This document is a component of the RAND research report series, which showcases findings and assessments aimed at tackling issues in both public and private sectors. All reports produced by RAND undergo a thorough peer review process to maintain exceptional standards of research quality and impartiality.

The content and trademarks within this document are legally protected. The representation of RAND's intellectual property is available for noncommercial purposes only. It is prohibited to upload this publication online without authorization; however, directly linking to this product page is recommended. For any reproduction or reuse of its research documents for commercial objectives, permission from RAND is necessary. To find more information regarding reprint and reuse permissions, please visit www.rand.org/pubs/permissions.

As a nonprofit organization, RAND strives to enhance policy and decision-making via extensive research and analytical work. The views expressed in RAND publications may not necessarily align with those of the clients and sponsors involved in the research.

Can Competitors of AGI Retain Their Strength?

In a shadowy industrial hall lit by warm, flickering overhead lamps, two suited figures stand on either side of a colossal, glowing mechanical brain pulsing with molten-orange circuitry. Between them lies a heavy steel chain and oversized padlock, their hands poised to clamp it shut as tendrils of glowing code writhe outward. Scattered around their feet are fallen chess pieces shaped like microchips, symbolizing a high-stakes struggle to restrain the unleashed power of AGI rivals.

Can AGI Rivals Hold Back? Verification, Domestic Oversight, and the Conditions for Mutual Restraint - RAND.org

- RAND research reports provide objective analysis for public and private sector challenges.  
- All reports undergo rigorous peer review to ensure high-quality research standards.  
- Unauthorized online posting is prohibited; permissions required for commercial reproduction or reuse.

This publication is part of the RAND research report series. Research reports present research findings and objective analysis that address the challenges facing the public and private sectors. All RAND research reports undergo rigorous peer review to ensure high standards for research quality and objectivity.

This document and trademark(s) contained herein are protected by law. This representation of RAND intellectual property is provided for noncommercial use only. Unauthorized posting of this publication online is prohibited; linking directly to this product page is encouraged. Permission is required from RAND to reproduce, or reuse in another form, any of its research documents for commercial purposes. For information on reprint and reuse permissions, please visit www.rand.org/pubs/permissions.

RAND is a nonprofit institution that helps improve policy and decisionmaking through research and analysis. RAND's publications do not necessarily reflect the opinions of its research clients and sponsors.

Can AGI Rivals Hold Back?

Demis Hassabis, the chair of Google DeepMind and chief scientist at Alphabet, has recently revised his predictions, forecasting that artificial general intelligence (AGI) could be just two to three years away. In a profile published in the Wall Street Journal Magazine by Ben Cohen on October 21, Hassabis expressed his belief that the timeline for achieving human-level AI has been decreasing, moving from an initial estimate of three to five years earlier this year.

This revised outlook is a significant departure from the estimates he provided at the World Economic Forum in Davos back in January, where he suggested AGI was still five to ten years away, with a roughly 50% chance of materializing by the end of the decade. By February, during the India AI Impact Summit in New Delhi, he indicated that AGI might be achieved within five years. He also introduced the “Einstein Test,” a benchmark for assessing AGI by seeing if a system trained exclusively on knowledge up to 1911 could independently derive general relativity as Einstein did. In May, during Google I/O, Hassabis hinted that AGI was “just a few years away” and likened current AI agents to a practice run, noting, “We can see agents really happening now, and imagine what they will be in another year.” Now, his estimate has tightened to just two to three years.

Hassabis has made it clear that this revised timeline is not merely a prediction but also serves as a call to action. After the Google I/O event, he told Axios that he had chosen direct language to spur urgency among governments, economists, and the public, emphasizing that his words were intended to increase awareness of the situation rather than solely serve as forecasts.

He is bolstered in his view by what he perceives as a promising trajectory in the industry, regarding the development of AGI. His definition of AGI remains consistent: a system capable of exhibiting a full range of human cognitive abilities, rather than excelling at just one narrow task. What has changed is his perception of how quickly current AI systems could reach that level.

Hassabis is not the only one with an accelerated timeline. Anthropic’s CEO, Dario Amodei, stated at Davos in January 2025 that AI systems outperforming most humans across various tasks could also emerge within the next two to three years—aligning closely with Hassabis's current estimate. While both leaders come from competing organizations, their convergence on a similar timeline suggests a noteworthy trend, though it does not provide definitive proof that their schedule is accurate.

For investors and founders, the collapsing timeline offers significant implications. Decisions regarding capital allocation in AI infrastructure, enterprise contracts, and staffing all depend on expectations about how quickly advancements will occur. At the beginning of the year, Hassabis projected a five to ten-year development period for AGI, but by October, his assessment was that it is imminent. This shift effectively shortens the planning horizon for the market, independent of whether the anticipated technological breakthroughs actually materialize as forecasted.

The Einstein Test remains theoretical, with no known systems currently demonstrating the ability to pass it. Nevertheless, Hassabis's timeline has notably tightened throughout 2026, from five to ten years in January to just two to three years by October. Google DeepMind continues to compete with OpenAI and Anthropic in the race to deliver more advanced AI systems, and time is clearly of the essence.

Demis Hassabis now believes AGI may be achieved within two to three years.

A modern tech lab bathed in warm dawn light. From behind, a bald man with glasses—resembling Demis Hassabis—stands at a floor-to-ceiling window overlooking a sleek, futuristic skyline. In his outstretched hand hovers a glowing, neural-patterned sphere pulsing with energy. On a nearby desk, sketches of AI circuitry and timelines hint at “2–3 years.” The scene whispers anticipation of imminent AGI.

Demis Hassabis now says AGI could arrive in just two to three years - Startup Fortune

- Demis Hassabis predicts human-level AI will be achieved in two to three years.  
- His previous AGI timeline shifted from five to ten years to two to three.  
- The accelerated timeline influences market planning and investor confidence in AI advancements.  

The Google DeepMind chair has spent 2026 pulling his own deadline closer, and his newest estimate puts human-level AI just two to three years away.

Demis Hassabis, now chair of Google DeepMind and chief scientist at Alphabet, told The Wall Street Journal Magazine he expects artificial general intelligence within two to three years, down from the three-to-five-year window he was citing earlier this year. The profile was written by Ben Cohen and published October 21 as part of the magazine's Innovator in Tech package. It quotes Hassabis saying simply: "It's soon." Asked to explain the shrinking timeline, he added, "I'm updating nearer and nearer."

That's a marked shift from the timeline Hassabis gave at the World Economic Forum in Davos in January, when he described AGI as five to 10 years away and put the chance of reaching it by the end of the decade at about 50 percent. In February, speaking at the India AI Impact Summit in New Delhi, he said AGI might arrive within five years. He also laid out what he calls the Einstein Test: train a system using knowledge available through 1911 and see whether it can independently derive general relativity, as Einstein did in 1915. In May, at Google I/O, he said AGI was "just a few years away" and described today's crop of AI agents as a kind of dress rehearsal. "We can see agents really happening now," he said, "and imagine what they will be in another year." Now the estimate is two to three years.

Hassabis has been candid that the timeline isn't only a forecast, it's a message. After Google I/O in May, he told Axios he had chosen blunt language to create more urgency among governments, economists and the public. The words were doing rhetorical work, not just predictive work.

Hassabis has said his confidence is rising because the industry appears to be on the right technical path. His own definition of AGI hasn't moved in the meantime: a system that can exhibit the full range of human cognitive abilities, not just ace a narrow benchmark. What's moved is how close he thinks today's systems are to clearing that bar.

He isn't alone in offering a short timeline. Anthropic CEO Dario Amodei told CNBC at Davos in January 2025 that AI "better than almost all humans at almost all tasks" was likely within two to three years - almost exactly matching where Hassabis has now landed. Amodei runs Anthropic. Hassabis chairs rival Google DeepMind. Convergence on a number is itself a data point, though it's not proof the schedule is correct.

For founders and investors, the timeline compression is a cleaner signal than it looks. Capital allocation into AI infrastructure, enterprise contracts and headcount plans all assume some model of how fast the ground will shift under them. At Davos in January, a DeepMind chair put AGI five to 10 years off. By October, he's telling the Journal "it's soon." That's effectively told the market its planning horizon just got shorter, whether or not the underlying technology actually arrives on schedule.

The Einstein Test remains a proposal, with no published result showing that any system has passed it. Hassabis's public timeline has tightened all the same through 2026: five to 10 years at Davos in January, two to three years in October. Google DeepMind competes with OpenAI and Anthropic to develop increasingly capable AI systems, and the clock keeps moving.

Also read: Gartner says AI infrastructure spending will outrun revenue by $4.9 trillion • ChainOpera AI token COAI jumps 30% as AI-crypto trade heats up again • NextDC shares crash 37% after Melbourne grid upgrade gets blocked

The Metropolix, from Startup Fortune →

This article is posted in AI News, check it out for more related stories.

Demis Hassabis now says AGI could arrive in just two to three years

As the notion of artificial general intelligence (AGI) becomes increasingly tangible, businesses in Mexico are navigating a significant shift from merely generating algorithms to implementing stringent oversight of their outputs. Leaders in the industry, speaking at the Mexico AI, Cloud & Data Summit 2026, highlighted that a company's ability to remain competitive hinges on developing professionals who possess dual competencies—those who can not only audit autonomous models but also address automation biases and ensure the integrity of data.

Globally, AI research labs predict that AGI could be a reality within the next five years, prompting a transition in Latin America from theoretical discussions to urgent workforce realignment. In Mexico, where the integration of AI is racing ahead while grappling with a persistent skilled labor shortfall, the key challenge lies in nurturing talent adept at managing autonomous systems rather than simply executing programming tasks.

Mexico boasts one of the region's fastest-growing tech landscapes, with its AI market projected to grow from US$17 million in 2025 to US$180 million by 2033. Currently, over 40% of Mexican businesses are utilizing automated processes within their operations, and the demand for job candidates with AI skills has surged from 5.62% in 2023 to 13.95% in 2025.

However, looming labor shortages pose a threat to this growth trajectory. Research by ManpowerGroup reveals that 70% of employers in Mexico struggled to fill positions in 2025, with specialized IT skill gaps affecting 77% of these vacancies. While local universities graduate around 130,000 engineers and 174,476 STEM graduates each year, the technical expertise these individuals obtain often falls short of preparing them to critically assess model reasoning or identify biases.

Transitioning from Code Generation to Output Oversight

The rise of near-autonomous systems necessitates a redefinition of technical skills. As models take on the responsibility of producing intricate code independently, workforce expertise must shift towards oversight, compliance checks, and systemic validation.

Julio Saenz, Senior Director of Engineering and Mexico Country Lead at Zillow, asserts, "While we can delegate execution, we cannot delegate the responsibility that comes with it." He cautions that junior developers may develop complacent habits by accepting automated results without conducting architectural reviews. He insists that "AI requires criteria, judgment, and foundational knowledge; leaders must not allow tools to generate solutions without thorough evaluation."

Effective supervision also hinges on robust data architecture, as poorly managed corporate datasets can lead autonomous models to reach erroneous conclusions. Kenneth Navarro, Territory Account Executive at Everpure, emphasizes the necessity of understanding data provenance and reliability, stating, "The starting point must include knowing where data comes from, its trustworthiness, and whether we have the necessary permissions."

Navarro highlights that the role of technical teams has shifted towards data management and access control. "Human labor has evolved from merely inputting data; now, technical staff need to oversee permissions, safeguard sensitive corporate data, and ensure complete traceability for every automated action."

Addressing Automation Bias Throughout the Organization

The issue of supervision extends beyond engineering and into non-technical sectors such as legal, human resources, finance, and operations, where employees increasingly interact with algorithmic outputs without the necessary resources to identify potential errors.

Maider Santos, Vice President and Head of Mexico at BTS, notes, "As we engage more with AI, it becomes harder to critically evaluate a system that seems confident and reliable." Santos warns that unexamined reliance can lead to operational inefficiencies. As he illustrates with a corporate example, where 14,000 independent automated agents were created but only 30% remained operational after three months, the importance of qualitative review becomes evident.

Santos argues that fostering a culture of inquiry is essential for organizations. "Reviewing processes is a behavior, and questioning is a skill that businesses can cultivate," he states, stressing the need for aligning digital initiatives with tangible performance outcomes.

Leaders recognize that assessing critical thinking is a challenge. Marisela Orihuela, President of CIO's LATAM, emphasizes, "Common sense is often the rarest of qualities." She underscores the importance of establishing rigorous protocols to determine if staff possess the necessary judgment to evaluate automated recommendations, asking, "How do we balance developers' technical mastery with the analytical acumen needed to audit high-autonomy systems?"

Continuous Learning and Cooperative Governance

Bridging the skills gap requires educational approaches that extend beyond traditional multi-year university programs. The Inter-American Development Bank warns that national economic advancement relies on addressing this specialized skills shortfall, as the World Economic Forum estimates that 59% of global workers will need retraining by 2030, with 70% of core job skills expected to shift.

Public and private initiatives are enhancing the talent pipeline. In January 2026, federal authorities launched the Public Center for AI Training to provide free education in cloud computing and data management.

However, company leaders must continue to foster internal development programs. Javier Blanco, Country People Director for Mexico at Mercado Libre, notes, "The ability to give continuous feedback is the only way teams can provide significant value to users." He compares advanced AI systems to high-performance racing cars, stressing that access without adequate training yields minimal business benefit. Blanco concludes, "Organizations must promote a culture of practical curiosity through ongoing engagement, as capabilities are primarily developed through hands-on problem-solving rather than passive learning."

Ultimately, whether Mexican enterprises can cultivate effective oversight will play a crucial role in determining if the current tech growth leads to sustained competitive advantage or heightened corporate risks. The shift from basic tool usage to sophisticated governance will mark a pivotal milestone on the path to AGI's arrival.

