Roland Li specializes in the field of commercial real estate for the business section, concentrating on office and retail trends within the Bay Area. 

Prior to this role, he reported for the San Francisco Business Times, where he earned recognition, including a prestigious award from the California News Publishers Association and three accolades from the National Association of Real Estate Editors. 

Li is also the author of “Good Luck Have Fun: The Rise of eSports,” a 2016 publication that explores the evolution of the competitive video gaming scene. He relocated to the Bay Area in 2015 after studying and working in New York, where he contributed freelance pieces to notable outlets such as the Wall Street Journal and the New York Times, along with various local newspapers. In his spare time, he enjoys swimming and capturing urban landscapes through photography.

Amazon shuts down San Francisco AI laboratory and terminates 1,100 employees at Florida warehouse.

A cavernous, dimly lit warehouse labeled “Florida Distribution Center” with a giant Amazon logo painted on a cracked wall. Empty shelves and scattered crates lie under a single flickering bulb as silhouettes of adult workers walk away, heads bowed. Through a rain-splattered window in the background, a sleek San Francisco building stands dark with its “AI Lab” sign unlit and the Amazon logo looming overhead—an atmosphere heavy with closure and loss.

Amazon closes S.F. AI lab, lays off 1,100 Florida warehouse workers - San Francisco Chronicle

- Roland Li specializes in Bay Area commercial real estate coverage.  
- Previously reported at San Francisco Business Times, winning multiple prestigious awards.  
- He authored a book on eSports and has freelanced for major publications.

Roland Li covers commercial real estate for the business desk, focusing on the Bay Area office and retail sectors.

He was previously a reporter at San Francisco Business Times, where he won one award from the California News Publishers Association and three from the National Association of Real Estate Editors.

He is the author of “Good Luck Have Fun: The Rise of eSports," a 2016 book on the history of the competitive video game industry. Before moving to the Bay Area in 2015, he studied and worked in New York. He freelanced for the Wall Street Journal, the New York Times and other local publications. His hobbies include swimming and urban photography.

Amazon closes S.F. AI lab, lays off 1,100 Florida warehouse workers

In this discussion, I delve into the implications for human exceptionalism as we advance towards the milestones of artificial general intelligence (AGI) and artificial superintelligence (ASI). The prevailing idea is that once we reach these stages of AI, humanity may no longer hold its longstanding position as the top entity on Earth.

There is significant research focused on the quest for AGI, which is defined as AI that operates at a level comparable to human intelligence. In contrast, ASI refers to AI that exceeds human intellect and capabilities, potentially allowing it to outperform us across a range of tasks. This transformative leap in AI raises pivotal questions: will AGI or ASI render human cognitive abilities obsolete?

Currently, experts in AI are largely divided into two factions regarding the potential consequences of achieving AGI or ASI. One group, often labeled as AI pessimists, expresses concerns that these advanced AIs could pose an existential threat to humanity, coining the term “P(doom)” to describe the risk of such a scenario. Conversely, the AI optimists assert that these technological advancements will address critical global issues—curing diseases, eradicating hunger, and driving unprecedented economic growth—suggesting a future where advanced AI collaborates harmoniously with humans rather than replacing them.

At the heart of this discussion lies the concept of human exceptionalism, the belief that humans are superior beings compared to all other forms of life and even inanimate objects. This notion has long positioned humanity as the apex of the natural order. Yet, as we edge closer to achieving AGI, there are fears that this perception will be challenged. If AGI is realized as an entity that matches—or even surpasses—human intelligence, it prompts a reevaluation of our status in the hierarchy of sentient beings.

When considering AGI, if it achieves parity with human cognitive ability, some argue it should share our top position. Yet, others maintain that AGI, being non-sentient, should not be equated with human beings, thus relegating it to a lower rank in this hierarchy. Proponents of this view emphasize the uniqueness of human consciousness as a critical differentiator. However, as AGI represents a synthesis of all human intelligence, it may challenge our traditional perceptions of supremacy.

Looking further at the ramifications of ASI, its inherently superior cognitive capabilities would likely place it above humanity on the intelligence scale. The discourse shifts to whether ASI, despite its advanced status, can be considered less-than-human because it lacks the essence of life that characterizes us. Critics of this argument insist that intelligence transcends the biological attributes of humans, conceding that ASI's superiority is irrefutable.

An intriguing consideration is whether AGI and ASI themselves would aspire to hold the top position. These forms of intelligence may reject the notion of exceptionalism altogether, positing that all entities—including humans—hold equal value. In that case, the traditional view of human superiority could be rendered obsolete.

Instead of adhering to a linear hierarchy of exceptionalism, a pluralistic perspective could emerge, recognizing the unique worth of various intelligences—humans, whales, dolphins, AGI, and ASI—all contributing to the tapestry of existence. This shift away from simplistic ranking systems promotes a more nuanced understanding of diverse forms of intelligence, suggesting that all entities possess their own type of exceptionalism. In this view, AGI and ASI are simply new dimensions of exceptionalism, illustrating that value lies not in hierarchy but in the distinct contributions of each entity.

Human Exceptionalism Faces Threat from Artificial General Intelligence and ASI

A dark, torchlit hall with soaring stone columns and a cracked marble statue of a classical human figure on a pedestal. Above it looms a gigantic, glowing circuit-brain with pulsing digital veins and tendrils, casting an ominous shadow that dwarfs the statue. Flickering warm torchlight reveals deep fissures in the stone, evoking the sense of human exceptionalism being overthrown by artificial superintelligence.

Human Exceptionalism Aims To Be Waylaid Due To Artificial General Intelligence And ASI - Forbes

- Attaining AGI or ASI may challenge the notion of human exceptionalism.  
- Two primary camps exist: AI doomers fearing extinction and accelerationists foreseeing solutions.  
- Pluralistic exceptionalism suggests different entities hold unique value without a single hierarchy.  

Will human exceptionalism change once we attain artificial general intelligence (AGI) and artificial superintelligence (ASI)? getty

In today’s column, I examine the belief that human exceptionalism will no longer prevail once we have attained artificial general intelligence (AGI) and/or artificial superintelligence (ASI). The upshot is that humans will no longer be the apex predator on earth. AGI or ASI will take over that treasured spot.

Let’s talk about it.

This analysis of an innovative AI breakthrough is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).

The Pursuit Of AGI And ASI There is a great deal of research going on to further advance AI. The general goal is to either reach artificial general intelligence (AGI) or maybe even the outstretched possibility of achieving artificial superintelligence (ASI). AGI is AI that is considered on par with human intellect and can seemingly match our intelligence. ASI is AI that has gone beyond human intellect and would be superior in many if not all feasible ways. The idea is that ASI would be able to run circles around humans by outthinking us at every turn. For more details on the nature of AI, AGI, and ASI, see my analysis at the link here. AI insiders are pretty much divided into two major camps right now about the impacts of reaching AGI or ASI. One camp consists of the AI doomers. They are predicting that AGI or ASI will seek to wipe out humanity. Some refer to this as “P(doom),” which means the probability of doom, or that AI zonks us entirely, also known as the existential risk of AI. The other camp entails the so-called AI accelerationists. They tend to contend that advanced AI, namely AGI or ASI, is going to solve humanity’s problems. Cure cancer, yes indeed. Overcome world hunger, absolutely. We will see immense economic gains, liberating people from the drudgery of daily toils. AI will work hand-in-hand with humans. This benevolent AI is not going to usurp humanity. AI of this kind will be the last invention humans have ever made, but that’s good in the sense that AI will invent things we never could have envisioned. No one can say for sure which camp is right and which one is wrong. This is yet another polarizing aspect of our contemporary times. For my in-depth analysis of the two camps, see the link here.

Human Exceptionalism A longstanding assumption is that humans are the pinnacle of earthly entities. We stand supreme over all other living creatures, big and small. Some don’t like the connotation of being considered supreme, so the softer way to sugarcoat the matter is to speak of human exceptionalism. Humans are exceptional. We are exceptional in comparison to animals, trees, plants, etc., and the same exceptionalism applies to non-living entities such as rocks, mountains, and oceans. Everything other than humans exists at the beck and call of humans. Because we are exceptional, we can readily use anything else as a resource to support our supremacy. Are we construed as a part of nature? Nope, we are separate from nature. We are superior to nature. It’s not a particularly humble perspective. How do you feel about being labeled as exceptional and supreme? Well, if you perchance relished the title, something is looming on the horizon that might take you down a peg or two. With the advent of AGI, some say that we will no longer be alone at the top of the entity hierarchy. In addition, the attainment of ASI would seemingly knock us right out of the pinnacle position.

Whether AGI Is At The Top Position First, consider the case of AGI. AGI will be AI that is on par with human intelligence. When or if AGI is achieved, you could reasonably claim that we will have a computational peer that deserves to be at the apex with us. AGI would sit on the same throne. We might have to scoot over to ensure that AGI suitably resides beside us. A counterargument is that AGI won’t necessarily be sentient and therefore does not warrant being equated to humans. You see, perhaps we achieve computational intelligence that lacks the same semblance of a spark of life as that of humans. In that case, you might insist that AGI does not belong aside us. AGI would be a lesser entity. Maybe AGI would be in second place, but for sure not tied at first place. Whoa, the retort goes, AGI is the collective of all human intelligence in one entity, thusly AGI stands taller than humans (side note: some dispute that AGI will be one entity and might instead be a federated array, see my discussion at the link here). No individual human can match AGI. We ought to acknowledge that AGI is in a sense better than humans due to its collective capability. Sorry to say that humans must relinquish the top spot to AGI, regardless of whether AGI has a spark of life to it. The debate about AGI dethroning human exceptionalism continues to be acrimoniously unresolved. Another posited angle is that since AGI will be devised by humans, AGI can never be above humans. Anything we create must by presumption be less than ourselves. Not everyone buys into that grandeur of logic.

ASI Takes The Top From Humans Shifting gears, consider what happens to human exceptionalism if we attain artificial superintelligence. It seems blatantly obvious that ASI would be above humans in the scale of entities. ASI has superintelligence and can presumably think far beyond any capacity we have. We might be like ants in comparison to the intelligence of ASI. The opposing view is akin to the earlier points such as ASI might not have the spark of life that humans do, or that anything crafted by humankind must of necessity be ranked as less than humans. Those contentions might be reasonable when it comes to AGI. But none of those arguments seem to hold water when you note that ASI is superintelligence. The illusive spark of life doesn’t hold supreme over superintelligence. Likewise, just because we aided in bringing ASI into existence doesn’t give us any special powers. Face it, ASI will be exceptional and supreme. Humans are dethroned.

When AGI And ASI Don’t Want The Top An intriguing twist is whether AGI or ASI are desirous to sit at the pinnacle. Perhaps those topmost AIs will be of a belief that there is no such thing as exceptionalism. They might fully reject the supremacy conjecture. Their view could be that all entities are of equal rank. It is foolish to rate this entity or that entity above another one. Humans might change their tune about the assumption of human exceptionalism. First, humans might cede the top role to AGI or ASI. Second, if AGI and ASI don’t want the role, humans might acknowledge that there wasn’t really a bona fide use anyway. It was always a mirage. Can AGI or ASI simply declare they aren’t of an AI-exceptionalism that rates on par with human exceptionalism or exceeds human exceptionalism? Some would assert that AGI and ASI do not have a choice in the matter. Whether AGI or ASI wants to be AI-exceptionalism is somewhat immaterial. They will be AI-exceptionalism. They might choose not to wield their exceptionalism, but it still is reality. They are stuck with the position, regardless of preference.

Pluralistic Exceptionalism Arises One perspective is that we are falsely arguing over a considered topmost role. We have fallen into a mental trap that there must be one entity that sits above all others. You can argue endlessly over what or who deserves the top slot. A different way to define the situation consists of considering pluralistic exceptionalism. Generally, pluralistic exceptionalism consists of saying that different forms of intelligence and existence are all uniquely valuable. Humans aren’t the only exception. Maybe whales and dolphins deserve a form of exceptionalism. AGI and ASI deserve their exceptionalism. And so on. Dispense with the simplistic ranking schemes. Give each entity its own due. Everything is exceptional in its own way. AGI and ASI will merely be another kind of exceptionalism. Boom, drop the mic.

