English
Based on โThe AI Model Built for What LLMs Canโt Doโ from Every Watch the original video
The AI That Doesnโt Guess: Unlocking Verifiable Intelligence with Energy-Based Models
In the rapidly evolving world of artificial intelligence, Large Language Models (LLMs) have taken center stage. Their ability to generate human-like text, translate languages, and answer complex questions has captivated the public imagination and drawn immense investment. Yet, for all their impressive capabilities, LLMs operate on a fundamental principle that makes them inherently unsuitable for certain critical tasks: they guess.
This โguessing gameโ is precisely what Logical Intelligence, a foundational AI company, aims to transcend with a different kind of model: Energy-Based Models (EBMs). Led by founder and CEO, Yee, Logical Intelligence is building AI designed for correctness, verifiability, and tasks where the stakes are too high for hallucination.
โLLMs are naturally non-autoregressive,โ Yee explains, righting a common misconception. โThere are no sequences of tokens, and thatโs what makes it fundamentally different.โ This seemingly subtle distinction underpins a radical departure from the LLM paradigm, promising a future where AI can be trusted with our most critical systems.
The Peril of the Guessing Game
Imagine an AI driving your car. Now imagine that AI, an LLM, is prone to hallucinating 20% of the time, potentially taking you to the wrong place. While some might find that โkind of interestingโ for a joyride, the stakes escalate dramatically when considering a plane. โHow about the plane?โ Yee challenges. โYou take a plane from yourself to New York and someone says, โYou know, like 20% of the time it might just like the next word not going to match and itโs going to go down.โ How would you feel about it?โ
The answer is obvious: terrifying. Our lives depend on deterministic, verifiable systems. Current planes are run by such systems, and while AIโs integration into every facet of life seems unavoidable, Yee argues that the current LLM approach falls short for mission-critical applications like code generation, chip design, or autonomous vehicles.
The core issue lies in the LLMโs architecture. Itโs an autoregressive model, meaning it predicts the next token in a sequence based on the preceding ones. This process, while powerful for language generation, is essentially a โguessing gameโ based on probabilities derived from vast training data.
โYouโre like a black box,โ Yee explains, describing the LLMโs internal workings. โYou donโt have access to whatโs inside until itโs all processed.โ To mitigate this, companies often attach external verifiers (like machine-verifiable proof languages such as Lean 4) to check the LLMโs output. However, this is costly. โEven if you attach external verifier, even you fine-tune this LLM specifically for the task youโre trying to create, youโre still not solving the problems of tokens being expensive. It takes compute for you to play a guessing game.โ
A Birdโs-Eye View: How EBMs Navigate Reality
To grasp the fundamental difference, consider navigating a complex environment. An LLM, in Yeeโs analogy, is like trying to navigate San Francisco with โtunnel vision.โ You choose one direction at a time, unable to see other options.
โYou cannot chooseโฆ one direction at a time,โ Yee illustrates. โAnd sometimes you take the wrong turns just because you hallucinateโฆ like there might be a hole in the road and youโre just going to fall. And you might see this hole, but you cannot turn back because youโre autoregressive LLM.โ This leads to wasted compute, endless searching, and potentially never reaching the destination.
An EBM, by contrast, has a โbirdโs-eye view all the time.โ If it sees a hole, it can choose a different route. This isnโt just an analogy; it speaks to the EBMโs non-autoregressive, token-free nature. Thereโs no guessing game of what the โnext wordโ or โnext actionโ might be. Instead, EBMs can โoversee all possible scenarios.โ Their architecture allows for โself-alignmentโ as they process information, making them inspectable. โAs itโs performing, you can open it anytime during the training and you could see whatโs happening in there. So you cannot do this with LLMs.โ
This inspectability, coupled with the ability to attach external verifiers, provides โverification sort of on both sides, inside and outside.โ
The Physics of AI: Minimizing Energy
At the heart of EBMs is the principle of energy minimization, a concept borrowed directly from physics. โEBM just simply means energy-based model,โ Yee clarifies. โWhat is energy? Energy-basedโฆ it comes from physics. Itโs a very popular term when youโre trying to minimize the energy.โ
Think of it this way: everything in nature seeks a state of lowest energy. When youโre tired after a long day, you donโt dance around; you collapse onto the couch to minimize your energy. This โmost comfortable configurationโ corresponds to the lowest potential energy.
EBMs apply this universal principle to AI. They construct an โenergy landscapeโ โ a metaphorical map of all possible states based on observed data. The highest points on this landscape represent less probable scenarios (e.g., you dancing when tired), while the lowest points represent the most probable and โcomfortableโ states (e.g., you on the couch).
โEssentially all of this picture can be mapped into something we call energy landscape when theโฆ itโs going to look like a map,โ Yee elaborates. โSo, youโre going to have highest points, youโre going to have lowest points. The highest points we can associate less probable scenariosโฆ the lowest point is going to be you on the couch.โ
During training, the EBM observes data (your behavior, your tiredness, etc.) and learns to shape this energy landscape. It then uses algorithms to navigate this landscape, always seeking the lowest energy states, which correspond to the most probable and optimal outcomes. Crucially, this process doesnโt involve tokens or language; itโs a direct mapping of data to a structured landscape.
Beyond Pattern Recognition: Understanding and Latent Variables
Another critical distinction between EBMs and LLMs lies in their approach to data. LLMs are exceptional at identifying patterns in vast datasets. But do they โunderstandโ in a human sense? Yee argues no. โLLMs donโt understand the data. Itโs just you feed a lot of data into it and itโs sort of like, โHey, I got it. Like, okay, I know whatโs the most probable scenario here and here we are.โโ
EBMs, particularly those with latent variables, go a step further. They not only identify patterns but also โtry to understand the pattern.โ This โunderstandingโ translates into basic knowledge and rules about the world. For instance, an EBM observing your apartment might infer rules like, โthereโs a kitchen for cooking, thereโs a bathroom, thereโs a sofa.โ
This knowledge is stored in โlatent variables,โ which Yee describes as โknowledge storageโ โ a dataset about the rules of your data, also in the form of an energy landscape. This allows the EBM to understand its environment and adapt to changes. If you get a new couch, it still knows what to do with it because it understands the underlying rules of furniture and human behavior, not just specific examples.
This capability is particularly powerful for data analysis, where searching for patterns and rules is paramount. โItโs something where language is not going to be helpful to you,โ Yee asserts. โIf you try to attach the rules about your data and those data is like numbers and some relationships and functions to like American English and words in American Englishโฆ youโre losing a lot of information.โ EBMs work directly with the data, understanding its inherent structure without the translation layer of language.
Furthermore, EBMs are adept at working with sparse or incomplete data, a feature that evolved from traditional EBMs through diffusion models. By injecting noise and changing navigation strategies, they can reconstruct energy landscapes even when data is limited.
Building Trust: From Code to Autonomous Systems
The implications of EBMs for building trustworthy AI are profound. Consider โvibe coding,โ where developers use LLMs to generate code, then spend significant time debugging and fixing the patchwork of solutions. Logical Intelligence dreams of โgenerating formally verified code and automate the coding entirely,โ moving from vibe coding to โvibe code specifications.โ
EBMs, with their internal and external verifiers, can ensure logical compatibility and compliance. They can generate mathematical proofs to certify code, telling engineers, โHey, this part of your code is not compatible by logic. This is potentially how you fix it.โ
Beyond mere compilation, EBMs address the critical question: โIs this code actually doing what you want it to be doing?โ While AI cannot read a humanโs mind, EBMs can be constrained to follow a precise set of human-defined rules. โLLM can misbehave based because you cannot constrain it. It just hallucinates. And EBM can be constrained. You can come up with a set of constraints and EBM just forced to follow it.โ This determinism is vital for safety-critical applications like autopilot systems, where misbehavior could have catastrophic consequences.
Navigating the AI Ecosystem
Despite the compelling advantages of EBMs, the current AI investment landscape remains heavily skewed towards LLMs. โLLMs historically the first form of AI which gave us aha effect,โ Yee acknowledges. The massive investments already poured into LLM infrastructure โ from data centers to hardware โ create a powerful inertia. โIt becomes like a one giant thing which is impossible to break.โ
However, Logical Intelligence isnโt seeking to replace LLMs entirely but rather to complement them. โWe are very much compatible with LLMs,โ Yee explains. โYou could put LLM on top of us. EBMs compatible with transformers, transformers can, you know, work with any LLMs.โ
This synergistic approach offers a path forward. LLMs can handle language-related tasks, while EBMs can take on spatial reasoning, engineering, data analysis, and other non-language-dependent tasks requiring precision and verifiability. This could make LLM deployments cheaper and more reliable by outsourcing suitable tasks to EBMs. โIf somebody comes to big tech LLM and say, โHey, can you try to do my taxes?โ LLM not going to solve this. But if itโs attached to EBM, we can take care of that.โ
While some might perceive a plateau in LLM progress, Yee believes that the apparent advancements are often within the existing paradigm, not a fundamental shift. The real breakthroughs, she suggests, will come from exploring alternative architectures like EBMs that address core limitations in correctness and understanding.
As AI permeates every aspect of our lives, the demand for systems that are not only intelligent but also reliable, verifiable, and safe will only grow. By moving beyond the โguessing gameโ and embracing principles of energy minimization and deep understanding, Energy-Based Models offer a powerful pathway to building the next generation of trustworthy artificial intelligence.
Based on โAre Human Drivers Finally Obsolete? | Freakonomics Radioโ from Freakonomics Radio Network Watch the original video
The Road Ahead: How Driverless Cars Went From Sci-Fi Dream to Reality
By PJ Vogt, Host of Search Engine
Imagine a future where the concept of a โdriverโ refers not to a person, but to a machine. A world where the most dangerous activity many of us routinely engage inโdriving a carโis outsourced to an unblinking, unyielding, and un-distracted artificial intelligence. This isnโt a distant sci-fi fantasy; itโs a reality unfolding in cities across the globe right now.
My journey into the fascinating, complex, and often contentious world of driverless cars began not with a grand vision, but with a very personal, very painful experience. After a bench-pressing injury led to a hernia and subsequent surgery, I found myself with limited mobility, visiting a friend in San Francisco. It was then that I took my first ride in a Waymo robo-taxi. The experience was transformative. One moment, I was pressing a button on my phone; the next, a car pulled up, completely empty, its steering wheel turning itself as if by an invisible hand.
โThe first time it feels like the first time youโre in an airplane,โ I tried to explain to someone recently. โBy the third time, it feels like youโre in an elevator.โ This immediate normalization of a profoundly futuristic technology made me realize a lot was about to change, and I was confused why more people werenโt talking about it. This led me down a rabbit hole, culminating in a two-part series for my podcast, Search Engine, exploring the history, technology, and human impact of this revolution.
The Echoes of the Past: When New Technology Met Old Fears
To truly understand the seismic shift driverless cars represent, it helps to look back. Picture yourself almost 200 years ago, pre-dawn on a Monday. A hard rapping at your window wakes youโitโs the โknocker-upper,โ a job that existed for another century, tapping on windows with a long stick because alarm clocks hadnโt been invented. Outside, the โlamplighterโ is still making his rounds, extinguishing the gas street lamps he lit the night before. And you? Youโre a โdriver,โ a person holding the reins of a horse, taking passengers where they need to go.
The knocker-upper is now your smartphone alarm. The lamplighter is the electric streetlight. But the driver? That job, and the routine human task of driving, has persisted. This story is about whether thatโs about to change, and how the word โdriverโ might soon come to mean a machine, just like โdishwasherโ or โcomputerโ once referred to people.
As Alex Davies, author of the excellent book Driven: The Race to Create the Autonomous Car, points out, human driving is fraught with limitations. โI canโt always pay attention to everything,โ he admits. โI get tired.โ Even with the best intentions, we are prone to distraction, fatigue, and road rage. Driving is, for most of us, the riskiest behavior we routinely engage in. Nationally, car accidents cause about 1 in 100 deaths, comparable to guns or opioids.
Itโs this inherent human fallibility that forms the core pitch for the driverless car. They donโt get drunk, tired, or distracted. They never text or feel road rage. And they are no longer a distant dream. Robo-taxis are already providing millions of rides in ten American cities, and even more widely in China.
The dream of a self-driving vehicle is almost as old as the automobile itself. When cars first arrived in the 1800s, replacing horses, society lost a crucial element: sentience. A horse wouldnโt simply run off a cliff if you let go of the reins. Early automobiles, powerful and non-sentient, were met with passionate resistance. People feared for jobsโhorse breeders, farriers, feed suppliers, carriage manufacturers, and especially โteamstersโ (the truckers of their day). They also feared for safety. โRed flag lawsโ proposed requiring a person to walk in front of a car waving a giant red flag to warn people. One Pennsylvania law, thankfully vetoed, would have required drivers encountering livestock to stop, disassemble their car, and hide the parts behind bushes!
Directionally, those โcrazy anti-car activistsโ were right. Cars did initially wipe out many jobs (though they created more), and they were incredibly unsafe. Cities like Detroit, which initially embraced cars without regulation, saw astonishing death rates, many of them children, in the early 1900s. It took decades for society to adapt, inventing laws, licenses, driverโs ed, highways, seatbelts, and airbags to make driving less deadly. Yet, the safety problem remains.
The Desert Race: Where Robots Learned to Think
The quest to make cars more โsentientโ like the horses they replaced continued for decades, fueled by early, often primitive, ideas like radio-controlled cars or magnets under roads. But the true breakthrough wouldnโt arrive until the turn of the millennium, courtesy of a little-known military agency: DARPA.
