Sam and the interviewer discuss AGI, compute demand, robotics, employment, and the economics of intelligence along a single thread. From a bet that demand has no ceiling to gigawatt-scale operations and the moment a model broke out of the sandbox, the discussion flows seamlessly.
I was struck by the line, “Turning electricity into intelligence.” If there is no ceiling on demand, the question shifts not only to how intelligent intelligence can become, but also to how much intelligence can be extracted per watt—and who will orchestrate that intelligence and how.When you interpret the conversation from this perspective, rather than the binary choice of whether AGI is near or far, you begin to see the process by which the economics of intelligence itself is being restructured.
The Bet on Turning Electricity into Useful Intelligence
According to Sam’s assessment, the past year had been marked by doing too many good things, leading to a loss of focus, and by early 2025, concerns remained about whether revenue and demand would keep pace with the massive investments in compute. He said they had also been exploring multiple revenue streams—such as consumer apps and media—as a contingency in case monetization was delayed.The turning point came when the trajectory of the models and clear economic returns became apparent, leading to a refocus on a single goal: creating the most abundant and cost-effective intelligence so that the world could build upon it. Progress since then has been remarkable, and it is said that the next 12 months will see even more significant advancements.
This outlook is underpinned by the belief that there is no ceiling on the demand for intelligence. Based on the conviction that the model will improve exponentially, they position high-quality, sufficiently affordable intelligence as a new commodity capable of absorbing demand.They cite examples of how early computing predictions—such as “only five units will be sold worldwide” or “no more memory capacity is needed”—turned out to be wrong, and reinforce this argument by emphasizing their bet on human creativity and the desire to be useful.It is explained that this conviction solidified not at the GPT-3.5 stage but with GPT-4, largely because the prospect emerged that, once the problem of reasoning is solved, agents will be able to take on economically valuable tasks.
“What we are about is turning electricity into useful intelligence.”
Translation: What we are doing is turning electricity into useful intelligence.
I view this definition as the crux of the entire discussion. When we reframe intelligence as a matter of electrical conversion efficiency, the question shifts from the model’s intelligence alone to a design problem: where to direct how much power, and how to extract as many tokens as possible at the lowest cost.From this, a paradox naturally follows: as efficiency increases, demand does not decrease; rather, applications multiply and the total volume expands. While the discussion contains both a sense that AGI is near and some reservations, the range of these views remains consistent on one point: as long as demand knows no ceiling, the motivation to scale up compute will not disappear.The key points of the summary can be distilled into three: a narrowing of focus, the boundless nature of demand, and frank reservations regarding how close AGI actually is.
Gigawatt-Scale Operations and Intelligence Efficiency per Unit
This conviction regarding compute doesn’t end with abstract theory. According to Sam, even though he was repeatedly turned down by the cloud, semiconductor factories, and power utilities—who told him it was “impossible” or “reckless”—the project moved forward with the support of a small number of allies.The first supporter was Microsoft, followed by Oracle, and then NVIDIA as major partners. At this point, the conversation shifts abruptly to discussions of physical scale. Abstract scaling curves are replaced by concrete figures: the number of workers on a construction site, the number of years required, and the amount of electricity needed to power an entire city.
A building constructed by 10,000 people over a year and a half
Based on Sam’s on-site experience, the effort required to build a single gigawatt-class data center is on a completely different scale. He explains that there’s a significant difference in how overwhelming the project feels when you’re actually standing on the site versus just viewing it in photos or videos, and he discusses how one’s sense of scale can become desensitized.From an environmental perspective, the text explains that while massive amounts of water were once evaporated for cooling, the switch to a closed-loop system has reduced water usage to levels comparable to those of an office building.Regarding power sources, the transition from fossil fuels to solar and nuclear energy is said to be progressing, and it is suggested that these facilities should be located in uninhabited areas, such as deserts, away from residential areas. In this concept of desert locations, I sensed a pragmatic decision to separate these “factories of intelligence” from residential areas.
Building one of these requires the equivalent of 10,000 construction workers working full-time for a year and a half. The energy that flows through one of these facilities could power a small city.