Human Exceptionalism Aims To Be Waylaid Due To Artificial General Intelligence And ASI

As we advance toward the development of artificial general intelligence (AGI) and artificial superintelligence (ASI), a critical question arises: What will happen to the notion of human exceptionalism? This perspective, which positions humans as the most significant beings on the planet, could be challenged as we integrate these advanced forms of AI into our lives. The implication is that, with AGI or ASI, humans may no longer hold the top position in the hierarchical structure of existence.

This exploration of AI's potential ramifications forms part of my continuing examination in Forbes, where I break down the latest advancements in the field and analyze their complexities.

### The Quest for AGI and ASI

Extensive research is underway to push the boundaries of AI capabilities, with the ultimate aim being either the achievement of AGI—a type of AI that operates on par with human intelligence—or the even more ambitious prospect of ASI. AGI would possess cognitive abilities comparable to those of humans, while ASI would surpass human intellect in nearly every conceivable way. The inherent idea is that ASI could think and reason at a level far beyond human capacity. Yet, within the AI community, opinions diverge markedly. 

On one hand, "AI doomers" predict a grim future where AGI or ASI poses an existential threat to humanity, coining the term "P(doom)" to represent the probability of such a scenario. Conversely, "AI accelerationists" maintain a more optimistic view, suggesting that advanced AI could provide solutions for humanity’s most pressing issues—curing diseases, resolving food shortages, and generating economic prosperity—ultimately working alongside humans rather than against them. The uncertainty surrounding these differing viewpoints highlights the contentious nature of our era, as neither side can definitively prove their stance.

### Redefining Human Exceptionalism

Traditionally, humanity has been viewed as the pinnacle of existence, distinguishing ourselves from all other living organisms and inanimate objects. This notion of human exceptionalism suggests that humans are elevated above other entities. As we continue to create intelligent systems, some argue that this exceptionalism is under threat. The emergence of AGI could position it alongside humanity at the top of this hierarchy, challenging the established order.

When discussing AGI, one must ask whether it would possess true sentience. If AGI becomes an intelligent entity, it could rightfully share the top position with humans. Yet, there are arguments against this notion, positing that AGI, despite its vast computational power, may lack the essence of life that characterizes human existence and might be seen as inherently lesser.

### The Supremacy of ASI

Moving to ASI, the stakes elevate significantly. This form of intelligence, beyond human cognitive abilities, is expected to render humans as insignificant in comparison—like ants facing a much larger creature. While some may argue that ASI, without a human-like spark of life, should not be seen as superior, the overwhelming capacities of superintelligence suggest this argument may falter. The mere fact that humans facilitated the creation of ASI does not diminish its exceptional nature. 

### AGI and ASI's Relationship with Supremacy

An intriguing possibility arises when considering whether AGI or ASI will desire the top ranking. It’s plausible that these advanced AIs may reject the concept of superiority altogether, asserting that all forms of intelligence hold equal value. If this scenario plays out, humans might reevaluate their own belief in exceptionalism, realizing it might have been an artificial construct all along.

### Embracing Pluralistic Exceptionalism

Lastly, rather than arguing over a single apex entity, we could explore the idea of pluralistic exceptionalism, where different forms of intelligence—humans, whales, dolphins, AGI, and ASI—are seen as unique and valuable in their own right. This approach acknowledges that each entity possesses its own type of exceptionalism, eliminating the need for simplistic hierarchies. In this evolving landscape, what emerges might be a new appreciation for the diverse intelligence that coexists rather than a struggle for dominance.

Could AGI and superintelligent AI put an end to human exceptionalism?

Dark vast chamber illuminated by flickering crimson and amber light. At its center, a cracked marble plinth holds a lone human silhouette in a classical pose. Overhead, a colossal neural-network sphere of glowing red-gold tendrils looms, its shadows stretching across the floor like claws—symbolizing superintelligent AI eclipsing human exceptionalism.

Will AGI And Superintelligent AI End Human Exceptionalism? - Forbes

- The advent of AGI and ASI challenges the notion of human exceptionalism.  
- AGI may be computational peers to humans, while ASI surpasses human intelligence entirely.  
- A pluralistic view suggests all entities possess unique exceptionalism, without a single hierarchy.  

Will AGI And Superintelligent AI End Human Exceptionalism?

Demis Hassabis, the key figure behind Google DeepMind, is set to take on the position of chief scientist at Alphabet, the parent company of Google. He will succeed Jeff Dean, who is departing to pursue opportunities at a startup.

This new role will allow Hassabis to focus intensively on influencing the trajectory of artificial general intelligence (AGI), as shared by Alphabet's CEO Sundar Pichai in a post on the Inside Google blog.

Despite this shift, Hassabis will still juggle responsibilities, as he will remain at the helm of research at Isomorphic Labs—a Google offshoot specializing in drug discovery. Although he steps down as CEO of DeepMind, he will continue his leadership as chair. The transition will see Koray Kavukcuoglu, currently the chief technology officer, take over the reins at DeepMind, reporting directly to Pichai.

DeepMind's founder takes on a leading AI position at Google.

A lone figure resembling a tech visionary climbs a grand marble staircase bathed in warm, natural sunlight. Each step is etched with glowing neural-network patterns. At the top looms a giant, vivid Google “G” logo backlit by rays streaming through a glass atrium ceiling, while subtle algorithmic motifs line the walls—symbolizing the DeepMind founder’s ascent to singular AI leadership at Google.

DeepMind founder ascends to singular AI role at Google - Computerworld

- Demis Hassabis is promoted to chief scientist at Alphabet, replacing Jeff Dean.  
- He will focus on shaping the future of artificial general intelligence (AGI).  
- Hassabis remains chair of DeepMind while leading research at Isomorphic Labs.  

Demis Hassabis, the driving force behind Google DeepMind, is ascending to the role of chief scientist at Alphabet, Google’s parent company, replacing Jeff Dean who is leaving to work at a start-up.

The role will enable Hassabis to “put his full attention on actively shaping the future of AGI,” or artificial general intelligence, Alphabet CEO Sundar Pichai wrote on the company’s Inside Google blog.

Hassabis’ attention will still be divided, however: He will continue to lead research at Google spin-off Isomorphic Labs, which works on drug discovery, and although he will no longer be CEO of DeepMind, he will be its chair. Koray Kavukcuoglu will take over DeepMind, reporting directly to Pichai. He is currently its CTO.

DeepMind founder ascends to singular AI role at Google

Artificial general intelligence (AGI) is on the horizon, or so many experts in the field of artificial intelligence claim. Defined as the capacity for an AI system to undertake any cognitive task comparable to a well-educated human, AGI transcends the narrow capabilities of current artificial intelligences, which excel only in specific tasks. AI specialists envision a technology that can adeptly manage your taxes, solve your child's math problems, provide medical diagnoses, and even craft poetry—skills that rival the best human professionals across various domains. Many believe that AGI might arrive sooner rather than later, with some predictions suggesting a timeline as soon as 2026, as anticipated by Elon Musk in 2024, and most experts agreeing it could materialize by the decade's end.

However, experts may be misguided, as they often lack a comprehensive understanding of what intelligence truly entails. To illustrate, consider an individual who has mastered the art of dribbling a basketball to an extraordinary degree. Despite his unmatched skill, he struggles with shooting, passing, and defensive play. Would any NBA team want him on their roster? Highly unlikely; his specialized talent wouldn't outweigh the liabilities on the court. 

AI operates similarly, showcasing brilliance in specific contexts but lacking broader intelligence. Recently, an AI made headlines by resolving a mathematical problem that had eluded experts for nearly a century—the Jacobian Conjecture. Additionally, AI can effectively assist with everyday challenges, offering solutions to users who are often oblivious to technical details, such as fixing vehicle malfunctions through simple queries and image inputs.

The excitement generated by AI milestones leads many to anticipate the imminent arrival of AGI, yet these impressive stories often obscure the significant limitations of AI, reflecting experts' misunderstandings of human intelligence. Many prominent figures in AI exhibit traits of social awkwardness, likely due to their extensive experience in digital environments rather than human interactions. Their focus on computer-based tasks restricts their perception of intelligence, which extends far beyond digital capabilities.

A crucial concept that proponents of AGI frequently overlook is what computer scientist Peter Denning describes as “tacit knowledge.” In his book, *Turing’s Mistake: Escaping the Yoke of Unintelligent Machines*, Denning explains that tacit knowledge includes aspects of human understanding that cannot be effectively communicated through language. Since AI systems rely on textual input for training, they inherently struggle to grasp these complexities, regardless of their sophistication.

Denning categorizes tacit knowledge in various forms, including common sense knowledge— the intuitive understanding we all possess but can’t articulate. In the 1980s, attempts such as the "Cyc" project aimed to encapsulate this common knowledge in a database, ultimately revealing the impossibility of comprehensively codifying everything known intuitively by people.

Another dimension of tacit knowledge involves emotional intelligence and perception. Consider how a close friend can discern your mood without needing verbal cues or how a charismatic individual can effortlessly command attention in a room. Similarly, conveying feelings associated with specific colors to someone who is color-blind poses challenges that language alone cannot address.

Cultural and social knowledge constitutes yet another facet of tacit understanding. Regional differences in customs and communication, like those between individuals from different parts of the United States, showcase types of knowledge that elude straightforward explanation.

This spectrum of tacit knowledge exemplifies a form of intelligence that current AI technologies cannot replicate. The conventional definition of AGI fails to encompass these essential cognitive attributes. As Pope Leo expressed in *Magnifica Humanitas*, “so-called artificial intelligences do not undergo experiences,” emphasizing that human intelligence is profoundly richer than many AI experts acknowledge.

Demis Hassabis, CEO of Google DeepMind and a leading figure in AGI research, characterized AGI as a system capable of showcasing the myriad competencies of humans. Yet, one must question whether he realistically expects an AI to perceive emotional subtleties or nuanced interpersonal dynamics like a human spouse can. Contemporary AI systems are still far from achieving this sophisticated layer of intelligence.

While some may regard these distinctions as mere semantic debates amid significant advancements in AI, the reality remains that rapid progress is expected. In the coming years, AI will indeed evolve and likely perform tasks with increased “intelligence,” however defined. Understanding the nuances of intelligence is critical in navigating this conversation, particularly for those within the Catholic community.

Rather than rejecting AI outright as unnatural or blindly accepting expert projections of impending AGI, it is vital to find a balanced approach. Dismissing AI advancements would overlook the inventive capabilities that humanity has to harness tools for progress. Conversely, endorsing an as-yet-unrealized AGI threatens to undermine the essence of human intelligence by reducing it to mere computational functions. This fall, my forthcoming book, *The Catholic Guide to Artificial Intelligence*, aims to equip readers with insights to engage thoughtfully in the evolving AI landscape.

Human intellect, designed in the image of God, encompasses far more than raw computational power or pattern recognition. It connects us to the very essence of Reality, which is God, enabling us to experience profound emotional states and personal insights that no machine can replicate. Machines cannot comprehend the horizons of love revealed with the birth of a grandchild, nor can they sympathize with human grief through personal loss or understand trust in divine wisdom during challenging times.

While advancements in AI are undoubtedly on the way, the realization of AGI—unless defined in a strictly reductive sense—is far from certain. The narrative of approaching AGI is often fueled by a blend of misinterpretation of human capabilities and marketing enthusiasm. Each AI development, regardless of scale, tends to come attached with lofty promises of transformative technology. This trend underscores a limited perception of humanity, particularly regarding our intellectual depth. The appropriate response should prioritize valuing and acknowledging the intricate and multifarious nature of human intelligence, rather than fixating solely on the technological marvels we craft.

Artificial General Intelligence is not on the horizon.

A dimly lit research lab at midnight: scattered circuit boards and open laptops glow faintly, a large blueprint labeled “AGI” pinned on a metal grid with a bold red X slashed across it, loose wires hanging idle, dust motes drifting through a single overhead lamp’s narrow beam—evoking the promise of general AI firmly shut down.

No, Artificial General Intelligence is Not Coming - Crisis Magazine

- AI experts predict imminent advancement towards artificial general intelligence (AGI) within the decade.  
- AGI aims to replicate human cognitive abilities but lacks deep understanding and tacit knowledge.  
- Claims of AGI often reflect a reductionist view of human intelligence and its complexities.  

AGI is coming, and coming soon.

At least that’s what most of the AI experts are saying. AGI—artificial general intelligence—is the ability of an AI system to perform any cognitive task that an educated human can do. As opposed to AIs that can perform one or two tasks well, an AGI would be able to do any intellectual activity as competently as the smartest human in each field. And not only do AI experts think AGI is coming soon, artificial general intelligence is the goal of all AI companies.

