesn't just match specialists in specific imaging tasks — in controlled conditions, it frequently outperforms them on speed, consistency, and detection of subtle findings at scale.

Cancer detection and genomic analysis: a pathway no single clinician can replicate

Where individual diagnostic accuracy is impressive, the more important capability is what happens when AI integrates multiple data streams simultaneously.

A 2025–2026 systematic review in ScienceDirect covering 123 papers on agentic AI in cancer detection found GPT-4 class models classifying skin lesions as accurately as dermatologists and staging ovarian cancer better than radiologists. But the more significant finding is what multimodal integration adds on top of image analysis.

AI systems can now combine imaging findings with genomic sequences, tumor biomarker panels, prior treatment history from EHRs, and patient demographics — in seconds. A peer-reviewed paper published in Frontiers in Artificial Intelligence in January 2026 described this directly: AI-enabled multimodal integration consolidating genomic, transcriptomic, proteomic, imaging, and EHR data into a unified analytical framework that enhances early disease detection and accelerates biomarker discovery in oncology, neurology, and cardiovascular medicine simultaneously.

No single clinician, no matter how skilled, can process those data streams in parallel during a clinical encounter. That's not a criticism — it's a structural reality. The volume and variety of data involved in a comprehensive diagnostic workup now exceeds what unassisted human cognition can reliably synthesize.

The leap from narrow AI to AGI-level reasoning in diagnostics

This is where the distinction between today's AI and Artificial General Intelligence in healthcare becomes meaningful.

Current AI diagnostic tools are narrow. Each system does one thing well — reads a specific type of scan, stages a specific cancer type, flags a specific genetic mutation. They don't connect findings across domains. A pulmonary nodule AI doesn't know what the patient's genomic sequencing showed, or what their EHR says about prior environmental exposures.

AGI-level diagnostic reasoning would. A true AGI system could synthesize findings across imaging, genomics, labs, social determinants of health, and clinical history — reasoning across all of them simultaneously, the way a multidisciplinary tumor board does, but faster, more completely, and at scale.

The diagnosis and detection segment is currently the fastest-growing AI subsegment in healthcare, at 39.8% CAGR from 2025 to 2030 according to MarketsandMarkets. That growth rate reflects real investment in exactly this direction: systems that move from isolated accuracy in narrow tasks toward integrated reasoning across diagnostic domains.

Personalized Medicine and AGI: From Population-Level Treatment to One-Patient Precision

The diagnostic accuracy story matters. But the bigger implication of where this technology is heading is what it means for treatment — specifically, for the shift from population-level protocols to truly individualized medicine.

How AI integrates genomics, biomarkers, EHR history, and social determinants today

Precision medicine has been a stated goal in healthcare for two decades. For most of that time, the tools available couldn't fully deliver on it. Genomic sequencing generated data that outpaced clinical interpretation. EHR histories were siloed. Social determinants of health were documented but rarely integrated into treatment planning in any systematic way.

Multimodal AI is changing that equation. Published in Frontiers in Artificial Intelligence (January 2026), peer-reviewed research confirmed that AI systems integrating genomic, transcriptomic, proteomic, imaging, and EHR data are already enabling personalized treatment recommendations that improve early disease detection and accelerate clinically actionable biomarker discovery — particularly in oncology.

This means a treatment recommendation that accounts for what a patient's tumor looks like, what their genome says about likely drug response, what their prior lab values show about organ function, and what their social history suggests about medication adherence — all at once, before the clinical team makes a decision.

The AI isn't replacing the oncologist. It's giving the oncologist a data synthesis that would otherwise require a team and a week.

The role of quantum computing in accelerating AGI-level medical diagnosis

One of the harder limits on current AI in diagnostics is processing speed at scale. Genomic datasets are enormous. Drug interaction modeling requires simulating molecular behavior across thousands of variables. Multimodal training requires handling data volumes that classical computing handles inefficiently.

Quantum computing addresses this directly. Research published in PMC in 2026 confirmed that quantum-assisted drug discovery is expected to revolutionize molecular simulations through algorithms like the Variational Quantum Eigensolver — improving molecular interaction predictions, streamlining structure-based drug design, and accelerating AI model training for clinical decision-making. Quantum machine learning techniques are already being applied to genomic analysis, improving biomarker identification and patient stratification in ways that classical ML cannot match at the same speed.

For AGI-level diagnostic systems — which will need to process patient-level multiomics data, imaging, and real-time monitoring simultaneously — quantum-enhanced processing isn't a future capability. It's the infrastructure that will make that scale practical.