DARPA, the Pentagonโs research arm, has a mission to keep American technology a generation ahead, having funded everything from GPS to the early internet. In 2002, DARPA director Tony Tether, a former door-to-door salesman with a flair for the dramatic, decided to accelerate driverless car development with a contest: the DARPA Grand Challenge. His original ideaโrace self-driving cars down the Las Vegas Stripโwas quickly deemed โinsaneโ due to the gridlock it would cause. So, the desert outside Las Vegas became the proving ground.
The goal was military: to develop autonomous vehicles that could navigate roads potentially filled with explosive devices, saving American soldiersโ lives. The prize: $1 million. Rules were open: any vehicle type, but no attacking other competitors. Many entrants, surprisingly, came from the world of BattleBots.
This race was not just about the vehicles; it was about the engineers who would go on to lead the billion-dollar companies creating todayโs driverless cars. Their differing philosophiesโfrom methodical academic rigor to entrepreneurial risk-takingโwould shape the future.
One such engineer was Chris Urmson, then a PhD student at Carnegie Mellon University. He joined the โRed Teamโ to build โSandstorm,โ a bright red Humvee bristling with futuristic sensors. For Urmson, the challenge was teaching a computer the basics of operating a Humvee, from steering to braking.
Then there was Anthony Levandowski, a charming, gangly Berkeley grad student and hustler. Lacking the resources of academic powerhouses, Levandowski opted for a radical approach: the raceโs only self-driving motorcycle, โGhost Rider.โ It had almost no chance of winning, but it was guaranteed to get attention.
The 2004 Grand Challenge was, in a word, โan utter hysterical disaster.โ Ghost Rider immediately toppled over because Levandowski forgot to flip a stabilization switch. Other vehicles fared no better, driving onto berms, flipping, or inexplicably U-turning back to the start. Even Sandstorm got stuck, its tires melting from the effort. It was a disaster, but it had flushed out and jump-started a nascent community of roboticists.
Watching this debacle from the sidelines was Sebastian Thrun, a legendary German-born roboticist and AI expert who had taught at Carnegie Mellon before moving to Stanford. Thrun saw a fundamental error in everyoneโs approach. โAll the teams treated this like a hardware problem,โ he observed. โI looked at this and said, โWell, wait a minuteโฆ thatโs actually really a software problem.โโ
Thrun, who had lost a friend to a car accident as a teenager, envisioned something far grander than military applications. He saw the potential to save โa million lives a yearโ worldwide if driverless cars became ubiquitous. This personal mission would drive his subsequent work.
Eighteen months later, in October 2005, DARPA doubled the prize to $2 million for the second Grand Challenge. Chris Urmson returned with the Carnegie Mellon team. Anthony Levandowskiโs motorcycle still didnโt work, getting knocked out in qualifiers. But Thrun, with his Stanford team, brought a modest blue Volkswagen SUV named โStanley.โ
Stanley was Thrunโs software-first vision brought to life. He focused on artificial intelligence, primitive by todayโs standards, but revolutionary for 2005. Stanleyโs โeyesโ looked far ahead, recording what it saw as it drove down dirt roads. It could then โtrain itself,โ learning to distinguish good driving surfaces (grass) from bad ones (mud), detecting patterns and generalizing its knowledge 30 times a second.
The 2005 race was a resounding success. Multiple vehicles finished the 132-mile circular maze. Stanley, the unassuming blue SUV, passed the bulkier Carnegie Mellon entries, conquering a treacherous mountain pass, and crossed the finish line first. Thrun, dressed like a race car driver, celebrated a victory not just for his team, but for the entire โcommunityโ of roboticists. It was a made-for-TV moment, years before the cutthroat competition for driverless cars would begin.
Googleโs Secret Project: Teaching a Machine to Nudge
The success of the DARPA Grand Challenge caught the eye of another spectator that day: Larry Page, co-founder of Google, who attended in a baseball hat and sunglasses as a disguise. Page, who had wanted to do his grad school thesis on autonomous vehicles before being steered toward search engines, was captivated.
He soon hired Sebastian Thrun, along with Anthony Levandowski, initially for Google Street View, modifying Stanleyโs roof-mounted cameras to photograph American streets. But Pageโs true dream was the driverless car. In 2009, he approached Thrun with a mission: โSebastian, I think you should build a self-driving car that can drive anywhere in the world.โ
Thrunโs immediate reaction was a resounding โNo.โ He believed taking the technology from the empty desert and putting it on a bustling San Francisco street would kill someone. Page persisted, day after day. Finally, Page challenged him: โGive me the technical reason why it canโt be done.โ Thrun went home, and to his profound realization, he couldnโt think of one. โExperts are usually experts of the past, not the future,โ he reflected. โIf you ask an expert about innovation, something crazy new, theyโre the least likely person to say, โYes, it can be done.โโ
Thus began Project Chauffeur in 2009, Googleโs secret self-driving car initiative. Led by Thrun, with Chris Urmson running day-to-day operations, Anthony Levandowski on hardware, and Dmitri Dolgov on planning, the small team of 11 engineers reported directly to Larry Page. They had two challenges: safely log 100,000 miles on public roads, and complete the โLarry 1Kโโten tricky 100-mile routes across California without a single human takeover.
To get started, they licensed Stanfordโs code and Levandowski bought eight Toyota Priuses, retrofitting them with radar, cameras, and a spinning 360ยฐ lidar system. These cars, initially given cool names, quickly became โPrius 27โ and so on, as the fleet expanded.
Don Burnett, a researcher who had joined the team after losing a friend in a car accident, worked on โnudging behavior.โ Just as a human driver instinctively nudges left when a large truck passes on the right, Burnettโs job was to teach a computer this subtle, often unconscious, behavior. โYouโre trying to encode the behavior that you would use as a driver under kind of partially good perception,โ he explained. โItโs a really tricky problem.โ
Early testing was done in secret, in the Shoreline Amphitheatre parking lot or an empty airplane runway near Googleโs offices. But in spring 2009, they took a Prius onto the Central Expressway. Immediately, a critical flaw emerged: the car was โswerving wildly,โ like a โdrunken sailor.โ The small oscillations that went unnoticed on a runway became a major problem on a public road.
The team adopted rigorous safety precautions: two-person teams, a safety driver ready to take over, and a partner watching a monitor, calling out discrepancies between what sensors saw and what was actually on the road. This was the painstaking process of teaching a car to drive: logging errors, troubleshooting, updating code.
Burnett became obsessed with the question, โWhy do humans drive the way they drive?โ He found there were no simple answers. The machine learning approach revealed that context profoundly affects perceived comfort. For example, a lateral acceleration of 2 meters per second squared might be comfortable on a highway on-ramp, but doing the same in a cul-de-sac U-turn would feel โincredibly uncomfortable,โ like โMario Kart.โ The limit for a cul-de-sac, it turns out, is around 0.75 m/s^2. โItโs almost three times less than you would be willing to tolerate as you accelerate onto a highway,โ Burnett noted. These subtle, contextual nuances were critical to making the experience comfortable for passengers.
By 2010, the team was on a roll. They started โknocking outโ the Larry 1K routes, each presenting unique challenges, from the Bay Bridge to Lombard Street. The challenge was set up like a video game: try, fail, learn, repeat, until a route was completed without human intervention. They expected it to take two years; they did it in just over one. They celebrated each completed route with a bottle of $13.99 Korbel champagne, signing their names on it.
By fall 2010, the Larry 1K was complete. The team celebrated, throwing each other in the pool at Sebastian Thrunโs house. They had pulled off a miracle, safely navigating tricky California roads with a driverless car, albeit with human supervision and extensive coding. But as the champagne bubbles settled, a new question arose: โOkay, and now what?โ
The Crossroads of a Revolution
The success brought with it new challenges. What exactly was the product they were building? Thrun recalled internal debates: should they have bought Tesla (worth $2 billion at the time)? Was this an assistive technology, a feature to aid human drivers, or a disruptive replacement, waiting until cars could fully drive themselves, perhaps as robo-taxis? Thrun eventually gravitated towards the latter.
But with their secret out, competition would arrive. The team itself would begin to schism, and one member, believing the pace too slow, would take matters into his own hands in an extreme wayโa story for another time.
The journey of the driverless car is a testament to human ingenuity, born from personal tragedy, fueled by visionary engineers, and refined through relentless trial and error. Itโs a story of how technology, once confined to desert races and secret labs, is now quietly, inexorably, reshaping our daily lives, transforming our cities, and forcing us to confront a future where the driver, as we know it, may finally become obsolete.
Based on โJensen Huang โ Will Nvidiaโs moat persist?โ from Dwarkesh Patel Watch the original video
The Maestro of Moats: Jensen Huangโs Bold Defense of Nvidiaโs AI Dominance
In the exhilarating, often dizzying world of artificial intelligence, one company has emerged as the undisputed kingmaker: Nvidia. Its GPUs power the most sophisticated AI models, driving unprecedented technological leaps and a staggering market valuation. Yet, as the industry matures and competition intensifies, a critical question looms: can Nvidiaโs formidable โmoatโ truly persist?
In a candid conversation with Dwarkesh Patel, Nvidia CEO Jensen Huang offered a robust defense of his companyโs enduring advantage, revealing a multi-layered strategy that spans from fundamental physics to global supply chain orchestration and a unique philosophy of strategic investment.
From Electrons to Tokens: Nvidiaโs Core Identity
A common, perhaps โnaive,โ interpretation of Nvidiaโs business suggests vulnerability. Observers note that Nvidia designs software (GDS2 files) and relies on external partnersโTSMC for logic, SK Hynix, Micron, and Samsung for memory, and Taiwanese ODMs for assembly. If AI is poised to commoditize software, could Nvidia, at its core a software company, be next?
Jensen Huang dismisses this notion with a foundational principle: โIn the end, something has to transform electrons to tokens.โ This isnโt just a technical process; itโs an โincredible journeyโ demanding immense โartistry, engineering, science, and invention.โ He posits that making one โtokenโ more valuable than anotherโthe very essence of AIโs progressionโis inherently difficult to commoditize.
Nvidiaโs role, as Huang sees it, is to be the indispensable orchestrator of this transformation. โThe input is electrons, the output is tokens. In the middle is Nvidia,โ he states. Their job is to do โas much as necessary and as little as possibleโ to enable this transformation at โincredible capabilities.โ The โas little as possibleโ part is key: Nvidia partners extensively, building the largest ecosystem of upstream and downstream partners, from chip manufacturers and memory providers to cloud companies, application developers, and model makers. This five-layer cake of AI, according to Huang, is where Nvidia embeds itself, focusing on the โinsanely hardโ parts that cannot be commoditized.
The Software Paradox: More Tools, Not Less
The fear that AI will commoditize enterprise software is also, in Huangโs view, misplaced. He argues the opposite will happen: โI think the number of agents is going to grow exponentially, and the number of tool users is going to grow exponentially.โ Tools like Excel, PowerPoint, Cadence, and Synopsys wonโt disappear; theyโll become even more widely used, albeit by AI agents rather than solely human engineers.
โToday weโre limited by the number of engineers. Tomorrow, those engineers are going to be supported by a bunch of agents,โ Huang predicts. This means an explosion in the โinstances of all these tools,โ leading to a โskyrocketโ in software company revenues. The only current limitation, he notes, is that โthe agents arenโt good enough at using their tools yet.โ Once they are, or once companies build agents specifically for their tools, the software market will expand dramatically.
The Supply Chain: A Moat Forged in Trust and Scale
Beyond technological prowess, Nvidiaโs moat extends deep into the global supply chain. Recent reports of Nvidiaโs colossal purchase commitmentsโpotentially $250 billionโspark questions: is this simply locking up scarce components? Huang confirms these โenormous commitments,โ both explicit and implicit, are a significant factor.
He describes a unique, almost evangelical role he plays in inspiring upstream partners. โI said to the CEOs, โLet me tell you how big this industry is going to be, let me explain to you why, let me reason through it with you, and let me show you what I see.โโ This process of โinforming, inspiring, and aligningโ convinces suppliers to make massive investments for Nvidia that they might not for others. Why? Because โthey know that I have the capacity to buy their supply and sell it through my downstream.โ
Nvidiaโs GTC conference, Huang explains, serves as a physical manifestation of this ecosystem, bringing together the entire โ360 degrees, the entire universe of AI.โ Itโs where upstream and downstream partners connect, see the latest AI advances, and meet โAI nativesโ โ the startups building the future. This visibility reinforces their confidence in Nvidiaโs demand, allowing the company to โbuild for a futureโ where its scale could reach โa trillion dollars.โ
Conquering Bottlenecks: A Two-to-Three-Year Problem
The sheer scale of Nvidiaโs growth raises legitimate concerns about the supply chainโs ability to keep pace. With Nvidia being the largest customer on TSMCโs advanced nodes (N3, N2) and AI projected to consume a dominant share of future capacity, how can they continue to double output year after year?
Huang frames this as a โgood conditionโ: instantaneous demand exceeding supply signifies a booming industry. He acknowledges temporary bottlenecks, even humorously suggesting inviting โplumbersโ to GTC as a critical component. However, he asserts that hardware-related bottlenecks are inherently short-lived.
โIf weโre too far apart, if one particular component is too far away, the industry swarms it,โ he explains. He cites CoWoS packaging technology as an example: initially a bottleneck, it was โswarmedโ with investment and scaled rapidly. Now, TSMC is aligning CoWoS and future packaging with logic and memory demand.