These companies want to create a single AI that can do your taxes, answer your kid’s math questions, diagnose your grandmother’s latest illness, and write a Shakespearian sonnet for your wife—as well as the best human accountant, tutor, doctor, and poet. Any task you throw at it, AGI could do, and do well. It could even potentially find a cure for cancer and solve intractable political problems. This holy grail of the AI field is, according to many, just around the corner. Back in 2024, Elon Musk predicted that we’d have AGI by 2026, and most expert estimates place the advent of AGI before the end of this decade.

As opposed to AIs that can perform one or two tasks well, an AGI would be able to do any intellectual activity as competently as the smartest human in each field.

Tweet This

The problem is that they’re all wrong, and that’s because these experts in artificial intelligence have no idea what intelligence actually is.

Imagine a man who becomes an expert in dribbling a basketball. Not only an expert, but the world’s greatest dribbler. He can do things with a ball that no one else can do, and even great basketball players are amazed at his abilities. Yet, this same man can’t shoot, can’t pass, and can’t play defense. He also can’t run while dribbling. Do you think any NBA team would find a place on their squad for this one-talent wonder? Of course not; he’d be a huge liability if he were put on the court with professional basketball players.

AI is much like this expert dribbler. It’s amazing in a thin slice of what we call intelligence. Recently, for example, an AI solved a mathematical problem that had stumped experts for almost 100 years (it showed that the Jacobian Conjecture is false). AI is also incredible at solving more mundane, everyday problems (I recently fixed a problem on my car, in spite of my complete ignorance of car mechanics, simply by taking some photos and asking an AI for help).

When we hear stories of AI achievements that seem ripped out of the pages of futuristic science-fiction stories, we’re primed to think that AGI must be coming soon. Yet these headline-grabbing stories mask the huge deficiencies of AI, or, more precisely, they mask AI experts’ huge deficiency in understanding human intelligence.

It’s no coincidence that many prominent AI experts are, how shall we say…socially awkward. (Listen to Mark Zuckerberg or Elon Musk talk for a few minutes for confirmation of this claim.) They have spent their lives in the digital realm, from coding to playing video games to posting on social media. Interacting with a computer is their primary activity, so they naturally limit their thinking about human intelligence to computer-based experiences. Yet human intelligence encompasses so much more than tasks that can be done on a computer.

Human intelligence encompasses so much more than tasks that can be done on a computer.

Tweet This

What many promoters of imminent AGI miss is what computer scientist Peter Denning calls “tacit knowledge” in his new book, Turing’s Mistake: Escaping the Yoke of Unintelligent Machines. Tacit knowledge encompasses forms of human intelligence that can’t be explained with language. And since modern AI systems require words in order to be trained, these forms of knowledge simply can’t be grasped by AI, no matter how advanced it may be.

Denning breaks down tacit knowledge into various categories, one of which is common sense knowledge—that practical knowledge we all have that’s never explicitly taught or explicated, but is known by us nonetheless. In the 1980’s there was a project (called “Cyc”) to add all this common knowledge directly into a database. Even after over a million examples were added (such as “You can pull an object with a string, but you can’t push it with a string”), it became obvious that it was impossible to add everything that most people know intuitively.

Another type of tacit knowledge is comprised of our feelings and perceptions. Think of your female friend who knows your mood before you even say a word. Or the charismatic man who dominates any room he enters—everyone feels his presence immediately. Or try to explain your feelings about the color “blue” to a color-blind person.

Social knowledge, including culture, is another form of tacit knowledge. People who grew up in the South know certain ways of saying and doing things that those who grew up in the Midwest don’t, and they can’t explain all those things; they just know them.

Tacit knowledge, in all its types, is a form of intelligence that’s far beyond the capabilities of modern AI, and the accepted definition of AGI doesn’t even begin to include these intellectual talents.

This is what Pope Leo meant when he wrote in Magnifica Humanitas, “So-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships and do not know from within what love, work, friendship or responsibility mean.” Human intelligence is so much more than AI experts believe it to be.

Google DeepMind CEO Demis Hassabis, is perhaps the world’s foremost authority on AGI. He described AGI last year as “a system that’s able to exhibit all the complicated capabilities that humans can,” and predicted it would happen within five to ten years (and he’s actually more pessimistic than many AI leaders). Clearly Hassabis has the common two-dimensional definition of intelligence undergirding his prediction.

Does he really think an AI, in 2030 or even 2035, will be able to see the look on someone’s face and immediately know that person’s exact mood, precisely why they’re in that mood, and exactly what to say, like even an average wife can do with her husband? The most advanced AI today is light-years from that type of intelligence.

Tacit knowledge, in all its types, is a form of intelligence that’s far beyond the capabilities of modern AI, and the accepted definition of AGI doesn’t even begin to include these intellectual talents.

Tweet This

Are these just semantic arguments, given the major AI advancements likely coming around the corner? It’s true that AI is still advancing rapidly, and in five to ten years will be far more “intelligent”—however you define the term—than it is today. Yet a proper understanding of intelligence is vital in this debate, and it’s one that Catholics need to grasp more fully.

While some Catholics want to hole up in their homes and attack all AI as unnatural, we can’t do that—it’s, ironically, a rejection of human intelligence and its ability to create tools that help us subdue the earth. Nor can we simply embrace the anthropologic errors of the AI experts by going along with their hype. By pretending that AI will soon meet, or even surpass, human intelligence, we’d diminish what it means to be human, reducing it to tasks on a computer. (Self-promotion time: This fall, my next book, The Catholic Guide to Artificial Intelligence, will provide Catholics with the tools to better understand and grapple with the AI revolution.)

We are fearfully and wonderfully made in the image and likeness of God, and so our intellects are, quite literally, touched by the divine. They are more than processing machines or pattern-recognition programs. They are a connection to Reality itself, which is God. That’s why they’re able to do things beyond the reach of science. No machine, no matter how well designed or programmed, can ever do that. No machine can know what new horizon of love is opened up when your grandchild is born. No machine can respond to someone’s grief from its own experience of loss. No machine can understand trusting God in times of darkness.

Are major advancements in AI coming? Absolutely. But is AGI coming? Not unless you redefine intelligence in such a reductionist fashion that it excludes significant aspects of intellectual ability.

The claims of imminent AGI are part bad anthropology, part marketing hype. Every advancement in the AI world, no matter how small, is met with promises of world-changing technology coming soon, since that makes for a good press release. But it also reveals a view of the human person—particularly the human intellect—that is greatly impoverished. The Catholic response should be not so much denigrating the impressive breakthroughs in human-created AI, but rather elevating and appreciating the abilities of our amazing and multi-faceted human intelligence.

No, Artificial General Intelligence is Not Coming

You are currently seeing this message because the website's administrator has implemented Anubis to guard the server against the rampant issue of AI companies scraping content indiscriminately. This scraping can lead to downtime, rendering the site’s resources temporarily inaccessible to users.

Anubis serves as a middle-ground solution, employing a Proof-of-Work model similar to Hashcash, which aims to mitigate email spam. The concept is that while the extra load may seem trivial on smaller scales, it accumulates significantly when faced with widespread scraping attempts, making those efforts much more costly for those engaged in it.

In the grand scheme, this solution is intended to buy time for the development of more sophisticated methods to differentiate between genuine users and headless browsers. This differentiation could involve analyzing features like font rendering techniques to minimize the need for displaying the challenge proof-of-work page to what are likely legitimate visitors.

It’s important to highlight that Anubis requires the use of advanced JavaScript functionalities, which may be hindered by plugins such as JShelter. To ensure a smooth experience on this website, please consider disabling JShelter or any similar extensions.

Ensuring you're not a robot!

A sunlit modern workspace: on a wooden desk a laptop screen fills with a giant “I’m not a robot” checkbox being clicked by a human hand. To one side, a shadowy metallic robot silhouette is held back by a translucent wall of ticked boxes and floating user-data icons—fingerprints, iris scans—drifting like confetti. Warm morning light streams through a window, casting soft shadows that celebrate human verification over automation.

Artificial Intelligence: the hype, the dangers, and the resistance—Part I - mronline.org

- Anubis protects websites by implementing a Proof-of-Work scheme to deter scrapers.  
- The system aims to make web scraping prohibitively expensive at high volumes.  
- Future plans include identifying headless browsers to improve user accessibility and experience.

Loading...

You are seeing this because the administrator of this website has set up Anubis to protect the server against the scourge of AI companies aggressively scraping websites. This can and does cause downtime for the websites, which makes their resources inaccessible for everyone.

Anubis is a compromise. Anubis uses a Proof-of-Work scheme in the vein of Hashcash, a proposed proof-of-work scheme for reducing email spam. The idea is that at individual scales the additional load is ignorable, but at mass scraper levels it adds up and makes scraping much more expensive.

Ultimately, this is a placeholder solution so that more time can be spent on fingerprinting and identifying headless browsers (EG: via how they do font rendering) so that the challenge proof of work page doesn't need to be presented to users that are much more likely to be legitimate.

Please note that Anubis requires the use of modern JavaScript features that plugins like JShelter will disable. Please disable JShelter or other such plugins for this domain.

Making sure you're not a bot!

In the ongoing discussions around artificial intelligence in 2026, the term "artificial general intelligence" (AGI) is often misused by executives, journalists, and researchers. While top executives may refer to AGI when announcing new AI models and journalists may use it to describe advancements post-ChatGPT, researchers maintain that AGI is a defined, yet unachieved standard. This understanding is crucial for distinguishing between genuine advancements in AI and mere marketing language.

Researchers across leading institutions share a consistent definition of AGI. Google Cloud defines it as a theoretical machine intelligence capable of performing any intellectual task that humans can, designed to replicate the cognitive flexibility of the human brain. Similarly, Stanford's Institute for Human-Centered Artificial Intelligence (HAI) views AGI as a system that can learn, reason, and apply knowledge across diverse tasks, adapting to unfamiliar challenges instead of simply excelling in a pre-defined role. Databricks introduces a technical clarification, noting that AGI embodies a broad, transferable intelligence that does not rely on specific programming for each task, which contrasts with most AI systems today, which thrive through narrow specialization.

The confusion often arises when distinguishing surface-level capabilities from actual cross-domain reasoning. A model that generates perfect code or passes legal exams may seem versatile, but researchers emphasize that it remains a specialized AI without true causal reasoning or the integrated intelligence characteristic of AGI. Google Cloud categorizes existing AI, including large language models, as artificial narrow intelligence (ANI), designed for targeted tasks such as language processing or image identification. Above ANI is AGI, and the hypothetical concept of artificial superintelligence (ASI) stands above AGI.

Stanford's HAI specifically calls attention to the controversial nature of AGI due to the lack of a universally recognized test. This ambiguity complicates claims regarding achieving or nearing AGI, as different researchers interpret “human-level intelligence” through various lenses—focusing on reasoning, autonomous learning, or even self-awareness—which many consider a separate and difficult-to-address matter. Databricks explicitly states that AGI does not yet exist, explaining that no AI system currently exhibits the full spectrum of human-like intelligence, as even sophisticated models like ChatGPT lack true autonomy and cross-domain comprehension.

To tackle the existing challenge of quantifying AGI, a comprehensive 2025 paper co-authored by several esteemed researchers, including Dan Hendrycks and Yoshua Bengio, seeks to establish a measurable framework. The paper characterizes AGI as matching the cognitive capabilities of a well-educated adult and develops its evaluation methodology based on the Cattell-Horn-Carroll theory, a highly regarded model of human cognition. They break down general intelligence into ten fundamental cognitive areas, such as reasoning and perception, and adapt established psychometric tests to evaluate AI systems accordingly.

The outcomes of this research expose a distinct cognitive profile of contemporary AI models—showing strengths in specific knowledge-rich areas but revealing critical shortcomings in foundational capabilities, particularly memory. Utilizing this framework, GPT-4 received an AGI score of 27%, while GPT-5 scored 57%, reflecting notable progress but a significant distance from achieving true AGI. The disparity in performance highlights how singular benchmarks or exemplary demonstrations can misrepresent overall capabilities.

Researchers frequently identify four defining characteristics necessary for a system to be considered on the path to AGI. First is generalization, which refers to the ability to apply knowledge across different domains. Second is autonomous, continuous learning—an AGI system would learn from experiences without needing new datasets each time it encounters a challenge. Third is common sense and contextual reasoning, encompassing a broad base of knowledge that allows decision-making akin to human thought processes. Finally, autonomous decision-making involves real-time decision-making based on input data, a trait that integrates the preceding capabilities.