Drug discovery at AGI scale: years instead of decades

Drug discovery is where the AGI trajectory has the most dramatic near-term implications.

Developing a new drug has historically taken over a decade and cost billions. AI is already cutting into that timeline by simulating trial outcomes, predicting compound toxicity, identifying drug-target interactions, and flagging approved molecules for repurposing — all tasks that previously required years of wet lab work.

The AGI-level version of this is a system that doesn't just accelerate individual steps but integrates across the entire discovery-to-trial pipeline. Mining multi-modal patient-derived data to identify disease targets. Designing candidate molecules. Predicting interaction profiles. Stratifying patients for trials by genomic subtype. These functions, currently handled by separate specialized teams working sequentially, become a continuous, connected workflow under AGI-adjacent AI architecture.

Building the Software Infrastructure That Powers AI-Enhanced Diagnostics in Clinical Environments

The clinical case for AI diagnostics is clear. The harder work is building systems that can be deployed safely in real healthcare environments — with regulatory approval, bias controls, and clinical validation that holds up when patients' lives depend on the output.

Why AI diagnostic tools require validated, bias-audited training data

AI diagnostic systems learn from historical data. When that data reflects historical disparities in healthcare — underrepresentation of certain populations in imaging datasets, inconsistent documentation for non-English-speaking patients, less comprehensive genomic sequencing for lower-income groups — the model learns those biases and encodes them into its outputs.

A 2025 systematic review of agentic AI in cancer detection found factual errors in 15% to 41% of model outputs across tested systems. That range is wide enough to matter clinically. Bias mitigation isn't a post-deployment concern — it must be addressed in how training datasets are assembled, how model performance is tested across demographic subgroups, and how ongoing monitoring catches drift.

Healthcare organizations deploying AI diagnostic tools need to ask their development partners a specific question: How was this model's performance validated across different patient populations, imaging equipment types, and clinical environments? If the answer is vague, the risk of bias is real.

FDA SaMD classification: what the regulatory pathway means for your build vs. buy decision

The FDA's January 2025 draft guidance on AI-enabled Software as a Medical Device (SaMD) introduced lifecycle management requirements that meaningfully changed the regulatory burden for AI diagnostic tools. The guidance covers marketing submission recommendations, lifecycle oversight through continuous monitoring, and change-control mechanisms for models that update over time.

In practical terms: any AI diagnostic tool that makes or significantly influences a clinical decision is likely to require FDA SaMD classification. That classification affects architecture decisions (static vs. adaptive models), documentation requirements (training data provenance, validation methodology, performance benchmarks by population), and post-market monitoring obligations.

For healthcare organizations deciding whether to buy an off-the-shelf AI diagnostic product or build a custom solution, this regulatory pathway is a central variable. Off-the-shelf products with existing FDA clearance offer faster deployment but less flexibility. Custom-built systems offer clinical fit and configurability — but require a development partner who understands the SaMD pathway from the first architectural decision, not as a regulatory retrofit at the end of the project.

How healthcare software development teams build to FDA 2025 standards from day one

The organizations that navigate this well share a common characteristic: they treat regulatory compliance as a design input, not a launch checklist.

That means data governance is mapped before model training begins — documenting what data was used, where it came from, and how demographic representation was assessed. It means validation methodology is defined before deployment, with acceptance criteria for performance across patient subgroups. It means the clinical evidence that supports the tool's safety claims is built during development, not assembled afterward when a regulatory reviewer asks for it.

A specialized healthcare software development team brings this orientation because it has worked through the FDA SaMD process before. It knows where reviewers focus on their scrutiny. It knows what documentation gaps create delays. And it knows that building AI diagnostic tools that survive a regulatory review and hold up in real clinical use requires the same rigor at the beginning of a project that most teams only apply at the end.

The 93% concordance figure from AI cancer detection is remarkable. The 98.88% X-ray classification accuracy is remarkable. But what these numbers really represent is the floor — the starting point of a diagnostic transformation that, as it moves toward AGI-level reasoning, will integrate every data source a clinician would want and synthesize them faster than any human team can.

The healthcare organizations that will benefit most from that transformation are the ones building the data infrastructure, governance frameworks, and compliant software architecture to receive it — right now, before the technology arrives at its full potential.