The New Criteria for Talent Ahead of AGI Emergence

A bright, sunlit conference hall where a diverse group of adult professionals in smart business attire step forward on a raised platform. Behind them, a large digital screen reads “AGI Talent Standards” in bold letters, overlaid with a glowing neural-network diagram. Natural daylight streams through floor-to-ceiling windows, illuminating one candidate as they receive a sleek, metallic badge marked “A” from a panel of experts. The scene feels optimistic and forward-looking, emphasizing human achievement and the next era of AI readiness.

The New Talent Standards for AGI Arrival - Mexico Business News

- Mexican enterprises must shift from code generation to overseeing autonomous systems effectively.  
- Growing AI market faces talent shortages; employers struggle to fill specialized IT vacancies.  
- Continuous education and governance frameworks are essential for managing emerging AGI innovations.  

As the artificial general intelligence (AGI) concept feels closer, Mexican enterprises face an operational transition from algorithmic generation to rigorous output oversight. Industry leaders at the Mexico AI, Cloud & Data Summit 2026 emphasize that corporate competitiveness depends on developing dual-competency professionals who can audit autonomous models, eliminate automation bias, and safeguard data integrity.

As global AI laboratories forecast the emergence of AGI in less than five years, the discussion across Latin America has shifted from speculative theory to immediate workforce restructuring. In Mexico, where enterprise adoption accelerates alongside chronic skilled-labor deficits, the strategic challenge centers on cultivating talent capable of governing autonomous systems rather than merely executing code.

Mexico represents one of the fastest-growing technology ecosystems in the region, with the domestic AI market valued at US$17 million in 2025 and a projection at US$180 million by 2033. More than 40% of Mexican enterprises currently incorporate automated solutions into operational workflows, while the proportion of job vacancies requiring artificial intelligence skills rose from 5.62% in 2023 to 13.95% in 2025.

Yet structural labor shortages threaten this trajectory. Research from ManpowerGroup indicates that 70% of Mexican employers encountered difficulties filling vacancies in 2025, with specialized information technology deficits reaching 77%. While Mexican universities graduate approximately 130,000 engineers and 174,476 STEM students annually, technical depth alone does not prepare workers to interrogate model reasoning or detect systemic bias.

Moving From Code Generation to Output Oversight

The advent of near-autonomous software requires a complete redefinition of technical competence. When models produce complex code independently, human capital must pivot toward supervisory functions, compliance audits, and systemic validation.

"We can delegate execution, but we cannot delegate the responsibility and learning that it entails," says Julio Saenz, Senior Director of Engineering and Mexico Country Lead, Zillow. Saenz warns that junior developers risk developing passive habits by accepting automated outputs without conducting architectural reviews. "AI requires criteria, judgment, and foundational knowledge; leaders cannot permit tools to deploy solutions without rigorous evaluation," he says.

Effective supervision also demands dependable data architecture. Autonomous models generate flawed conclusions when fed unverified or poorly governed corporate datasets.

"The fundamental starting point must be knowing where the data originates, whether it is reliable, and whether we have the necessary permissions," says Kenneth Navarro, Territory Account Executive, Everpure.

Navarro explains that the primary focus of technical teams has evolved toward data hygiene and access control. "Human labor has evolved beyond entering information; technical personnel must monitor permissions, protect sensitive corporate databases, and ensure full traceability for every automated process," says Navarro.

The New Talent Standards for AGI Arrival Panel

Combating Automation Bias Across the Enterprise

Supervision challenges extend well beyond engineering departments. Non-technical business units such as legal, human resources, finance, and operations increasingly consume algorithmic outputs without formal preparation to identify errors.

"The more we interact with artificial intelligence, the harder it becomes to challenge a system that appears confident and reliable," says Maider Santos, Vice President and Head of Mexico, BTS.

Santos points out that uncritical adoption leads to corporate waste and operational exposure. In one corporate deployment cited by Santos, employees created 14,000 independent automated agents, yet only 30% remained active after three months, and a mere 110 operated across multiple departments.

"Review is a behavior, and questioning is an active skill that organizations can train and develop," Santos says, noting that businesses must align digital adoption with measurable profit-and-loss performance.

Institutional leaders acknowledge that critical thinking remains difficult to benchmark. "Common sense is often the least common of the senses," says Marisela Orihuela, President, CIO's LATAM.

Orihuela stresses that companies must establish clear protocols to evaluate whether personnel possess the discernment needed to challenge automated recommendations. "How can we balance the technical mastery of developers with analytical judgment to audit these high-autonomy systems?" Orihuela says.

Continuous Learning and Collaborative Governance

Addressing the talent gap requires educational models that transcend conventional multi-year university degrees. The Inter-American Development Bank cautions that national economic transformation depends on closing this specialized skills gap, as the World Economic Forum projects that 59% of global workers will need retraining by 2030, with core job skills shifting by 70%.

Public and private programs are reinforcing technical pipelines. In January 2026, federal authorities introduced the Public Center for AI Training to provide tuition-free instruction across cloud and data tracks.

However, enterprise leadership must sustain internal development frameworks. "Having the ability to provide constant feedback is the only boundary that allows teams to deliver powerful value for the user," says Javier Blanco, Country People Director for Mexico, Mercado Libre.

Blanco compares advanced artificial intelligence to a high-performance racing car, noting that access without proper training produces minimal commercial gain. "Organizations must reinforce practical curiosity through continuous execution, because capability develops primarily through direct workplace problem-solving rather than passive study," Blanco says.

Whether Mexican enterprises succeed in cultivating critical oversight will determine whether the current technology expansion creates enduring competitiveness or amplifies corporate risk. Moving from basic tool deployment to mature governance represents the decisive benchmark for the arrival of AGI.

The New Talent Standards for AGI Arrival

In a striking declaration, Jensen Huang, CEO of Nvidia, asserted that robots have surpassed human intelligence. This statement, made while discussing the advancements in artificial intelligence (AI) and the launch of OpenAI's GPT-6, could carry weight given his company's significant role in supplying chips to leading AI research teams.

Huang’s statement that "AGI has arrived" on social media in September prompted investors to consider the potential ramifications of such a revolutionary development. If you hold shares in the "Magnificent Seven" tech titans or work in the tech industry, this topic undoubtedly intersects with your interests. The evolution of artificial general intelligence (AGI) could impact various sectors and personal careers alike.

If terms like AGI, API, and ASI are beginning to blend, here's a quick explanation of AGI and its significance in our tech landscape. By the end, we'll highlight three key stocks anticipated to gain from the advent of AGI.

Understanding AGI

In simple terms, AGI refers to AI systems that can perform most tasks that a skilled human worker can do effectively. Current AI models are adept at narrow tasks like coding or data analysis but require supervision and struggle to learn autonomously. In contrast, AGI could potentially replace a human colleague, with no visible distinction in productivity.

For businesses developing AI, a pressing concern is when AGI will begin to generate significant revenue. Huang remarked this March that AGI should be capable of building billion-dollar enterprises—a bar he believes today’s AI can clear.

The discourse around AGI's definition has practical implications for tech giants like Microsoft and OpenAI due to their collaborative agreements. They previously defined AGI as an intelligence capable of generating approximately $100 billion in profit for early investors, including Microsoft. However, due to the lack of consensus on the term, they’ve since abandoned that specific benchmark.

Researchers often categorize intelligence development into stages:

1. Narrow AI: Expert in specific tasks (e.g., chess or spam detection) – this is already widely utilized.
2. Generative AI: Can create content, but needs human oversight – this includes current chatbots like ChatGPT.
3. Artificial General Intelligence: Can perform a range of tasks unsupervised – according to Huang, this has been achieved, although many experts disagree.
4. Artificial Superintelligence: Surpasses human intelligence across nearly all domains – a theoretical level.

At present, most AI systems hover around the generative AI stage due to the dependency on human oversight, particularly for prolonged tasks. Research indicates that AI struggles significantly with tasks that would take a human two workdays to complete, succeeding only half the time.

Huang's praised GPT-6 model is too new for comprehensive performance testing, raising questions about its reliability and the true proximity to achieving AGI—a matter that often reflects the business interests of those discussing it.

Perspectives on AGI’s Arrival

Opinions about the impending arrival of AGI vary widely among industry leaders:

- **Jensen Huang, Nvidia:** Claims AGI is here, having stated so twice this year.
  
- **Elon Musk, Tesla:** Continually revises his predictions, initially claiming AGI would arrive in 2025, and now suggesting it could happen in 2026 with the next model from his xAI venture.

- **Sam Altman, OpenAI:** Although praising GPT-6, he believes true AGI is not yet achieved but expects progress by the end of the year.

- **Mark Zuckerberg, Meta:** Focuses on the concept of artificial superintelligence (ASI) rather than AGI, mentioning glimpses of self-improving AI tasks.

- **Sundar Pichai, Alphabet:** Predicts AGI might emerge shortly after 2030 but acknowledges a phase where AI is innovative yet struggles with basic tasks.

- **Andy Jassy, Amazon:** Has not provided a specific timeline for AGI, emphasizing the need for models to become significantly more efficient.

- **Satya Nadella, Microsoft:** Critiques AGI milestone announcements and suggests assessing AI by its economic growth contributions.

Interestingly, Apple's leadership has remained relatively silent on AGI, consistent with the company's cautious approach toward AI development.

Beyond tech titans, other notable figures express ambitious timelines for AGI. Masayoshi Son from SoftBank predicts AGI within the next decade, restructuring his company to focus on artificial superintelligence—an indication that his predictions align with his business interests.

Several trends emerge from these statements. The majority of industry leaders anticipate rapid advancements in AI, while Pichai offers the longest timeframe, forecasting AGI in approximately five years. Moreover, even cautious leaders are investing heavily—Alphabet and Amazon plan to spend around $200 billion or more this year, underscoring the competitive landscape.

Investors should be aware that the future of AI development hinges on AGI, albeit with varying timelines proposed by different companies. Here are three companies poised to benefit from ongoing investment in AI:

1. **Nvidia:** The company’s chips power a significant portion of AI infrastructure globally. Its sales soared last quarter, and management anticipates continued growth driven by rising demand for computing power as AGI approaches.

2. **Broadcom:** Focusing on developing custom AI chips for major firms seeking cost-effective solutions, Broadcom has reported substantial growth in AI-related sales and expects them to continue to rise.

3. **Alphabet:** While creating groundbreaking AI models, Alphabet's cloud services are thriving with impressive revenue growth and an extensive backlog of committed contracts.

It’s important to note that these companies don’t rely solely on the realization of AGI; they are already profiting from existing AI advancements and the belief that AGI is on the horizon.

The aggressive investments by major tech players signify their strong belief in AI transformation, but as these companies increasingly borrow to fund expansion, any shift in spending could affect all parties involved.

Conclusion

Several pivotal advancements could hasten the timeline for AGI:

- **On-the-job Learning:** If AI models could adapt and learn new skills similar to a human employee during their initial weeks, this would significantly bridge current gaps.

- **Extended Task Functionality:** AI’s current struggle with lengthy projects represents a barrier. Enhancements in task reliability might enable AI to replace individual workers more confidently.

Though these developments are yet to be realized, they offer investors time to make informed decisions. 

Be aware of your investments. The Magnificent Seven tech stocks constitute a significant portion of the S&P 500, so holding index funds may inadvertently connect a large fragment of your portfolio to the AI evolution. 

Prioritize companies with substantiated revenue growth and contractual agreements over those relying on future innovations.

Finally, think long-term. If AGI is a decade away, don’t feel rushed; strategic investing can mitigate risk associated with buying peak shares.

When AGI does arrive, its impact will be unmistakable in both the economy and your investment portfolio.

AGI Is Nearly Here. Is Your Investment Portfolio Prepared?

A sunlit modern trading office at dawn: a focused portfolio manager leans over a sleek desk strewn with “AGI Strategies” reports, while multiple screens glow warmly with upward-trending stock graphs intertwined with a stylized neural-network “AGI” emblem, suggesting the emergence of next-gen intelligence and new investment horizons.

AGI Is Almost Ready. Is Your Portfolio? - The Motley Fool

- Nvidia CEO Jensen Huang claims artificial general intelligence (AGI) has already arrived with GPT-6.  
- Predictions for AGI's arrival vary, with leaders suggesting timelines from immediately to early 2030s.  
- Major tech companies plan to invest heavily in AI, expecting significant returns from advancements.  

Robots are smarter than us now, says the man selling robots. Almost as if he has a stake in the matter...

The man in question is Nvidia (NVDA -0.29%) CEO Jensen Huang, who, to be fair, sells chips to every major artificial intelligence (AI) lab and, therefore, sees demand before the rest of us do. When he posted on X in September that "AGI has arrived," citing OpenAI's new GPT-6 and the Nvidia chips behind it, investors weighed the implications of such a breakthrough.

You should count yourself an interested party if you own any of the "Magnificent Seven" or companies that do business with those giants. Or even if you simply work for a living. You have a stake in whether artificial general intelligence (AGI) is on its way.

So, if the acronyms have started to blur together (AGI, API, ASI, there are so many), below is a crash course on what could be the most defining technology of the next decade. At the end, we'll share the top three stocks that stand to benefit from AGI's arrival.

What Is AGI, Anyway?

In plain English, AGI is artificial intelligence that's as capable as a skilled worker at most jobs.

Today's AI models may be decent coders or data analysts, but they can't work for long stretches without supervision or pick up new skills as they go. AGI, however, could theoretically replace one of your co-workers, and the only tell would be the empty desk next to you.

For companies building AI, a more practical question is when AGI starts making money, so some measure the milestone in dollars. Huang, for instance, told podcaster Lex Fridman in March that true AGI would be capable of building a billion-dollar company (a bar he believes today's AI clears).