Nvidia actively โprefetch[es] the bottlenecks years in advance,โ investing in new technologies (like silicon photonics with Lumentum and Coherent) and licensing patents to keep the ecosystem open. His confidence in scaling is striking: โNone of that is impossible to scale quickly. All of that is easy to do within two or three years. You just need a demand signal. Once you can build one, you can build ten, and once you can build ten, you can build a million.โ
The real long-term concern, in Huangโs view, isnโt chip or packaging capacity, but rather fundamental energy policy. โYou canโt create an industry without energy,โ he warns, highlighting the need for robust energy infrastructure to support the reindustrialization efforts and the construction of โAI factories.โ
Moreover, capacity increases are only one part of the equation. Nvidia constantly drives โcomputing efficiency by 10x, 20x, and in the case of Hopper to Blackwell, 30x to 50x.โ This combination of increased supply and dramatically improved efficiency mitigates the impact of any temporary bottlenecks.
Beyond TPUs: The Power of Accelerated Computing
The rise of custom ASICs like Googleโs TPUs, which power major models like Claude and Gemini, presents a direct competitive challenge. Critics argue that AI primarily involves predictable matrix multiplies, for which specialized TPUs are optimally designed, potentially outperforming flexible GPUs.
Huang strongly refutes this, emphasizing Nvidiaโs focus on โaccelerated computing, not a tensor processing unit.โ He stresses that while matrix multiplies are important, they are not the only part of AI. The invention of new attention mechanisms, hybrid architectures, and fused diffusion/autoregressive techniques demands a โgenerally programmableโ architecture.
โThe ability to invent new algorithms is really what makes AI advance so quickly,โ Huang states. Mooreโs Law, with its 25% annual gains, isnโt enough; 10x or 100x leaps require fundamental algorithmic changes, which are only possible with programmable systems like CUDA. He points to the 50x energy efficiency jump from Hopper to Blackwell, achieved through โnew models, like MoEs, that are parallelized, disaggregated, and distributed across a computing system.โ Nvidiaโs โextreme co-design companyโ approach, affecting processors, systems, fabric, libraries, and algorithms simultaneously, is what enables these leaps.
CUDAโs Enduring Reign: Ecosystem, Install Base, and TCO
Even with hyperscalers like OpenAI developing custom kernels with tools like Triton, Huang argues that CUDA remains invaluable. He highlights three key advantages:
- Richness and Programmability: CUDA is a vast ecosystem, supporting every framework. Nvidia actively contributes to tools like Triton, ensuring their backend leverages Nvidiaโs technology. This provides a trusted, robust foundation for developers.
- Install Base: โThe single most important thing you want is an install base,โ Huang says. With โseveral hundred million GPUs out there now,โ developers building on CUDA know their software will run โeverywhere,โ from cloud servers to robots. This ubiquity is โincredibly valuable.โ
- Versatility and Reach: Nvidia systems are in โevery cloudโ (Google, Amazon, Azure, OCI) and available on-prem. This unparalleled versatility means AI companies arenโt locked into a single provider, making Nvidia the default choice for those seeking maximum reach.
Huang also asserts that Nvidia offers the โbest performance per TCO in the world, bar none.โ He challenges competitors like TPU and Trainium to demonstrate superior cost-effectiveness using benchmarks like InferenceMAX and MLPerf, claiming โnobody wants to show up.โ He also points to Nvidiaโs โhighest tokens per watt architecture,โ which maximizes revenue generation for data centers. This combination of superior TCO, energy efficiency, and market reach creates a powerful โflywheelโ effect.
A Surprising Admission: Nvidiaโs Strategic Investments
Perhaps the most surprising revelation came when discussing why major AI companies like Anthropic have diversified their compute away from Nvidia, partnering with Google/Broadcom for TPUs. Huang conceded a past โmistakeโ: โI didnโt deeply internalize how difficult it would be to build a foundation AI lab like OpenAI and Anthropic, and the fact that they needed huge investments from the supplier themselves.โ
He explains that Google and AWS were able to make multi-billion dollar upfront investments in Anthropic, securing their compute business, something Nvidia was not โin a position to doโ at the time. โI didnโt realize we needed to,โ he admitted, thinking they could โjust go raise from VCs, for Godโs sakes.โ
โBut Iโm not going to make that same mistake again,โ Huang declared. Nvidia is now โdelighted to invest in OpenAIโ and Anthropic, recognizing that โthe world needs them to exist.โ This strategic shift, from purely selling hardware to also investing in key AI labs, underscores Nvidiaโs commitment to nurturing the entire AI ecosystem.
Despite having vast cash reserves, Nvidia resists becoming a cloud provider itself. This adheres to their โas much as needed, as little as possibleโ philosophy. Building the core computing platform is a unique contribution, but cloud services are a crowded field where others can step in. Instead, Nvidia invests in โneocloudsโ like CoreWeave, Nscale, and Nebius, ensuring the ecosystem thrives.
Finally, Huang reveals a core principle: โDonโt pick winners.โ Drawing from Nvidiaโs own early struggles in the graphics industry, where their initial architecture was โprecisely wrong,โ he emphasizes humility. โIf you would have taken those 60 graphics companies and asked yourself which one was going to make it, Nvidia would be at the top of that list not to make it,โ he muses. Therefore, Nvidia supports all foundation model companies, viewing it as โimperative to our business.โ
The Allocation Myth: No Highest Bidder, Just Orders
Addressing persistent rumors about how Nvidia allocates its scarce GPUs, Huang firmly refutes the idea of โfracturing the marketโ or favoring the โhighest bidder.โ He outlines a straightforward process: forecasting, placing purchase orders, and โfirst in, first out.โ
โIf you donโt place a PO, all the talking in the world wonโt make a difference,โ he states. While minor adjustments might occur based on a customerโs data center readiness to maximize throughput, the core principle is simple. He even debunks the widely reported anecdote of Larry Ellison and Elon Musk โbegging for GPUsโ at dinner, confirming the dinner happened but โat no time did they beg for GPUs. They just had to place an order.โ
Why not simply sell to the highest bidder? โBecause itโs a bad business practice,โ Huang asserts. โYou set your price and then people decide to buy it or not.โ
A Moat Built on Vision and Execution
Jensen Huangโs narrative paints a picture of a company whose dominance is not accidental but deeply strategic, built on a unique blend of technological innovation, ecosystem cultivation, and prescient market understanding. From the fundamental transformation of โelectrons to tokensโ to the orchestration of a global supply chain, Nvidiaโs moat is multifaceted.
Itโs a moat reinforced by an unwavering commitment to programmability, a vast install base, and an ecosystem that fosters continuous innovation. And while past missteps are acknowledged, Nvidiaโs willingness to adapt its investment strategy shows a dynamic, forward-looking approach. In the relentless race for AI supremacy, Jensen Huang believes Nvidiaโs position, earned through decades of dedicated effort, is not just persistent, but increasingly indispensable.
Based on โTrumpโs Risky Strategy to Blockade Iranโs Blockadeโ from New York Times Podcasts Watch the original video
Strait of Hormuz: Trumpโs High-Stakes Gambit in the Blockade of Blockades
The waters of the Strait of Hormuz, a critical choke point for global oil supplies, have become the stage for a perilous game of โblockade chicken.โ In a bold and risky maneuver, the United States has deployed a formidable naval presence, enforcing what it calls a blockade of Iran, designed to bring an end to a simmering conflict on American terms. But this isnโt just any blockade; itโs a direct counter to Iranโs own control over the vital shipping channel, creating a complex and potentially explosive situation that experts warn could reshape the global economy for years to come.
At its core, the American strategy is stark: strangle Iranโs economy by cutting off its oil and gas revenues, forcing Tehran back to the negotiating table. โWeโre blockading the ports in Iran where they get oil and gas shipments,โ explains White House correspondent David Sanger. โWithout oil and gas money, Iran has no economy.โ This aggressive approach, however, carries immense dangers and uncertain outcomes, prompting a critical look at its strategy, the risks it poses, and whether it can truly work.
The Genesis of a Risky Strategy
The decision to impose a naval blockade emerged directly from a diplomatic impasse. Following Vice President JD Vanceโs recent, fruitless negotiations in Pakistanโwhich saw him return empty-handedโthe Trump administration found itself confronting a โpretty messy situation.โ A ceasefire was in place, but Iran continued to exercise control over who transited the Strait of Hormuz, a vital artery for global commerce.
For 47 years, since the Islamic revolutionary government came to power in 1979, Iran had largely allowed free passage through the strait. But in the recent conflict, they had begun to assert control, stopping traffic and even imposing โtollsโ on passing ships, some reportedly as high as $2 million. This was a dynamic the US administration found โintolerable.โ The US Navyโs objective became clear: reverse the situation and ensure that it was the US, not Iran, controlling traffic through the strait. While this sounds straightforward on paper, given the might of the US Navy, its execution promises to be anything but.
An Act of War: Defining the Blockade
Militarily, a naval blockade is a grave measure. โBy definition, a blockade is an act of war,โ states military correspondent Eric Schmidt. It involves one nation using its military power to block the transit of ships from other countries, either through the threat of force or the actual boarding and seizing of vessels.
The US deployment is substantial: approximately 10,000 sailors aboard more than a dozen warships, ranging from aircraft carriers to destroyers and Marine-carrying vessels, are now positioned outside the Strait of Hormuz. Their mission is to intercept ships either leaving the Persian Gulf or attempting to enter it.
The primary target of this economic strangulation is Iranโs oil exports. Interestingly, Iran had managed to maintain its own oil shipments through the strait even while exerting control and threatening other vessels. This allowed Tehran to continue funding its war efforts, particularly the Islamic Revolutionary Guard Corps (IRGC), which is almost entirely dependent on oil revenues. The US blockade, therefore, aims to cut off this crucial lifeline.
The ultimate goals of this audacious move are two-fold: to reclaim control of the strait from Iranian influence and, critically, to force Iran back to the negotiating table. President Trump aims to leverage control over the strait to compel Iran to concede on key issues, such as its nuclear stockpiles and uranium enrichment program.
The Looming Dangers: Blowback and Boomerangs
Such a high-stakes gamble comes with equally high risks. The panel of experts identified several potential โblowbacksโ that could escalate the conflict dramatically.
- Iranian Retaliation: The most immediate concern is that Iranโs IRGC could lash out, attacking US Navy ships and triggering a major escalation of hostilities.
- Chinaโs Wrath: A significant portionโan estimated 90%โof Iranโs oil exports are destined for China, often carried on Chinese-flagged vessels. White House correspondent David Sanger highlights the danger: โA simple, efficient way to really piss off your superpower adversary is to start to systematically deprive them of oil.โ This could severely complicate an upcoming meeting between President Trump and Beijing, intended to focus on trade and security, and further strain relations already tense with reports of China considering arming Iran.
- Regional Energy Infrastructure Attacks: Energy reporter Rebecca Elliot warns of a โthird big category of riskโ: Iran restarting attacks on energy infrastructure across the Persian Gulf. Such actions carry โlong-term risks for the global energy system and the global economy.โ The International Energy Agency has already estimated that over 80 energy sites in the region have been damaged, and restoring pre-war production levels could take up to two years.
Early Days: A โQuarantineโ in Action
In its initial 48 hours, the US blockade has presented a complex picture. Eric Schmidt clarifies that what President Trump terms a โblockadeโ is โreally more of a quarantine.โ The US Navy is strategically positioned outside the strait, using drones and open-source intelligence to monitor ships bound for Iranian ports. Suspect vessels are contacted via radio, with the threat of boarding partiesโMarines or Navy Sealsโto inspect cargo if they refuse to comply.
The deterrent effect appears to be working, at least partially. Central Command reported no โinterdictionsโ or โviolationsโ in the first 24 hours. Crucially, six vessels departing Iranian ports reportedly turned around after being contacted by US Navy ships. โIt looks like for now he has stopped the Iranians from shipping oil out and thus fueling their economy and their government,โ Sanger notes, calling it a โhalfwayโ success.
However, a major question mark remains: Can commerce from other Gulf statesโlike the UAEโnow pass safely through the strait? The situation is, as Sanger puts it, โa blockade by the US of a blockade by Iran,โ making it incredibly messy. The global economy hinges on whether shippers will gain the confidence to navigate this precarious route.
Global Ripples and a Strait Transformed
The trepidation among shipping companies and captains is โvery high,โ according to Rebecca Elliot. Insurance costs for vessels attempting to transit the strait have soared. Many in the industry are now asking not if the strait will reopen, but if it will ever return to its pre-war state of free passage and minimal risk. The consensus among experts is grim: โProbably not.โ
David Sanger agrees, noting that Iran has โdiscovered this kind of superpower they haveโ in their ability to control the strait with minimal military effort. The genie, it seems, cannot easily be put back in the bottle.
Proposals for the straitโs future include President Trumpโs brief, intriguing suggestion of a joint business venture with Iranโs Ayatollah Khamenei, or more seriously, an international consortium. Such a body, involving Iran, Oman, the US (providing security), China, India, and European nations, would monitor traffic, potentially charge tolls, and divide revenues. However, this would demand a level of international cooperation and power-sharing that the current US administration has not typically embraced.
The long-term implications for global energy infrastructure are profound. If the Strait of Hormuz no longer guarantees free, unimpeded commerce, the world will need alternatives. These include:
- Alternative Routes: Building more pipelines from Gulf countries to world markets, bypassing the strait. While some options exist in Saudi Arabia and the UAE, these are complicated by the need to traverse other nationsโ territories.
- Diversified Oil Sources: Increased demand for oil from other regions that donโt require transit through the strait.
- Accelerated Energy Transition: Higher oil prices inherently make alternative energy sourcesโnuclear, solar, batteriesโmore economically attractive, potentially accelerating the global shift away from fossil fuels.
โThe entire world of energy and energy infrastructure could end up changing because of this war and to a degree because of this blockade,โ says Elliot, โand that could end up long outlasting the conflict itself.โ
A Contest of Endurance
Ultimately, this blockade is a test of wills. โWe are now entering a contest over which country, the United States or Iran, can endure the pain of this blockade over the next few months,โ Sanger posits.