Artificial General Intelligence in Healthcare Diagnostics: What 93% Accuracy in Cancer Detection Tells Us About What's Coming Next

In Anchorage, Alaska, a group of concerned residents recently took to the streets to voice their opposition to the rapid advancements in artificial intelligence (AI), urging the establishment of an international agreement to pause AI development until its safety can be assured.

The demonstration was spearheaded by PauseAI Anchorage, a local chapter of a nationwide grassroots organization advocating for a more cautious approach to AI. Alongside the protest, three Alaskans previously traveled to Washington, D.C., in April to discuss AI with Alaska's congressional representatives. Among them was Maxim Ishchuk, a computer science student at the University of Alaska Anchorage, who participated in last week’s protest.

Ishchuk expressed concern after realizing that many experts are raising alarms regarding AI progress and that there is no scientific consensus on achieving safe Artificial General Intelligence (AGI). “At some point, I noticed a lack of initiatives within our community to address this issue,” he remarked.

Public opinion reflects a similar sentiment; according to PEW Research Center, both experts and a majority of U.S. adults believe increased regulation of AI is essential. Ishchuk noted that during his visit to Washington, D.C., he discovered that many lawmakers were not well-versed in AI technology. “Congress is currently quite uninformed about AI,” he stated, emphasizing the need to educate officials on the significant risks posed by AI.

Some observers, like Alex Jacquez, are doubtful that a treaty, such as the one proposed by PauseAI, is feasible. Jacquez, who is the Senior Vice President for Policy Advocacy and Research at Groundwork Collaborative and has previously worked in the White House National Economic Council, expressed skepticism about the government’s ability to manage AI regulation effectively. “There's a genuine doubt about the government’s capacity to handle this matter, even with public consensus on the need for more oversight,” he noted.

Moreover, PEW's data reveals that many American workers are worried about AI's impact on their jobs, with 52% of employed individuals expressing concerns about AI in the workplace. Jacquez highlighted that managerial pressure to adopt new AI tools is contributing to workplace anxiety, as employees increasingly encounter AI features in their everyday software applications.

As Alaska prepares for a closely watched U.S. Senate race, the topic of AI might influence voter decisions, especially in light of growing interest in establishing data centers in the state. Jacquez mentioned that if data centers prove financially viable and regulatory hurdles are managed effectively, they'll likely emerge despite local resistance. 

There will inevitably be opposition, similar to the protests in Anchorage. While Ishchuk supports the use of AI in specific, limited contexts, he emphasizes the need to slow down unrestricted advancements. “Narrow applications of AI can enable us to achieve our innovative goals,” Ishchuk stated, “but we can do this without creating autonomous beings capable of surpassing human intelligence.”

As activists displayed their signs along the Seward Highway, passersby might consider a larger question regarding the trajectory of AI technology. “Alaskans should reflect on the implications of supporting the race towards artificial superintelligence and AGI that might exceed human capabilities,” Ishchuk concluded.

Alaskans push for a pause on the progress of AI technology.

A group of rugged Alaskan adults in heavy parkas stand atop a rocky ridge under a stormy, brooding sky, clutching handmade “Slow AI” and “Pause the Machine” signs. Behind them looms a dark silhouette of server racks with faint red warning lights, and an Alaska state flag waves at the scene’s edge.

Alaskans advocate for slowing AI advance - Alaska's News Source

- Anchorage locals protested for an international treaty to pause AI development for safety.  
- Experts believe more government regulation on AI is necessary, amid public concern.  
- The protest highlights local opposition to rapid AI deployment and data center expansion.  

ANCHORAGE, Alaska (KTUU) - A group of Anchorage locals took to the streets last week to protest the quickening pace of AI development, calling for “an international treaty to pause AI development until it can be made safely.”

The protest was organized by PauseAI Anchorage, a local chapter of a national grassroots advocacy organization. In addition to the protest, 3 Alaskans flew to D.C. with PauseAI in April, to meet with Alaska’s congressional delegation on the topic of AI. One of those three was Maxim Ishchuk, a computer science major at UAA, who was also at last Monday’s protest.

“I noticed a lot of experts were sounding the alarm on AI development and that there is no scientific consensus to make AGI [Artificial General Intelligience] safely,” Ishchuk said. “And at some point, I realized that there weren’t many people in the community doing anything about it.”

When it comes to government regulation on AI, experts and U.S. adults think more regulation is needed, according to PEW Research Center. For his part, Ishchuk said what he encountered in D.C. is that many elected officials were lacking in knowledge on AI.