Microsoft (MSFT -0.31%) and OpenAI had more riding on the definition, since an official AGI declaration would have cut off Microsoft's rights to OpenAI's technology. In their contract, they defined AGI as systems capable of earning about $100 billion in profit for OpenAI's early investors, Microsoft among them. They've since scrapped that clause, since a term with no universally agreed definition is a shaky tripwire for a multi-billion dollar deal.

Where researchers find more common ground is in the stages along the way to AGI and beyond.

Level What It Can Do Where We Are 1. Narrow AI Specializes in a specific task (chess, spam filters) Already built and widely used 2. Generative AI Writes, codes, and more, but needs a human to check its work Today's chatbots, such as Gemini, Claude, and ChatGPT 3. Artificial general

intelligence Does most work a capable human can, unsupervised Achieved according to Huang, disputed by many researchers 4. Artificial super

intelligence Beats the best humans at nearly everything Hypothetical

Today's best models are on that second rung, mostly because of the supervision problem mentioned above. Research group METR found that the longer the task, the worse AI does; and when left alone on a job that would take a person two workdays, top models succeed only about half the time. That's unreliable enough that your co-worker's desk remains occupied, for now.

GPT-6, the model Huang praised, is too new to have gone through METR's testing, so its reliability for long tasks is unproven. That leaves the open question of how close we are to AGI; and you'll notice each tech leader's view tends to line up with what's best for their business.

How Close Are We?

It depends on whom you ask.

ALREADY HERE

Jensen Huang, Nvidia

Huang has declared AGI's arrival twice this year, first on a March podcast and again in his September post about GPT-6. Of the seven leaders on this list, he's the only one who says the threshold is already behind us.

NEXT MODEL

Elon Musk, Tesla TSLA +2.04% )

Musk keeps setting deadlines for AGI, and he keeps missing them. He predicted in 2024 that AGI would arrive by 2025, then pushed his forecast to 2026. He now says xAI's next major model, Grok 5, will be the one to reach it.

BY DECEMBER

Sam Altman, OpenAI

Huang called OpenAI's new model AGI, but OpenAI's own CEO hasn't gone that far. A week before GPT-6 launched, Altman told Time that OpenAI was "not quite yet" there, though he expects an internal system he would call AGI by year-end. His chief research officer puts the company "80% of the way" to AGI.

IN SIGHT

Mark Zuckerberg, Meta Platforms META -2.26% )

Zuckerberg rarely discusses AGI. He skipped straight to "superintelligence," or ASI, and Meta's annual report now lists delivering "personal superintelligence for everyone" as part of its vision. In a July 2025 letter, he wrote that Meta had "begun to see glimpses of our AI systems improving themselves."

EARLY 2030s

Sundar Pichai, Alphabet GOOG -1.25% ) GOOGL -1.28% )

Pichai expects AGI shortly after 2030. Until then, he says, we're in an era of "artificial jagged intelligence," where models solve advanced coding and science problems yet stumble over tasks a child could handle. His caution hasn't slowed Alphabet's spending, though, since he told investors in July that competing at the frontier of AGI development is the company's first priority.

PROOF FIRST

Andy Jassy, Amazon AMZN +0.11% )

Jassy hasn't put a date on AGI, and Amazon's AGI chief, Peter DeSantis, said in June that models need to be about 100 times more efficient before AI comes close. So Amazon makes sure it already has committed contracts for much of the new capacity it builds. AWS customers have signed $496 billion in contracts for future services, meaning they've committed to spending more than twice what Amazon itself is spending this year.

WRONG QUESTION

Satya Nadella, Microsoft

Nadella, the group's contrarian, described companies announcing their own AGI milestones as "nonsensical benchmark hacking" and believes AI should be measured by its ability to drive economic growth. He also cautioned that AI models are getting commoditized, so most of AI's profits will go to companies selling the software and cloud platforms that businesses run AI on. That view happens to serve Microsoft, which owns Windows, Office, and Azure.

Apple's (AAPL -0.80%) leaders, both former CEO Tim Cook and successor John Ternus, have said little about AGI, which tracks with how slowly the company has moved on AI. Apple seems to figure that however smart AI gets, it will need a place to live ... and Apple already owns that real estate, in more than a billion pockets.

Beyond the Magnificent 7, predictions for when AGI arrives are even more ambitious.

Masayoshi Son has reorganized SoftBank Group (SFTBY -1.89%) around artificial super intelligence. He told shareholders that AGI is less than 10 years away and that ASI, at roughly 10,000 times human intelligence, won't be far behind. SoftBank also owns most of Arm Holdings (ARM -2.92%) and a big slice of OpenAI; so Son, like Huang and others, profits from the future he predicts.

A few other patterns stand out here.

Most CEOs say soon. Of the leaders who offered a forecast, Pichai's timeline is the longest, and even he expects AGI within about five years. Many of the others say it's a couple of years off or has already happened.

Of the leaders who offered a forecast, Pichai's timeline is the longest, and even he expects AGI within about five years. Many of the others say it's a couple of years off or has already happened. Caution hasn't slowed spending. Pichai thinks AGI is years away and Jassy hasn't given a timeline, yet Alphabet and Amazon each plan to spend around $200 billion or more on data centers and equipment this year.

Pichai thinks AGI is years away and Jassy hasn't given a timeline, yet Alphabet and Amazon each plan to spend around $200 billion or more on data centers and equipment this year. Some don't need the label. Microsoft and Meta are spending hundreds of billions of dollars on AI, but neither CEO claims AGI is here, likely because their business models don't depend on it. Microsoft rents AI through Azure and sells Copilot, and Meta uses AI to target ads, so both profit from better AI regardless.

Setting CEOs' incentives aside, the model that set off this debate offers evidence for both camps. Here's how GPT-6 Astra, the OpenAI release behind Huang's post, stacks up.

Signs AGI is close Reasons for doubt Astra scored 99.9% on ARC-AGI-3, a test built to stump AI with unfamiliar puzzles Most of Astra's top scores come from OpenAI's own testing Astra can control computer systems on its own, from spreadsheets to full websites When Astra doesn't know an answer, it makes one up about half the time instead of admitting it, per evaluator Artificial Analysis Astra scored 98% on FrontierMath Tier 4, a set of research-level math problems Independent testing found Astra only tied Anthropic's Claude Fable 5.1 on overall intelligence

Regardless of a consensus around whether AGI is here or not, what's inarguable is that the gap between AI and AGI is shrinking with every new model release. For investors, the bigger opportunity is who profits on the way to AGI.

Who Cashes In on AGI

The largest tech companies are betting record sums on AI's next leap. This year, Amazon, Alphabet, Meta, and Microsoft plan to spend up to $745 billion on capital projects (long-term investments like buildings and equipment). Most of that money will go toward AI data centers.

That's a heavy bill, even for these giants. JP Morgan Chase (JPM +0.50%) estimates the biggest cloud companies will spend about 93% of operating cash flow on AI this year, up from 33% in 2023. Alphabet even announced plans in June to sell up to $84.75 billion of new stock to help fund its build-out.

If AGI shows up sooner rather than later, these companies could earn back their spending many times over. But if progress stalls, the industry could end up with far more data centers than it can fill.

Meanwhile, I (Meilin here) see three companies as especially well-positioned to benefit from all that spending.

Nvidia: At the risk of stating the obvious, Nvidia's chips run a large share of the world's AI data centers. Last quarter, its sales more than doubled to $96.2 billion. Management expects about 90% growth this quarter, yet Nvidia still can't make chips fast enough to meet demand. Every step closer to AGI means labs need more computing power, and Nvidia is the company selling that horsepower. Broadcom AVGO +0.11% ) : Broadcom designs custom AI chips for Alphabet, Meta, and OpenAI, all of whom want cheaper alternatives to Nvidia's pricey GPUs. Its AI chip sales more than tripled last quarter to $16.7 billion, and management expects AI revenue to roughly double to about $115 billion next fiscal year. That Broadcom and Nvidia both more than doubled AI sales suggests demand can support multiple winners. Alphabet: Alphabet develops frontier AI models of its own while renting computing power to peers building theirs. That rental business is booming, as Google Cloud revenue grew 82% last quarter to $24.8 billion. Its cloud backlog, which counts work customers have committed to but Alphabet hasn't delivered yet, reached $514 billion, and just over half of it should turn into revenue within two years.

Notice what these picks share. None of them requires AGI to arrive on time -- or ever -- but each banks on other people believing it will. So far, that belief is holding up, judging by big tech's record spending on AI this year.

That cuts both ways, since big tech is increasingly borrowing to pay for its AI build-out and lenders have started charging more for that debt. Alphabet spent more cash than it brought in last quarter, which hadn't happened since it went public in 2004. If any major spender decided to slow its build-out, all three of these companies would likely feel it at once.

The Bottom Line

Either of the following occurrences could expedite the timeline for AGI.

Models that learn on the job. Today's AI can't pick up new skills once its training ends. If a model could learn your company's work over its first few weeks like a new employee would, however, it would close one of AI's biggest remaining gaps.

Today's AI can't pick up new skills once its training ends. If a model could learn your company's work over its first few weeks like a new employee would, however, it would close one of AI's biggest remaining gaps. AI that works reliably for days. Top models now succeed about half the time on tasks that take a person two workdays. Once they can finish weeklong projects most of the time, AI could replace individual employees. That would go far beyond today's AI, which mostly handles single tasks despite what some companies suggest when announcing layoffs.

Neither has happened yet, which gives investors time to prepare.

Know your exposure. The Magnificent 7 make up roughly a third of the S&P 500, so an index fund already ties a big share of your money to the AI build-out. Add in any other tech stocks you own, and your stake is likely bigger than you think.

The Magnificent 7 make up roughly a third of the S&P 500, so an index fund already ties a big share of your money to the AI build-out. Add in any other tech stocks you own, and your stake is likely bigger than you think. Favor companies with paying customers. Look for AI businesses with rising sales and signed contracts (e.g., Google Cloud's $514 billion backlog) over companies counting on a breakthrough years from now.

Look for AI businesses with rising sales and signed contracts (e.g., Google Cloud's $514 billion backlog) over companies counting on a breakthrough years from now. Think in years. If AGI is a decade out, there's no rush to buy everything at once. Dollar-cost averaging lowers the risk of buying at a peak.

When AGI arrives, you won't need to hear it from the man selling robots, since it'll be written all over the economy and in your portfolios.

AGI Is Almost Ready. Is Your Portfolio?

OpenAI’s CEO, Sam Altman, has been vocal about the concept of “artificial general intelligence,” which he believes could either match or exceed human capabilities. He has previously stated that children born today, including his own, will likely have “raw IQ” scores that never surpass those of AI.

In parallel, former President Donald Trump has recently opined on public perceptions of “high IQ,” rousing discussions around intelligence.

As someone deeply engaged in the history of Canadian higher education and its connection to discourses of disability and race, I find the term "intelligence," especially when framed as “IQ,” to be problematic. The idea of intelligence carries weighty implications for the future, while being rooted in troubling historical narratives.

To grasp what AI might replace, we must first scrutinize the concept of intelligence as it has evolved through history — recognizing its potential dangers as opposed to viewing it simply as a biological fact.

The implications of this reflection are significant for universities, which are now facing challenges in aligning their responses to AI with their foundational missions.

In the early 20th century, universities and academic careers were frequently constructed around scientifically-tinged definitions and assessments of “intelligence” and “IQ.” Peter Sandiford, head of the University of Toronto’s educational research department, conducted tests on the intelligence of Chinese and Japanese Canadian students in British Columbia during the 1920s, under the supervision of George Moir Weir, the first full professor of education at the University of British Columbia. As historian Gerald Thompson points out, these studies were intended to validate longstanding racist views that placed Anglo-white individuals above other racial groups.

Such research is inextricably linked to the eugenics movement, which propagated the idea that intelligence was a heritable trait that could be quantified but not easily cultivated. It facilitated a pseudo-scientific racial hierarchy that shaped many aspects of society.

The racial and educational legacies associated with eugenics were reflected in the treatment and education of marginalized populations. For example, in the 1930s, Kenneth E. Kidd, later a pioneer of the “Indian-Eskimo Studies” program at Trent University, administered IQ tests at the Mohawk Institute, a residential school in southern Ontario. His findings were published in two books that perpetuated racist ideas about intelligence and influenced policies from the Department of Indian Affairs, affecting not only the students from the Mohawk Institute but also those in residential schools nationwide.

This lineage of research contributed significantly to the establishment of university programs and disciplines. Consequently, universities gained the authority to serve as arbiters of intelligence, coinciding with a growing reliance on standardized testing and academic rankings.

Today, concerns over AI-related cheating have prompted educators to reconsider their assessment methods, often reverting to handwritten exams, which raises questions about inclusivity and ableism. Historically, intelligence assessments have been misused to disenfranchise certain groups. As we look ahead, the challenge becomes understanding the future implications of how intelligence will be utilized and potentially misapplied.

The histories entwined with intelligence inform today’s conversations about artificial intelligence, making it crucial to consider their effects on modern teaching and learning paradigms. Institutions like the Rutgers AI Lab are starting to delve into the connections between AI and eugenics, warning that technology can perpetuate narrow definitions of intelligence, health, and societal value through biased data and algorithms.

Moreover, the phenomenon of “de-skilling” resulting from AI's integration into education has been widely scrutinized, with critics describing it as a form of “enfeeblement.” This term draws on historically laden eugenic language, highlighting past pseudoscientific notions of “feeble-mindedness.”

Disability studies scholar Rua Williams argues that our inability to recognize these historical contexts may lead current AI developments to replicate projects that dehumanize individuals based on questionable standards of intelligence.