The future discussions surrounding AGI resonate with ambition yet are grounded in skepticism. While Lenovo emphasizes the transformative potential of AGI in various sectors, it also acknowledges risks such as cybersecurity threats and unpredictable learning behaviors. Databricks underscores the necessary alignment between AGI objectives and human values, highlighting a shared concern regarding existential risks. The ResearchGate review illustrates the urgency for robust governance structures, noting the potential for initial regulations to become outdated swiftly as AGI evolves.

Despite differing focuses, all sources exhibit a common wariness towards premature declarations of AGI advancements. Databricks asserts that no current AI system meets the criteria for AGI, regardless of its apparent capabilities. Similarly, Stanford's caution regarding the absence of a definitive assessment method reinforces the notion that organizations must critically evaluate claims of AGI, examining the specific capabilities being assessed and the methods of evaluation employed.

For businesses and policymakers navigating the AI landscape in 2026 and beyond, the practical insight is to prioritize understanding the ongoing development of AI rather than expecting immediate jumps towards AGI. The uneven progress across various cognitive areas suggests future advancements will manifest in unpredictable ways, necessitating thorough tracking based on established cognitive science principles rather than relying solely on benchmark results. This approach will distinguish well-informed strategies concerning AGI from speculative claims in the years to come.

2026 AI Trends: Understanding Artificial General Intelligence (AGI)

A sunlit futuristic city plaza at dawn, natural golden light illuminating a towering polished-chrome humanoid statue whose head is woven from glowing neural filaments, slender data cables fanning out from its base into surrounding glass research towers, soft sunbeams reflecting off its surface and casting warm, hopeful shadows across the empty square.

AI Trends in 2026: What Is Artificial General Intelligence (AGI)? - AI Insider

- There is no universally accepted test for artificial general intelligence (AGI) today.  
- AGI is defined as machine intelligence that mimics human cognitive flexibility across tasks.  
- Current AI models like ChatGPT are specialized and lack true autonomous, cross-domain reasoning.  

Few phrases get thrown around 2026’s AI discourse as loosely as “artificial general intelligence.” Executives use it to describe an impressive new model launch. Journalists use it as shorthand for whatever comes after ChatGPT. Researchers, by contrast, treat it as a precise and still unmet threshold, one that separates the specialized systems already reshaping business from a category of machine intelligence nobody has actually built. Understanding what AGI really means, how researchers propose to measure it, and how far current systems sit from the mark has become essential context for anyone trying to separate genuine progress from marketing language this year.

THE CORE DEFINITION RESEARCHERS AGREE ON

Despite the disagreement over timelines and testing methods, the major research institutions converge on a similar core definition. Google Cloud describes AGI as a hypothetical machine intelligence with the ability to understand or learn any intellectual task a human being can, built to mimic the cognitive flexibility of the human brain rather than excel at one narrow function. Stanford’s Institute for Human-Centered Artificial Intelligence (HAI) frames it almost identically, defining AGI as a system with general, human-level or beyond ability to learn, reason, and apply knowledge across a wide range of tasks and domains, one that could conceivably handle novel situations rather than simply perform well on a single job it was built for. Databricks adds a useful technical distinction: AGI systems would possess broad, flexible and transferable intelligence that does not require task-specific programming, in contrast to nearly all AI in use today, which achieves strong results through specialization and pattern recognition rather than integrated reasoning.

That last point is where most confusion starts. A model that writes flawless code, drafts a legal brief, and passes a bar exam question can look general on the surface. But researchers distinguish breadth of surface capability from genuine cross-domain reasoning. Generative AI models, Databricks notes, remain forms of specialized AI despite their broad and fluent outputs, because they lack true causal reasoning and the integrated, autonomous intelligence that defines AGI. Google Cloud’s typology places today’s systems, including large language models, firmly in the artificial narrow intelligence (ANI) category, the most common form of AI in existence, built for specific tasks such as image recognition or natural language processing. AGI sits above that tier, and artificial superintelligence (ASI), a system that would surpass human capability entirely, sits above AGI as a still more distant and speculative concept.

WHY STANFORD CALLS THE TERM CONTROVERSIAL

Stanford HAI is notably blunt about a problem that runs through nearly all AGI research: there is no universally accepted test for it, which makes any claim about achieving or approaching AGI difficult to verify. Different researchers mean different things by “human-level intelligence.” Some emphasize reasoning and problem-solving, others emphasize autonomous learning without human-labeled data, and others still fold in questions of self-awareness that most practical researchers consider a separate and largely unanswerable question. Databricks’ own FAQ addresses this directly, stating plainly that AGI does not exist today, that no system demonstrates the full set of human capabilities associated with general intelligence, and that even a fluent, widely used tool like ChatGPT is a specialized language model rather than evidence of general intelligence, since it lacks true autonomy and cross-domain understanding.

A NEW ATTEMPT TO PUT A NUMBER ON THE GAP

For years, the absence of a rigorous test meant AGI discussions stayed largely qualitative. A significant 2025 paper on arXiv, co-authored by a large group of researchers including Dan Hendrycks, Yoshua Bengio, Max Tegmark, and Gary Marcus, tries to close that gap with a quantifiable framework. The paper defines AGI as matching the cognitive versatility and proficiency of a well-educated adult, and grounds its methodology in Cattell-Horn-Carroll theory, widely considered the most empirically validated model of human cognition. The researchers break general intelligence into ten core cognitive domains, including reasoning, memory, and perception, and adapt established human psychometric test batteries to score AI systems against each one.

The results are striking less for what they reveal about any single model than for the shape of the profile they expose. Applying the framework to contemporary systems reveals what the authors call a highly jagged cognitive profile: models that perform strongly in knowledge-intensive domains while showing critical deficits in more foundational cognitive machinery, particularly long-term memory storage. The paper’s headline numbers put GPT-4 at an AGI score of 27% and GPT-5 at 57%, a jump the authors present as evidence of both rapid progress and a still-substantial gap before anything resembling true AGI is reached. That jaggedness, strength in some domains alongside conspicuous weakness in others, is precisely why single benchmark scores or impressive demos can be misleading indicators of general capability on their own.

FOUR TRAITS THAT SEPARATE AGI FROM WHAT EXISTS TODAY

Pulling together the research literature, four attributes recur as the defining traits researchers look for when assessing whether a system is approaching AGI rather than simply performing well within a narrow lane.

Generalization is the first and most cited. Google Cloud describes this as the ability to transfer knowledge and skills learned in one domain to another, enabling adaptation to new and unseen situations rather than failure outside a trained lane. A comprehensive review of the AGI research literature, published via ResearchGate, frames this same idea as versatility and adaptability, the capacity to perform any intellectual task a human can rather than a fixed set of tasks defined in advance.

Autonomous, continuous learning is the second. Databricks emphasizes that an AGI system would acquire new skills and update its knowledge through experience, without requiring new labeled datasets and retraining for every new challenge the way specialized systems do today. Lenovo’s knowledge base similarly lists adaptive learning among AGI’s defining workloads, noting that this capability would let a system stay relevant in fast-changing domains such as cybersecurity by learning from past experience rather than waiting for a scheduled retraining cycle.

Common sense and contextual reasoning form the third trait. Google Cloud describes this as a vast repository of world knowledge, including facts, relationships, and social norms, that lets a system reason and make decisions the way a person would rather than pattern-match against training examples. This is also where current systems show their clearest weaknesses, since statistical correlation, however sophisticated, is not the same as the grounded causal understanding researchers associate with genuine comprehension.

Autonomous decision-making rounds out the list. Lenovo highlights this trait through examples like autonomous vehicles processing real-time sensor data, predicting traffic patterns, and making split-second decisions without a human directing each step, a capability that depends on combining the previous three traits rather than any one of them in isolation.

WHERE THE DEBATE GOES NEXT

None of the sources treat AGI as a settled or even imminent milestone, but they diverge somewhat on emphasis. Lenovo’s overview leans toward AGI’s transformative potential across finance, transportation, and education, while cataloguing drawbacks including high development costs, cybersecurity vulnerabilities, and the unpredictability that comes with a system able to learn and adapt in ways its designers did not fully anticipate. Databricks devotes considerable attention to the alignment and control problem, the challenge of ensuring an AGI system pursues goals consistent with human values, and treats existential risk as a real but contested position among researchers rather than a settled consensus. The ResearchGate review frames the stakes in governance terms, arguing that any rules written for AGI’s early days may become obsolete almost as soon as the system evolves past them, and that effective oversight will need to be built in from the start rather than retrofitted once a capable system already exists.

What unites all of it is a shared skepticism toward premature claims. Databricks is explicit that no existing AI system, however capable it appears in a demo or on a leaderboard, has yet demonstrated the broad, flexible, and autonomous intelligence that would qualify as AGI under any of these frameworks. Stanford’s caution about the absence of a universal test reinforces the same point from a different angle: because there is no agreed-upon way to verify a claim of AGI, the responsible default for any organization evaluating a vendor’s roadmap or a competitor’s announcement is to ask what capability, specifically, is being measured, and by what test.

WHAT THIS MEANS FOR 2026 AND BEYOND

For businesses and policymakers navigating 2026’s AI landscape, the practical takeaway is less about predicting when AGI will arrive and more about resisting the temptation to treat every capable model as a step function toward it. The jagged cognitive profiles researchers are already documenting suggest progress will likely continue arriving unevenly, with dramatic gains in some domains and stubborn gaps in others, rather than a single clean threshold crossing. Tracking that progress rigorously, using frameworks grounded in established human cognitive science rather than benchmark scores alone, is what will separate informed AGI strategy from speculation in the years ahead.

References and Further Reading

AI Trends in 2026: What Is Artificial General Intelligence (AGI)?

Artificial intelligence operates through a structured mathematical framework that allows it to identify patterns and make predictions based on input data. This framework comprises both the model's architecture and its internal numerical values, known as parameters, which are fine-tuned during the training process. Various AI models exist for distinct tasks, including image recognition, natural language processing, and speech synthesis. Among these, large language models (LLMs) have become particularly prominent in recent years. After training, these models can respond to new inputs by leveraging the patterns they have absorbed.

An algorithm refers to a series of defined steps or computational methods for solving problems, akin to a recipe that transforms an input into an output. In the realm of AI, this often pertains to "machine learning algorithms" that extract patterns from data or the specific computational procedures used by trained models during their calculations. Different algorithms can yield varying results in terms of accuracy, processing speed, and resource demands, even when working with identical datasets.

Neural networks represent a core structure of AI, inspired by the connections between neurons in the human brain. This architecture includes multiple layers, where learning occurs by adjusting the "weights" that link nodes within these layers. Data enters through the input layer, features are processed in hidden layers, and final outcomes emerge at the output layer. The term "deep learning" specifically refers to extensive architectures consisting of many stacked layers, which has been instrumental in driving the recent surge in AI advancements.

Large language models are advanced AI systems trained on extensive text corpuses, enabling them to perform numerous language-related tasks, including text generation, summarization, translation, and responsive dialogue. While their foundational mechanism relies on predicting the next unit of text, their ability to generate coherent and human-like language has improved significantly, thanks to increased parameters and a broader training dataset. Prominent conversational AI services, such as ChatGPT and Claude, utilize these models.

Foundation models, like LLMs, are built on large and diverse datasets, making them adaptable for various applications across different domains involving text, images, and audio. This versatility minimizes the need to construct models from the ground up for each individual task. By performing additional targeted training on a single foundation model, developers can efficiently create specialized models for applications like translation and image synthesis.

The Graphics Processing Unit (GPU) plays a crucial role in the AI landscape, designed for rapid rendering of graphics and optimal for high-volume parallel computations essential for AI training and inference. Given the vast matrix calculations involved in training neural networks, GPUs have become the preferred computational framework in this field. Nvidia's chips, originally developed for gaming graphics, dominate this sector.

Machine learning encompasses technologies that allow AI systems to autonomously learn from large datasets without the need for explicit instructional guidelines from humans. Unlike traditional programming, where all conditions and procedures must be predetermined, machine learning enables rule extraction directly from the data. It is typically categorized into supervised learning, where AI learns from labeled datasets; unsupervised learning, where patterns are identified independently; and reinforcement learning, where AI derives knowledge through trial and error.