“Congress is pretty uneducated on AI right now,” Ishchuk said. “And so a lot of it was basically just educating them about the existential risk and just major risks that AI poses in general.”

Some are skeptical that any such treaty like the one PauseAI is calling for will come, among them is Alex Jacquez,

“I think it’s very hard to imagine something right now that looks like a multilateral agreement on AI,” Jacquez said.

Jacquez is currently Senior Vice President for Policy Advocacy and Research at Groundwork Collaborative, prior to that, he previously served on the White House National Economic Council.

“There’s just real doubts, I think, that the government is up to the task, even though people do believe that we need more regulation,” Jacquez said.

Further data from PEW suggests American adults are concerned what AI might mean for their jobs, with 52% of employed adults saying they are “worried” about AI use in the workplace.

“It is largely, I think, being driven from managers, from executives, who are basically being told that they have to deploy these new tools,” Jacquez said. “I think that’s actually causing some of the backlash that we’re seeing amongst people at work who are wondering why there is suddenly an AI button next to every piece of software that they use.”

Amidst a nationally watched race for Alaska’s U.S. Senate Seat, the topic of AI might factor into voter’s decisions, especially as interest in data centers increases in the 49th state.

“If there’s a viability for a data center in Alaska with permitting and the regulatory measures sorted out and it’s going to be cost effective for them, they’ll do it,” Jacquez said. “I think they are trying to deploy as many data centers as possible as quickly as possible and really struggling up against some local opposition in a lot of places.”

Opposition, such as the protest in Anchorage. Ishchuk said he supports the “narrow application” of AI, but is still seeking a to slow unchecked development.

“Narrow applications still allow us to do a lot of the cool things that we want to do with AI,” Ishchuk said. “But we can do that without making beings that can think on their own and, you know, overpower us eventually.”

So as signs spin along the Seward highway, passerby may be prompted to ask a larger question about the future of AI technology.

“I think that Alaskans should really just ask the question,” Ishchuk said. “Do we want to support this push for artificial super intelligence and artificial general intelligence that is smarter than humans?”

See a spelling or grammar error? Report it to web@ktuu.com

Alaskans advocate for slowing AI advance

achieve is the extensive engineering experience gained from applying theoretical research in real-world scenarios. This distinguishes Dunxu Technology from early AI startups, as the company has a solid foundation built on rigorous data collection and practical application.

Third, Dunxu Technology has established a clear vision for how to turn collective intelligence into a tangible product. The process is not straightforward; it requires careful consideration of how to transform this abstract concept into operational systems that deliver results in various environments. 

By focusing on the orchestration of different intelligent agents, Dunxu Technology aims to bridge the gap between individual capabilities and collective outcomes. The company is not just developing algorithms but is crafting a framework that allows different devices and systems to collaborate effectively, resulting in enhanced performance that surpasses what each could achieve independently.

Ultimately, the challenge lies in building systems that can adapt to shifting conditions while maintaining the cohesion necessary for effective teamwork. This adaptability is crucial as it drives the development of solutions that can enhance how various technologies interact, offering businesses innovative ways to optimize their operations.

As collective intelligence principles gain traction within the broader AI landscape, Dunxu Technology finds itself in a unique position. By leveraging its extensive background in this field, the company aims to lead the charge in defining how organizations can harness the power of collaboration among intelligent agents. 

In conclusion, the implications of Dunxu Technology's approach could be transformative, changing the dynamics of the AI industry by highlighting that true intelligence does not solely originate from powerful models but also from how these models and systems can be organized and integrated to create a higher level of collective functionality.

Collective AGI Company: Introducing an "Organization" Layer to AI for the Advancement of Collaborative Intelligence

A modern glass-walled boardroom bathed in warm daylight, a large interactive touchscreen shows five concentric rings labeled “Data,” “Learning,” “Reasoning,” “Organization,” “Collaboration,” with the “Organization” ring softly glowing in amber. Around a sleek wood table, a diverse team of professionals leans in, pointing and animatedly discussing the new layer. The Collective AGI logo is prominently displayed on the screen, sunlight casting dynamic patterns across the polished floor.

Collective AGI Company: Adding an "Organization" Layer to Artificial Intelligence for Collaborative Intelligence Evolution - 36Kr

- Dunxu Technology aims to enhance AI by organizing collective intelligence among various models.  
- The company introduces an "Organization Layer" to improve collaboration between intelligent agents and devices.  
- Collective AGI seeks emergent intelligence through proper organization, not just powerful individual models.