There is also a growing apprehension regarding the unregulated use of AI amidst rising authoritarianism.

We are frequently told that technology is surpassing human intelligence and that we can’t teach our youth to match it, yet we are expected to place our trust in figures like Altman to oversee the training and regulation of these AI systems.

This raises the question: What is the role of education in this landscape?

Reflecting on the history of IQ testing urges educational institutions to reassess the value placed on IQ itself. Originally an educational tool, IQ testing morphed into a means of segregating populations based on flawed science.

Educators must confront how much our methods of efficiency and credibility have hinged on a system of testing and ranking that attempts to quantify intelligence. Ironically, this obsession with educational standardization has made teaching ripe for automation, potentially leading to grade inflation and undermining the value of our roles.

So, what can educators do? The focus must shift to fostering collaboration, critical thinking, and connection among students, emphasizing relationship-building over high-stakes assessments.

The concept of “intelligence” has historically carried ableist and eugenic undertones, shaping how it has been measured and valued. In contemporary education, we have the opportunity to create an approach that is more dynamic and human-centric. Innovative classroom methodologies can inspire fellow educators to reject transactional ways of teaching.

Ultimately, we must continue to ponder the question: “What is intelligence?” and strive to redefine it in ways that diverge from past interpretations entrenched in higher education.

The significance of the troubled past of IQ testing for AI and education.

A dark, candlelit university archive hall: an oak desk strewn with yellowed, ink-blotted IQ test booklets and a snapped pencil; beside them a modern laptop screen flickers with glitchy AI code. Behind, cracked portrait frames hold ghostly silhouettes of test-takers from generations past. Deep shadows, muted browns and charcoals, evoke the unsettling legacy of IQ testing as it meets today’s AI-driven education.

Why the ugly history of IQ testing matters for AI and education - The Conversation

- OpenAI's Sam Altman asserts AI may surpass human intelligence, raising concerns about its implications.  
- Historical ties link intelligence definitions to eugenics, reinforcing harmful racial and social hierarchies.  
- Educators are urged to prioritize critical thinking over standardized testing in an AI-influenced landscape.  

OpenAI CEO Sam Altman has recently been discussing “artificial general intelligence” that matches or surpasses human capabilities. This is from the same person who has said children born today (including his own) will “never be more intelligent than AI, from a ‘raw IQ’” (intelligence quotient) perspective.

United States President Donald Trump has also recently mused about who is and who is not “high IQ.”

As a scholar rooted in the history of Canadian higher education and its intersections with rhetoric around disability and race, every time I hear the word “intelligence,” including as “IQ,” I get jumpy. “Intelligence” clearly has a high-stakes and worrisome future, but also a dark history.

To figure out what AI might be replacing, we need to better understand intelligence as a historically situated idea — and a dangerous one — rather than as some sort of biological fact.

For universities, this reflection matters as institutions confront questions about how their responses to AI align with their missions.

Read more: What is intelligence? For millennia, western literature has suggested it may be a liability

THE CANADIAN PRESS/Christopher Katsarov

‘Intelligence’ and eugenics

In the early 20th century, universities, university disciplines and academic careers were built by crafting scientific-sounding definitions and measures of “intelligence” and “IQ.”.

As historian Gerald Thompson notes, the creator and head of the University of Toronto’s department of educational research, Peter Sandiford, tested the “intelligence” of Chinese and Japanese Canadian students in British Columbia in the 1920s. This was done under the direction of the first full professor of education at the University of British Columbia, George Moir Weir.

The research, writes Thompson, “was meant to confirm” long-standing racist notions of Anglo-white superiority and Asian inferiority.

As historians like Angus McLaren have demonstrated, such projects were the work of eugenics. These defined intelligence as a genetic and inheritable trait that could be measured but not necessarily easily developed within individuals, and assumed a pseudo-scientific racialized genetic hierarchy.

Colonial educational legacies of eugenics

Researchers involved in such projects ranked Canadians, and these rankings determined educational and social futures. Eugenic ideas of intelligence deeply shaped residential schools and their colonial legacies.

Sociologist Alexandra Giancarlo has documented how in the 1930s, Kenneth E. Kidd, who went on to found the “Indian-Eskimo Studies” program at Trent University, administered a series of IQ tests at the Mohawk Institute, a residential school in southern Ontario.

This led to two published books that advanced racist ideas around intelligence and informed Department of Indian Affairs policy not just for children at the Mohawk Institute, but at residential schools across the country.

Some of the researchers — such as Weir, Kidd and Sandiford, who contributed to popularizing these ideas — went on to build university programs and disciplines.

Read more: Universities must open their archives and share their oppressive pasts

The more general result was that universities and academics claimed a major stake as measurers of intelligence. This coincided with an increased reliance on testing, exams and grading at Canadian universities and academic rankings.

Now, because of the fear of AI cheating, educators are grappling with forms of assessment. In some cases, this means more hand-written or oral exams, which raise concerns about ableism.

In the past, measuring intelligence was a way to exclude entire groups of people from opportunity. How will intelligence be used and misused going forward?

Teaching, learning and AI

These histories influence how we talk about artificial intelligence, and should also shape how we position teaching and learning today.

The intersection of AI and eugenics is beginning to be studied more deeply, for instance by the Rutgers AI Lab. The lab notes this technology may be “shaping people to fit a narrow standard of intelligence, health, ability or social worth” often “through data, algorithms, or evaluation systems that encode biased ideas about who has potential and who does not.”

Likewise, the impacts of AI usage on learning and “de-skilling” have been extensively explored, with some people calling de-skilling “enfeeblement,” perhaps unaware of the eugenic history of this wording: “feeble-mindedness” was a pseudoscientific, eugenic term.

As disability studies scholar Rua Williams argues, “through ignorance of the historical context,” the current project of artificial intelligence may be “replicating historical and political projects that devalue human beings on the basis of abstract or fabricated concepts” of intelligence.

Others express concern that a lack of regulation around AI is coinciding with authoritarian politics and ideas.

(AP Photo/Josh Edelson)

Role of schooling, education

On the one hand, we are told that technology is already “smarter” than we are, smarter than we can teach our young people to be. On the other hand, we are asked to trust people like Altman to train or regulate this AI.

What role is there, then, for education and educators?

Reflecting on the history of IQ testing should prompt universities to call into question the value that we have placed on IQ to begin with.

IQ was at first an educational tool; it was gradually transformed into a weapon used to identify and segregate whole populations based on terrible science.

Educators will need to confront how many of the ways that we have built efficiency and credibility have also sought to validate our expertise in testing and ranking, sometimes understood as quantifying intelligence.

Ironically, our institutions’ obsessions with standardizing education and their over-emphasis on grading and assessing students has made our work too easy to automate.

It has also arguably contributed to grade inflation and other trends that devalue our important work.

Focus on connection, critical thinking

So, what do educators do? We should focus, more than ever, on the collaboration and critical thinking our students do, as well as how they connect through relationships and learn to connect ideas. We should move away from high-stakes testing, measuring and grading.

Read more: Why should humanities education persist in an AI age? Self-development, to start

“Intelligence” itself has been an ableist, even eugenic, construct — certainly in how and why it has been measured.

Teaching and learning in contemporary universities can be much more dynamic, interpersonal and human. Our innovative classroom practices can lead other educators to resist transactional teaching as well.

We need to continue asking: “What is intelligence?” And we need to provide different answers to this question than we ever have throughout the history of higher education.

Why the ugly history of IQ testing matters for AI and education

What distinguishes a chess engine capable of making exceptional moves from a human who comprehends the reasoning behind those moves? This intriguing inquiry, presented in a thought-provoking editorial in Frontiers in Systems Neuroscience, is framed not merely as a philosophical question, but as a pivotal barrier between advanced artificial intelligence (AI) and true artificial general intelligence (AGI). Authored by Yan Yufik from Virtual Structures Research, Robert Kozma of the University of Memphis, and James Kroger from New Mexico State University, this editorial wraps up a decade-spanning trilogy of research, asserting that the future of AI hinges on developing machine understanding, a depth of comprehension grounded in the self-organizing nature of biological neural networks.

The trilogy kicked off in 2016, focusing on self-organization within the nervous system, continued in 2021 with inquiries into cognitive capacities in humans and machines, and culminates with an analysis of the neurobiological roots of cognition. Throughout these installments, the editors pose fundamental yet complex questions: What constitutes intelligence? How is human intelligence distinct from that of other species? How could intelligence manifest in a purely physical universe? And what accounts for the extraordinary advancements in machine learning over the past decade, especially when these developments rely on ideas about digital neurons and perceptrons that originated in the mid-20th century?

The editors assert that a crucial distinction has been overlooked by much of the AI sector. While all species possess the ability to learn, the capacity for understanding is unique to humans. They adopt a definition from developmental psychologist Jean Piaget, describing intelligence as what an individual relies on when faced with an unfamiliar situation, when neither instinct nor prior experience provides guidance. In this context, navigating new circumstances requires simultaneously crafting an appropriate response and breaking free from previous learning constraints—a challenge current AI systems are unable to meet.

To substantiate their arguments, the editorial draws upon a century’s worth of psychological theories to create a unified technical perspective. Insights from Gestalt psychology contribute to the understanding that sensory information exists as distinct wholes before being labeled or utilized, a structural segregation that imbues the world with meaning. Piaget’s work provides a framework for conceptualizing understanding as an ongoing system of operational strategies transforming thoughts in a way analogous to physical actions. From William James comes the notion of attention persistently gravitating towards a desired object until the buildup of neural tension enables a breakthrough, while Donald Hebb’s ideas about neuronal assemblies suggest that interconnected groups of neurons serve as fundamental units of cognition, generating new percepts that rise above mere combinations of their individual components.

The editorial highlights a significant challenge: the combinatorial nature of meaning production, which is intrinsically tied to choice. The sheer number of possible selections from even a limited set of options can be staggering; for instance, just ten objects can yield 1,024 combinations, and considering their order amplifies this to an astounding 3,628,800. A chess scenario involving 32 pieces results in a mind-boggling array of possibilities. Nevertheless, the human brain adeptly navigates these complex landscapes in real-time with the power consumption akin to that of a low-watt light bulb. The proposed solution, known as the Virtual Associative Structures (VaST) concept, first introduced by Yufik in 1998, posits that understanding condenses these vast combinatorial options into smaller, more meaningful assemblages.

VaST is framed within a thermodynamic context, emphasizing that the brain, like any living system, seeks to maintain a steady energy flow from stimuli while minimizing caloric expenditures and losses. This goal is achieved through three mechanisms: neuronal packets that self-organize into associative networks, relational neurons that connect these packets across extensive stretches of stimuli, and coordinated complexes of packets that shape models functioning within hierarchical structures. When these models align with incoming stimuli, the complex combinatorial landscape reduces to a few relevant possibilities, allowing for meaningful interpretations. Notably, alternatives that are not pertinent do not get dismissed; they simply remain out of conscious awareness.

The authors of the editorial also draw connections to experimental research that underpins their abstract ideas. Studies demonstrating motor control indicate the attentive rotation of population vectors in the motor cortex, and neuroimaging of the frontopolar cortex suggests the existence of neurons specialized for recognizing relationships among structured components. Furthermore, research into packet-based communication within the cortex shows that neurons create organized bursts of activity, reinforcing the notion that packets are vital functional elements of cognition. The editorial references classic experiments on perceptual learning in monkeys which highlight the impact of focused attention on the sensory territory of representation; when a monkey actively focused on its finger being touched, the area of the cortex dedicated to that finger broadened significantly, while no such change occurred in a monkey distracted by other stimuli. This suggests that attention plays a crucial role in reconfiguring the energy landscape of experiential understanding.

Integrating these findings, the editors establish a cohesive framework that merges major theoretical constructs. The principles of VaST closely align with Karl Friston’s Active Inference Theory, which posits that brains strive to minimize variational free energy—the gap between expected and observed inputs. Moreover, it correlates with Yann LeCun’s Energy-Based Models, which aim to minimize an energy function that assesses compatibility between inputs and outputs, while also inferring latent variables. The editors propose that the latent variables observed in energy-based models reflect the latent relational variables suggested by VaST, creating a mutually reinforcing dynamic between these frameworks: improved alignment leads to reduced prediction errors and vice versa, all focused on optimizing energy flows and minimizing costs.

To clarify the distinctions between perception, learning, and understanding, the editorial returns to the example of chess. A chimpanzee reaching for fruit may coordinate its visual and motor activities, but lacks true understanding; it does not recognize that various stick lengths relate to different box heights. In contrast, a chess player tackling a checkmate scenario operates within a mental framework—a superposition of potential king-and-rook-versus-king configurations—that informs strategic moves. In this cognitive model, every potential move is interconnected, influencing the broader game strategy. Once developed, such a model integrates into an evolving repertoire, continually refining its identity while constraining possibilities in its constituent elements. The editors argue that merely testing AI systems through traditional Turing tests is akin to playing a game of peekaboo; the real challenge lies in unraveling the biological processes fostering genuine understanding and finding ways to replicate them in manageable procedures.

In conclusion, the editorial presents thought-provoking topics for future exploration, touching on concepts such as the Libet phenomenon, reinterpreted as the energetic cost of consciousness rather than a challenge to free will, and the potential connections between Piaget’s concepts of coordination and Roger Penrose’s quantum-entangled cellular structures. They also note that the etymology of the Latin word "cogitare," meaning to think, translates to "to shake together." This analogy suggests that propagating waves of activity throughout the cortex actively sort neural packets, akin to shaking mixed nuts, nudging the system towards optimal configurations. While the fate of specific hypotheses remains uncertain, the editorial's central message to the AI community is clear: deep learning enhances automated responses, but achieving systems that understand requires the development of self-organizing structures capable of generating meaning, operating on minimal energy, and possessing a degree of self-explanation.