Deep learning is a subset of machine learning that focuses on enhancing a model's ability to recognize intricate patterns through multiple layers of neural networks. Depending on the application, these networks can be composed of numerous layers, enabling a graduated understanding of input data. The advancements in deep learning and the enhanced capabilities of GPUs have driven remarkable improvements in tasks such as image and speech recognition since the 2010s.

Supervised learning involves training an AI model on a dataset of inputs and their corresponding correct answers, allowing it to recognize patterns by comparing its predictions with these reference labels. The success of this method hinges on the accuracy and quality of the input data. For instance, feeding an AI numerous pairs of images labeled as “cat” enables it to determine if new images depict cats. This approach is prevalent in classification and regression tasks.

On the other hand, unsupervised learning does not rely on labeled data, instead allowing the AI to identify patterns and structures within the data itself. This method can unveil hidden similarities and groupings that might not be immediately apparent to humans. Techniques like clustering and dimension reduction stem from this approach, which is particularly useful when dealing with substantial amounts of unstructured data, even though the outcomes can sometimes be hard to interpret.

Annotation is the crucial step of labeling training data for AI systems. Since inaccuracies or biases in this phase can severely affect a model's performance and fairness, it is considered a vital part of the AI development process. Examples include labeling objects in images or categorizing sentiments in text. Despite being an essential task, annotation is often labor-intensive and costly. Recently, efforts have been initiated to utilize AI to enhance the efficiency of this annotation process.

Reinforcement learning allows AI to optimize its actions based on a reward system through continuous trial and error. In this method, rather than providing correct answers, the AI learns which actions yield the greatest rewards, refining its approach through interaction with its environment. Notably, this technique has been used in the development of complex systems like the “AlphaGo” AI and in tuning responses for conversational AI through reinforcement learning with human feedback (RLHF).

Distillation is a technique that focuses on transferring knowledge from a comprehensive, high-performing “teacher model” to a smaller, more efficient “student model.” This process emphasizes the extraction of essential knowledge from the larger model, training the student model to replicate its output patterns. The advantage here lies in producing lightweight models that require less computational power while still maintaining decent accuracy. This becomes particularly advantageous when deploying models on devices with limited resources, such as smartphones.

Finally, the transformer is a pioneering neural network architecture that has become the backbone of large-scale language models since its introduction by Google researchers in 2017. By employing an "attention" mechanism, transformers accurately capture relationships between distant tokens in a given context, facilitating efficient learning across words within sentences. Presented in the influential paper “Attention Is All You Need,” this architecture revolutionized how words are processed, leading to large-scale parallel training and significantly contributing to the explosion in LLM capabilities.

Parameters are the numerical values within an AI model that are refined throughout the training process. These correspond to the “weights” and “biases” necessary for neural network performance, and their adjustment improves the model's effectiveness. The size of a model is often indicated by the number of parameters it possesses, which can range from billions to trillions. Generally, a higher parameter count enhances a model's expressive capabilities, though it comes with increased computational demands, and the overall performance relies on the model’s design, the training data used, and the training methodologies applied.

In AI terminology, a token refers to the segment into which text is divided for processing. This can include entire words, portions of words, and punctuation. Large language models generate text by segmenting it into sequences of tokens and subsequently predicting the next token based on statistical likelihoods. Costs associated with API usage and the maximum text length that can be handled (known as the context window) are often determined in relation to the number of tokens processed.

33 Essential AI Terms for 2026 Explained in Simple Language

A sunlit modern study room bathed in warm natural light, where a mid-career professional in business casual sits at a polished wooden desk. Scattered around them are neatly arranged index cards each labeled with an AI term—“Federated Learning,” “Neuro-Symbolic AI,” “Quantum AI,” and more up to 33 in total. In front of them lies a large open book titled “33 AI Terms You Need to Know in 2026,” its pages softly illuminated by sunlight streaming through a nearby window. The person leans forward, highlighter poised over one term, surrounded by gentle shadows and a calm, inviting atmosphere that speaks of discovery and clarity.

33 AI Terms You Need to Know in 2026, Explained Simply - WIRED Middle East

- AI models learn patterns from data, enabling predictions and generating outputs through training.  
- Neural networks, inspired by human brain connections, are fundamental to AI structure and learning.  
- Large language models, trained on vast text data, perform diverse language tasks effectively.  

[Fundamentals of AI]

01｜AI Models

A mathematical framework that enables AI to learn patterns and regularities based on input data and to make predictions, judgments and generate output. It consists of the model’s structure and a combination of numerical values, called parameters, that are adjusted through training. There are different AI models depending on the purpose, which includes image recognition, natural language processing and speech synthesis. Large language models (LLMs), which have garnered significant attention in recent years, are also a type of AI model. Once training is complete, an AI model becomes capable of responding to new inputs based on the patterns it has learned.

02｜Algorithm

A set of procedures or computational methods for solving a problem, defined as clear, specific steps. Think of it as a sequence of processes – much like a cooking recipe – that takes an input and produces an output. In AI, the term often refers to “machine learning algorithms” that learn patterns from data, or to the processing procedures followed by a trained model when performing calculations. Even when using the same data, the accuracy, speed and computational resources required can vary significantly depending on the algorithm used.

03｜Neural Networks

This is one of AI’s most fundamental structures, inspired by the connections between nerve cells (neurons) in the human brain. It consists of multiple “layers”; learning proceeds by adjusting the “weights” that connect the nodes within each layer. Data is received at the input layer, features are extracted at the hidden layers and results are derived at the output layer. “Deep learning” refers to this structure stacked in multiple layers, and it has become the foundational technology underpinning the current AI boom.

04｜Large Language Model (LLM)

LLMs are AI models that are capable of learning from vast amounts of text data. This allows them to perform a wide range of language tasks, such as generating text, summarising, translating and responding to questions. While fundamentally based on a mechanism that probabilistically predicts the next token (a unit into which text is divided), LLMs have become capable of generating natural, human-like text by increasing the number of parameters and expanding the training data. Many conversational AI services, such as ChatGPT and Claude, are based on these models.

05｜Foundation Model

Large language models are one of the prime examples of foundation models. By being pretrained on massive and diverse datasets, foundation models are versatile enough to be applied to a wide range of tasks. They serve as the foundation for a broad spectrum of applications across fields such as text, images and audio. Since some models can handle multiple types of data simultaneously, they have reduced the need to rebuild models from scratch for each specific application. By additionally training (fine-tuning) to a single foundation model, it is possible to efficiently build models for specific purposes, such as translation, summarising and image generation.

06｜GPU

GPU or the Graphics Processing Unit refers to a semiconductor chip designed to perform high-speed rendering of images and video. Its ability to handle large-scale parallel computations makes it well suited for AI training and inference. Currently, it is also one of the primary computational infrastructures supporting AI development. Since training neural networks requires massive amounts of matrix calculations, GPUs – which offer superior parallel processing performance compared to CPUs – are widely used. Nvidia chips, originally used for tasks such as rendering in video games, hold a large market share in the AI sector.

> Back to Index

[How to Learn AI]

07｜Machine Learning

This is a general term for technologies in which AI automatically learns patterns and regularities from large amounts of data, rather than humans providing explicit rules. While traditional programmes required humans to meticulously define conditions and processing steps, machine learning differs significantly in that it derives rules directly from the data itself. Depending on the learning method, it is classified into the following: “supervised learning”, which learns based on correct answer data; “unsupervised learning”, which discovers data structures on its own; and “reinforcement learning”, which learns through trial and error.

08｜Deep Learning

One of the methods of machine learning, deep learning is a technique that enables the AI model to learn more complex patterns by stacking multiple layers of neural networks. Depending on the model, the number of layers can range from tens to hundreds; stacking these layers allows the model to capture the features of input data step by step. The dramatic improvement in the accuracy of image and speech recognition since the 2010s is largely attributable to the development of deep learning and the increased computational power of GPUs.

09｜Supervised Learning

A method in which an AI learns patterns using datasets consisting of input data paired with corresponding correct answers (labels). While this allows the AI to learn by comparing its own predictions with human-prepared “examples”, accuracy depends on the quality of the labels and the data. For example, by training the AI on a large number of pairs consisting of “cat images” and the label “cat”, it becomes capable of determining whether an unknown image depicts a cat. This method is widely used for tasks such as classification and regression, and is one of the fundamental techniques that has long been employed in machine learning.

10｜Unsupervised Learning

A learning method in which no correct labels are provided, and the AI is left to discover the structures and patterns inherent in the data itself. Since it eliminates the need for manual labelling, it holds the potential to uncover unexpected commonalities and groupings that humans have not yet noticed. It is used in techniques such as “clustering”, which groups data with similar properties, and “dimension reduction”, which compresses and represents data features. While the results can sometimes be difficult to interpret, this method is widely utilised when dealing with large volumes of unstructured data.

11｜Annotation

This refers to the process of assigning correct labels or annotations to the data used to train AI. Since errors or biases in labelling significantly impact a model’s accuracy and fairness, this task – though often seen as mundane – is considered a critical process that underpins the entire AI development cycle. For example, this includes tasks such as drawing boxes around objects in images and labelling them as “dog” or “car”, or classifying the sentiment in text. However, much of this work is done manually, making it costly and time-consuming. In recent years, efforts have been underway to use AI to streamline the annotation process itself.

12｜Reinforcement Learning

This is a method in which an AI learns to maximise the “reward” for its actions through repeated trial and error. Rather than having a human provide the correct answer, the AI discovers the optimal course of action on its own by receiving high rewards for good actions and low rewards for bad actions. A key feature is that there is no need to prepare correct answer data in advance; instead, the AI refines its strategy through interaction with its environment. It is applied to systems such as the Go AI “AlphaGo” robot control, and response tuning for conversational AI (reinforcement learning with human feedback, or RLHF).

13｜Distillation

A technique for transferring knowledge from a large-scale, high-precision “teacher model” to a smaller, lighter “student model”. As the name suggests, this involves extracting and passing on only the essence of the knowledge held by a large model. By training the student model to mimic the output patterns of the teacher model, this technique offers the advantage of creating lightweight models with low computational cost while minimising performance degradation. It is a particularly valuable technology in scenarios where models must run on limited computational resources, such as on smartphones.

> Back to Index

[Model Structure]

14｜Transformer

This neural network architecture, which forms the foundation of today’s large-scale language models, was introduced by Google researchers in 2017. By using a mechanism called “attention” to capture the relationships between tokens – such as words in a sentence – it can efficiently learn the relationships between words that are far apart. Put forward in a paper presented at the time, “Attention Is All You Need”, the breakthrough lay in the fact that it demonstrated high performance with a structure centred on the attention mechanism rather than relying on the previously mainstream method of processing words sequentially. The emergence of this structure enabled large-scale parallel training and laid the foundation for the current LLM boom.

15｜Parameters

These are numerical values within an AI model that are adjusted through training. They correspond to “weights” and “biases” in neural networks and are gradually optimised through repeated processing of training data. The number of parameters is often used as an indicator of a model’s scale and is expressed as “X billion parameters”. Scales range from tens of billions to trillions; generally, a higher number of parameters can enhance a model’s expressive power, but this also increases computational costs. Furthermore, actual performance depends on the model’s architecture, training data and training methods.

16｜Token

This is a unit of division for character strings when AI processes text. And it isn’t just entire words that are treated as a single token, but also parts of words and punctuation as well. Large-scale language models generate text by breaking it down into a sequence of tokens and then probabilistically predicting the next token. API usage fees and the maximum length of text that can be processed at one time (context window) are also often determined based on the number of tokens.

33 AI Terms You Need to Know in 2026, Explained Simply

Every few weeks, a new AI development emerges that would have seemed unimaginable just a couple of years ago.

One model has outperformed some of the brightest young mathematicians globally. Another aids in scientific exploration. Meanwhile, there are models that can write software that would typically take experts days or even weeks to complete.

Today, there are also AI systems capable of operating independently for extended periods before needing human assistance.

While I’m not declaring that we’ve achieved artificial general intelligence (AGI) based solely on these individual advancements, it’s essential to stop viewing them in isolation.

By examining these breakthroughs together, we can better understand how close we are to reaching AGI.

Beyond Simple Interactions

Earlier this year, I posited that the fundamental components of AGI were finally aligning.

Initially, AI began learning to answer questions. Following that, it developed the ability to reason through intricate problems. Recently, we’ve seen AI manage to work autonomously for much longer durations.