In the past few years, the clearest development path in the AI industry has been to continuously strengthen the performance of individual models.

With larger parameters, more data and stronger reasoning capabilities, a single model can handle increasingly complex tasks. However, as AI starts to expand from chat interfaces to robots, drones, industrial and urban systems, the problems to be solved have also changed accordingly.

Real-world systems rarely operate relying on only one single model.

In a robot fleet, each machine is equipped with different sensors, computing power and assigned tasks; a low-altitude security system may integrate cameras, radars, drones, edge devices and different algorithms at the same time. No matter how intelligent a single model is, it cannot natively know how these devices should divide work, when to collaborate, and who should take over the work once a failure occurs.

When there are more and more intelligent elements, how to organize them properly? — This is exactly the question that Dunxu Technology is committed to answering.

In April this year, Dunxu Technology was officially established. Derived from the State Key Laboratory of Complex System Cognition and Decision-Making under the Institute of Automation, Chinese Academy of Sciences, this company has a somewhat unexpected goal.

As an artificial intelligence enterprise, Dunxu Technology does not focus on large models and related tracks. Instead, it tries to add a new capability layer — the Organization Layer — between existing intelligent units such as large models and end applications.

This is no small undertaking. In the computer era, the emergence of the new Operating System Layer gave birth to Microsoft and Apple. In the AI era, the emergence of the new Intelligence Layer spawned OpenAI and Anthropic. It can even be said that every leap of the information revolution originates from the emergence of a new tier in the computing architecture.

In the technical logic of Dunxu Technology, models are responsible for generating individual intelligence, while organization is responsible for generating collective intelligence: connecting agents with different capabilities, costs and forms, enabling them to collaborate, give feedback and adjust, so that the entire system can finally obtain capabilities that a single individual does not originally have.

Simply put, large models solve the problem of whether an agent is "intelligent enough". What Dunxu Technology wants to solve is: When more and more individuals with basic intelligence appear, how to organize them — and even make them generate higher-level intelligence through emergence.

The company calls this goal Collective AGI, Collective Artificial General Intelligence. This is also the problem that Pu Zhiqiang, founder and chairman of Dunxu Technology, has studied for more than ten years. The 38-year-old man is the deputy director of the State Key Laboratory of Complex System Cognition and Decision-Making, a researcher at the Institute of Automation of Chinese Academy of Sciences, and a recipient of the National Science Fund for Excellent Young Scholars.

In 2024, Pu Zhiqiang discussed distributed collective intelligence with his team

More than ten years ago, it did not have such a cutting-edge name as Collective AGI. At that time, what attracted the 20-something Pu Zhiqiang was an ant.

01

Another Source of Intelligence

A single ant can do very few things. But a group of ants, without a central "commander" issuing orders, can complete complex behaviors such as foraging and nesting that far exceed the capabilities of a single individual — and guiding this process only requires one or two simple rules.

Relying on collective intelligence, ants that do not know how to build bridges can still put up a bridge (AI-generated image)

The summary of complex science for this phenomenon is very simple: the essence of collective intelligence is to generate advanced intelligence by properly organizing simple rules. "That is to say, it is not 1+1=2, but 1+1>2." What really interests Pu Zhiqiang is exactly that "extra" part. In the research of complex science, this phenomenon is often called "emergence".

Although this research field won the Nobel Prize in Physics in 2021, it was a very unpopular direction more than a decade ago. Pu Zhiqiang later recalled that if a direction was already very hot and the path was very clear, he might not be willing to enter it.

"There is nothing to hesitate about." He thinks collective intelligence is cutting-edge enough, but it is not cutting-edge for the sake of being cutting-edge. In today's artificial intelligence era, it touches a more essential problem: does intelligence necessarily come from increasingly powerful individuals? Or can intelligence be generated through the organizational relationship between individuals themselves?

He has been working on this problem for more than ten years. Pu Zhiqiang has always emphasized that research should solve "real problems", which made his research move from algorithms and simulations to the real world very early.

Around 2020, he began to apply collective intelligence to the research of real football matches. Football is actually a very intuitive collective intelligence system. No player in a team has all the information. Everyone can only see part of the scene, but the team must constantly make overall decisions: who to pass the ball to next, who should run for position, and how to adjust after the opponent changes the playing style.

What Pu Zhiqiang wants to solve is not just to do data statistics after the game, but the problems "during the game" and "before the game": when the game reaches this point, how to play next? What will happen if another player is replaced? Is it possible to deduce in advance when facing opponents of different styles?