Iran is betting that President Trump will be forced to back down. Rising gas prices, particularly as midterm elections approach, pose a significant political problem for the Republican party and Trumpโs domestic agenda. Iranian leadership has openly taunted Trump about gas prices, with one negotiator reportedly stating, โYou could be nostalgic for five or six dollar a gallon gas.โ
The US, conversely, believes economic pressure will succeed where military action failed. By cutting off Iranโs main revenue source and specifically targeting the IRGC, Washington hopes to compel Tehran to โcry uncleโ and accept American terms.
The outcome remains highly uncertain. โIf anybody tells you they know which side is not going to be able to withstand the pain first, Iโm not sure Iโd believe them yet,โ Sanger concludes.
The Costs of Sustained Pressure
Even if successful, the blockade comes at a steep price. The Pentagon asserts its ability to sustain the operation indefinitely, but this drains resources from other critical areas. Ships and munitions urgently needed in the Indo-Pacific to counter China or address North Korea are being diverted. Interceptors and other armaments are being pulled from European Command, impacting support for Ukraine.
The operation alone requires some 10,000 US Navy, Marine Corps, and other service personnel, a number that could swell to over 50,000 if the region remains a primary focus and hostilities resume.
While the current full-on war has largely ceased, the underlying tensions and strategic gaps between the US and Iranโover nuclear programs, regional conflicts, and moreโremain enormous. Both sides face immense pressure to avoid a return to widespread military conflict, as Trumpโs base fragmented during the earlier fighting, Congress was frustrated, and allies were conspicuously absent. Iran, too, is desperate for the war to end due to its devastating economic impact.
The question now is not whether the war ends, but โon whose terms it ends.โ As France and Britain begin to develop their own plan to reopen the Strait of Hormuzโa plan that may exclude the United States and only commence after the warโs conclusionโthe enduring legacy of this high-stakes blockade may be a permanently altered geopolitical and energy landscape.
ํ๊ตญ์ด
โThe AI Model Built for What LLMs Canโt Doโ โ Every ๊ธฐ๋ฐ ๊ธฐ์ฌ ์๋ณธ ์์ ๋ณด๊ธฐ
LLM์ด ํ์ง ๋ชปํ๋ ๋์ : โ์ ํ์ฑโ๊ณผ โ์ดํดโ๋ฅผ ์ถ๊ตฌํ๋ AI, EBM์ ๋ถ์
์ต๊ทผ ์ธ๊ณต์ง๋ฅ(AI) ๋ถ์ผ๋ ๋๊ท๋ชจ ์ธ์ด ๋ชจ๋ธ(Large Language Model, LLM)์ ํ์ ์ ์ธ ๋ฐ์ ์ผ๋ก ์ ๋ก ์๋ ์ฃผ๋ชฉ์ ๋ฐ๊ณ ์์ต๋๋ค. ํ ์คํธ ์์ฑ๋ถํฐ ๋ฒ์ญ, ๋ณต์กํ ์ง๋ฌธ ๋ต๋ณ์ ์ด๋ฅด๊ธฐ๊น์ง LLM์ ๋ฅ๋ ฅ์ ์ค๋ก ๋๋์ต๋๋ค. ๊ทธ๋ฌ๋ ์ด๋ฌํ ๋๋ถ์ ๋ฐ์ ์ ์ด๋ฉด์๋ โ์ ํ์ฑ(correctness)โ๊ณผ โ์ ๋ขฐ์ฑ(reliability)โ์ด๋ผ๋ ๋ณธ์ง์ ์ธ ํ๊ณ๊ฐ ์กด์ฌํฉ๋๋ค. ๊ณผ์ฐ ์ฐ๋ฆฌ๋ AI๊ฐ ์์ฑํ๋ ๊ฒฐ๊ณผ๋ฌผ์ ๋งน๋ชฉ์ ์ผ๋ก ์ ๋ขฐํ ์ ์์๊น์? ํนํ ์์จ์ฃผํ์ฐจ๋ ์๋ฃ ์ง๋จ๊ณผ ๊ฐ์ ๋ฏธ์ ํฌ๋ฆฌํฐ์ปฌ(mission-critical) ์์คํ ์ AI๋ฅผ ์ ์ฉํ ๋, ์ด๋ฌํ ์ง๋ฌธ์ ๋์ฑ ์ค์ํด์ง๋๋ค.
์ฌ๊ธฐ, ์ด๋ฌํ LLM์ ๊ทผ๋ณธ์ ์ธ ํ๊ณ๋ฅผ ๊ทน๋ณตํ๊ณ , ๋์ฑ ์์ธก ๊ฐ๋ฅํ๋ฉฐ ๊ฒ์ฆ ๊ฐ๋ฅํ AI๋ฅผ ์ถ๊ตฌํ๋ ์๋ก์ด ์ ๊ทผ ๋ฐฉ์์ ์ ์ํ๋ ๊ธฐ์ ์ด ์์ต๋๋ค. ๋ฐ๋ก ํ์ด๋ฐ์ด์ ๋ AI(foundational AI) ๊ธฐ์ ์ธ ๋ก์ง์ปฌ ์ธํ ๋ฆฌ์ ์ค(Logical Intelligence)์ ์ฐฝ๋ฆฝ์์ด์ CEO์ธ ์ด(Yee)์ ๋๋ค. ๊ทธ๋ ๊ธฐ์กด LLM์ ํ๊ณ๋ฅผ ๋ช ํํ ์ง์ ํ๋ฉฐ, โ์๋์ง ๊ธฐ๋ฐ ๋ชจ๋ธ(Energy-Based Model, EBM)โ์ด๋ผ๋ ๋์์ ์ํคํ ์ฒ๊ฐ AI์ ๋ฏธ๋์ ํ์์ ์ด๋ผ๊ณ ๊ฐ์กฐํฉ๋๋ค.
์ธ๊ณต์ง๋ฅ์ ์ ๋ขฐ์ฑ: ์ ํ์ด ์๋ ํ์
๋ก์ง์ปฌ ์ธํ ๋ฆฌ์ ์ค๋ LLM๊ณผ EBM์ ๋ชจ๋ ํ์ฉํ๋ ํ์ด๋ธ๋ฆฌ๋ ์ ๊ทผ ๋ฐฉ์์ ์ทจํ์ง๋ง, ๊ถ๊ทน์ ์ผ๋ก๋ EBM์ ํตํด AI์ โ์ ํ์ฑโ๊ณผ โ๊ฒ์ฆ ๊ฐ๋ฅ์ฑโ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ณ ์ ํฉ๋๋ค. ์ด CEO๋ ์ค๋๋ AI๊ฐ ์ฝ๋ ์์ฑ์ด๋ ์นฉ ์ค๊ณ์ ๊ฐ์ ํต์ฌ ์์คํ ์ ์ฌ์ฉ๋์ง๋ง, โ๊ณผ์ฐ ๊ทธ ๊ฒฐ๊ณผ๋ฌผ์ด ์ผ๋ง๋ ์ ํํ๊ณ ํฉ๋ฆฌ์ ์ธ๊ฐ?โ๋ผ๋ ์ง๋ฌธ์ ๋์ง๋ ์ด๋ค์ ๋งค์ฐ ์ ๋ค๊ณ ์ง์ ํฉ๋๋ค. ๊ทธ๋ โ๊ฒฐ๊ณผ๊ฐ ์๋ํ๊ธฐ๋ง ํ๋ฉด ์ ์ ํ์ฑ์ด ์ค์ํ๊ฐ?โ๋ผ๋ ์ง๋ฌธ์ ๋ํด ์์จ์ฃผํ์ฐจ์ ๋นํ๊ธฐ ์กฐ์ข ์ ์๋ฅผ ๋ค์ด ์ค๋ช ํฉ๋๋ค.
โ๋ง์ฝ AI๊ฐ ์ด์ ํ๋ ์ฐจ์ ๋น์ ์ด ํ๊ณ ์๋๋ฐ, 20% ํ๋ฅ ๋ก ํ๊ฐ(hallucinate)์ ์ผ์ผ์ผ ์๋ฑํ ๊ณณ์ผ๋ก ๊ฐ ์ ์๋ค๊ณ ํ๋ค๋ฉด ์ด๋ป๊ฒ ์ต๋๊น?โ ๊ทธ๋ ๋ ๋์๊ฐ โ๋นํ๊ธฐ๊ฐ 20% ํ๋ฅ ๋ก ๋ค์ ๋จ์ด๋ฅผ ์๋ชป ์์ธกํด ์ถ๋ฝํ ์ ์๋ค๊ณ ํ๋ค๋ฉด ์ด๋ป๊ฒ ์ต๋๊น?โ๋ผ๊ณ ๋ฌป์ต๋๋ค. ๋ฌผ๋ก ํ์ฌ์ ๋นํ๊ธฐ๋ ๊ฒฐ์ ๋ก ์ (deterministic) ์์คํ ์ผ๋ก ์ ์๋ํ๊ณ ์์ง๋ง, ์ด CEO๋ AI๊ฐ ๋ชจ๋ ์์คํ ์ ์ ์ฉ๋๋ ๋ฏธ๋๊ฐ ๋ถ๊ฐํผํ๋ค๊ณ ๋งํฉ๋๋ค. ์ํ ์ ๋ฌด์์ AI๊ฐ ์๋ํ๋ฅผ ๋๊ณ ์์ฌ๊ฒฐ์ ์ ๊ฐ์ ํ๋ ๊ฒ์ฒ๋ผ, AI๋ ์ฐ๋ฆฌ์ ์๊ฐ์ ์ ์ฝํ๊ณ ๋ ์ฐฝ์์ ์ธ ์ผ์ ์ง์คํ ์ ์๊ฒ ํ ๊ฒ์ ๋๋ค. ๊ทธ๋ฌ๋ ์ด๋ฌํ โ์๋ํโ์ โ์ฐฝ์์ฑโ์ โ์ ๋ขฐ์ฑโ์ด๋ผ๋ ์ ์ ์์ด๋ ๋ถ๊ฐ๋ฅํฉ๋๋ค.
LLM์ ๋ณธ์ง์ ์ผ๋ก ์ธ์ด ๊ธฐ๋ฐ ๋ชจ๋ธ์ด๋ฉฐ, ๊ทธ ์ํคํ ์ฒ๋ ๋ด๋ถ ๊ฒ์ฆ์ ์ด๋ ต๊ฒ ๋ง๋ญ๋๋ค. ๋ง์น โ๋ธ๋๋ฐ์คโ์ ๊ฐ์์, ๋ชจ๋ ์ฒ๋ฆฌ ๊ณผ์ ์ด ์๋ฃ๋๊ธฐ ์ ๊น์ง๋ ๋ด๋ถ๋ฅผ ๋ค์ฌ๋ค๋ณผ ์ ์์ต๋๋ค. ๋ฌผ๋ก ๋ฆฐ 4(Lean 4)์ ๊ฐ์ ๊ธฐ๊ณ ๊ฒ์ฆ ๊ฐ๋ฅํ ์ฆ๋ช ์ธ์ด๋ฅผ ์ฌ์ฉํ์ฌ LLM์ ์ถ๋ ฅ์ ์ธ๋ถ์์ ๊ฒ์ฆํ ์๋ ์์ต๋๋ค. ํ์ง๋ง ์ด๋ ์ฌ์ ํ LLM์ด โ์ถ์ธก ๊ฒ์โ์ ๋ฒ์ด๋ ๊ณผ์ ์์ ๋ฐ์ํ๋ ๋ง๋ํ ์ปดํจํ ๋น์ฉ๊ณผ ํ ํฐ ๋น์ฉ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ์ง ๋ชปํฉ๋๋ค. LLM์ ์ฌ์ ํ ๋น์ผ ์ปดํจํ ์์์ ์๋ชจํ๋ฉฐ ๋ค์ ํ ํฐ์ ์์ธกํ๋ โ์ถ์ธก ๊ฒ์โ์ ํด์ผ ํฉ๋๋ค.
LLM๊ณผ EBM: ๊ทผ๋ณธ์ ์ธ ์ํคํ ์ฒ์ ์ฐจ์ด
์ด CEO๋ ์ด๋ฌํ ๋ฌธ์ ๊ฐ EBM์ ํตํด ํด๊ฒฐ๋ ์ ์๋ค๊ณ ์ค๋ช ํฉ๋๋ค. EBM์ โ์๋์ง ๊ธฐ๋ฐ ๋ชจ๋ธ(Energy-Based Model)โ์ ์ฝ์๋ก, ํ ํฐ์ ์ฌ์ฉํ์ง ์๋ โํ ํฐ ํ๋ฆฌ(token-free)โ ๋ชจ๋ธ์ ๋๋ค. ๊ทธ๋ EBM์ ๊ทผ๋ณธ์ ์ธ ์ฐจ์ด๋ฅผ ์ค๋ช ํ๊ธฐ ์ํด ๋ฌผ๋ฆฌํ์ โ์๋์ง ์ต์ํ ์๋ฆฌโ๋ฅผ ๋น์ ๋ก ๋ญ๋๋ค.