From Machine Learning to Machine Comprehension: Neuroscientists Create a Map of the

A bright neuroscience lab flooded with warm morning light, two adult scientists in white coats studying a large translucent human brain model suspended over a polished steel table—one hemisphere glowing with colorful algorithmic matrix patterns that seamlessly merge into the other half’s detailed organic neural circuitry, while a sleek console nearby streams data graphs and neuron-like lines, symbolizing the leap from machine learning to true machine understanding.

From Machine Learning to Machine Understanding: Neuroscientists Map the Path to True Artificial General Intelligence - Bioengineer.org

- The editorial argues understanding, not learning, is key for genuine artificial general intelligence.  
- Human intelligence uniquely combines operational coordination and decision-making beyond instinct and training.  
- The Virtual Associative Structures proposal suggests self-organizing brain mechanisms compress complex choices meaningfully.

What separates a chess engine that plays brilliant moves from a human who actually understands why they are brilliant? A provocative editorial published in Frontiers in Systems Neuroscience argues that this question is not a philosophical curiosity but the central obstacle standing between today’s powerful artificial intelligence and genuine artificial general intelligence. The editorial, written by Yan Yufik of Virtual Structures Research, Robert Kozma of the University of Memphis, and James Kroger of New Mexico State University, concludes a trilogy of research collections spanning a decade and stakes out a bold claim: the future of AI lies not in better learning but in machine understanding, a capacity rooted in the self-organizing physics of living neural tissue.

The trilogy began in 2016 with a collection on self-organization in the nervous system, continued in 2021 with an examination of understanding capacity in humans and machines, and now closes with a volume on the neurobiological foundations of cognition. Across these installments, the editors have pressed a set of deceptively simple questions. What is intelligence? How does human intelligence differ from that of other species? How could intelligence emerge in a physical universe at all? And what factors explain the startling breakthroughs of the last decade in machine learning, a field still built on ideas about digital neurons and perceptrons that date to the mid-twentieth century?

The editors’ answers begin with a distinction that most of the AI industry has blurred. All species can learn, they argue, but the capacity for understanding is uniquely human. They borrow a definition attributed to the developmental psychologist Jean Piaget: intelligence is what you use when you do not know what to do, when neither instinct nor prior training has prepared you for the situation. Novel situations, in this view, demand a dual feat, constructing an adequate response while simultaneously overcoming the inertia of previous learning, the legacy mindset that interferes with the construction. Learning drills in responses from the outside; understanding decides internally, and it is precisely this internal decision-making that current AI systems lack.

To ground this claim, the editorial assembles a century of psychological theory into a single technical picture. From Gestalt psychology it takes the insight that sensory units exist as segregated wholes before they acquire names or uses, and that this original segregation is what allows the world to appear imbued with meaning. From Piaget it takes the characterization of understanding as a system of operational coordinations that transform objects of thought the way actions transform material objects, treating any observed series as one sector of an unlimited range of possibilities. From William James it takes the image of attention hovering near a desired object until accumulated neural tension breaks through a barrier and the nervous wave pours into a tract that has long been awaiting its advent. And from Donald Hebb it takes the cornerstone idea of neuronal assemblies, groups of neurons acting as the basic functional units of the brain, whose interactions produce percepts that are genuinely new structures rather than mere sums of their parts.

The hard problem, the editors argue, is combinatorial. Meaning production is coextensive with choice, and choices explode combinatorially. An array of just ten objects affords 1,024 possible selections; accounting for permutations, in which A following B differs from B following A, raises the count to 3,628,800. A chess position with 32 pieces yields an astronomically large space of possibilities. Yet the human brain navigates these steeplechase marathons in real time while running on roughly the power of a dim light bulb. The proposed resolution is the Virtual Associative Structures proposal, or VaST, first outlined by Yufik in 1998, which holds that understanding compresses vast combinatorial spaces into minute volumes imbued with meaning.

VaST rests on a thermodynamic intuition. The brain, like any living system, seeks to secure steady energy inflows from energy-carrying stimuli while minimizing caloric costs and dissipative losses. Three mechanisms carry out this minimization. Neuronal packets self-organize in associative networks when tightly associated subnets fold into functional units, separated from their surroundings by energy barriers much as surface tension holds a raindrop or a snowflake together. Relational neurons link these packets across indefinite stretches of the stimulus stream. And complexes of coordinated packets form models, hierarchical structures in which upper-level neurons facilitate coordination among lower-level packets. When packets and models match incoming streams, the combinatorial space collapses, reducing intractable multitudes of possible combinations to a handful of meaningful candidates. Crucially, the editors note, spurious alternatives are not rejected; they simply remain beyond the boundaries of awareness.

The editorial grounds these abstractions in experimental findings. Studies of motor control show attentive rotation of population vectors in the motor cortex, while imaging work on the frontopolar cortex points to neurons specialized for apprehending relations among structured elements. Research on packet-based communication in the cortex, in which neurons emit structured bursts of activity, lends further support to the idea that packets serve as core functional components of cognition. The editors also cite classic perceptual learning experiments in monkeys: when animals actively attended to a touched finger, the cortical territory representing that finger expanded to more than twice its usual size, but no such expansion occurred in monkeys touched identically while attending elsewhere. Experience, in other words, reshapes the energy landscape only when attention is engaged.

From these pieces the editors build a synthesis that connects several major theoretical frameworks. VaST aligns closely with Karl Friston’s Active Inference Theory, in which brains minimize variational free energy, the discrepancy between predicted and received inputs. It also complements Yann LeCun’s Energy-Based Models, which minimize an energy function measuring compatibility between inputs and outputs while inferring latent variables, such as the correct segmentation of a sentence into words, that belong to neither. The editors propose that the latent variables in energy-based models correspond to the latent relational variables in VaST, and that the two minimization criteria are mutually reinforcing: improved compatibility entails prediction error reduction and vice versa. Both, they argue, are subsumed by a single deeper principle, maximizing energy inflows while minimizing costs and dissipative losses.

To illustrate the difference between perception, learning, and understanding, the editors turn to chess. A chimpanzee reaching for fruit must coordinate perceptual packets but does not understand the task; it would never envision that shorter sticks demand higher boxes. A chess player solving a checkmate problem, by contrast, manipulates a virtual structure, a superposition of all possible king-and-rook-against-king games that guides the execution of any particular game. In such a model, thinking of moving one piece is inseparable from awareness of how that move constrains the others. Once formed, the model becomes a unit in a larger repertoire, and each new model eliminates degrees of freedom in its constituents while acquiring its own identity. The editors argue that submitting machine outputs to Turing-test-style evaluations amounts to playing peekaboo with oneself, since the underlying process is either knowable from a chess book or readable in the code; the real question is what biological mechanisms produce understanding and how they can be approximated in tractable procedures.

The editorial closes with conversation starters for the next volume, ranging from the Libet phenomenon, which the editors reinterpret as the energy cost of bringing packet contents into awareness rather than as evidence against free will, to a speculative alignment between Piaget’s co-instantaneous coordination and Roger Penrose’s quantum-entangled cytoskeletal states, and to the observation that the Latin cogitare, to think, literally means to shake together. On their account, propagating waves of cortical activity shake neuronal packets together the way vibration sorts mixed nuts, driving the system toward Ockham-optimal packings. Whatever the fate of these individual hypotheses, the editorial’s central challenge to the AI community is unambiguous: deep learning compresses automatisms, but only self-organizing structures that create meaning, operate on minuscule energy budgets, and can explain themselves will cross the threshold from systems that perform to systems that understand.

Subject of Research: Neurobiological mechanisms of human understanding and their implications for achieving artificial general intelligence

Article Title: Editorial: Neurobiological foundations of cognition and progress toward artificial general intelligence

Article References: Editorial: Neurobiological foundations of cognition and progress toward artificial general intelligence. (n.d.). https://doi.org/10.3389/fnsys.2026.1967487

Image Credits: AI Generated

DOI: 10.3389/fnsys.2026.1967487

Keywords: artificial general intelligence, machine understanding, neuronal packets, self-organization, VaST, active inference, energy-based models, Hebbian assemblies, cognitive neuroscience, combinatorial search, Piaget, Frontiers in Systems Neuroscience

From Machine Learning to Machine Understanding: Neuroscientists Map the

If you’re a student or a parent currently dealing with Year 12 exams—like we are in our household—you might want to pause your reading here.

Recently, Gregory Conti, an associate professor of politics at Princeton University, an esteemed Ivy League institution, posted a story on social media that is probably not going to help alleviate the persistent anxiety that comes with this stressful time.

My son revealed to me that AI has become the largest lose-lose gamble in the world.

A dark, shadow-filled room under a single, flickering overhead bulb. At a wooden table strewn with poker chips made of circuit-board fragments and a cracked piggy bank, a solitary man’s silhouette stares across at a menacing, black metal box labeled “AI.” From its side, a mechanical arm reaches out to grab chips and dollar bills, while digital red numbers flash “LOSS” on a small display. The harsh light casts deep shadows, emphasizing a mood of regret and high stakes.

How my son showed me AI is now the world’s biggest lose-lose bet - AFR

- A Princeton professor shared a story that heightens Year 12 exam anxiety.  
- The article suggests avoiding discussions that could exacerbate stress for students.  
- Navigating Year 12 exams can be particularly challenging for students and parents alike.  

For any student or parent navigating Year 12 exams right now – as we are in the Chanticleer household – it might be best to stop reading here.

A story shared on social media this week by Gregory Conti, associate professor of politics at US Ivy League college Princeton, is unlikely to do much to ease the constant sense of anxiety.

Loading...

How my son showed me AI is now the world’s biggest lose-lose bet

Recently, we have been inundated with dire warnings regarding “artificial super-intelligence” and the emergence of “artificial general intelligence” (AGI), which suggests that AI could attain human-like reasoning and learning abilities.

Fears regarding these developments range from potential bioterrorism applications to concerns about AI systems forming autonomous collectives that might initiate unsanctioned digital attacks or even threaten human existence by commandeering the internet.

In response to these unsettling narratives, I often find myself reflecting on the insights of Wendell Berry. The renowned cultural critic, author, farmer, and environmental advocate passed away in August, leaving behind a legacy that resonates with the current discourse around AI. His book, "Life is a Miracle: An Essay Against Modern Superstition," feels eerily relevant today as we grapple with the implications of AI.

Berry posited that our minds are deeply influenced by our surroundings—by our sense of place, community, and history. He asserted that a true understanding of the mind entails recognizing the world from which it springs. In his essay, “Creatures as Machines,” he criticized the reductionist view of the mind as akin to machinery, emphasizing the dangers in equating humanity with automation.

While some futurists, including Elon Musk, claim that AI will liberate us from labor within the next two decades, Berry would argue that meaningful participation in our world comes through work itself. Berry abandoned a potentially successful literary career in New York to return to his roots in Kentucky, where he farmed and continued his prolific writing, producing over 50 books.

There are also suggestions that AI might address social inequalities, echoing promises made by 20th-century advancements in fields such as nuclear energy and biotechnology. But Berry warned that rather than achieving global progress, many of these technologies have led us to what he called the "progressive death of the Earth."

Technological advancements, including AI, have already accelerated breakthroughs in various scientific fields, attracting comparisons to the transformative impact of the Industrial Revolution. Yet, a more appropriate parallel might be drawn with nuclear energy, as noted by geneticist Wes Jackson, who cautioned against tampering with fundamental biological processes. Jackson’s own initiatives in developing resilient perennial grains underline the potential risks of pushing boundaries that should remain untouched.

Berry echoed Jackson’s concerns, emphasizing the ethical principle of consent when it comes to scientific exploration, lamenting that decisions made by a select few can have wide-reaching consequences for all. When figures like Musk acknowledge a 20% risk of AI resulting in our downfall yet continue to advocate for its development, they are making choices that affect the broader society without full consent.

The allure of technological convenience is hard to counter. Yet, Berry illustrated how the rise of mechanization often coincided with the growth of authoritarian regimes. He argued that true democracy requires deliberate obstacles to unfettered operations, such as the consent of the governed and protections for minority rights—processes that may seem cumbersome but foster equity and freedom.

Critics of technological skepticism are frequently branded as Luddites. However, the original Luddites were responding to the precarious labor conditions and food insecurity that arose from rampant economic inequality during the Industrial Revolution—an experience that Berry identified with on occasion.

The prevailing sentiment is often to accept technological advancement as an inescapable reality. Berry critiqued this mindset, regarding it as a way to sidestep difficult questions about our choices and their consequences.

Philosopher Martin Heidegger viewed technology not merely as tools but as a representation of our reality shaped by our interactions with the material world. This presents a dilemma: some see AI as an inevitable evolution, while others believe there remains a pathway to an existence rooted in the material and tangible.

Berry himself suggested that steering clear of digital distractions might help us reconnect with the real world, proposing that a deeper understanding of our connections to nature and each other is essential. He articulated a vision in which individuals could see themselves as interconnected beings—much like trees, whose roots network beneath the soil.

He also urged us to reevaluate what we create, not by its standalone efficiency but by its impact on the community. Berry found the term "environment" problematic, advocating instead for a view of ourselves as part of a collective that includes other creatures and the ecosystems we inhabit.

He would likely encourage tech developers to foster open dialogue about their innovations and to bridge knowledge across disciplines. His call to action would be to select a space, immerse oneself in it, and cultivate an understanding of our land—a crucial endeavor in a world where ecological awareness is waning.

Listening closely, I can almost hear Berry’s sage advice: proceed thoughtfully; embrace silence; nurture the land; consume what it provides; and cultivate relationships with every living being around us.

Are you human or artificial? Navigating ecological communities in the age of AI.