The team even conducted counterfactual deduction and tried to turn this system into a kind of "virtual coach". In 2025, in a specific complex decision-making experiment, the team tried to make two self-built models with 100-million-level parameters work collaboratively. Under the same conditions, compared with the general large model of 100-billion-level parameters at that time, the experimental results of this task were significantly better.

For Pu Zhiqiang, this is not just a story of "small models defeating large models", but a more plain evidence: Individuals do not need to be omnipotent. By properly organizing rules, they can also generate more advanced capabilities.

Using real UEFA Champions League match data and collective intelligence, the 100-million-parameter model "defeats" the 100-billion-parameter model

However, the players on the football field have gradually become drones, cameras, radars and different algorithms, and then robot dogs and robots. The problem has not actually changed. Each individual has limited information and different capabilities. The real difficulty is to make them complete division of labor, collaboration and adjustment in a constantly changing real environment.

When these laboratory researches enter the real world, they also start to leave another thing for this research: data and engineering experience.

In the first real-world application scenario of low-altitude security, Pu Zhiqiang's team has accumulated about 50TB of multi-modal drone detection data, covering different modalities such as visible light, near-infrared, thermal infrared and radar; at the same time, there are 15 million frames of effectively annotated target samples and more than 2000 related experiments.

The significance of these numbers is not only "large". The fact that an algorithm works in simulation does not mean it still works after changing the weather, background or camera; the fact that drones can collaborate in an ideal environment does not mean they can still operate normally when encountering communication delay, equipment error and node failure.

Real scenarios test technology on the one hand, and generate data required for the next round of technology on the other hand.

From early cluster management and multi-agent game to UAV swarm and heterogeneous management of low-altitude security, the same problem has been repeatedly verified in different systems. Algorithms, data, dedicated models and software and hardware collaboration experience are also retained in this process.

This constitutes the difference between Dunxu Technology and many early AI startups: The company is new, but the capabilities it needs to productize are not new. Theoretical construction, data accumulation and engineering foundation have already been completed before the company was founded.

02

Turning "Relationship" into Products

Having accumulated enough technology does not mean that a company must be established. In fact, in the path Pu Zhiqiang designed for this matter, the company is placed at the very end.

Around 2023, he began to consider the industrialization of collective intelligence more seriously. The team formed an internal sequence: Real application — Real system — Product — Company. First enter the real environment to confirm that the technology can really work; then move from a single algorithm to a complete system; after the system runs, judge how many capabilities can be precipitated into products. Only at the last step is a company needed.

When Pu Zhiqiang judges "whether the time has come", he does not look at whether there is a beautiful Demo, but at another more practical indicator — reusability.

At the beginning, the team was still doing specific projects one by one. When a customer came with a demand, they solved the corresponding problem. Unconsciously, the team completed projects with a total value of over 100 million yuan in a few years. But after doing so many projects, Pu Zhiqiang gradually found a change: Some capabilities that originally needed to be redeveloped can now be reused between different projects.

"It's only the last kilometer left." He described that stage like this. The last kilometer is no longer suitable to be completed in the research institute.

The market then accelerated this process. An appearance at an industry exhibition attracted more than 30 institutions to express their willingness to cooperate, and the market demand of tens of millions of yuan suddenly appeared in front of him, which made Pu Zhiqiang realize: "it's time".

Pu Zhiqiang later called it "order-driven entrepreneurship": It is not to first establish a company with the mission of Collective AGI, and then find a scenario that can verify the business logic; but that the technology is already in demand, and the company has conversely become a necessary part to be completed.

In April this year, Dunxu Technology was established. The May and June after the establishment of the company were called "Market Months" internally. Several co-founders in charge of different fields had to meet customers, sort out demands and seek cooperation.

Two months later, no one in the team worried about the market anymore, "The demands are all piling up!"

At this time, he began to worry about products.

Because there are enough demands, sometimes it will drag a technology company back to the most familiar and dangerous path: this customer needs a set of solutions, and the next customer needs a customized set. There are more and more projects, and the revenue may also grow, but the underlying capabilities can never be precipitated.

Pu Zhiqiang chose to exercise restraint. He did not accept every demand immediately, but judged first: after completing this project, what on earth can be retained?

This is also the meaning of "reusability" after it really enters the business world. For Dunxu Technology, what needs to be reused is not just a certain model, but whether the "cooperation mode" between different models and devices can be reused.