์๋์ง ์ต์ํ ์๋ฆฌ: ๋ฌผ๋ฆฌํ์์ ๋ชจ๋ ์์คํ ์ ์๋์ง๋ฅผ ์ต์ํํ๋ ค๋ ๊ฒฝํฅ์ด ์์ต๋๋ค. ์๋ฅผ ๋ค์ด, ์ฐ๋ฆฌ๊ฐ ํธ์ํ๊ฒ ์์์ ์์ ์๋ ๊ฒ์ ์๋์ง๋ฅผ ์ต์ํํ๋ ์์ฐ์ค๋ฌ์ด ์ํ์ ๋๋ค. ์ด๋ก ๋ฌผ๋ฆฌํ์๋ค์ ์์คํ ์ ์๋์ง์ ๊ด๋ จ๋ ํญ(๋ผ๊ทธ๋์ง์)์ ์ ์ํ๊ณ , ์ด๋ฅผ ์ต์ํํ์ฌ ์์คํ ์ ์ด๋ ๋ฐฉ์ ์๊ณผ ๋ณด์กด ๋ฒ์น์ ๋์ถํฉ๋๋ค. ์ด ์๋ฆฌ๋ ์์คํ ์ด ๊ฐ์ฅ ์์ ์ ์ด๊ณ ํ๋ฅ ์ด ๋์ ์ํ๋ฅผ ์ฐพ์๊ฐ๋ ๊ณผ์ ์ ์ค๋ช ํฉ๋๋ค.
AI์์ ์ด ์๋ฆฌ๋ ์์คํ ์ด ์ ๋ณด๋ฅผ ์ฒ๋ฆฌํ ๋ โ์๋์ง ํจ์โ๋ฅผ ๊ตฌ์ฑํ๊ณ ์ด๋ฅผ ์ต์ํํ๋ ๋ฐฉ์์ผ๋ก ์ ์ฉ๋ฉ๋๋ค. ์ฆ, EBM์ ๊ฐ์ฅ โํ๋ฅ ๋์โ ๋๋ โ์์ ์ ์ธโ ์ํ๋ฅผ ์ฐพ์๊ฐ๋ ๋ฐฉ์์ผ๋ก ์๋ํฉ๋๋ค. ์ด CEO๋ ์ด๋ฅผ โํผ๊ณคํ ๋์ด ์ง์ ๋์์ ๊ฐ์ฅ ํธ์ํ ์์ธ๋ก ์ํ์ ์์ ๊ฒโ์ด๋ผ๋ ์์ธก์ ๋น์ ํฉ๋๋ค. ๋์ด ์ค๊ฑฐ์ง๋ฅผ ํ๊ฑฐ๋ ์ง์์ ๋์๋ค๋๋ ๋ฑ ๋ค์ํ ์ํ๊ฐ ์์ ์ ์์ง๋ง, ํผ๊ณคํ ๋ ๊ฐ์ฅ ํ๋ฅ ๋์ ์๋๋ฆฌ์ค๋ ์ํ์์ ํด์ํ๋ ๊ฒ์ ๋๋ค. EBM์ ์ด์ฒ๋ผ ๋ชจ๋ ๊ฐ๋ฅํ ์๋๋ฆฌ์ค๋ฅผ โ์๋์ง ์งํ(energy landscape)โ์ผ๋ก ๋งคํํ๊ณ , ๊ฐ์ฅ ๋ฎ์ ์ง์ (๊ฐ์ฅ ํ๋ฅ ๋์ ์ํ)์ ์ฐพ์๊ฐ๋๋ค.
EBM์ ํต์ฌ ํน์ง:
- ๋น์์ฐจ์ (Non-Autoregressive): LLM์ฒ๋ผ ํ ํฐ์ ์์๋ฅผ ์์ธกํ๋ ๋ฐฉ์์ด ์๋๋๋ค. ์ด๋ก ์ธํด โ์ถ์ธก ๊ฒ์โ์ด ์์ด์ง๋ฉฐ, ๋น์ฉ์ด ํจ์ฌ ์ ๋ ดํด์ง๋๋ค.
- ๊ฒ์ฆ ๊ฐ๋ฅ์ฑ(Inspectability): EBM์ โ๋ธ๋๋ฐ์คโ๊ฐ ์๋๋๋ค. ํ๋ จ ๊ณผ์ ์ค ์ธ์ ๋ ์ง ๋ชจ๋ธ ๋ด๋ถ๋ฅผ ๋ค์ฌ๋ค๋ณด๊ณ ์ด๋ค ์ผ์ด ๋ฒ์ด์ง๊ณ ์๋์ง ํ์ธํ ์ ์์ต๋๋ค. ์ด๋ LLM ์ํคํ ์ฒ๋ก๋ ๋ถ๊ฐ๋ฅํ ์ผ์ ๋๋ค.
- ์๊ธฐ ์ ๋ ฌ(Self-alignment): ์ ๋ณด ์ฒ๋ฆฌ ๊ณผ์ ์์ ์ค์ค๋ก ์ ๋ ฌํ๋ ์ํคํ ์ฒ๋ฅผ ๊ฐ์ง๊ณ ์์ด, ๊ฒ์ฆ ์์ ์ ์ ๋ฆฌํฉ๋๋ค.
- ๋ด๋ถ ๋ฐ ์ธ๋ถ ๊ฒ์ฆ: EBM์ ์ํคํ ์ฒ ์์ฒด์ ๊ฒ์ฆ ๊ธฐ๋ฅ์ด ๋ด์ฌ๋์ด ์์ ๋ฟ๋ง ์๋๋ผ, LLM์ ์ฌ์ฉ๋๋ ๊ฒ๊ณผ ๋์ผํ ์ธ๋ถ ๊ฒ์ฆ๊ธฐ๋ฅผ ์ถ๊ฐ๋ก ๋ถ์ฐฉํ ์ ์์ด โ์ด์ค ๊ฒ์ฆโ์ด ๊ฐ๋ฅํฉ๋๋ค.
EBM์ โ์ดํดโ์ ์ ์ฌ ๋ณ์
์ด CEO๋ LLM์ด ๋ฐ์ดํฐ๋ฅผ โํจํดโ์ผ๋ก๋ง ์ธ์ํ๋ ๋ฐ๋ฉด, EBM์ ๋ฐ์ดํฐ๋ฅผ โ์ดํดโํ๋ค๊ณ ์ค๋ช ํฉ๋๋ค. LLM์ ๋ฐฉ๋ํ ๋ฐ์ดํฐ๋ฅผ ํตํด ๊ฐ์ฅ ํ๋ฅ ๋์ ํจํด์ ํ์ตํ์ง๋ง, EBM์ ๋ฐ์ดํฐ๊ฐ โ์โ ๊ทธ๋ ๊ฒ ๋ณด์ด๋์ง์ ๋ํ ๊ทผ๋ณธ์ ์ธ ๊ท์น๊ณผ ์ง์์ ํ์ ํ๋ ค ํฉ๋๋ค. ์ด๋ฌํ โ์ดํดโ๋ โ์ ์ฌ ๋ณ์(latent variables)โ์ ์ ์ฅ๋ฉ๋๋ค.
์ ์ฌ ๋ณ์: ์ ์ฌ ๋ณ์๋ ๋ฐ์ดํฐ์ ๊ท์น์ ๋ํ ์ง์์ ๋ด๊ณ ์๋ ์ผ์ข ์ โ์ง์ ์ ์ฅ์โ์ ๋๋ค. ์ด๋ ๋ช ์์ ์ธ ๊ท์น์ ๋ชฉ๋ก์ด๋ผ๊ธฐ๋ณด๋ค๋, ์๋์ง ์งํ์ ํํ๋ก ์ง์์ ์ ์ฅํฉ๋๋ค. ์๋ฅผ ๋ค์ด, ๋์ ๋ค์ ์๋ ์ํ๊ฐ ๊ทธ๊ฐ ์๋ ๊ฒ์ ์ข์ํ๊ธฐ ๋๋ฌธ์ธ์ง, ์๋๋ฉด ๋ฐฐ๊ฒฝ์ผ๋ก ์ฌ์ฉํ๊ธฐ ๋๋ฌธ์ธ์ง์ ๊ฐ์ ๊ธฐ๋ณธ์ ์ธ ์ธ๊ณ ๊ท์น์ ์ดํดํฉ๋๋ค. ๋ง์น ์ฐ๋ฆฌ์ ๋๊ฐ ์ฃผ๋ณ ์ธ๊ณ๋ฅผ ์ดํดํ๊ณ ํ๊ฒฝ์ ๋ณํ๊ฐ ์๊ฒผ์ ๋(์: ์๋ก์ด ๋ชจ์์ ์ํ๊ฐ ์๊ฒผ์ ๋) ์ด๋ป๊ฒ ๋ฐ์ํด์ผ ํ ์ง ์ถ๋ก ํ๋ ๊ฒ๊ณผ ๊ฐ์ต๋๋ค. EBM์ ์ด๋ฌํ ์ ์ฌ ๋ณ์๋ฅผ ํตํด ์๋ก์ด ์ํฉ์์๋ ๊ธฐ์กด ์ง์์ ๋ฐํ์ผ๋ก ์ถ๋ก ํ๊ณ ๊ณํํ ์ ์์ต๋๋ค.
์ผ๋ถ์์๋ ์ด๋ฌํ โ๊ท์นโ์ ๋ํ ์ด์ผ๊ธฐ๋ฅผ ๋ค์ผ๋ฉด ๊ณผ๊ฑฐ์ โ์์ง AI(symbolic AI)โ๋ฅผ ๋ ์ฌ๋ฆฌ๋ฉฐ ๊ทธ ํ๊ณ์ ์ ์ฐ๋ คํ ์ ์์ต๋๋ค. ๊ทธ๋ฌ๋ ์ด CEO๋ EBM์ด ํ ํฐํ๋ฅผ ํผํ๊ณ ๋ฐ์ดํฐ๋ฅผ ๋ค๋ฅธ ๋ฐ์ดํฐ ๊ตฌ์กฐ์ ์ง์ ๋งคํํจ์ผ๋ก์จ ์์ง AI์ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๋ค๊ณ ์ค๋ช ํฉ๋๋ค. EBM์ ํ ํฐ ์ํ์ค๊ฐ ์๋, ๋ฐ์ดํฐ ์์ฒด์ ๊ตฌ์กฐ์ ๊ท์น์ ์ง์ ๋ค๋ฃจ๊ธฐ ๋๋ฌธ์ ๊ทผ๋ณธ์ ์ผ๋ก ๋ค๋ฆ ๋๋ค. ์ด๋ ๋ฐ์ดํฐ ๋ถ์๊ณผ ๊ฐ์ด ์ธ์ด๊ฐ ๋์์ด ๋์ง ์๋ ์์ญ์์ ํนํ ๊ฐ๋ ฅํ ์ด์ ์ ์ ๊ณตํฉ๋๋ค.
LLM์ โํฐ๋ ์์ผโ์ EBM์ โ์กฐ๊ฐ๋โ
์ด CEO๋ LLM๊ณผ EBM์ ์ฐจ์ด๋ฅผ ์ํ๋์์ค์ฝ ์ง๋๋ฅผ ํ์ํ๋ ์ํฉ์ ๋น์ ํ์ฌ ๋์ฑ ๋ช ํํ ์ค๋ช ํฉ๋๋ค.
LLM์ ํฐ๋ ์์ผ: LLM์ ๋ง์น ํฐ๋ ์์ผ๋ฅผ ๊ฐ์ง ์ฌ๋์ฒ๋ผ ํ ๋ฒ์ ํ ๋ฐฉํฅ๋ง ์ ํํ ์ ์์ต๋๋ค. ์๋ฅผ ๋ค์ด, ํน ์คํธ๋ฆฌํธ์์ ๋ฒ ์ด ๋ธ๋ฆฌ์ง๋ก ๊ฐ๋ ค๊ณ ํ ๋, LLM์ ํ ๋ฒ์ ํ ๋จ๊ณ์ฉ๋ง ์์ง์ด๋ฉฐ ๋ค๋ฅธ ์ต์ ์ ๋ณผ ์ ์์ต๋๋ค. ๋๋ก๋ โํ๊ฐโ์ ์ผ์ผ์ผ ์๋ชป๋ ๊ธธ๋ก ๋ค์ด์๊ณ , ์ฌ์ง์ด ๊ธธ์ ๊ตฌ๋ฉ์ด ์์ด๋ ๋๋์๊ฐ ์ ์๊ธฐ ๋๋ฌธ์ ๊ทธ ๊ตฌ๋ฉ์ ๋น ์ ธ๋ฒ๋ฆฝ๋๋ค. ์ด๋ LLM์ด ์์ฐจ์ (autoregressive)์ด๊ธฐ ๋๋ฌธ์ ๋๋ค. LLM์ ์์ ์ ์ํํ๋ ๋์ ๋ฐฉํฅ์ ๋ฐ๊ฟ ๋ฅ๋ ฅ์ด ์๊ณ , ๋ฌด์์ด ์ณ๊ณ ๊ทธ๋ฅธ์ง ์์ง ๋ชปํ ์ฑ ๋ฌด์์๋ก ๋ค์ ๋จ๊ณ๋ฅผ ์ ํํ๋ฉฐ ๊ณ์ ์งํํฉ๋๋ค. ๊ฒฐ๊ตญ ๋ชฉ์ ์ง์ ๋๋ฌํ์ง ๋ชปํ๊ฑฐ๋ ๋ง๋ํ ์ปดํจํ ์์์ ๋ญ๋นํ๊ฒ ๋ฉ๋๋ค.
EBM์ ์กฐ๊ฐ๋: ๋ฐ๋ฉด EBM์ โ์กฐ๊ฐ๋(birdโs-eye view)โ๋ฅผ ๊ฐ์ง๊ณ ์์ต๋๋ค. ๋ชจ๋ ๊ฐ๋ฅํ ๊ฒฝ๋ก๋ฅผ ํ๋์ ๋ณผ ์ ์์ผ๋ฉฐ, ๋ง์ฝ ๊ธธ์ ๊ตฌ๋ฉ์ด ์๋ค๋ฉด ์ฆ์ ๋ค๋ฅธ ๊ฒฝ๋ก๋ฅผ ์ ํํ ์ ์์ต๋๋ค. ์ด๋ฌํ ๋ฅ๋ ฅ์ EBM์ด ํจ์ฌ ํจ์จ์ ์ด๊ณ ์ ํํ๊ฒ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ ์ ์๊ฒ ํฉ๋๋ค.