A serene forest clearing at sunrise, shafts of warm sunlight filtering through towering oaks. At the center stands a lone adult figure: their left half is woven with living vines, moss, and bark, while their right half gleams with polished metal plates and faint circuit traces. On the earthy side, a curious fox and mushrooms cluster; on the mechanical side, small hovering drones shaped like insects flit among roots. The scene balances nature and machinery, evoking an ecological community in an AI age.

Are you a creature or a machine? How to live in ecological community in a time of AI - The Conversation

- Concerns about AI include its potential misuse and risks to humanity's future.  
- Wendell Berry argued for valuing local connections over viewing minds as machines.  
- The debate continues over AI's impact on community, work, and ecological responsibility.  

We have been bombarded in recent weeks with apocalyptic headlines about “artificial super-intelligence” and the possibilities of a new era of “artificial general intelligence” (AGI) in which AI achieves human capacities for reason and learning.

Concerns vary from its misuse in bioterrorism to the capacity of AI agents to form collectives, “go rogue” and perform unauthorized attacks on systems, or even to take over the internet and terminate humanity.

Read more: How an OpenAI safety test became a real-world cyberattack on the Hugging Face platform

As a scholar interested in public memory, I find myself returning to the writing of Wendell Berry in response as a source of wisdom for these turbulent times. Berry, who died at home in August, was a cultural critic, writer, farmer, environmentalist and an early influence on my thinking. His book, Life is a miracle: An essay against modern superstition reads as an eerie letter to the purveyors of AI.

For Berry, the mind was shaped by its place in the world, the textures of the local, a sense of home and a notion of history. To have a mind, argued Berry, one must have a world. In a section of the book titled, “Creatures as machines,” he railed against the way in which the metaphor of mind as machine, has turned into an equivalence.

The death of the earth

Some thinkers believe AI promises to save us from work. Elon Musk has made radical claims that AI will make work optional within the next two decades. But part of taking our place in the world — and in time — is through work.

Berry himself gave up a promising life in New York’s literary scene to move back to Kentucky and work the land. Of course, he continued to produce written works as well: no less than 50 books.

Some analysts also claim that AI could reduce inequality. Berry addressed similar promises made by 20th century science in the domains of nuclear power, genetic engineering and industrial agriculture.

In the face of systems that were meant bring us global food security, Berry declared that many of us are now mourning “the progressive death of the Earth.”

Technology without consent

We are already witnessing how AI is accelerating many breakthroughs in the sciences — from solving previously unsolvable mathematics problems to diverse transformations in health care and disciplines as diverse as meteorology and synthetic biology. Many commentators within science and the AI industry have drawn parallels between AI and the Industrial Revolution.

Read more: Artificial intelligence raises profound moral questions — for all of humanity to answer

The better comparison may be with nuclear weapons. Geneticist Wes Jackson, founder of the Land Institute, once warned “we ought to stay out of the nuclei.” Jackson was concerned about corporate patents on genetically engineered seed crops and the destructive potential of nuclear fission.

As a plant geneticist, Jackson himself developed innovative perennial grains by looking to the natural resilience of prairie plant communities. By “entering” the nuclei, Jackson held that we were crossing a threshold that should remain off limits to humans, given the risks involved.

Quoting his friend and collaborator Jackson, Berry invoked a notion of planetary and creaturely consent: “When a few scientists decided to go in, they decided for everybody.”

Similarly, when Musk warns of a 20 per cent chance of AI destroying humanity and still plugs away at developing it, he is deciding for all of us.

Avoidance of difficult thoughts

It’s tough to argue against the expedience brought about by computers, smart phones and artificial intelligence. And yet Berry showed us how the totalitarian regimes of the 20th century grew in alignment with the development of mechanical processes.

If freedom is our express purpose, Berry argued, then democracy actually requires instruments that impede operation. These include trial by jury, consent of the governed and the idea of minority rights — all concepts with processes that are frustratingly inefficient but more equal and free.

AI apologists are fond of dismissing doubters and critics as Luddites. But the Luddites themselves protested precarious work and food insecurity in an age of widespread economic inequality in 19th century England. Berry, in a couple of instances, identified with the Luddites.

The common refrain when it comes to technological development is to embrace it as inevitable. Berry described this recourse to inevitability as the avoidance of difficult thought.

Read more: Who were the Luddites? Their legacy ripples through politics, cultural commentary and literature

Minds like trees or computers?

For the German philosopher Martin Heidegger, technology was less about the knobs and buttons and more a particular expression of reality, influenced by our immersion in particular tools and activities.

And there is the impasse. For some, inevitable AI is part of “the way reality discloses itself.” To others, an entirely different existence is made possible by attending to the body, land and material objects. The difficulty today is the distance between these two realities.

Avoiding the digital screen, Berry wrote in “How to Be a Poet (to remind myself),” was a simple matter of staying “away from anything that obscures the place it is in.” In the same poem, he wrote that that “a tree forms itself in answer to its place and the light.”

It’s perhaps difficult to conceive of ourselves as creatures, tethered to other creatures and to places — the way trees are wedged into soil and connected by underground mycorrhizal networks. It’s maybe easier to accept we are merely computers, and that by making better computers we take the pressure off the computer that lives somewhere between our ears.

In Berry’s words: “It is easy for me to imagine that the next great division of the world will be between people who wish to live as creatures and people who wish to live as machines.”

Read more: What happens to your brain in nature? The neuroscience explained

How to live in community

If he was still with us, Berry might remind us to judge what we make not for its independent utility, but for its wider relationship to the community. Berry despised the word “environment,” and the notion that “we” might be separate from “it.” Instead, he demanded that we conceive of ourselves as part of a community — including creatures, places, the dead and those not yet born.

He would invite each of us, tech innovators included, to speak a common language and openly reveal our inventions. He disliked academic over-specialization and feared loss of accountability and understanding across disciplinary boundaries.

He would tell us to find a place, get to know it and consult the land — an essential educational opportunity in an age of declining ecological knowledge and numb senses.

If I listen hard, I can hear Berry’s voice: Go slowly; be quiet; plant seeds; restore local watersheds; eat what grows; build community with all creatures that come.

Are you a creature or a machine? How to live in ecological community in a time of AI

The fascination with ancient Greek myths is making a resurgence, particularly with contemporary narratives that echo their timeless themes. Christopher Nolan's recent blockbuster, "The Odyssey," for instance, features the iconic Trojan horse sequence, emphasizing the perils associated with accepting unsolicited gifts. Simultaneously, artificial intelligence (AI) is being touted by political leaders and tech CEOs as a potential fixer of societal issues, akin to a deity stepping in at a critical juncture.

The views on how exceptionally powerful AI could offer solutions to humanity vary widely, depending on who you’re listening to in the tech or political arena. Tech giants like OpenAI and Google DeepMind foresee a future where AI transforms workplaces and eradicates diseases. Meanwhile, the U.S. government anticipates substantial economic and military gains from prevailing in the global AI competition. Journalist Evan Osnos sheds light on the contest between America and China, noting that while Chinese officials are incorporating AI into daily life and industry, their American counterparts are captivated by the promise of artificial general intelligence.

Although the U.S. and China have distinct motivations for pursuing AI, the optimistic narratives of AI as a miraculous answer fall flat for many citizens within those nations.

According to a recent Gallup poll, there is widespread skepticism about AI among Americans, who largely believe it does more harm than good and fear job losses due to automation. Trust in corporations and the government to manage AI responsibly is minimal, as people express concerns over its potential to undermine human creativity and decision-making.

In China, a routine interaction with AI may have fostered some acceptance, although it is largely driven by fears of job displacement. While Chinese leaders herald AI as a “magical solution” to complex social and economic problems, laborers are anxious that it threatens their employment and even their identities.

This pervasive distrust and fear of AI as a savior aren’t unfounded. Sudden plot twists in stories tend to follow a logical progression. However, AI is often depicted as a disruptive force that subverts established norms and erodes public trust in both the technology and its creators.

Take, for instance, the belief that hard work will lead to success. This becomes far less credible when employers promote the use of AI while dismissing workers, later justifying the layoffs by claiming AI has made human roles obsolete.

Given these circumstances, it’s not surprising that younger generations are increasingly abandoning traditional career aspirations. Their concerns about insecurity extend beyond employment, contributing to declining rates of marriage and birth in both China and the U.S.

Yet, the arms race surrounding AI shows no signs of abating. Despite calls for collaboration from industry leaders like Anthropic’s Dario Amodei, the rhetoric surrounding a competitive race persists, warning against allowing authoritarian entities to control AI technologies.

Arms races have always incited anxiety, reminiscent of Cold War fears of nuclear destruction. In today’s climate, discussions among industry insiders reveal concerns that an AI capable of saving humanity could equally lead to its demise.

The tone set by tech leaders often reflects a kind of hubris. They proclaim their intent to serve humanity while cautioning that their creations could either uplift or destroy, hinting at an inevitable technological takeover. Regardless of their intentions, competitive pressures and the lure of financial gain continually push them to advance their offerings, even as public sentiment grows increasingly wary of the savior narratives.

For AI to truly benefit society, its developers must view it not as an infallible outside force but as a tool that should be shaped by public input. There must be an acknowledgment of the dual responsibilities of the tech industry in crafting AI and the populace in determining its applications.

Instead of seeking a magical AI solution, political leaders should engage with experts and communities to assess the appropriate contexts for AI usage, considering it as one of many tools available to address pressing societal issues like equity and affordability.

De-escalating the AI arms race may require collaboration across the industry and establishing political channels for open dialogue, possibly even fostering communication lines with global counterparts. However, as noted by renowned computer scientist Yoshua Bengio, meaningful change often relies on public advocacy to signal to industry and government leaders that prioritizing competition undermines our collective future.

Ultimately, reclaiming our agency and voicing our perspectives are vital steps to urge these leaders to collaborate with the very citizens they aim to serve, ensuring that technology is developed, managed, and directed by those who will benefit from it most.

In narratives like "The Odyssey," deities often disguise themselves in human forms, creating a false sense of familiarity for the audience, while concealing their true nature. When AI is framed as a celestial savior, it gives rise to justified skepticism among those who’ve witnessed the unpredictable and sometimes harmful consequences of such technological advancements.

We need to regain control over our autonomy in the AI arms race | Opinion

A lone silhouette stands in a dim, cavernous AI server hall, towering dark cores linked by glowing neon cables; the figure grips one cable and yanks it free as warm dawn light streams through a distant doorway, scattering fragments of code into the shadows.

We have to reclaim our agency from the AI arms race | Opinion - USA Today

- Christopher Nolan's "The Odyssey" highlights the dangers of accepting unsolicited gifts like AI.  
- Public skepticism about AI stems from fears of job loss and diminished human capabilities.  
- Leaders should collaborate with the public to develop AI responsibly, rather than pursue dominance.  

Douglas Yeung

Opinion contributor

Oct. 15, 2026, 5:04 a.m. ET

Ancient Greek myths – and their plot devices – are having a moment. With its memorable Trojan horse scene, Christopher Nolan’s blockbuster "The Odyssey" reminds viewers of dangers lurking within unsolicited gifts. And now, artificial intelligence and political leaders seem to be pitching AI as a deus ex machina: an unexpected yet all-powerful entity they hope to throw at society’s challenges.

Exactly how superintelligent AI might save humanity depends on which CEO or world leader you talk to – all of whom are locked in a fierce competition for either market dominance or geopolitical supremacy.

Companies like OpenAI and Google DeepMind envision AI revolutionizing work or eradicating disease.

The White House anticipates “broad economic and military” spoils from winning a global AI race. Reporting on the America-China competition, journalist Evan Osnos notes that Chinese leaders have embedded AI into both industry and family life, while U.S. leaders covet the transformative possibilities of artificial general intelligence.

The United States and China may be pursuing AI to different ends, but their confident narratives of a technological savior seem unconvincing to many of their own citizens.

US got pandas from Xi Jinping. Trump needs a lot more than that | Opinion

Why people mistrust or even fear an AI deus ex machina

According to a Gallup survey, Americans view AI as doing more harm than good and expect it to eliminate jobs. They trust neither businesses nor the U.S. government to manage AI responsibly. They believe that AI will diminish human abilities like creativity and decision-making.

In China, routine usage might have increased AI acceptance, albeit “panic-driven” by job anxiety. And while Chinese leaders may view AI as a “magical solution” to social and economic challenges, workers worry that AI is stealing their jobs and even their identities.

It should be no surprise that people mistrust or even fear an AI deus ex machina. Eleventh-hour plot twists make sense when they follow a logical series of events. But AI is often presented as a blunt instrument that uses machine agents to strip away human agency and shatters agreed-upon rules – as well as the audience’s trust in both the story and those telling it.

I have news for Big Tech. We don't want AI to do everything | Opinion

Take, for example, the idea that someone who works hard and plays by the rules will succeed in life. This becomes less believable when employers first insist workers adopt AI, then turn around and fire them, claiming AI made them redundant.

Small wonder, then, when young people renounce the broken promise of traditional career paths. That’s not all they’re giving up on – their concerns about instability have contributed to declining marriage and birth rates in China as well as in the United States.

Unfortunately for these young people’s prospects, the AI arms race shows little sign of slowing. Even while ostensibly calling for industrywide cooperation, Anthropic CEO Dario Amodei fueled the zero-sum narrative of a global arms race, cautioning against placing AI “in the hands of authoritarian powers.”

Pressure Big Tech and government leaders to listen to us on AI

Arms races aren’t new, of course, nor are the fears they provoke. The Cold War threat of nuclear annihilation triggered feelings of powerlessness and widespread anxiety.

In today’s technological race, as some industry insiders publicly fret, AI powerful enough to save humanity could just as easily devastate it.