Low-altitude security has become the earliest test field. In a real low-altitude system, there may already be cameras, radars, drones, edge devices and disposal terminals from different manufacturers. Each of them can complete part of the tasks, but in the past, people still needed to connect these links in most cases.

Dunxu Technology adds a "Collective Intelligence Brain" in the middle. The front end is connected to perception devices, the middle end carries out information fusion, judgment and scheduling, and the back end is connected to execution devices; different modules can also be deployed separately according to the existing systems of customers.

From this perspective, what Dunxu Technology really wants to productize is not a radar or a drone, but the organizational relationship of collective intelligence between them.

There is a somewhat technical term inside Dunxu Technology called "organizational distillation". Simply understood, whether an effective collaboration can be turned into an empirical rule that can be used again next time, and even eventually become a set of formulas.

If several devices have gone through many rounds of operation and formed an effective division of labor and collaboration mode, the system hopes to precipitate it. Next time when replacing devices or changing tasks, there is no need to start from scratch; the new operation results will continue to feed back, so that this organizational mode can be adjusted continuously.

Therefore, what a project leaves behind is not only data and models, but also "how to organize these intelligences". This is also the most critical step from artificial intelligence to Collective AGI — Collective Artificial General Intelligence.

The realization of this matter will have great industrial significance. It may directly change the cost structure and market rules of the AI industry: it is not only the strongest model that can create powerful AI.

03

Why Dunxu Technology?

What Dunxu Technology faces is not an uncharted territory. On the contrary, "how to organize more and more agents" is becoming a problem of common concern in the AI industry.

In the software world, large model manufacturers are making multi-agent collaboration and workflow orchestration basic capabilities; in the physical world, autonomous system companies such as Shield AI and Anduril are also enabling different platforms, sensors and unmanned systems to work collaboratively.

Even in 2026, the boundary of similar products is still expanding. Applied Intuition launched the Agentic Platform for Physical AI, hoping to connect AI, data, development tools and real physical systems.

These adjacent players in the market start from two directions: one starts from large models and Agents to solve workflow orchestration in the software world; the other starts from physical systems such as drones, robots and autonomous driving, and continuously abstracts autonomous decision-making software to a higher level. Dunxu Technology starts from a completely different paradigm — collective intelligence.

Collective intelligence is a completely different intelligence paradigm

This also determines that its relationship with large model companies is not simple competition — the Organization Layer needs models. In Dunxu Technology's system, more powerful general models are responsible for complex reasoning, professional models are responsible for vertical tasks, and small models and end-side devices handle low-cost and low-latency demands; the Organization Layer is responsible for organizing these different capabilities according to tasks.

This is a path that can cooperate with basic model manufacturers, rather than rebuilding a set of AI outside large models. The continuous enhancement of model capabilities is not necessarily bad news for Dunxu Technology — the stronger and more diverse the nodes that can be organized, the more important the organizational problem will be.

However, only seeing this demand is not enough to explain why it is Dunxu Technology. What it really has unique identification at present is that it superimposes several capabilities at the same time.

First, more than ten years of collective intelligence research. Dunxu Technology did not start studying how multiple agents collaborate after the Agent concept became popular. From cluster management, multi-agent game to heterogeneous collaboration and real unmanned systems, this research line is far earlier than the establishment of the company, and the team has a deep understanding of the underlying interdisciplinary mechanism of collective intelligence.

Second, data and engineering experience accumulated through long-term practice in real physical systems. Dozens of TB of data is only the easiest part to quantify. What is more difficult to

Collective AGI Company: Adding an "Organization" Layer to Artificial Intelligence for Collaborative Intelligence Evolution

Large language models (LLMs) are designed to learn patterns from vast datasets, enabling them to generate responses. However, as noted by Goertzel, many of these models are unable to continuously adjust their fundamental parameters based on new experiences once their initial training is completed. While they can access previously stored information or reference earlier interactions, they lack a comprehensive memory that tracks their entire "life."

“They lack a long-term memory similar to that of a human,” he explained. “This limitation means they do not fully understand their identity, capabilities, or how they connect with the world.”

According to Goertzel, for a machine to exhibit general intelligence, it must possess a self-model—an internal representation that reflects its history, skills, constraints, and aspirations. Moreover, it should be capable of retaining useful insights and applying them to future choices.

Participants at a recent conference explored predictive coding as a potential method for advancing AI intelligence. This technique involves having a model forecast an outcome, assess its predictions against reality, and learn from the discrepancies.