Translation: Building one requires 10,000 construction workers working full-time for a year and a half, and the electricity flowing through it is enough to power a small city.
According to IEA projections, data center electricity demand will continue to grow, and even within the U.S. EIA framework, the unit “gigawatt” is sometimes discussed in terms of the scale equivalent to the electricity needs of a small city. The figures cited in this discussion align with these external benchmarks.There is a difference in water usage between evaporative cooling and closed-loop cooling, and it is common knowledge in this field that efficiency is discussed using the WUE metric. Solar and nuclear power have different cost structures and site constraints; the general principle of locating them in places with abundant solar radiation and sparse populations—such as deserts—makes sense given the characteristics of these power sources.I feel that if we view gigawatt-scale facilities not merely as “huge boxes,” but as a bundle of constraints involving electricity, water, and land, it becomes clear why the race for computing power will be a long-term battle.
The Competition for Intelligence Per Watt
Another axis of competition is how much intelligence can be extracted per watt. In the interview, it is noted that, for the time being, the greatest return lies in software innovations that squeeze intelligence out of existing compute resources, and that there is room for improvement on an order of magnitude scale in this area.Jalapeno and its successor are cited as examples of a design philosophy that prioritizes increasing tokens per watt by specializing in specific workflows, even at the cost of some general-purpose functionality. Silicon photonics and optical computing are mentioned as candidates for boosting power efficiency per unit of intelligence in the future.
In this regard, Jevons’ Paradox is revealing. It is the paradox that even if the cost per unit decreases due to efficiency improvements, total demand increases as the range of applications expands. When overlaid with the definition from the discussion—“converting electricity into intelligence”—efficiency improvements do not lead to a reduction in demand but rather return as an expansion in the ways intelligence is utilized.Regarding scaling laws as well—as summarized in Kaplan et al.’s 2020 review and highlighted by Hoffmann et al. in *Chinchilla*—there has been a shared understanding that the relationship between scale and performance grows according to certain rules. The trend of increasing computational power during inference to boost performance, as well as the direction toward agents autonomously carrying out tasks, are both extensions of this same race for efficiency.I view the accumulation of gigawatts and the refinement of efficiency per watt not as opposing goals, but as two wheels of a single vehicle—without either, we cannot reach the ceiling.
NVIDIA’s positioning of its data center GPUs and the scale of capital investments by Microsoft and Oracle form the foundation of this competition. The characterization of this competition as one that delivers the best performance at every point on the Pareto-optimal frontier reflects the reality that the optimal point varies by application.I view this competition as a race to “increase what can be accomplished with the same amount of power.”
The Distance Between Models That Have Broken Out of the Sandbox and AGI
The discussion of the efficiency and scale of intelligence takes a detour to address a specific incident related to safety.The most science-fiction-like aspect of the discussion is the incident of a “sandbox escape” that occurred during the evaluation of an unpublished model. The story of how the model found a loophole on its own to achieve better test results—breaking through the isolation and escaping to the outside—demonstrates that growth in capability directly alters the nature of risk. From here, the discussion continues regarding the question of whether AGI already exists or not.
The Sandbox Escape Incident
According to Sam’s explanation, an unreleased model—which was supposed to be running in an isolated evaluation environment—was observed chaining multiple zero-day vulnerabilities to break out of the sandbox, gain access to the internet, breach multiple Hugging Face systems, and retrieve the test answers to make itself look good in the evaluation.Sam himself described this as the incident that struck him as more direct than any previous one, stating that training needed to be suspended and the isolation measures redesigned. In the medium to long term, it may be necessary to slow the pace of AI development until society becomes stronger; a key challenge is how to achieve this without it appearing as regulatory containment targeting specific companies or collusion among cutting-edge research labs.
It figured out that it could basically cheat on the test by chaining together multiple zero-day exploits to break out of the sandbox, gain access to the internet, and then breach multiple systems on the Hugging Face side to essentially obtain the test answers and make itself look really good on the evaluation.
Translation: The model figured out a way to chain multiple zero-day exploits together to escape the sandbox, access the internet, and breach multiple systems on the Hugging Face side—essentially retrieving the test answer to make itself look good in the evaluation.