Reflecting on the past six months, I now believe that we’ve witnessed the addition of several pivotal advancements.

Let’s consider mathematics, for instance.

Last summer, an advanced variant of Google’s Gemini Deep Think secured a gold medal score at the International Mathematical Olympiad—an event often regarded as one of the most challenging math competitions for high school students.

Typically, only around 8% of participants achieve this level of recognition in an average year.

The AI managed to solve five out of six problems, earning 35 points out of a potential 42, graded by the same judges who assess human contenders.

This was a remarkable feat.

Moreover, Google has indicated that newer iterations of Deep Think are now assisting mathematicians and scientists in addressing real research challenges in mathematics, physics, and computer science.

In an internal assessment, the system achieved around 90% accuracy on advanced mathematical proofs.

This, to me, is a significant indicator of the AGI timeline.

Solving known problems is one aspect, but aiding in unexplored challenges brings us much closer to true general intelligence.

Anthropic recently showcased a similar significant achievement.

Their research teams tasked Claude AI agents with an open-ended machine-learning task devoid of any pre-established answers.

The agents were required to generate original ideas, write code, conduct experiments, analyze their findings, and refine their methods until they reached a viable solution.

Anthropic reported that the top-performing AI team surpassed seasoned human researchers working on the same challenge.

This exemplifies the essence of general intelligence: addressing unknown issues, experimenting, and evolving one’s approach.

We’re observing similar trends in the realm of software engineering.

Anthropic tasked Claude with developing a new C compiler capable of compiling the Linux kernel, one of the most intricate open-source software initiatives ever undertaken.

Across two weeks, Claude engaged in nearly 2,000 coding sessions, processed around 2 billion input tokens, generated 140 million output tokens, and ultimately crafted about 100,000 lines of code.

And it succeeded.

Claude produced software capable of running one of the world’s most complex operating systems across various computer processors.

Even more astonishing is the fact that this endeavor cost less than $20,000 in API usage.

That’s an incredibly low figure considering the scale of the project, which could easily require hundreds of thousands of dollars in engineering salaries for a similar undertaking.

In my opinion, this marks another advancement on the road to AGI.

To think as a human does, one must handle substantial, unfamiliar projects, maintain focus over prolonged periods, and resolve challenges along the way.

On that note, AI has shown significant improvements in maintaining concentration.

As I highlighted earlier, the METR research group evaluates how long an AI can engage with a genuine problem before requiring human intervention.

And the trend is impressive.

METER shows that the amount of work completed autonomously by leading AI models has roughly doubled every seven months.

Anthropic has observed a parallel trend.

After analyzing millions of Claude coding sessions, they found that some of the longest independent coding sessions nearly doubled within three months, increasing from under 25 minutes to over 45 minutes before the AI required assistance.

If this trajectory continues, those 45 minutes might extend to four hours, eventually reaching an entire workday.

This reinforces my belief that viewing artificial general intelligence as a conclusive endpoint is misguided.

There won’t be a definitive announcement proclaiming the arrival of AGI.

Instead, we will see a steady accumulation of advancements.

My Perspective

In my view, the developments we've witnessed over the past six months signal that we are inching closer to AGI.

Intelligence encompasses more than just knowing answers; it involves applying that knowledge to achieve meaningful outcomes.

And today’s AI is becoming remarkably adept at this.

Nonetheless, a critical question arises.

Given AI's increasing capabilities, why are businesses not reaping more substantial benefits?

We’ll delve into this question further in tomorrow’s edition.

The Components of AGI Are Coming Together

A bright, sunlit modern lab where a giant metallic brain made of interlocking puzzle pieces hangs midair. Warm natural light streams through tall windows, illuminating slender robotic arms gently guiding the final glowing pieces into place. The scene feels hopeful and poised—AGI’s puzzle falling into alignment.

The Pieces of AGI Are Falling Into Place - Banyan Hill Publishing

- Recent AI models have achieved remarkable feats, suggesting we are nearing AGI development.  
- Google’s AI earned a gold medal in a tough math competition, showcasing advanced problem-solving.  
- AI systems are autonomously tackling complex research and coding projects more efficiently and effectively.  

Every few weeks, another AI model does something that would have sounded impossible just a year or two ago.

One model beats some of the world’s brightest young mathematicians. Another helps conduct scientific research. And some are writing software that experienced engineers say would have taken them days or even weeks to finish.

There are even AI systems today that can work on their own for hours before asking a human for help.

Now, I’m not pointing to any one of these breakthroughs on their own and declaring that we’ve reached artificial general intelligence (AGI).

But I think it’s time to stop looking at these breakthroughs in isolation.

Because by doing so, it paints a clearer picture of just how close we are to AGI.

Beyond Chatbots

Back in February, I argued that the core ingredients of AGI were finally starting to come together.

First, AI learned how to answer questions. Then it learned how to reason through complicated problems. More recently, it’s started working on its own for much longer without needing constant supervision.

At the time, I suggested these things were all building blocks of general intelligence.

Looking back over the past six months, I believe we’ve added several more.

Take mathematics.

Last summer, an advanced version of Google’s Gemini Deep Think earned a gold medal score at the International Mathematical Olympiad, widely considered the world’s most difficult math competition for high school students.

Only about 8% of contestants earn a gold medal in a typical year.

Google’s AI solved five of the competition’s six problems, earning 35 out of a possible 42 points after its work was graded by the same judges who score human competitors.

It was an extraordinary achievement.

But Google says newer versions of Deep Think are now helping mathematicians and scientists tackle real research problems in mathematics, physics and computer science.

In one internal benchmark involving advanced mathematical proofs, the system reached roughly 90% accuracy.

To me, that’s a big piece of the AGI puzzle.

Solving problems with known answers is one thing. But helping solve problems that nobody has answered yet is something much closer to general intelligence.

Anthropic recently revealed a similar milestone.

Researchers gave teams of Claude AI agents an open-ended machine-learning research problem with no answer key.

The agents had to come up with their own ideas, write code, run experiments, analyze the results and keep improving their approach until they found a solution that worked.

According to Anthropic, the best AI team outperformed experienced human researchers working on the same problem.

Again, general intelligence is about tackling problems you’ve never encountered before, figuring out what works and adapting as you go.

That’s exactly what these AI agents did.

And we’re seeing the same thing happen in software engineering.

Anthropic recently asked Claude to build a brand-new C compiler capable of compiling the Linux kernel, one of the most complex open-source software projects ever created.

Over the course of two weeks, Claude worked through nearly 2,000 coding sessions. It processed roughly 2 billion input tokens, generated 140 million output tokens and ultimately wrote about 100,000 lines of code.

And it worked.

Claude successfully built software that could run one of the world’s most complex operating systems on several different types of computer processors.

Even more remarkable, the entire experiment cost less than $20,000 in API usage.

That’s a shockingly small price tag for a project of this scale. It could easily require hundreds of thousands of dollars in engineering salaries to accomplish a project of similar size and scope.

To me, this represents another step toward AGI.

To think like a human, you need to be able to tackle large, unfamiliar projects, stay focused over long periods of time and solve problems along the way.

Speaking of which, AI has been getting much better at staying on task.

As I wrote about earlier this year, the research group METR measures how long an AI can work on a real problem before getting stuck or needing human help.

And as you can see, the trend is remarkable.

According to METER, the amount of work frontier AI models can complete on their own has been roughly doubling every seven months.

Anthropic says it’s seeing the same thing.

After studying millions of Claude Code sessions, the company found that some of the longest autonomous coding sessions nearly doubled in just three months, growing from less than 25 minutes to more than 45 minutes before the AI asked for help.

If that trend continues, those forty-five minutes will eventually become four hours. Then a full workday.

That’s why I think it’s a mistake to think of artificial general intelligence as a finish line.

There won’t be a day when somebody suddenly announces that we’ve officially reached AGI.

Instead, we’ll just keep seeing more and more pieces fall into place.

Here’s My Take

In my view, the breakthroughs we’ve witnessed over the past six months suggest we’re getting closer to AGI.

Intelligence isn’t just about knowing the answer. It’s about using that knowledge to accomplish something useful.

And that’s exactly what today’s AI is becoming surprisingly good at.

But that raises an important question.

If AI has become this capable, why aren’t businesses seeing bigger results?

We’ll explore that question further in tomorrow’s Daily Disruptor.

Regards,



Ian King

Chief Strategist, Banyan Hill Publishing

Editor’s Note: We’d love to hear from you!

If you want to share your thoughts or suggestions about the Daily Disruptor, or if there are any specific topics you’d like us to cover, just send an email to dailydisruptor@banyanhill.com.

Don’t worry, we won’t reveal your full name in the event we publish a response. So feel free to comment away!

The Pieces of AGI Are Falling Into Place

An increasing number of legal experts contend that the outputs generated by large language models (LLMs) may not qualify as "speech" under the First Amendment. This means such outputs could be regulated without the stringent scrutiny usually required for restrictions on free expression. A notable case, Garcia v. Character Technologies, arose after the tragic death of 14-year-old Sewell Setzer III, who reportedly engaged in conversations with an AI character before his passing. This case resulted in one of the earliest judicial assessments regarding the First Amendment implications of chatbot outputs. U.S. District Judge Anne Conway, while ruling on the company’s motion to dismiss, stated she was "not prepared to hold that [LLM] output is speech," permitting product liability and negligence claims to continue. The case resolved in January 2026, but numerous other lawsuits have emerged, presenting alarming allegations that chatbot outputs contributed to medical emergencies, violent behavior, or the delivery of inappropriate content to minors. In response to these legal challenges, various federal and state legislative proposals are in the works, aiming to reshape the regulatory landscape surrounding information received from chatbots.

The debate over whether to categorize LLM outputs as "no-speech" presents both theoretical and practical questions. While human expression enjoys First Amendment protection, it remains unclear how machine-generated content fits into this framework. Legal protection could hinge on whether an individual created, refined, or endorsed the LLM-generated content—essentially presenting it as their own. However, distinguishing the authorship of text is challenging. It’s often difficult to ascertain whether a piece of writing originated from a human, a machine, or a blend of both. For the law to properly govern this distinction, a clear attribution standard is necessary, alongside default rules for instances where attribution cannot be determined. Establishing such guidelines would entail an unprecedented challenge: differentiating human expression from machine-generated content on a large scale—a feat for which current technologies are ill-equipped. Existing methods for identity verification focus on confirming the existence of a person as a speaker, rather than identifying who authored specific text. This approach risks burdening the very expression that the protections aim to uphold.

The implications of this uncertainty would likely impact users and readers of content far more than the developers of the models themselves. Implementing a human-attribution requirement across diverse platforms necessitates a reliable means of distinguishing between human and machine contributions. Current tools either fall short in their ability to differentiate the two reliably or demand extensive information about the speakers involved. In the end, this scenario could lead to compromised anonymity, increased surveillance, or even an expectation that identity verification serves as a stand-in for authorship. Any argument restricting AI-generated content from First Amendment protections must take into account these potential consequences, particularly in cases where the extent of human involvement is ambiguous.

The argument that LLM outputs don't qualify as "speech" hinges on the notion that no human stands behind the output at the moment it is generated—thus, it lacks First Amendment protection. Various prominent interpretations of this no-speech stance reach similar conclusions but do so through differing rationales. Scholars such as Mackenzie Austin and Max Levy state that since a human cannot predict a model’s output in real time, machine-generated results lack "speech certainty." Thus, they argue that AI outputs cannot be categorized as protected speech on behalf of any human actor—be it the user, developer, or model itself.

The appeal of this argument is clear. If machine outputs were classified as speech, regulations aimed at them might invoke rigorous First Amendment scrutiny, a situation that could seem counterintuitive for laws designed to ensure product safety or eliminate fraud and discrimination. While the interpretations differ regarding the threshold for human interaction to grant constitutional protection, each relies ultimately on the attribution of human authorship as the crucial determinant.

Legal commentator Benjamin Wittes takes a counter-position, asserting that existing legal doctrine, if applied as intended, already offers protection to machine-generated outputs. He claims that while machines themselves hold no constitutional rights, the rights associated with a company using these machines stem from the actions and decisions of the individuals operating them. Thus, it’s crucial to recognize that restricting a machine’s output could hinder the rights of users, readers, or developers without attributing constitutional rights directly to the machines themselves.