Goertzel shared that his team has successfully implemented predictive coding within transformers at a scale similar to GPT-2, an earlier model by OpenAI. However, they have yet to scale this approach to match the capabilities of today's most advanced models.

While continual learning is a step forward, Goertzel argues that it addresses only a portion of the challenges. He advocates for a neural-symbolic AI approach, which merges pattern-recognizing neural networks with symbolic techniques that articulate rules, relationships, and objectives more transparently.

In addition, he emphasizes the importance of integrating ethical reasoning, self-awareness, and goal-oriented pursuits directly into the learning process, rather than retrofitting safeguards after the fact.

Establishing that an AI has achieved general intelligence may also pose a challenge. Companies can often train models to excel in popular tests without truly demonstrating the broader competencies those evaluations are meant to assess.

To tackle this issue, Goertzel suggests an ambitious measure he calls the "MIT test." In this scenario, an AI-driven robot would need to earn a doctorate from the university while adhering to the same regulations as human students. This would require the machine to navigate the campus, attend lectures, pass written and oral examinations, collaborate with professors, and generate original research. Achieving this would necessitate skills in communication, physical navigation, social reasoning, long-term planning, and knowledge creation.

“If I could develop a robot capable of attending MIT and obtaining a PhD while complying with all the same rules as human students, I would be genuinely impressed,” Goertzel remarked. “It needs to navigate hallways, perform an oral exam, complete a written test, and produce a thesis that represents an original contribution to knowledge.”

Researcher claims AGI may be achieved within a year.

A sunlit high-tech lab at dawn: a lone researcher seen from behind stands before a large glass wall covered in neural network diagrams and a wall calendar flipping to “12 months,” while a translucent 3D digital brain model glows softly on a nearby monitor, the scene bathed in warm natural light to convey anticipation of AGI’s imminent arrival.

AGI could arrive within a year, researcher says - IBM

- Large language models lack long-term memory and self-modeling for personal context understanding.  
- Goertzel supports neural-symbolic AI, integrating pattern recognition with explicit goal representation.  
- He proposes the "MIT test" for AI, requiring robot PhDs while following human rules.  

Large language models (LLMs) learn patterns from enormous collections of data and use them to generate responses. Most deployed models cannot continually update their central parameters from experience after training, Goertzel said. They can retrieve saved information or refer to earlier conversations, but those tools do not give them a continuous personal history.

“They don’t have a long-term memory of their whole life like a person does,” he said. “The result of that is they don’t know who and what they are and how they relate to the world around them.”

In his view, a generally intelligent machine would need a self-model, meaning an internal understanding of its history, abilities, limits and goals. It would also have to be able to preserve useful lessons and connect them to future decisions.

Conference participants discussed predictive coding as one possible route. The method asks a model to predict an outcome, compare its expectation with what actually occurs and learn from the difference.

Goertzel said his team has made predictive coding work in transformers at roughly the scale of GPT-2, an early OpenAI language model. The group hasn’t demonstrated the approach at the size of today’s leading models.

Continual learning addresses only part of the problem. Goertzel favors neural-symbolic AI, which combines neural networks that recognize patterns with symbolic methods that represent rules, relationships and goals more explicitly.

He also says that developers should place ethical reasoning, self-understanding and goal pursuit inside the core learning process. Adding safeguards after training doesn’t provide the same foundation, he said.

Proving that an AI has reached general intelligence could present another obstacle. Companies can train models to perform well on popular evaluations without demonstrating the broader skill each one aims to measure.

Goertzel proposes a more demanding alternative, which he calls the “MIT test.” An AI-controlled robot would have to earn a doctorate at the university while following the same rules as human students. The robot would need to navigate the campus, attend classes, pass written and oral exams, work with professors and produce original research. Success would require communication, physical competence, social judgment, long-term planning and the ability to develop new knowledge.

“If I could make a robot that could go to MIT and get a PhD without breaking any of the rules humans follow, I would be very impressed,” Goertzel said. “It’s got to roll through the hallways. It’s got to do an oral exam with someone, do a written exam and write a PhD thesis that passes the criteria of an original contribution to knowledge.”

At AI-focused bar in Beijing, DeepSeek tokens are included with the drinks.

More from Artificial General Intelligence

A Single Question to Govern Them All: The Source of Functional Information

Are We Properly Understanding AI Intelligence?

Understanding AGI: The journey to achieving human-like Artificial General Intelligence and its societal implications

Employing A.I. May Diminish Your Mental Agility