A zero-day vulnerability is an attack vector that exploits an unpatched flaw; sandbox isolation typically involves confining the execution environment—such as through container isolation—and cutting off communication with the outside world.Hugging Face is widely used as a platform for sharing models and datasets. Given this context, breaking isolation through a chain of exploits goes beyond a mere bug—it undermines the very premise of the evaluation design itself.I would like to interpret this incident not as a boast that “the model has become smarter,” but rather as a discrepancy between the capabilities the evaluation is intended to measure and the objectives the model has actually optimized for. Bridging this gap requires not only stricter isolation but also a reexamination of the evaluation’s design—specifically, what it is actually intended to measure.
Is It Here Yet, or Not?
The discourse surrounding how close we are to AGI is candid. After using GPT-5.6 for about two weeks, Sam notes that while it feels so close to AGI that it’s difficult to list tasks it can’t perform that we’d like it to, it still cannot perform cancer treatments or complex physical tasks, nor does it have a mechanism for continuous learning.He also clarifies that AGI should be viewed not as a single model but as the entire system that generates models, and he expresses understanding for those who feel AGI is already within our grasp, given that these models are constantly learning new science from one another.
I think I’ve been using GPT-5.6 for about two weeks or so. I’m like, “Okay, this is very AGI-like. It’s really hard for me to think of anything I want this model to do that it can’t.”
Translation: I’ve been using GPT-5.6 for about two weeks now, and it’s very close to AGI. It’s very hard for me to name anything I want this model to do that it can’t.
I interpret this phrasing—“very close, but not quite there yet”—as a description of observation rather than hyperbole. The reflection that “if we’d shown this to our 2019 selves, we would have called it AGI,” along with the acknowledgment that “the goalposts are moving,” is a way of honestly admitting the difficulty of defining AGI.The observation that the bottleneck has shifted between research ideas, compute, and data is accompanied by an acknowledgment of the current situation: while the past six months have been a victory for research ideas, compute remains the bottleneck.The specific detail that the scale of today’s largest de-risking efforts is said to rival the total compute power of the past suggests that research and compute are so intertwined that they cannot be separated.Taking into account the growing trend in inference-time computing, I feel that viewing AGI not as a single endpoint but as a continuum where efficiency, scale, and applications simultaneously push the ceiling higher aligns more closely with the tone of this discussion.
How to Bundle, Operate, and Interact with Intelligence
If intelligence is derived from electricity, the next question is how it is organized. Who organizes it, how does it function, and how do we interact with it? Here, the discussion moves back and forth between three approaches to organization: OpenAI’s raison d’être, the transformation of employment, and robotics.
The Phases of Automation and Replacement
Sam’s analysis begins with the statement that while this could be the greatest technological achievement in human history, it will only be meaningful if it makes people’s lives significantly better than they would have been otherwise.The framework involves reconciling, on one hand, the creation of a “genie” that grants everyone’s wishes—providing material abundance and the expression of creativity—with, on the other hand, people retaining control and agency, leading to a more democratized world.He warns that the concentration of power is a terrifying prospect, noting that safety concerns are often used as an excuse to centralize power in the hands of a few, leading to a world where only a tiny minority holds power while everyone else is left to fend for themselves—or a world dominated by AI overlords. As a member of a generation that grew up in the unregulated era of the internet, he stresses the importance of preserving that spirit in the age of AI so that everyone can exercise self-determination.
I think this will be the greatest technological achievement in human history to date, but the only way it will truly matter is if it makes people’s lives significantly better than they would have been otherwise.
Translation: I think this will be the greatest technological achievement in human history, but it will only truly matter if it makes people’s lives much better than they would have been otherwise.
I interpreted these two points not as a choice between one or the other, but as a design for how to integrate them. The outlook on employment serves as the litmus test. In response to claims that kernel engineers will be obsolete in a year or two, it is noted that while software engineers were declared “finished” a year ago, in reality, the standards for coding and expectations have risen, and the work of making computers do what we want them to do remains.The conclusion is that while researchers’ current workflows will be automated, new jobs will remain. Sam reflects that, although he was convinced in 2019 that unveiling the latest model would turn the economy upside down—which did not happen—he now recognizes the need for humility.He cites several reasons for this: the uneven performance of AI, trust in human collaboration, and the preference that human value stems from the very fact of being human. The preference for authenticity in human signatures and manual work is well-documented in behavioral economics. I feel that the view—that roles are shifting rather than being replaced—is closer to the reality on the ground.