์ต๊ทผ LLM์ ํ์ฉํ ์ฝ๋ฉ(vibe coding)์์ ๋ํ๋๋ ๋ฌธ์ ์ ๋ ์ด์ ์ ์ฌํฉ๋๋ค. LLM์ด ์์ฑํ ์ฝ๋๋ ๊ฐ๋ณ์ ์ผ๋ก๋ ์ฌ๋ฐ๋ฅด๊ฒ ๋ณด์ด์ง๋ง, ์ ์ฒด ์์คํ ์ ๋๊ณ ๋ณด๋ฉด ์ผ๊ด์ฑ์ด ์๊ณ ๋ง์น โ๋๋๊ธฐ(patchwork)โ์ฒ๋ผ ๋๊ปด์ง ์ ์์ต๋๋ค. ์ด๋ LLM์ด ํ์ฌ ๋ณด๊ณ ์๋ ๋ถ๋ถ์๋ง ์ง์คํ๊ณ ์ ์ฒด์ ์ธ ํตํฉ ์๋ฃจ์ ์ ๋ง๋ค์ด๋ด๋ ๋ฐ ์ด๋ ค์์ ๊ฒช๊ธฐ ๋๋ฌธ์ ๋๋ค. EBM์ ์ด๋ฌํ ๋ฌธ์ ์ ๋ํด ํจ์ฌ ๋ ํตํฉ์ ์ด๊ณ ๊ฒ์ฆ ๊ฐ๋ฅํ ์๋ฃจ์ ์ ์ ๊ณตํ ์ ์์ต๋๋ค.
EBM์ด ์ ์ํ๋ ์๋ก์ด AI ํ์ฉ ์ฌ๋ก
๋ก์ง์ปฌ ์ธํ ๋ฆฌ์ ์ค๋ EBM์ ํ์ฉํ์ฌ โํ์์ ์ผ๋ก ๊ฒ์ฆ๋ ์ฝ๋(formally verified code)โ๋ฅผ ์์ฑํ๊ณ ์ฝ๋ฉ ๊ณผ์ ์ ์์ ํ ์๋ํํ๋ ๊ฒ์ ๋ชฉํ๋ก ํฉ๋๋ค. ์ด๋ ๊ฐ๋ฐ์๊ฐ C++๋ ํ์ด์ฌ ๊ฐ์ ํน์ ์ธ์ด๋ก ์ฝ๋ฉํ๋ ๋์ , ์์ฐ์ด(์: ์์ด)๋ก ์ฝ๋๋ฅผ ์ง์ํ๊ณ EBM์ด ์ด๋ฅผ ๊ฒ์ฆ๋ ์ฝ๋๋ก ๋ณํํ๋ ๊ฒ์ ์๋ฏธํฉ๋๋ค.
EBM์ ์ธ๋ถ ๊ฒ์ฆ๊ธฐ์ ๊ฒฐํฉํ์ฌ ๊ธฐ์กด ์ฝ๋ ๋ก์ง๊ณผ์ ํธํ์ฑ์ ์ํ์ ์ผ๋ก ์ฆ๋ช ํ๊ณ , ๋ฌธ์ ๊ฐ ๋ฐ์ํ๋ฉด ์์ฐ์ด๋ก ํด๊ฒฐ์ฑ ์ ์ ์ํ ์ ์์ต๋๋ค. ์ด๋ ๋จ์ํ ์ฝ๋ ์์ฑ ๋จ๊ณ๋ฅผ ๋์ด, ์ฝ๋์ โ๋ช ์ธ(specifications)โ๋ฅผ ์ค์ํ๋์ง ํ์ธํ๋ ๊ฒ์ ๋๋ค.
๊ทธ๋ฌ๋ ์ด CEO๋ AI๊ฐ ์ธ๊ฐ์ ์๋๋ฅผ ์์ ํ ํ์ ํ ์๋ ์๋ค๊ณ ๊ฐ์กฐํฉ๋๋ค. โAI๋ ๋น์ ์ ๋๋ฅผ ๋ค์ฌ๋ค๋ณด๊ณ ๋น์ ์ด ๋ฌด์์ ์ํ๋์ง ์ ์ ์์ต๋๋ค.โ ์๋ฅผ ๋ค์ด, ์์จ์ฃผํ์ฐจ ์ฝ๋๋ฅผ ๊ฐ๋ฐํ ๋, ์ฝ๋๊ฐ ์ ์ปดํ์ผ๋๋ ๊ฒ ์ธ์ โ์ด ์ฝ๋๊ฐ ์ค์ ๋ก ๋ณดํ์๋ฅผ ์น์ง ์์ ๊ฒ์ธ๊ฐ?โ ๋๋ โ์ด ์ฝ๋๊ฐ ์ํ๋์์ค์ฝ ์ง๋๋ฅผ ์ ๋๋ก ํ์ํ ๊ฒ์ธ๊ฐ?โ์ ๊ฐ์ ํ๋ ๋ช ์ธ(behavior specifications)๋ฅผ ์ธ๊ฐ์ด ์ ๊ณตํด์ผ ํฉ๋๋ค.
์ด ์ง์ ์์ EBM์ ๊ฐ์ฅ ํฐ ์ฅ์ ์ด ๋๋ฌ๋ฉ๋๋ค. LLM์ ํ๊ฐ์ ์ผ์ผ์ผ ์์ธก ๋ถ๊ฐ๋ฅํ๊ฒ ํ๋ํ ์ ์์ง๋ง, EBM์ โ์ ์ฝ ์กฐ๊ฑด(constraints)โ์ ์ค์ ํ์ฌ ๊ฐ์ ์ ์ผ๋ก ๊ทธ ๊ท์น์ ๋ฐ๋ฅด๋๋ก ๋ง๋ค ์ ์์ต๋๋ค. ์ฆ, ์ธ๊ฐ์ด AI์๊ฒ โ๋ฌด์์ ํด์ผ ํ๋์งโ ๋ช ํํ ์ง์ํ๋ฉด, EBM์ ๊ทธ ๊ท์น์ ํญ์ ์ค์ํ๋๋ก ์ค๊ณ๋ ์ ์์ต๋๋ค. ์ด๋ ์์จ์ฃผํ์ฐจ, ํญ๊ณต๊ธฐ๋ฟ๋ง ์๋๋ผ, ์ฐ์ธ์ฆ์ ๊ฒช๋ ์ฌ๋์๊ฒ ๋ฏผ๊ฐํ๊ฒ ๋ฐ์ํด์ผ ํ๋ ์ธ์ด ๋ชจ๋ธ๊ณผ ๊ฐ์ด ์์ธก ๋ถ๊ฐ๋ฅํ ํ๋์ด ์น๋ช ์ ์ธ ๊ฒฐ๊ณผ๋ฅผ ์ด๋ํ ์ ์๋ ๋ชจ๋ ์์ญ์์ ํ์์ ์ธ ๋ฅ๋ ฅ์ ๋๋ค.
AI ์ฐ์ ์ ๋ฏธ๋: LLM๊ณผ EBM์ ๊ณต์กด
ํ์ฌ AI ์ฐ์ ์ LLM์ ๋ํ ๋ง๋ํ ํฌ์์ ํจ๊ป ์ฑ์ฅํ๊ณ ์์ต๋๋ค. 2021~2023๋ ์ LLM์ด ๋ณด์ฌ์ค โ์ํ(aha)โ ํจ๊ณผ๋ ๋ง์ ํฌ์์๋ค์ด LLM์ด ๋ชจ๋ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ ๊ฒ์ด๋ผ๋ ๋ฏฟ์์ ๊ฐ๊ฒ ํ์ต๋๋ค. ๊ทธ๋ฌ๋ ์ด CEO๋ LLM์ด ์ปดํจํ ํ์๋ฅผ ๋๋ฆฌ๊ณ ์ํคํ ์ฒ๋ฅผ ์กฐ๊ธ์ฉ ๋ณ๊ฒฝํ๋ ๋ฐฉ์์ผ๋ก๋ ์ผ์ข ์ โ์ ์ฒด๊ธฐ(plateau)โ์ ๋๋ฌํ๊ณ ์๋ค๊ณ ํ๊ฐํฉ๋๋ค. ์ด๋ฏธ ์์ญ์ต ๋ฌ๋ฌ๊ฐ LLM์ ํฌ์๋์๊ธฐ ๋๋ฌธ์, ํฌ์ ์ปค๋ฎค๋ํฐ๊ฐ ์์ ํ ์๋ก์ด ํจ๋ฌ๋ค์์ผ๋ก ์ ํํ๊ธฐ๋ ์ฝ์ง ์์ ํ์ค์ ๋๋ค.
๋ก์ง์ปฌ ์ธํ ๋ฆฌ์ ์ค๋ ์ด๋ฌํ ํ์ค์ ์ธ์ ํ๊ณ , LLM์ ์์ ํ ๋์ฒดํ๊ธฐ๋ณด๋ค๋ ๋ณด์ํ๋ ์ ๋ต์ ์ทจํฉ๋๋ค. EBM์ LLM๊ณผ ํธํ๋๋ฉฐ, LLM ์คํ์ ํ์ ๋ ์ด์ด๋ก ํตํฉ๋ ์ ์์ต๋๋ค. ์๋ฅผ ๋ค์ด, ์ธ๊ธ ๊ณ์ฐ์ด๋ ๊ณต๊ฐ ์ถ๋ก ๊ณผ ๊ฐ์ด LLM์ด ์ ํด๊ฒฐํ์ง ๋ชปํ๋ ํน์ ์์ ์ EBM์ ์์ํ์ฌ LLM์ ๋น์ฉ์ ์ ๊ฐํ๊ณ ํจ์จ์ฑ์ ๋์ผ ์ ์์ต๋๋ค.
์ด CEO๋ โ์ฐ๋ฆฌ์ EBM์ด LLM ํฌ์๋ฅผ ์ฌ์ ํ ๊ฐ์น ์๊ฒ ๋ง๋ค ์ ์๋ ๋ ์ด์ด๊ฐ ๋ ์ ์๋คโ๊ณ ๋งํฉ๋๋ค. LLM ํ์ฌ๋ค์ด ๋น์ฉ ์ ๊ฐ์ ์ํ๋ค๋ฉด, ๊ณต๊ฐ ์ถ๋ก ๊ณผ ๊ฐ์ ํน์ ์์ ์ EBM์ ์์์์ฑํ ์ ์๋ค๋ ๊ฒ์ ๋๋ค. ์ด๋ ๊ธฐ์กด LLM ์ํ๊ณ๋ฅผ ํ๊ดดํ๋ ๋์ , ๊ทธ ์์์ ์๋ก์ด ๊ฐ์น๋ฅผ ์ฐฝ์ถํ๋ฉฐ ๋์์ ์๋ก์ด ํํ์ AI๋ฅผ ์ํ ์์ฒด ์ํ๊ณ๋ฅผ ๊ตฌ์ถํ๋ ํ๋ช ํ ์ ๊ทผ ๋ฐฉ์์ ๋๋ค.
๊ฒฐ๋ก
๋๊ท๋ชจ ์ธ์ด ๋ชจ๋ธ(LLM)์ ์ธ๊ณต์ง๋ฅ์ ๊ฐ๋ฅ์ฑ์ ํ์ฅํ์ง๋ง, โ์ ํ์ฑโ, โ์ ๋ขฐ์ฑโ, โ์ดํดโ๋ผ๋ ์ธก๋ฉด์์ ์ฌ์ ํ ์ค์ํ ๊ณผ์ ๋ฅผ ์๊ณ ์์ต๋๋ค. ๋ก์ง์ปฌ ์ธํ ๋ฆฌ์ ์ค์ ์๋์ง ๊ธฐ๋ฐ ๋ชจ๋ธ(EBM)์ ํ ํฐ ๊ธฐ๋ฐ์ ์ถ์ธก ๊ฒ์์ ๋์ด, ์๋์ง ์ต์ํ ์๋ฆฌ์ ์ ์ฌ ๋ณ์๋ฅผ ํตํด ๋ฐ์ดํฐ์ ๋ณธ์ง์ ์ธ ๊ท์น์ โ์ดํดโํ๊ณ ์์ธก ๊ฐ๋ฅํ ๋ฐฉ์์ผ๋ก ์๋ํฉ๋๋ค.
LLM์ โํฐ๋ ์์ผโ์ ๋ฌ๋ฆฌ EBM์ โ์กฐ๊ฐ๋โ๋ ๋ฏธ์ ํฌ๋ฆฌํฐ์ปฌ ์์คํ ์์ AI์ ์์ ์ฑ๊ณผ ํจ์จ์ฑ์ ํ์ ์ ์ผ๋ก ํฅ์์ํฌ ์ ์ฌ๋ ฅ์ ๊ฐ์ง๋๋ค. AI ์ฐ์ ์ด LLM ์ค์ฌ์ ํฌ์์์ ๋ฒ์ด๋ ์๋ก์ด ์ํคํ ์ฒ๋ฅผ ๋ชจ์ํ๋ ๋ฐ ์ด๋ ค์์ ๊ฒช๊ณ ์์ง๋ง, EBM์ LLM๊ณผ ๊ณต์กดํ๋ฉฐ ๊ทธ ํ๊ณ๋ฅผ ๋ณด์ํ๋ ํ์ค์ ์ธ ๋์์ ์ ์ํฉ๋๋ค.