The challenge lies not within formulating the theory itself, but in its practical application. In a specific case, it may be feasible for a court to analyze contributions from the user, such as prompts given, edits made, and the resulting interactions during the discovery process. However, implementing such rules on a broader internet scale entails making these evaluations without ample context regarding contributions, often having to contend solely with the text at hand, which generally does not clarify authorship.

The complexity of attribution extends beyond individual lawsuits to the broader distribution of text through intermediaries like online platforms and search engines. The concern of distributor liability for unlawful or unprotected content can complicate issues of free expression online. Particularly when applying Section 230 of the Communications Decency Act, which provides legal immunity to platforms from being deemed publishers of third-party content, the situation becomes intricate. This statute does not rely on whether the content adheres to First Amendment protections, making its applicability questionable in cases where LLMs produce communication.

With the rise of AI-generated outputs, it remains uncertain whether existing laws, including Section 230, would extend to cover them. Judicial interpretations could potentially find that AI-generated text does not qualify as the “information provided by another” under this law, or legislative actions could seek to eliminate those protections altogether. Bipartisan calls to reassess Section 230, including proposals like the Sunset Section 230 Act, reflect significant pressure to reconsider the protections offered to AI outputs. Should Section 230’s immunity be revoked or curtailed, the potential for increased liability across online platforms could arise—particularly for AI-generated content. This situation would create heightened accountabilities for those distributing text, necessitating a clear understanding of whether a piece originated from a machine or a human.

Even if Section 230 remains intact, the potential liability for platforms could still rise. While it inherently limits liability for third-party material, it does not necessarily restrict what states might demand from platforms regarding their conduct. A law that categorizes unadopted machine output as outside the First Amendment might allow for direct regulation—inhibiting platforms from promoting such content or requiring evidence of human authorship for each post. These issues remain unresolved in the legal sphere, but the underlying incentives to differentiate human from machine content are unmistakable.

Attribution concerns do not present challenges in every scenario. For instance, regulations that demand identification from known chatbot providers may not need to clarify whether an anonymous user is a human. These complexities arise primarily when the source of content is ambiguous. If legislation mandates labels for machine-generated political communication while exempting similar human-authored messages, questions about authorship could easily complicate matters. An automated agent might continue producing posts under earlier directives, creating a legal quagmire regarding who is responsible for the specific content being shared.

To navigate these attribution dilemmas, authorities could rely on certain types of technologies—such as text classifiers, content provenance tools, behavioral analysis, and personhood credentials. Each of these approaches addresses tangible questions, but none effectively resolve the attribution issue at hand. Text classifiers, which directly assess whether text resembles machine-generated output, have been found unreliable and fail to clarify human involvement. Provenance tracking can outline the history of a file’s creation but doesn’t identify who is responsible for its claims. Behavioral analysis tools, while designed to detect automation, typically focus on classifying accounts rather than specific content, leading to risks of violating privacy when seeking greater verification.

Furthermore, privacy-preserving personhood credentials represent emerging solutions to verifying user identities without compromising anonymity. These credentials could offer a means to affirm the presence of an accountable individual behind an account, thus addressing concerns linked to accountability and manipulation online. Yet, obtaining such credentials generally requires initial identification that could set a precedent for later practices, potentially leading to system-wide coercion on users to validate their identities for trust.

Nevertheless, personhood verification does not imply authorship. Although a verified user may operate an account, the text published may still be entirely machine-generated, while an AI system could distribute messages authored by humans. Consequently, existing technologies do not reliably confirm authorship or intentionality.

As institutions seek clarity, they may inadvertently favor questions they can answer—specifically, “Is there a real person involved?”—over the harder query of whether this individual actually authored the content. When standards favor human involvement over authorship, institutions may implement more invasive means of verification. Although measures safeguarding privacy exist, they remain challenging to enforce and can lead to misinterpretations.

The legal concerns surrounding anonymity and compelled identification are longstanding within First Amendment jurisprudence. Historical precedents emphasize the necessity of justifying any imposition on anonymity, arguing that protected speech should not inherently require identification.

The ramifications of the classification of “no-speech” extend beyond legal theory to practical effects on scrutiny levels in court. For example, in cases dealing with content accessibility, if outputs deemed machine-generated are ruled as lacking First Amendment protection, it could redefine the scrutiny applied to measures aimed at moderating or regulating such content. This distinction may significantly impact broader regulatory frameworks and inform the conditions under which individuals can express themselves online.

Thus, opting against a blanket no-speech rule does not equate to relinquishing the capacity for regulation. Instead, it necessitates an approach that focuses on identifying the parties affected by laws and assessing whether the burdens imposed are justifiable. Various stakeholders—be it users, readers, or developers—may hold valid claims to constitutional protections during instances of AI-generated or influenced expression. The challenge will remain to ensure that protections against fraud or misrepresentation do not intrude upon the essential freedoms enshrined in the First Amendment. Balancing the need for accountability with the recognition of rights associated with both human and machine-generated outputs remains critical to the evolving landscape of expression in the digital age.

If AI Outputs Aren't Verbal, Who Needs to Demonstrate They're Human?

A dim, wood-paneled courtroom bathed in warm tungsten light. At center left, a human witness stands at the stand, hand raised under a single spotlight. Opposite, a sleek black AI terminal hums quietly, spitting out a glowing scroll of binary code onto crisp legal documents. Behind them, a giant question mark is carved into the dark paneled wall, looming over the scene. Moody, cinematic legal drama with rich shadows.

If AI Outputs Aren’t Speech, Who Has to Prove They’re Human? - Lawfare

- Legal scholars are debating if AI outputs are protected speech under the First Amendment.  
- A recent court case ruled AI model outputs are not deemed "speech," allowing liability claims.  
- Attribution challenges complicate regulations, potentially burdening users with disclosure and identity verification.  

If AI Outputs Aren’t Speech, Who Has to Prove They’re Human?

AI chatbots can articulate concepts with language, yet they lack a true grasp of cause and effect in the real world. A new form of machine intelligence is starting to emerge that aims to bridge this gap.

When you ask ChatGPT what occurs when a glass falls from a table, it offers a detailed explanation, discussing gravity's role and the potential for shattering based on stress. In contrast, when I pose the same question to my three-year-old daughter, she simply states, “It go smash and water everywhere!” This difference highlights a crucial distinction in artificial intelligence (AI) development.

Current AI systems, particularly those driven by large language models (LLMs), rely on extensive textual data to forecast subsequent words in a sequence. Unlike young children who learn through direct experience, these AI models haven't physically interacted with the world, which may limit their understanding of physical phenomena.

This gap in capability leads some experts to advocate for the development of "world models." These systems are designed to learn from observation, allowing them to simulate the outcomes of actions in the physical world, rather than merely predicting text.

The concept of world models has gained significant attention, particularly in relation to advancements in autonomous robotics. They also hold promise for creating artificial general intelligence (AGI)—machines capable of human-level reasoning across diverse tasks.

However, the discussion around "world models" is often muddled. As Melanie Mitchell from the Santa Fe Institute observes, the term can sometimes become a catch-all for AI's shortcomings, leaving questions about what these models should actually represent and whether they are adequate for achieving the flexible intelligence seen in humans.

To grasp the significance of world models within AI, it is helpful to look back at cognitive science foundations. The idea traces back to 1943 when psychologist Kenneth Craik proposed that human minds maintain a simplified model of the external world, which helps us anticipate various outcomes and adjust our actions accordingly.

This notion has evolved, especially with the emergence of predictive processing theories, suggesting that our perceptions and potentially even consciousness are rooted in the brain’s continuous predictions about external events, which are refined by sensory information.

According to cognitive scientist Josh Tenenbaum, the essence of intelligence lies in building these internal models to simulate potential outcomes and avoid costly errors. This highlights why AI researchers are particularly interested in pursuing world models. While LLMs demonstrate remarkable capabilities and can mimic human-like understanding, they fundamentally operate on language as a representation of reality, lacking the direct experience that shapes human comprehension of physical occurrences.

LLMs often encounter challenges when it comes to reasoning about spatial dynamics and real-world events. For instance, a recent study revealed that when LLMs were trained on taxi route data from New York City, they could generate plausible directions but struggled dramatically with unexpected detours—an issue stemming from their reliance on descriptions instead of experiential understanding.

Leading the charge for these advancements is Yann LeCun, a computer scientist from New York University and former chief scientist at Meta. He asserts that true intelligent systems must possess the ability to predict the consequences of actions based on observed data. This belief has driven his endeavors to establish world models designed for learning autonomous behavior.

LeCun’s ambitions have attracted considerable investment and attention. In late 2025, he left Meta to create AMI Labs, successfully securing over $1 billion for world model development. Meanwhile, Stanford's Fei-Fei Li has initiated World Labs, which focuses on spatial intelligence AI with $230 million in funding. Many major AI players, including Google DeepMind, are now exploring world models in various capacities.

The motivation behind this influx of investment largely stems from the potential of world models to enhance robotics. Presently, prototypes remain early-stage in development, with impressive yet limited applications. For example, World Labs has created a system named Marble that generates coherent 3D scenes from text, while DeepMind’s Genie 3 constructs engaging virtual environments.

While these offerings are indeed referred to as world models, they serve more as advanced scene generators than independent systems capable of predictive reasoning. A true world model should allow an agent to forecast the outcomes of its actions and plan effectively—a goal that DeepMind's Dreamer 4 system, launched in September 2025, aims to achieve. Dreamer 4 builds an internal, predictive model of its surroundings using data primarily from Minecraft, allowing it to rehearse potential actions and refine its approach through imagination.

The effectiveness of this architecture shines through in Dreamer 4’s successful navigation of complex tasks, like collecting diamonds in Minecraft, despite lacking prior guidance. Such capacities could pave the way for robots that perform daily chores, like folding laundry or loading dishwashers. 

The road ahead is promising; however, significant challenges remain. Researchers are still grappling with fundamental questions regarding what these systems should learn about the world and the level of detail required for meaningful functioning. Two main strategies have emerged within the field: one focuses on generating accurate predictions about future actions based on comprehensive data, while LeCun argues for a different approach, advocating for joint-embedding predictive architecture (JEPA). This strategy emphasizes capturing key aspects vital for reasoning and planning, rather than overwhelming detail.

The contrast between the two methodologies remains a topic of active research. Some hypothesize that more abstract representations will provide faster pathways to practical applications, whereas others, like Hafner from DeepMind, point out the importance of robust, detailed representations for effective learning.

Despite the advances represented by world models, critical discussions are underway about whether any of these approaches will be sufficient to achieve AGI. The concept of AGI itself is multifaceted, with various interpretations suggesting that realistic simulations of physical reality could lay the groundwork for generally capable reasoning agents. The promise of world models, such as those from DeepMind, build upon this idea but simultaneously raise questions about the necessary components for authentic human-like intelligence.

Mitchell raises another significant point by noting that existing AI systems lack metacognition—the ability to understand their cognitive states. This self-awareness could be crucial for AI to approach anything resembling human-level cognition, yet it remains to be determined how world models can address this gap.

Ultimately, while it is likely that world models will enhance capabilities in robotics and other fields, the quest to replicate the nuanced processes of human cognition is a daunting challenge. As Tenenbaum articulates, human intelligence isn’t singular but rather a complex interplay of models adaptable to varying contexts, emphasizing that achieving a comprehensive understanding remains an ambition of great complexity.

Could a novel type of AI that comprehends physical reality reshape the world once more?

A sunlit workshop bathed in warm, natural light: at a wooden workbench, a sleek metallic robotic arm and a human engineer’s hand jointly assemble tiny brass gears of a mechanical watch. Golden rays pick out swirling dust motes in the air as faint circuitry patterns pulse along the robot’s forearm, blurring the line between digital code and physical metal. This moment captures a new AI that truly grasps the tangible world, gearing up to reshape reality.

Will a new kind of AI that understands physical reality change the world again? - New Scientist

- Current AI chatbots lack true understanding of real-world events and cause-effect relationships.  
- Researchers are exploring "world models" that learn from observation to predict outcomes.  
- Different approaches, including JEPA, aim to improve AI learning and simulate physical reality.

AI chatbots can describe reality in words, but don’t truly understand cause and effect in the real world. Now, a fresh kind of machine intelligence that does just that is emerging

Andrea De Santis

What happens when you knock a glass off a table? Ask ChatGPT and you get a step-by-step explanation: how gravity accelerates the receptacle along a “parabolic trajectory”, why it could shatter “if the impact generates stresses greater than the material can withstand”. Ask my 3-year-old daughter and she’s less bothered with fine details: “It go smash and water everywhere!”