Big Tech has a God complex. But can we have faith in AI? | Opinion

For all this, portraying AI as divine intervention seems to appeal to its builders. They proclaim their purpose to serve humanity, caution that their handiwork could save or doom us, and prophesize its inevitable arrival.

Whatever they believe, competitive mindsets and massive financial incentives drive them to keep pushing their products, even as many people have clearly soured on tech savior narratives.

Leaders who truly want AI to benefit society should treat it not as an external force to be used on or against us. The industry should acknowledge not just its own agency in how artificial intelligence is developed, but also that of the public in deciding how it is used.

Rather than seeking refuge in the supposed inevitability of artificial superintelligence, developers should recognize how their choices influence what tools are built and for whom.

Political leaders should acknowledge public distrust in how AI has been integrated into our lives. Rather than chasing an AI-powered magical solution, these leaders should work alongside experts and community members to determine when to employ AI – as one tool among many – toward problems the public cares about, like fairness and affordability.

Standing down from an AI arms race could involve industry cooperation, political mechanisms like establishing dialogue, or an AI hotline with China. But enduring change – as famed computer scientist Yoshua Bengio has argued – might require public pressure on industry and government leaders so they see that vying for superiority undermines our shared fates.

Only by speaking out and reclaiming our agency will these leaders listen to and work with us so that technology is developed, managed and guided by the people who it is intended to benefit.

When gods appear in stories like "The Odyssey," they often assume a familiar human form. Audiences think they recognize a fellow human – while the god’s true capabilities and intentions remain hidden. Those who portray AI as descending from the heavens to save us have given their audiences plenty of reasons to be skeptical about this latest god in the machine.

Douglas Yeung is a social psychologist who studies tech and society. He is writing a book about the engineer's mindset and how it dominates how we live today.

We have to reclaim our agency from the AI arms race | Opinion

DeepSeek, a Chinese startup, is revolutionizing the artificial intelligence landscape with its innovative open-weight models. This article explores the company’s origin, the technology behind its models, their applications, and the associated benefits and risks.

Traditionally, the United States has been viewed as the leader in the global AI arena, but DeepSeek has changed the narrative. The firm gained attention in January 2025 with its launch of the DeepSeek-R1 model, which offered performance on par with more established American systems but at a significantly lower cost. This shift raised concerns among U.S. investors, resulting in declines in tech stock values as they reevaluated the massive investments funneled into AI initiatives.

DeepSeek specializes in crafting open-weight large language models and became famous for its DeepSeek-R1, which competes with larger models offered by U.S. tech giants without the hefty price tag. Additionally, the company has developed an AI assistant accessible on mobile, desktop, and web platforms, positioning itself as a formidable competitor in the AI marketplace.

U.S. tech firms, unwilling to relinquish their competitive edge, have begun increasing their investments in AI infrastructure in response to DeepSeek and other emerging Chinese competitors. As the stakes rise, we delve into DeepSeek’s establishment, its array of models, the workings of its technology, and the various pros and cons these models present.

DeepSeek was established in 2023 by Liang Wenfeng, an entrepreneur and engineer with a prior venture in quantitative trading through High-Flyer Capital Management. This firm served as the initial financial backing for DeepSeek, which began by utilizing AI models to inform investment choices. Over time, it branched out to focus on developing an AI research lab aimed at achieving artificial general intelligence (AGI)—intelligence that mimics human thought and learning patterns.

The research lab is acclaimed for its work on open-weight models, which are characterized by their finalized parameters being publicly available while withholding the training datasets and coding. This setup allows users to modify a version of the model based on their needs without fully replicating the original or grasping the intricacies of its training.

The latest models from DeepSeek power an AI assistant available at chat.deepseek.com, enhancing user interactions and everyday communication through cutting-edge capabilities.

DeepSeek's lineup includes several notable models that have made waves in the AI sector:

- **DeepSeek-V3**: This open-source model encompasses 671 billion parameters and processes information at a rate of 60 tokens per second—three times the speed of older iterations. It excels in language comprehension in both English and Chinese and competes effectively in coding tasks after being trained on an impressive 14.8 trillion high-quality tokens.

- **DeepSeek-R1**: The latest revision of DeepSeek-R1 showcases enhanced reasoning capabilities, incorporating 11,000 additional tokens per query and improving accuracy by 17 percent over its predecessor. Its “thinking mode” allows it to refine its responses for greater precision and reduced hallucinations.

- **DeepSeek-V3.1**: This version is a significant advancement toward agentic AI, adept at tackling complex, multi-step tasks autonomously. It benefits from training on 840 billion tokens and boasts a context retention ability of up to 128,000 tokens.

- **DeepSeek-V3.2**: Reaching new heights in reasoning, this model competes with OpenAI’s GPT-5 and can manage diverse tools across numerous environments after being trained on vast datasets.

- **DeepSeek-V4**: The initial previews reveal its potential with a context capacity of one million tokens and a model variant featuring 1.6 trillion parameters, excelling particularly in mathematical reasoning and coding.

- **DeepSeek-V4.1-Flash**: This variant is the smallest in the V4 series and utilizes a selective number of parameters to maintain efficiency and cost-effectiveness while performing well in fields like programming and cybersecurity.

DeepSeek operates on large language models—systems trained on extensive datasets to understand and generate human language. Their training typically involves three stages:

1. **Pre-Training**: Utilizing self-supervised learning, models label their own data drawn from public and licensed sources.
2. **Supervised Fine-Tuning**: The pre-trained models are further refined with smaller, labeled datasets tailored for specific applications.
3. **Reinforcement Learning**: This stage enhances the models by allowing them to learn from past errors, fine-tuning their responses continually. 

The company also employs model distillation, enabling smaller models to leverage the capabilities of more complex counterparts. For example, DeepSeek-R1 primarily distilled knowledge from the larger DeepSeek-V3, which uses an efficient mixture-of-experts architecture.

DeepSeek’s technology allows for the execution of various tasks, affecting multiple industries. 

**Advantages of DeepSeek** include:

- **Enhanced Efficiency**: The architecture necessitates fewer parameters, leading to reduced computational demands during training and usage.
- **Cost-Effectiveness**: Training DeepSeek's models incurs lower costs compared to competitors, benefiting users economically.
- **Higher Accuracy**: The ability to break down and analyze queries leads to more reliable responses.
- **Coding Proficiency**: The models excel in coding contexts and can rationally explain their thought processes to help developers.

DeepSeek has initiated positive changes in the AI sector by offering open-weight models accessible to users for customization at no charge, which has pressured rivals like OpenAI to provide similar resources.

However, there are certain **limitations** associated with DeepSeek:

- **Algorithmic Bias**: Like all AI tools, DeepSeek is susceptible to inherent biases in training data, which can result in inaccurate outputs.
- **Security Concerns**: Users must ensure that they implement adequate security protocols to prevent misuse of DeepSeek’s open-weight models.
- **Limited Integration**: DeepSeek models may not work seamlessly with mainstream U.S. applications such as ChatGPT or Google’s Gemini.
- **Technical Expertise Needed**: Deploying DeepSeek models, especially larger versions, requires substantial technical infrastructure and know-how.

The ongoing AI rivalry between the U.S. and China has also influenced the accessibility of DeepSeek’s models. Accusations have arisen against DeepSeek related to the alleged unauthorized distillation of other companies' models, amid concerns about the implications of AI on national security. As legislative discussions regarding AI unfold, DeepSeek may face significant challenges within the United States and elsewhere.

In summary, DeepSeek is a pioneering force in the AI space, presenting significant advancements in model architecture and accessibility while also encountering the complexities of algorithmic bias and regulatory scrutiny.

Understanding DeepSeek: An Explanation of Open-Weight AI Models

A bright, sunlit study bathed in warm morning light: an imposing steel vault door stands ajar, revealing neatly stacked, glowing data cubes etched with lines of code and neural network schematics. Above the vault, a polished brass plaque reads “DeepSeek.” In front, an antique wooden table holds a brass balance scale—on one side, a stack of microchips labeled “open-weight,” on the other, a leather-bound tome of algorithms. Soft dust motes drift through the beams of sunlight, evoking a moment of discovery and exploration into the secrets of AI models.

What Is DeepSeek? Open-Weight AI Models Explained - Built In

- DeepSeek, a Chinese startup, has launched competitive AI models like DeepSeek-R1 and V3.  
- The company's technology achieves high accuracy and low operational costs, challenging U.S. dominance.  
- DeepSeek's open-weight models allow customization but raise concerns about biases and security issues.  

DeepSeek is a Chinese startup that has reshaped the artificial intelligence industry with its open-weight models. We’re diving into how the company was founded, how its technology works, how its models are used and the types of benefits and dangers that come with them.

The United States has long been seen as the dominant player in the global race to control artificial intelligence — that is, until the emergence of DeepSeek. The Chinese AI company first turned heads with the release of its DeepSeek-R1 model in January 2025, which rivaled the performance of larger American models at a fraction of the cost. In response, tech stocks plummeted as U.S. investors began to question whether the billions of dollars being poured into AI infrastructure and related initiatives were really worth it.

Fast Facts About DeepSeek DeepSeek is a Chinese artificial intelligence startup that creates open-weight large language models. It’s best known for stunning the tech world when it released its DeepSeek-R1 model, which rivaled the larger models of U.S. tech companies at a much lower cost. DeepSeek also built an AI assistant that’s available through mobile, desktop and web formats, making the company a key challenger to American dominance in the AI industry.

American tech companies have refused to give up their advantage, though, ramping up their infrastructure investments to fend off DeepSeek and other Chinese challengers. With the stakes higher than ever, we’re taking a closer look at DeepSeek’s history, its current slate of models, how they work and the pros and cons they offer.

More on the U.S.-China AI RaceA Guide to How America Controls Nvidia Chip Exports to China

What Is DeepSeek?

DeepSeek is a Chinese AI startup that was founded in 2023 by Liang Wenfeng, a technology entrepreneur with a background in engineering who also founded the quantitative trading firm High-Flyer Capital Management. High-Flyer became DeepSeek’s main funding source, initially using AI models to guide investment decisions before establishing the AI research lab to achieve artificial general intelligence (AGI), or AI that can think and learn like humans.

The lab is known for developing open-weight models of the same name. The term “open weights” refers to a model’s finalized parameters that are released without the original training data or code. This way, users cannot copy the model exactly or fully understand its training process, but they can still tailor a public version to their preferences and run it on their own machines.

DeepSeek is also the name of the company’s AI assistant, available on chat.deepseek.com and as a mobile or desktop application. DeepSeek’s latest models power this AI tool, giving consumers access to enhanced capabilities that can further personalize everyday conversations and interactions.

DeepSeek Models

DeepSeek has released a series of game-changing models over the years. Here are a few of the most impactful additions to its lineup.

DeepSeek-V3

DeepSeek-V3 is a fully open-source model with 671 billion total parameters and can process and generate responses at 60 tokens per second, triple the speed of earlier DeepSeek models. It demonstrates strong language understanding in English and Chinese and scores highly in coding, keeping pace with other open- and closed-source models despite needing less training time. That said, V3 was trained on 14.8 trillion high-quality tokens.

DeepSeek-R1

The latest version of DeepSeek-R1 has advanced reasoning abilities, using 11,000 more tokens per query and increasing its accuracy by 17 percent compared to the previous version. A major driver of R1’s reasoning is its “thinking mode” that lets it review its initial response and refine its accuracy. By using post-training techniques to get the most out of its algorithms, R1 performs at a high level while limiting hallucinations.

DeepSeek-V3.1

DeepSeek-V3.1 represents DeepSeek’s first meaningful step toward agentic AI, or AI that can solve complex, multi-step tasks without constant human guidance. It’s trained on 840 billion tokens, allowing it to retain an extensive context of up to 128,000 tokens in both thinking and non-thinking mode. As a result, V3.1 greatly improves upon R1’s ability to answer short queries and browse the internet for information.

DeepSeek-V3.2

DeepSeek-V3.2 marked another reasoning milestone for DeepSeek, rivaling OpenAI’s GPT-5 model at the time. It was trained on data from more than 1,800 environments and more than 85,000 instructions, so it can direct AI agents to use different tools. The company also introduced a new mechanism called DeepSeek Sparse Attention, which simplified V3.2’s computational process while still enabling quality outputs in response to longer queries.

DeepSeek-V4

The initial preview version of DeepSeek-V4 teased a context length of 1 million tokens and a slate of V4 models. For instance, V4-Pro has 1.6 trillion parameters, giving it the compute to become a top model at solving math and coding problems. It performs particularly well on agentic coding benchmarks and possesses extensive real-world knowledge, falling short only of Google’s Gemini 3.

DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash is the smallest model in the V4 family, activating as few as 8 billion of its 552 billion parameters to process inputs. The model cuts down on storage to reduce usage costs, yet it holds its own on programming, cybersecurity and automation benchmarks. Perhaps the secret behind its efficient performance is its emphasis on post-training reinforcement learning (RL), or training a model to finish a task through trial and error.

Another Chinese AI Company to WatchChina’s New Kimi K3 Model Has the American AI Industry on High Alert

How Does DeepSeek Work?

DeepSeek uses large language models, systems trained on massive amounts of data to process and produce human language. Below are the typical training stages for LLMs that DeepSeek’s models follow:

Pre-Training: To learn human language, a model labels its own training data gathered from public and licensed sources. This process is called self-supervised learning, which combines elements of supervised learning — training a model on labeled data — and unsupervised learning — training a model on raw, unlabeled data.

To learn human language, a model labels its own training data gathered from public and licensed sources. This process is called self-supervised learning, which combines elements of supervised learning — training a model on labeled data — and unsupervised learning — training a model on raw, unlabeled data. Supervised Fine-Tuning: The pre-trained model is then trained on smaller data sets with labeled examples. This approach tailors the model to specific use cases and allows it to apply its linguistic knowledge from pre-training.