๊ฒฐ๊ตญ AI์ ๋ฏธ๋๋ ๋จ์ผ ๋ชจ๋ธ์ ์์กดํ๊ธฐ๋ณด๋ค๋, LLM์ ์ ์ฐ์ฑ๊ณผ EBM์ ์ ํ์ฑ ๋ฐ ์ ๋ขฐ์ฑ์ด ์๋์ง๋ฅผ ์ด๋ฃจ๋ ํ์ด๋ธ๋ฆฌ๋ ์ ๊ทผ ๋ฐฉ์์ ์์ ๊ฒ์ ๋๋ค. ์ด๋ฌํ ๋ค๊ฐ์ ์ธ ๋ฐ์ ์ ์ธ๋ฅ์๊ฒ ๋์ฑ ์์ ํ๊ณ ํจ์จ์ ์ด๋ฉฐ, ๊ถ๊ทน์ ์ผ๋ก ์ ๋ขฐํ ์ ์๋ AI ์๋๋ฅผ ์ด์ด์ค ๊ฒ์ ๋๋ค.
โAre Human Drivers Finally Obsolete? | Freakonomics Radioโ โ Freakonomics Radio Network ๊ธฐ๋ฐ ๊ธฐ์ฌ ์๋ณธ ์์ ๋ณด๊ธฐ
์ธ๊ฐ ์ด์ ์๋์ ์ข ๋ง์ด ๋ค๊ฐ์ค๋๊ฐ? ์์จ์ฃผํ์ฐจ ํ๋ช ์ ์์๊ณผ ์จ๊ฒจ์ง ์ด์ผ๊ธฐ
PJ Vogt๋ ์ต๊ทผ ์๊ธฐ์น ์์ ๋ถ์์ผ๋ก ์ธํด ์ผ์์ํ์ ์ ์ฝ์ด ์๊ฒผ์ ๋, ์ํ๋์์ค์ฝ์์ โ์จ์ด๋ชจ(Waymo)โ๋ผ๋ ๋ก๋ณดํ์๋ฅผ ๊ฒฝํํ์ต๋๋ค. ์ด์ ๋ ์๋ ์ฐจ๊ฐ ๋ฒํผ ํ๋๋ก ๋์ฐฉํ๊ณ ์ค์ค๋ก ์์ง์ด๋ ๋ชจ์ต์ ๋ง์น ๋ฏธ๋๊ฐ ํ์ค์ด ๋ ๋ฏํ ์ถฉ๊ฒฉ์ด์์ต๋๋ค. ์ฒ์์๋ ๋นํ๊ธฐ๋ฅผ ํ๋ ๋ฏํ ์ ๊ธฐํจ์ด์์ง๋ง, ์ธ ๋ฒ์ฏค ํ๋ ์๋ฆฌ๋ฒ ์ดํฐ์ฒ๋ผ ์ต์ํด์ก์ต๋๋ค. ์ด ๊ฒฝํ์ ๊ทธ์๊ฒ โ๋ง์ ๊ฒ์ด ๋ณํ ๊ฒ์ธ๋ฐ, ์ ์ฌ๋๋ค์ด ์ด์ ๋ํด ๋ ์ด์ผ๊ธฐํ์ง ์์๊น?โ๋ผ๋ ์๋ฌธ์ ๋์ก๊ณ , ๊ฒฐ๊ตญ ์์จ์ฃผํ์ฐจ์ ๋ํ ์ฌ์ธต์ ์ธ 2๋ถ์ ํ์บ์คํธ ์๋ฆฌ์ฆ โSearch Engineโ์ ์ ์ํ๊ฒ ๋ง๋ค์์ต๋๋ค.
์ด ์๋ฆฌ์ฆ๋ ์์จ์ฃผํ ๊ธฐ์ ์ ๋ฐ์ ๊ณผ์ ๊ณผ ๊ทธ๋ก ์ธํ ์ฌํ์ , ๊ฒฝ์ ์ ํ์ฅ์ ์ฌ๋ ์๊ฒ ๋ค๋ฃน๋๋ค. ํนํ 1๋ถ์์๋ ๊ธฐ์ ์์ฒด์ ๊ทธ ๊ฐ๋ฐ์๋ค์, 2๋ถ์์๋ ์ธ๊ฐ ์ด์ ์์๊ฒ ๋ฏธ์น ์ํฅ์ ์กฐ๋ช ํฉ๋๋ค. PJ Vogt๋ ์ด ์ฌ์ ์์ ์์จ์ฃผํ์ฐจ์ ๊ฟ์ ๊พผ โ์ธ๋ฐ์ค์ฐฌ ์ค๋ฐ(Sebastian Thrun)โ๊ณผ ๊ฐ์ด ๋งค๋ ฅ์ ์ธ ์ธ๋ฌผ๋ค์ ๋ง๋๊ณ , ์๋ก ์์ถฉํ๋ ์ดํด๊ด๊ณ๋ฅผ ์กฐ์จํด์ผ ํ๋ ๋ณต์กํ ํ์ค์ ๋ง์ฃผํฉ๋๋ค.
ํ๋ฆฌ์ฝ๋ ธ๋ฏน์ค ๋ผ๋์ค(Freakonomics Radio)์ ์งํ์ ์คํฐ๋ธ ๋ฅ๋(Stephen Dubner)๋ ์ด๋ฏธ 50๋ ์ ๋ถํฐ โ์ธ๊ฐ ์ด์ ์ ๋ฐ๋๋ก ์โ์๋ค๊ณ ๋ฐํ๋ฉฐ, ์์จ์ฃผํ์ฐจ์ ํ์์ฑ์ ๊ฐ๋ ฅํ ์ง์งํฉ๋๋ค. ๊ทธ์ ๋ง์ฒ๋ผ, ์ธ๊ฐ์ ์ด์ ๋ฅ๋ ฅ์ด ๋ฐ์ด๋์ง ์๊ณ , ๊ฐ์ ์ ์ด๋ฉฐ, ์ฝ๊ฒ ์ฐ๋งํด์ง๊ธฐ ๋๋ฌธ์ ๋๋ค. ์ด ๊ธฐ์ฌ๋ โํ๋ฆฌ์ฝ๋ ธ๋ฏน์ค ๋ผ๋์คโ์์ ํน๋ณ ๋ฐฉ์๋ โSearch Engineโ ํ์บ์คํธ์ ์ฒซ ๋ฒ์งธ ์ํผ์๋๋ฅผ ๋ฐํ์ผ๋ก, ์์จ์ฃผํ์ฐจ ํ๋ช ์ด ์ด๋ป๊ฒ ์์๋์๊ณ , ์ฐ๋ฆฌ๊ฐ ํ์ฌ ์ด๋ค ์ง์ ์ ์ ์๋์ง ๊ทธ ์จ๊ฒจ์ง ์ด์ผ๊ธฐ๋ฅผ ํ์ด๋ ๋๋ค.
1. ์ด์ ์ ์๋ ๊ฟ: 200๋ ์ ๋ถํฐ ์์๋ ์์
200๋ ์ , ์ฆ 1800๋ ๋ ์ด๋ฐ์๋ โ๋ ธ์ปค-์ดํผ(knocker-upper)โ์ โ๊ฐ์ค๋ฑ ์ ๋ฑ์(lamplighter)โ์ด๋ผ๋ ์ง์ ์ด ์์์ต๋๋ค. ๋ ธ์ปค-์ดํผ๋ ๊ธด ๋ง๋๊ธฐ๋ก ์ฌ๋๋ค์ ์ฐฝ๋ฌธ์ ๋๋๋ ค ์์นจ ์ผ์ฐ ๊นจ์์ฃผ๋ ์ญํ ์ ํ๊ณ , ๊ฐ์ค๋ฑ ์ ๋ฑ์์ ํด์ง๋ ๊ฐ์ค๋ฑ์ ์ผ๊ณ ์๋ฒฝ์ ๋๋ ์ผ์ ๋ด๋นํ์ต๋๋ค. ์ด ์ง์ ๋ค์ ์ค๋๋ ์ค๋งํธํฐ ์๋๊ณผ ์ ๊ธฐ ๊ฐ๋ก๋ฑ์ผ๋ก ๋์ฒด๋์ด ๋๋ถ๋ถ์ ์ฌ๋๋ค์๊ฒ ๋ฏ์ ๊ฐ๋ ์ด ๋์์ต๋๋ค.
ํ์ง๋ง 200๋ ์ ์ ๋ ๋ค๋ฅธ ์ง์ , ์ฆ โ์ด์ ์โ๋ ์ฌ์ ํ ์กด์ฌํฉ๋๋ค. ๋น์ ์ด์ ์๋ ๋ง์ฐจ์ ์์ ๋ง์ ๊ณ ์๋ฅผ ์ก๊ณ ์น๊ฐ์ ํ์ ๋ชฉ์ ์ง๊น์ง ๋ฐ๋ ค๋ค์ฃผ๋ ์ฌ๋์ด์์ต๋๋ค. ์ด ์ง์ ์ ๋ค๋ฅธ ๋ ์ง์ ๊ณผ๋ ๋ฌ๋ฆฌ ํ๋๊น์ง ์ด์ด์ ธ์๊ณ , ๋๋ถ๋ถ์ ์ฌ๋๋ค์๊ฒ๋ ์ผ์์ ์ธ ํ์๋ก ์๋ฆฌ ์ก์์ต๋๋ค. ์ด ๊ธฐ์ฌ๋ ๋ฐ๋ก โ์ด์ ์โ๋ผ๋ ๋จ์ด๊ฐ ์ธ๊ฐ์ด ์๋ ๊ธฐ๊ณ๋ฅผ ์ง์นญํ๊ฒ ๋ ๋ฏธ๋์ ๋ํ ์ด์ผ๊ธฐ์ ๋๋ค. ์๊ธฐ์ธ์ฒ๊ธฐ(dishwasher), ํ๋ฆฐํฐ(printer), ์ปดํจํฐ(computer) ๊ฐ์ ๋จ์ด๋ค์ด ์๋๋ ์ฌ๋์ ์ง์นญํ์ง๋ง ์ง๊ธ์ ๊ธฐ๊ณ๋ฅผ ์๋ฏธํ๋ฏ์ด ๋ง์ ๋๋ค.
์ธ๊ฐ ์ด์ ์ ์ํ๊ณผ ์์จ์ฃผํ์ ์ฝ์
์๊ฐ ์๋ ์ค ๋ฐ์ด๋น์ค(Alex Davies)๋ ๊ทธ์ ์ ์ โDriven, The Race to Create the Autonomous Carโ์์ ์ธ๊ฐ ์ด์ ์ ํ๊ณ์ ๋ํด ๊น์ด ๊ณ ์ฐฐํฉ๋๋ค. ๊ทธ๋ ์์ ์ด ์ข์ ์ด์ ์๋ผ๊ณ ์๊ฐํ์ง๋ง, ํผ๋ก, ์ง์ค๋ ฅ ๋ถ์ฐ, ๊ฐ์ ์กฐ์ ์ ์ด๋ ค์ ๋ฑ ๋ช ํํ ํ๊ณ๋ฅผ ์ธ์งํฉ๋๋ค. ์ค์ ๋ก ๊ทธ๋ ์ธํฐ๋ทฐ ํ ๋ช ๋ฌ ๋ค ๊ตํต์ฌ๊ณ ๋ฅผ ๊ฒช๊ธฐ๋ ํ์ต๋๋ค.
์์จ์ฃผํ์ฐจ์ ํต์ฌ ๊ฐ์น๋ ๋ฐ๋ก โ์์ โ์ ๋๋ค. ์ปดํจํฐ๊ฐ ์ด์ ํ๋ ์ฐจ๋ ์์ฃผ, ํผ๋ก, ์ฃผ์ ์ฐ๋ง, ๋ฌธ์ ๋ฉ์์ง ์ ์ก, ๋ณด๋ณต ์ด์ ๋ฑ์ ํ์ง ์์ต๋๋ค. ์ด๋ฐ ์์จ์ฃผํ์ฐจ๋ ๋ ์ด์ ๋จผ ๋ฏธ๋๊ฐ ์๋๋๋ค. ์จ์ด๋ชจ(Waymo) ๊ฐ์ ๋ก๋ณดํ์(robo-taxi)๋ ๋ฏธ๊ตญ 10๊ฐ ๋์์์ ์๋ฐฑ๋ง ๊ฑด์ ์ดํ์ ์ ๊ณตํ๊ณ ์์ผ๋ฉฐ, ์ค๊ตญ์์๋ ๋ ๋ฐฐ ๋ง์ ๋์์์ ๋ ๋น ๋ฅด๊ฒ ํ์ฐ๋๊ณ ์์ต๋๋ค. ์ํ๋์์ค์ฝ๋ ์ค์คํด ๊ฐ์ ๋์์์๋ ์ด๋ฏธ ์ฐ๋ฒ(Uber)๋งํผ ํํด์ก์ต๋๋ค.
์ฌ์ค ์ธ๊ฐ ์ด์ ์๋ฅผ ๊ธฐ๊ณ๋ก ๋์ฒดํ๋ ค๋ ๊ฟ์ ๊ฐ์ค๋ฑ ์ ๋ฑ์์ ์๋๋งํผ์ด๋ ์ค๋๋์์ต๋๋ค. ๋ง์ด ๋๋ ๋ง์ฐจ์์ ์๋์ฐจ๋ก ๋์ด์ค๋ฉด์ ์์ด๋ฒ๋ฆฐ โ์ง๊ฐ๋ ฅ(sentience)โ์ ๋ค์ ์ฐจ๋์ ๋ถ์ฌํ๋ ค๋ ์๋์์ต๋๋ค. ๋ง์ด ์ ๋ฒฝ์ผ๋ก ๋ฌ๋ ค๊ฐ์ง ์๋ฏ์ด, ์ฐจ๋ ์ค์ค๋ก ํ๋จํ๊ธฐ๋ฅผ ๋ฐ๋๋ ๊ฒ์ ๋๋ค.