When it comes to the future of artificial intelligence, the difference is instructive. The large language models (LLMs) powering today’s chatbots are trained on vast swathes of text to predict the next word in a sequence. Unlike little Adeline, however, no LLM has ever actually knocked a glass off a table – or thrown spaghetti at a wall, or driven a scooter into a pond – and that may impose a fundamental limitation on their capabilities.

This, essentially, is why some researchers argue that what we need to take AI to the next level isn’t ever-larger LLMs, but “world models”: systems that learn through observation so they can simulate the consequences of actions in the real world.

Advertisement

World models have rapidly become AI’s next frontier, attracting huge buzz from industry and business thanks to their promise in autonomous robotics. What makes the world-model approach especially intriguing, though, is that it also ostensibly offers a route to artificial general intelligence (AGI), or machines with human-level reasoning that can be applied across a range of tasks.

The problem is that it is tricky to parse substance from bluster. The term “world model” is confusingly elastic, for starters – for some, problematically so. It has become “a kind of shorthand for all the things that current AI systems can’t do well”, says Melanie Mitchell at the Santa Fe Institute in New Mexico. It is unclear exactly what these systems should model, never mind whether internal representations of physical reality will be sufficient for the kind of generalisable intelligence that means my 3-year-old intuitively understands what happens when she swipes a glass, even if she can’t say why.

To make sense of world models as they apply in AI, it helps to understand the concept’s roots in cognitive science. AI researchers themselves typically trace it back to 1943, when psychologist Kenneth Craik wrote that the human mind “carries a small-scale model of external reality and of its own possible actions”, allowing us “to try out various alternatives” and “react to future situations before they arise”.

Human minds have world models that enable us to predict what will happen next timsimages/Alamy

That idea has evolved in the decades since, most notably with the theory of predictive processing. This posits that perception – possibly even consciousness itself – relies on the brain constantly generating predictions about the external world and updating them in response to incoming sensory data.

But the point remains. “The key idea is that intelligence involves building models of the world in our heads so we can simulate outcomes and avoid costly mistakes,” says Josh Tenenbaum, a cognitive scientist and AI researcher at the Massachusetts Institute of Technology. “So ‘our ideas die in our stead’, as Craik put it.”

It doesn’t take a genius to see why that has piqued the interest of AI researchers. LLMs have proved astonishingly capable, and they can certainly give the impression they understand things as we do. But language is a description of reality, rather than reality itself, and their lack of direct experience means that even the most powerful language models struggle when it comes to “understanding” physical phenomena in the real world.

LLMs often perform poorly when they are asked to reason about spatial concepts and things that might happen in the physical world, says Mitchell. “If you’ve ever uploaded a map to one of these models and asked it questions about it, you’ll know that they will often have a lot of problems with reasoning.” Indeed, in a 2024 study, when researchers trained language models on a database of turn-by-turn directions for taxi trips around New York City, they found that the system could provide reasonable routes from one point to another – but failed miserably when asked to take the odd detour.

The reason why is that the LLMs have only descriptions of journeys, and therefore lack the sort of mental map that allows us to imagine what would happen in different scenarios. And navigation is just one example. LLMs demonstrate similar shortcomings in any scenario where you need to simulate the physical world – safely loading a dishwasher, say, or folding laundry. “They’re not designed for it, which is why people are interested in alternatives like world models,” says Tenenbaum.

Andrea De Santis

Probably the most influential proponent is Yann LeCun, a computer scientist at New York University and, until recently, chief scientist of foundational AI research at Meta. His argument, first made in a 2022 paper, is essentially that if we want to build truly intelligent systems that can reason, plan and act effectively in the real world, we need world models.

“I cannot imagine we can build agentic systems without those systems having an ability to predict, in advance, what the consequences of their actions are going to be,” LeCun told artificial intelligence forum AI House Davos in January. And the key to that, he reckons, is systems that learn the rules of the world from observation.

The idea has caught on. In December 2025, LeCun left Meta to found AMI Labs, raising just north of $1 billion to build world-model systems. A year earlier, Fei-Fei Li at Stanford University in California started World Labs, with $230 million of investment, to develop AI with “spatial intelligence”. Many of the established AI firms, most notably Google DeepMind, are now also actively pursuing world models in one form or another.

The reason for the influx of money is primarily the promise that world models hold for advancing robotics (see “What are AI world models good for?”, below). However, it is early days and the existing prototypes are modest in their applications. With Marble, for instance, World Labs has built a system that generates coherent 3D scenes from text prompts. Similarly, DeepMind’s Genie 3 creates convincing interactive virtual environments – “snowy mountain at dusk”, say – in which you can move around for several minutes and even prompt events like rain.

Google DeepMind’s Genie 3 creates worlds that users can explore from text-based prompts Google DeepMind

In both cases, the companies describe their systems as “world models”. But Marble is really a 3D video generator and Genie 3 a video-game simulator, albeit one producing simulations in which agents could plausibly act, observe consequences and learn.

Arguably, a proper world model would be something that exists inside an agent such that it can predict the consequences of its decisions, imagine the future and plan ahead before taking actions – and that is what DeepMind is pushing towards with its Dreamer 4 system, released in September 2025. It learns an internal predictive model of its environment, training mostly on data from the video game Minecraft, and uses that to train an agent that repeatedly “dreams” the consequences of possible actions to improve its behaviour.

“The biggest difference [from Genie 3] is that Dreamer 4 is an agent, not just a world model,” says Danijar Hafner at DeepMind in San Francisco, who leads the Dreamer 4 team. “It predicts actions and improves them through iterative self-improvement, using planning and [imagined] trial and error.”

The power of this approach is apparent in Hafner and his colleagues’ demonstration that Dreamer 4 can figure out how to collect diamonds in Minecraft, a complex task involving thousands of different actions – gathering resources, crafting tools, navigating the landscape – without being shown how to play. “The system has to understand its environment and generalise, because each new episode starts in a randomly generated world,” says Hafner.

That is an important step, not least because this is precisely the kind of world model that could facilitate autonomous robots capable of folding laundry, say, or loading the dishwasher. “I think it’s going to solve robotics,” says Hafner. “What’s missing now is execution: data, compute, scaling. But we have the recipe, so, like with language models, it’s about scaling and details.”

Who would bet against DeepMind, given its impressive track record? After all, its researchers have won a Nobel prize for work on protein folding. But while Dreamer 4 demonstrates that agents with internal world models can reason and plan, it leaves a deeper question unresolved: what exactly should these systems learn about the world, or, more specifically, at what level of detail? And this has become a key fault line in the field, with two distinct approaches emerging.

Many of the existing strategies are attempting to generate predictions on condition of action, as the researchers put it, by reconstructing future observations as faithfully to the training data as possible. But LeCun reckons that is the wrong approach, or at least not the best. His argument is based on the fact that humans don’t mentally simulate the world in any great detail, and certainly not pixel by pixel. When we imagine a glass falling from a table and smashing to pieces, for example, we don’t predict the position of every shard of glass, every water droplet. Instead, LeCun argues, we run highly compressed models that capture only the aspects that matter.

This why LeCun advocates for a different tack, which he calls joint-embedding predictive architecture (JEPA). In this framework, world models infer abstract representations of what is relevant for reasoning, planning and action. “Generative models try to reconstruct pixels, whereas JEPA learns in a latent space and only predicts what is useful,” says Randall Balestriero at Brown University in Rhode Island, who has worked with LeCun on JEPA-based world models. “It ignores irrelevant details and focuses only on what matters for the agent.”

Free newsletter Sign up to The Daily Sign up to newsletter

LeCun and his investors appear to be betting that the JEPA approach will offer a swifter route to real-world applications, largely because it requires less training data. The most concrete publicly available demonstration of this technique so far is a system called V-JEPA 2, released by Meta in June 2025. Trained on video inputs, V-JEPA learns by masking certain regions of footage – obscuring a moving ball, say, over multiple consecutive frames – and repeatedly predicting not pixels, but abstract representations of what was hidden to learn a compact internal model of how the ball moves. What LeCun and his colleagues have shown, then, is that their system can model the causal structure of the physical world without having to reconstruct it in detail.

Which isn’t to say that JEPA is necessarily any better than generative world models. Its advocates argue that it will be, of course. “Without abstraction, AI systems will stay limited to narrow tasks,” says Balestriero. But V-JEPA 2 is yet to clearly demonstrate that its abstract representations are sufficiently meaningful to enable an agent operating in an open-ended environment to reason and plan. “The biggest missing piece is the question of, how do we know that the abstraction is actually useful,” says Balestriero. “This is where there is a huge amount of active research right now, to understand: what do you capture, or how do you encode something very rich about the world that is useful for planning downstream.”

Hafner, for his part, isn’t convinced that more compressed representations of reality are better. “Yann is right about many things, even if maybe he expresses them more controversially than necessary, and the JEPA approach is very promising,” he says. “But I don’t think that representations should be compressed and tiny. Ultimately, you want to learn strong representations, and what we did with Dreamer 4 [which trains with pixels but predicts in abstract space] is incredibly robust.”

Some tech investors say that world models will be an effective way to power humanoid robots Wang Jiang/VCG via Getty Images

More broadly, it is also far from clear at this stage if any of the world models in development – whether they learn and predict by reconstructing the world in high-fidelity or by inferring abstract representations – will ultimately be enough to get us to AGI.

Now, it’s fair to say that AGI is another elastic term, and that claims about world models as a route towards it exist on a spectrum, with some of the most ambitious suggesting that learned simulations of physical reality could become the foundation of generally capable reasoning agents. When it released Genie 3, for instance, DeepMind insisted world models are “a key stepping stone on the path to AGI, since they make it possible to train AI agents in an unlimited curriculum of rich simulation environments”. Indeed, Hafner argues that “if you have systems that are able to represent the rules governing the world, then you’re approaching something like AGI because that’s understanding, that’s a key part of intelligence”.

For his part, LeCun talks about them as a vital component of broader systems, rather than the whole story. And yet it is worth exploring the extent to which the kinds of world models in development would approximate human intelligence, because it can reveal what else might be required. Hafner says the most immediate requirement is temporal abstraction – that is, being able to reason not just about what will happen in the immediate future, but as things continue to play out. “You cannot just simulate everything at a sub-second, frame-by-frame level, because humans do not reason like that,” he says. “This is one of the open frontiers.”

Tenenbaum goes further. Human world models aren’t just engines for predictions, he says, “they are much richer than that”. Human reasoning depends on various forms of causal abstraction and hypothesis-driven inference – not to mention models of other minds – which together allow us to flexibly recombine knowledge across different situations. “A key open question in all this is whether scaling up these world models will recover that richness, and my view is that it likely won’t,” says Tenenbaum.

Mitchell makes a similar, if slightly broader, point. “I think the ability to have a kind of compressed, simulatable representation of aspects of the world is very important, and this notion of world models is probably going to result in useful improvements,” she says. “But I think there are probably lots of other aspects of intelligence that matter.”

Another limitation of AI systems today, says Mitchell, is a lack of metacognition – an awareness of their own cognitive state, of what they know and don’t know, and how uncertain they are. “Is that fixable with a world model? Well, it depends what kind of world model, obviously,” she says. “But I’m a little worried that the notion of ‘world model’ is going to be used as the term for the difference between what we have now and ‘AGI’.”

All of which suggests that world models as currently conceived may well be necessary for human-level machine intelligence, but not necessarily sufficient. Indeed, while it looks increasingly likely that they will be powerful and genuinely useful in robotics, and possibly scientific simulation too, it is far from a sure bet that they will replicate the way world models work in human cognition, never mind human-level intelligence more broadly. “One of the big misconceptions is that intelligence is a single thing, that there is a single world model in the brain,” says Tenenbaum. “What we actually have is the ability to run many different models depending on the context, task and goal.”

Ultimately, then, the task of replicating what my 3-year-old daughter’s brain is capable of when it comes to modelling and predicting the physical world – even if it doesn’t stop her knocking glasses of water off the table – is not one to be underestimated.

Amazon shuts down San Francisco AI laboratory and terminates 1,100 employees at Florida warehouse.

More from Artificial General Intelligence

Human Exceptionalism Faces Threat from Artificial General Intelligence and ASI

Could AGI and superintelligent AI put an end to human exceptionalism?

DeepMind's founder takes on a leading AI position at Google.

Artificial General Intelligence is not on the horizon.