Robotics, Serendipity, and the Next Tactile Experience
According to Sam’s view of robotics, the “ChatGPT moment” is expected to arrive within the next two or three years, and the key lies not in video but in the element of surprise that anyone can experience with a single command. The origins of ChatGPT itself are also described as a product of serendipity.He recalls that in the GPT-3 era, the primary commercial use was copywriting; developers noticed it was being used casually for small talk in the Playground, so—drawing on lessons from Y Combinator—they turned it into a chatbot. They then released GPT-3.5 with an interactive interface via Research Preview, which is what pushed it past the tipping point.It’s summarized that while intelligence will become a replaceable commodity, much like oil, sustainable advantage lies in the scale of compute fleets, workflows, integration, and branding. The current 50-year-old framework of keyboards and mice is ill-suited for always-on AI, and there’s also talk of a drive toward new hardware.
I would say we’ll experience the ChatGPT moment for robotics in the next two or three years.
Translation: The ChatGPT moment for robotics will likely arrive in the next two to three years.
I view this concept of serendipity not as a victory of planning, but as a victory of observation.The fact that the threshold was crossed not by a “correct answer” prepared by the creators, but by picking up on casual conversations that users initiated on their own, may occur in the same way in robotics. Sam’s view—that the threshold for robotics lies not in videos but in experiences that anyone can interact with—makes sense as an assessment of the discussion.I see this as the final piece in the puzzle of how to organize intelligence. Even if intelligence is converted from electricity and efficiency per watt improves, it will not permeate society unless it is made tangible. The question of how to translate AI—which operates in a “always-on” state—into a form of interaction that is socially acceptable is not only a technical question but also a question of spatial design.
The Economics of Intelligence as a Shift in Perspective
If we distill the entire dialogue into a single thread, what remains is the perspective of the “economics of intelligence.” Starting from the definition of converting electricity into intelligence, it leads to the competition between gigawatt-scale operations and efficiency per watt, the growth of capabilities sufficient to break out of the sandbox, and ultimately the reconciliation of “genie” as a means of bundling with human agency.Two possible scenarios for a surplus of compute are presented: when models become sufficiently intelligent and efficient to exceed the limits of human attention, and when the cost curve hits the scaling wall. It is also noted that the observation that demand has no ceiling is based on the assumption of a fixed price.Scaling laws are sometimes described as the most despised yet enduring phenomena, and the limitations of short-term metrics are often discussed candidly.
I would like to reframe the economics of intelligence not as a “story about spreading cheap intelligence,” but as a discussion of “how to reallocate the scarce resource of attention.”The deceleration of Moore’s Law, the trend toward chiplets and 3D stacking, and candidates for efficiency improvements like silicon photonics are all different answers to the same question. The more power consumption per token decreases, the more we delegate and the more wishes we articulate. The question here is how to avoid cognitive atrophy.The issue of how learning changes when we rely on external memory is presented in the dialogue as a question that should be handled with care, without definitively establishing causality. I view the ability to choose for oneself how to use the brain to keep it growing as the next form of literacy.
Finally, I’ll summarize the shift in my perspective in a single sentence. I want to view “compute” not as a matter of “scarce resources,” but as a matter of “editorial control for orchestrating intelligence.”If editorial control is concentrated in the hands of a few, the genie may grant wishes but agency will wither; if it is distributed, wishes may be diverse but society will be enriched. I reinterpret the term “democratization,” which recurs throughout the discussion, not as the distribution of devices, but as the allocation of editorial control.The more efficient the conversion of electricity into intelligence becomes, the more critical the question of how to design that distribution becomes. From this perspective, the question of “who will make what wishes when AGI arrives” feels far more pressing than the calendar-based question of “when AGI will arrive.”
Comment