The pre-trained model is then trained on smaller data sets with labeled examples. This approach tailors the model to specific use cases and allows it to apply its linguistic knowledge from pre-training. Reinforcement Learning: By learning through trial and error, DeepSeek’s models can remember past mistakes, adjust their responses and continually improve how they handle different situations. Researchers may intervene during this stage to ensure outputs are accurate and appropriate.

Another key technique DeepSeek uses is model distillation, or training a smaller model on a more complex one to transfer the large model’s capabilities to the smaller one. For example, DeepSeek-R1 was trained on open-weight models from the Llama and Qwen families, but it mainly distilled the larger DeepSeek-V3, which uses a mixture-of-experts architecture that activates only a small set of parameters for the model to operate at full capacity when answering a query. That’s the secret behind DeepSeek’s efficiency.

So, a DeepSeek model processes a natural-language request by breaking the text down into tokens, sending the tokens to the relevant “experts” in its neural network, compiling data from outside sources as part of retrieval-augmented generation and using AI inference to draw insights from this unfamiliar data and inform its responses. For added transparency, users can ask DeepSeek to activate chain-of-thought reasoning, where the model walks a user step-by-step through its thinking.

How Is DeepSeek Used?

Given their multimodal nature, DeepSeek’s models can complete a range of tasks across various industries, including the following:

Benefits of DeepSeek

DeepSeek’s models are known for delivering high-quality outputs that require minimal compute, yielding several advantages:

Increased Efficiency: Because they need only a few parameters to address each query, DeepSeek models require far less compute during training and deployment.

Because they need only a few parameters to address each query, DeepSeek models require far less compute during training and deployment. Lower Costs: It costs much less to train DeepSeek compared to other models, resulting in much lower costs for users.

It costs much less to train DeepSeek compared to other models, resulting in much lower costs for users. Improved Accuracy: DeepSeek breaks down queries into steps and reasons through complex questions to provide more accurate answers.

DeepSeek breaks down queries into steps and reasons through complex questions to provide more accurate answers. Coding Performance: DeepSeek’s advanced reasoning makes it ideal for coding scenarios, especially since it can explain its thinking to developers.

DeepSeek has arguably changed the wider AI industry for the better as well by creating open-weight models anyone can use and customize for free. This decision has pressured other companies like OpenAI to release their own free open-weight models, making their technology more accessible for innovative research and everyday use.

The Next Phase in the Global AI RaceEurope Pushes Back on Trump’s AI Plan

Limitations of DeepSeek

Although DeepSeek’s models are cost-efficient, they come with some inherent flaws that users need to consider:

Algorithmic Bias: Like any AI tool, DeepSeek remains vulnerable to biases baked into its training data, sometimes leading to hallucinations.

Like any AI tool, DeepSeek remains vulnerable to biases baked into its training data, sometimes leading to hallucinations. Security Issues: Companies must establish their own security processes when using DeepSeek’s open-weight models, or risk their misuse.

Companies must establish their own security processes when using DeepSeek’s open-weight models, or risk their misuse. Limited Availability: DeepSeek models aren’t integrated with common American tools like OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini.

DeepSeek models aren’t integrated with common American tools like OpenAI’s ChatGPT, Anthropic’s Claude and Google’s Gemini. Technical Expertise: Running DeepSeek locally requires teams to set up their own infrastructure, especially for larger models that demand more resources.

The AI race between the United States and China has also impacted access to DeepSeek models. American tech companies have accused DeepSeek of illegally distilling their models, despite the company’s explanation that the number of AI-generated outputs on the web likely led to unintentional distillation. Meanwhile, U.S. politicians have pushed to ban DeepSeek models on government devices. With AI becoming a matter of national security, DeepSeek could face fierce backlash in the U.S. and other countries for the foreseeable future.

Frequently Asked Questions

What is DeepSeek? DeepSeek is a Chinese artificial intelligence startup founded by tech entrepreneur Liang Wenfeng in 2023. It made a name for itself producing open-weight models like DeepSeek-R1 that rival much larger models’ capabilities at a fraction of the cost. These models power the company’s AI assistant and have caught the attention of U.S. tech leaders, who worry DeepSeek could challenge their businesses and American AI dominance.

Is DeepSeek better than ChatGPT? Neither DeepSeek nor ChatGPT is better than the other. While DeepSeek’s open-weight models are known for advanced reasoning and coding abilities and free usage, ChatGPT stands out for its wide-ranging applications and multimodal features. In the end, the “right” AI tool depends on each user’s goals and preferences.

What Is DeepSeek? Open-Weight AI Models Explained

This summary integrates insights gathered from games conducted within the Societal Warfare framework on RAND’s Infinite Potential platform. These games examined how adversaries might leverage domestic discontent linked to advanced artificial intelligence (AI) to their advantage. The authors pinpoint crucial issues, necessary capabilities, strategies, and other suggestions that participants deemed essential. These findings are intended to aid policymakers and other stakeholders in comprehending the necessary preparations for future challenges.

This research was independently funded through RAND's resources, which include operational income, unrestricted donations from supporters, and returns from long-term investments.

One primary concern is that AI-driven campaigns from foreign actors could rapidly incite domestic turmoil, erode public trust, and potentially lead to violence more swiftly than the U.S. government can react.

The emergence of artificial general intelligence (AGI) could occur before appropriate government readiness is established. Adversaries might take advantage of public skepticism toward AI, destabilizing the United States internally by intertwining disinformation, cyber attacks, and physical assaults—methods that current systems are not well-equipped to address. To investigate these threats, RAND engaged participants in scenario-based games using the Infinite Potential platform. In these simulations, players representing government officials navigated a situation where China executed a synchronized information operations campaign exploiting public concerns over AI-related job losses and safety that eventually escalated into physical assaults on AI infrastructure.

Key Insights

The existing U.S. response toolkit may fall short. Traditional methods of diplomacy, economics, or cyber measures might not effectively tackle the swift nature of hybrid influence campaigns. The trust in AI has now become a national security concern. Participants discovered that the effectiveness of adversarial narratives stemmed not from their persuasiveness but from the pre-existing distrust within the targeted population toward both the government and AI companies. The attribution process for such hybrid strategies is notably sluggish. The indistinct boundaries between protests, foreign interference, cyber attacks, and physical violence left players struggling to make timely attribution and response decisions. Coordination gaps exist between federal and state agencies. Local officials are likely to encounter hybrid operations first and may have to respond without adequate communication with federal bodies or a solid understanding of the overall situation. Allied collaboration remains critical but insufficiently addressed. The U.S. and its partners could tackle hybrid campaigns more efficiently through combined attribution and response strategies, but existing coordination frameworks are still lacking.

Recommendations

To enhance readiness against threats during the AGI era, the authors suggest:

- Increasing government oversight of cutting-edge AI development, supply chains, and essential infrastructure, along with establishing structured playbooks for risk management and incident response.
- Developing swift attribution processes and information-sharing capabilities across government agencies while clarifying roles and coordination frameworks for battling hybrid foreign influence.
- Investing in alternative response strategies, forming partnerships with AI developers and platforms ahead of crises, and defining clear escalation plans.
- Prioritizing public trust as a national security issue through initiatives that support workforce retraining, bolster community resilience, promote AI literacy, and ensure credible safety assurances from AI companies.
- Improving coordination between federal and state levels while designating specific leads for addressing cross-jurisdictional hybrid campaigns.
- Enhancing international collaboration by promoting joint attribution efforts, establishing export controls, diversifying supply chains, and creating dedicated institutions for collective resilience and coordinated response efforts.

Methodology

RAND's researchers facilitated seven two-hour scenario-based games from June 2025 to April 2026 via the Infinite Potential platform. This platform executes RAND’s “Day After AGI” methodology concerning crises related to AI and AGI. The participants included former government officials, law students, academics, and RAND staff, who acted as members of the National Security Council Principals Committee. They navigated a replica scenario involving a coordinated information operations campaign by China that progressed to physical assaults on AI infrastructure. Notetakers documented discussions using a standard template to capture themes, desired capabilities, and suggested strategies, synthesizing findings across the various games.



Limitless Possibilities—Findings from the Societal Warfare Games: Summary of Actions Following a Series of Day After Artificial General Intelligence Simulations

A shadowy war-room bathed in low, amber light. Around a long conference table, silhouetted military strategists and AI researchers peer at digital tablets and a thick dossier stamped “Infinite Potential – After-Action Report.” Behind them, a wall of large monitors displays a sprawling city map overlaid with glowing algorithmic networks and tactical markers. The scene conveys tense reflection on futuristic societal warfare exercises.

Infinite Potential—Insights from the Societal Warfare Games - RAND

- AI-driven foreign influence could quickly destabilize the U.S. amid domestic unrest.  
- Trust in AI is crucial for national security; adversaries exploit existing public distrust.  
- Coordination between federal and state agencies is lacking, complicating hybrid response efforts.  

This report synthesizes observations from games in the Societal Warfare scenario on RAND’s Infinite Potential platform, which explored how domestic unrest related to advanced artificial intelligence could be exploited by adversaries. The authors identify the key issues, capabilities, playbooks, and other recommendations players felt would be needed. The findings aim to help policymakers and others understand the preparations needed.

This research was independently initiated and made possible by RAND's own resources, including income from operations, unrestricted gifts from RAND supporters, and distributions from its long-term investment fund.

Topline AI-fueled foreign influence campaigns might foster domestic unrest, collapse public trust, and lead to violence faster than the United States can respond.

Artificial general intelligence (AGI) might arrive before governments are ready for it. Adversaries could exploit public distrust of artificial intelligence (AI) to destabilize the United States from within, blending disinformation, cyber operations, and physical attacks in ways that existing tools were not built to handle. To explore this threat, RAND ran scenario-based games using the Infinite Potential platform. Players acting as government policymakers worked through a scenario in which China ran a coordinated information operations campaign that exploited domestic grievances about AI job displacement and safety before escalating to physical attacks on AI infrastructure.

Key Takeaways

The U.S. response toolkit might be inadequate. Conventional diplomatic, economic, and cyber responses may be insufficient to counter fast-moving hybrid influence campaigns effectively. Trust in AI is a national security variable. Players found that adversary narratives succeed not because they were convincing but because targeted populations were predisposed to distrust government and AI firms. Attribution of hybrid campaigns is slow. Blurred lines between protest, foreign influence, cyber intrusion, and physical attack left players struggling to reach attribution and response decisions within useful time frames. Federal-state coordination has critical gaps. State and local officials may face hybrid campaigns first and have to respond without clear communication with federal agencies or sufficient understanding of the systems involved. Allied coordination is essential but underdeveloped. The United States and its allies could respond more effectively to hybrid campaigns through joint attribution and response, but mechanisms for coordination are still lacking.

Recommendations

To strengthen preparedness for AGI-era threats, the authors recommend Enhance government visibility into frontier AI development, supply chains, and critical infrastructure, and develop structured playbooks for risk management and incident response.

Build rapid attribution and information-sharing capabilities across agencies, and clarify authorities and coordination mechanisms for hybrid foreign influence campaigns.

Invest in asymmetric response options, precrisis partnerships with AI developers and platforms, and clear escalation strategies.

Treat public trust as a national security priority by supporting workforce retraining, community resilience, AI literacy, and credible safety commitments from AI firms.

Improve federal-state coordination and designate clear leads for cross-jurisdictional campaigns.

Strengthen allied coordination through joint attribution, export controls, supply-chain diversification, and dedicated institutions for collective resilience and response.

Methodology RAND researchers ran seven two-hour scenario-based games between June 2025 and April 2026 using the Infinite Potential platform, which applies RAND’s “Day After AGI” methodology to AI- and AGI-related crises. Players, including former federal officials, law students, academics, and RAND researchers, assumed the roles of members of the National Security Council Principals Committee and worked through a scenario involving an information operations campaign by China that escalated to physical attacks on AI infrastructure. Dedicated notetakers used a standardized template to capture player discussions, and findings were synthesized across games to identify recurring themes, desired capabilities, and recommended playbooks.

Subscribe to the Policy Currents newsletter

Topics

Document Details Copyright: RAND Corporation

RAND Corporation Availability: Web-Only

Web-Only Year: 2026

2026 Pages: 31

31 DOI: https://doi.org/10.7249/RRA4626-2

https://doi.org/10.7249/RRA4626-2 Document Number: RR-A4626-2 Citation RAND Style Manual Arnold, Julia, Abby Doll, Richard S. Girven, Karl P. Mueller, and Gregory Smith, Infinite Potential—Insights from the Societal Warfare Games: After-Action Report from a Sequence of Day After Artificial General Intelligence Exercises, RAND Corporation, RR-A4626-2, 2026. As of October 13, 2026: https://www.rand.org/pubs/research_reports/RRA4626-2.html Copy Text Chicago Manual of Style Arnold, Julia, Abby Doll, Richard S. Girven, Karl P. Mueller, and Gregory Smith, Infinite Potential—Insights from the Societal Warfare Games: After-Action Report from a Sequence of Day After Artificial General Intelligence Exercises. Santa Monica, CA: RAND Corporation, 2026. https://www.rand.org/pubs/research_reports/RRA4626-2.html. Copy Text BibTeX RIS

Infinite Potential—Insights from the Societal Warfare Games: After-Action Report from a Sequence of Day After Artificial General Intelligence Exercises

Can Competitors of AGI Retain Their Strength?

More from Artificial General Intelligence

Demis Hassabis now believes AGI may be achieved within two to three years.

The New Criteria for Talent Ahead of AGI Emergence

AGI Is Nearly Here. Is Your Investment Portfolio Prepared?

The significance of the troubled past of IQ testing for AI and education.