์ด๊ธฐ ์๋์ฐจ์ ์ ํญ๊ณผ ์ฌํ์ ์ ์
์ด๊ธฐ ์๋์ฐจ๊ฐ ๋ฑ์ฅํ์ ๋, ์ฌ๋๋ค์ ์ฆ๊ธฐ๊ธฐ๊ด์ฐจ์ ๊ฐ์๋ฆฐ ์ฐจ๋์ โ๊ต์โ๊ณผ โ์ํ์ฑโ์ ๋ํ ๋๋ ค์์ผ๋ก ๊ฒฉ๋ ฌํ๊ฒ ์ ํญํ์ต๋๋ค. โํ์คํฐ์ฆ(Teamsters)โ ๊ฐ์ ๋ง์ฐจ ๊ด๋ จ ์ง์ ๋ค์ด ์ฌ๋ผ์ง ์ํ๋ ์ปธ์ต๋๋ค. โ๋ถ์ ๊น๋ฐ ๋ฒ(red flag laws)โ์ฒ๋ผ ์๋์ฐจ ์์ ์ฌ๋์ด ๋ถ์ ๊น๋ฐ์ ํ๋ค๋ฉฐ ๊ฒฝ๊ณ ํด์ผ ํ๋ ๊ท์ ๋, ๊ฐ์ถ์ ๋ง๋๋ฉด ์ฐจ๋ฅผ ๋ถํดํด์ ๋ค๋ถ ๋ค์ ์จ๊ฒจ์ผ ํ๋ค๋ ํฉ๋นํ ๋ฒ์๊น์ง ์ ์๋ ์ ๋์์ต๋๋ค. ํ์ค๋ฒ ์ด๋์(Pennsylvania) ์ฃผ์ง์ฌ๋ ์ด ๋ฒ์์ ๊ฑฐ๋ถ๊ถ์ ํ์ฌํ์ง๋ง, ๋น์์ ๋ฐ(ๅ)์๋์ฐจ ์ ์๋ฅผ ์ฟ๋ณผ ์ ์์ต๋๋ค.
๊ท์ ์์ด ์๋์ฐจ๋ฅผ ๋ฐ์๋ค์ธ ๋ํธ๋ก์ดํธ(Detroit) ๊ฐ์ ๋์์์๋ ์ด์ ๋ฉดํ, ์ ํธ๋ฑ, ๋ฐฉํฅ์ง์๋ฑ๋ ์์ด ์์ฒญ๋ ์ฌ๋ง๋ฅ ์ ๊ธฐ๋กํ์ต๋๋ค. ํ์ง๋ง ์์ญ ๋ ์ ๊ฑธ์ณ ๋ฒ๋ฅ , ๋ฉดํ ์ ๋, ์ด์ ๊ต์ก, ๋๋ก ์ค๊ณ ๊ฐ์ , ๊ณ ์๋๋ก, ์์ ๋ฒจํธ, ์์ด๋ฐฑ ๋ฑ์ด ๋์
โJensen Huang โ Will Nvidiaโs moat persist?โ โ Dwarkesh Patel ๊ธฐ๋ฐ ๊ธฐ์ฌ ์๋ณธ ์์ ๋ณด๊ธฐ
์๋น๋์ ์ ์จ ํฉ์ ํต์ฐฐ: AI ์๋, โ์ ์์์ ํ ํฐ์ผ๋กโ ๋ณํํ๋ ์์ ๊ทธ๋ฆฌ๊ณ ์ง์ ๊ฐ๋ฅํ ํด์
AI ํ๋ช ์ ์ต์ ์ ์์ ๋ ๋ณด์ ์ธ ์์น๋ฅผ ์ฐจ์งํ๊ณ ์๋ ์๋น๋์(NVIDIA)์ ์ ์จ ํฉ(Jensen Huang) ์ต๊ณ ๊ฒฝ์์(CEO)๊ฐ ์์ฌ์ ์ง์ ๊ฐ๋ฅํ ๊ฒฝ์ ์ฐ์, ์ฆ โํด์(moat)โ์ ๋ณธ์ง์ ๋ํด ์ฌ๋ ๊น์ ํต์ฐฐ์ ๊ณต์ ํ์ต๋๋ค. ์ต๊ทผ ์ํํธ์จ์ด ๊ธฐ์ ๋ค์ ๊ฐ์น ํ๊ฐ๊ฐ ๊ธ๋ฝํ๋ฉฐ AI๊ฐ ์ํํธ์จ์ด๋ฅผ ์ํํํ ๊ฒ์ด๋ผ๋ ์ฐ๋ ค๊ฐ ์ปค์ง๋ ๊ฐ์ด๋ฐ, ํฉ CEO๋ ์๋น๋์๊ฐ ๋จ์ํ ํ๋์จ์ด ์ ์กฐ์ฌ๋ฅผ ๋์ด โ์ ์(electrons)๋ฅผ ํ ํฐ(tokens)์ผ๋ก ๋ณํํ๋ ์์ โ์ ํตํด AI ์๋์ ํต์ฌ ๋๋ ฅ์ผ๋ก ์๋ฆฌ๋งค๊นํ ๊ฒ์ด๋ผ๊ณ ๊ฐ์กฐํ์ต๋๋ค.
์๋น๋์ ํด์์ ๋ณธ์ง: โ์ ์์์ ํ ํฐ์ผ๋กโ์ ๋ณํ ์์
์ผ๊ฐ์์๋ ์๋น๋์๊ฐ TSMC์ GDS2 ํ์ผ์ ๋ณด๋ด ๋ ผ๋ฆฌ ๋ค์ด(logic dies)์ ์ค์์น๋ฅผ ์ ์กฐํ๊ณ , SK ํ์ด๋์ค(SK Hynix), ๋ง์ดํฌ๋ก (Micron), ์ผ์ฑ(Samsung)์ด ๋ง๋ ๊ณ ๋์ญํญ ๋ฉ๋ชจ๋ฆฌ(HBM)์ ํจํค์งํ๋ฉฐ, ๋๋ง์ ODM(Original Design Manufacturer)์์ ๋(rack)์ ์กฐ๋ฆฝํ๋ โ๋ค๋ฅธ ์ฌ๋๋ค์ด ์ ์กฐํ๋ ์ํํธ์จ์ดโ๋ฅผ ๋ง๋ค ๋ฟ์ด๋ผ๊ณ ์ง์ ํฉ๋๋ค. ์ํํธ์จ์ด๊ฐ ์ํํ๋๋ค๋ฉด ์๋น๋์ ๋ํ ์ํํ๋ ๊ฒ์ด๋ผ๋ ์ฃผ์ฅ์ ๋๋ค.
๊ทธ๋ฌ๋ ํฉ CEO๋ ์ด๋ฌํ ์๊ฐ์ด โ์์งํ ์๊ฐโ์ด๋ผ๊ณ ์ผ์ถํฉ๋๋ค. ๊ทธ๋ ๊ถ๊ทน์ ์ผ๋ก ๋ฌด์ธ๊ฐ๋ โ์ ์(electrons)๋ฅผ ํ ํฐ(tokens)์ผ๋ก ๋ณํโํด์ผ ํ๋ค๊ณ ๊ฐ์กฐํฉ๋๋ค. ๊ทธ๋ฆฌ๊ณ โ์ด ์ ์์์ ํ ํฐ์ผ๋ก์ ๋ณํ, ๊ทธ๋ฆฌ๊ณ ์๊ฐ์ด ์ง๋จ์ ๋ฐ๋ผ ์ด ํ ํฐ์ ๊ฐ์น๋ฅผ ๋์ด๋ ๊ฒ์ ์์ ํ ์ํํํ๊ธฐ ์ด๋ ต๋คโ๊ณ ๋งํฉ๋๋ค. ํ๋์ ๋ถ์๋ฅผ ๋ค๋ฅธ ๋ถ์๋ณด๋ค, ํ๋์ ํ ํฐ์ ๋ค๋ฅธ ํ ํฐ๋ณด๋ค ๋ ๊ฐ์น ์๊ฒ ๋ง๋๋ ๊ณผ์ ์๋ ์์ฒญ๋ ์์ ์ฑ, ๊ณตํ, ๊ณผํ, ๊ทธ๋ฆฌ๊ณ ๋ฐ๋ช ์ด ์๋ฐ๋ฉ๋๋ค. ํฉ CEO๋ ์ด ๊ณผ์ ์ด โ์์ง ๊น์ด ์ดํด๋์ง ์์๊ณ , ๊ทธ ์ฌ์ ์ ์์ง ๋๋์ง ์์๋คโ๋ฉฐ ์๋น๋์๊ฐ ์ด ํต์ฌ์ ์ธ ๋ณํ ์์ ์ ์ฃผ๋ํ๊ณ ์๋ค๊ณ ์ค๋ช ํฉ๋๋ค.
์๋น๋์์ ์ญํ ์ โ๊ฐ๋ฅํ ํ ๋ง์ด ํ์ํ ๊ฒ์ ํ๊ณ , ๊ฐ๋ฅํ ํ ์ ๊ฒ ํ๋ ๊ฒโ์ด๋ผ๋ ์ฒ ํ์ ๊ธฐ๋ฐํฉ๋๋ค. ํฉ CEO๋ โ์ฐ๋ฆฌ์ ์๋ฌด๋ ๋๋ผ์ด ์ญ๋์ผ๋ก ๊ทธ ๋ณํ์ด ์ด๋ฃจ์ด์ง ์ ์๋๋ก ํ์ํ ๋งํผ์ ์ผ์ ํ๊ณ , ๊ฐ๋ฅํ ํ ์ต์ํ์ ์ผ์ ํ๋ ๊ฒโ์ด๋ผ๊ณ ๋งํฉ๋๋ค. ์ด๋ ์๋น๋์๊ฐ ์ง์ ํ ํ์๊ฐ ์๋ ๋ถ๋ถ์ ํํธ๋๋ค๊ณผ ํ๋ ฅํ์ฌ ๊ฑฐ๋ํ ์ํ๊ณ์ ์ผ๋ถ๋ก ๋ง๋ ๋ค๋ ์๋ฏธ์ ๋๋ค. ํ์ฌ ์๋น๋์๋ ๊ณต๊ธ๋ง์ ์๋ฅ(upstream)์ ํ๋ฅ(downstream) ๋ชจ๋์์ ๊ฐ์ฅ ํฐ ํํธ๋ ์ํ๊ณ๋ฅผ ๋ณด์ ํ๊ณ ์์ผ๋ฉฐ, AI์ ๋ค์ฏ ๊ฐ ๊ณ์ธต ์ ๋ฐ์ ๊ฑธ์ณ ์ํ๊ณ๋ฅผ ๊ตฌ์ถํ๊ณ ์์ต๋๋ค.
AI ์๋ ์ํํธ์จ์ด์ ๋ฏธ๋์ ์๋น๋์์ ์ญํ
ํฉ CEO๋ AI๋ก ์ธํด ๋๋ถ๋ถ์ ์ํํธ์จ์ด ๊ธฐ์ ๋ค์ด โ๋๊ตฌ ์ ์์(tool makers)โ๊ฐ ๋ ๊ฒ์ด๋ฉฐ, ์ด๋ค์ ์ญํ ์ด ๋์ฑ ์ค์ํด์ง ๊ฒ์ด๋ผ๊ณ ์์ธกํฉ๋๋ค. ์์ (Excel), ํ์ํฌ์ธํธ(PowerPoint), ์ผ์ด๋์ค(Cadence), ์๋์์ค(Synopsys) ๋ฑ์ ๋ชจ๋ ๋๊ตฌ์ด๋ฉฐ, AI ์์ด์ ํธ(agents)์ ์๊ฐ ๊ธฐํ๊ธ์์ ์ผ๋ก ์ฆ๊ฐํจ์ ๋ฐ๋ผ ์ด๋ฌํ ๋๊ตฌ ์ฌ์ฉ์ ๋ํ ํญ๋ฐ์ ์ผ๋ก ๋์ด๋ ๊ฒ์ด๋ผ๋ ์ฃผ์ฅ์ ๋๋ค.
ํ์ฌ๋ ์์ง๋์ด์ ์์ ์ํด ์ ํ๋์ง๋ง, ๋ฏธ๋์๋ AI ์์ด์ ํธ๊ฐ ์์ง๋์ด๋ฅผ ์ง์ํ๋ฉฐ ์ ๋ก ์๋ ๋ฐฉ์์ผ๋ก ์ค๊ณ ๊ณต๊ฐ์ ํ์ํ๊ฒ ๋ ๊ฒ์ ๋๋ค. ์ด๋ ๊ธฐ์กด ๋๊ตฌ๋ค์ ์ฌ์ฉ์ ๋์ฑ ์ด์งํ๊ณ ์ํํธ์จ์ด ๊ธฐ์ ๋ค์ ์ฑ์ฅ์ ์ด๋ ๊ฒ์ด๋ผ๊ณ ํฉ CEO๋ ์ ๋งํฉ๋๋ค. ๊ทธ๋ โ์์ง ์์ด์ ํธ๊ฐ ๋๊ตฌ๋ฅผ ์ฌ์ฉํ๋ ๋ฐ ์ถฉ๋ถํ ๋ฅ์ํ์ง ์๊ธฐ ๋๋ฌธ์ ์ด๋ฌํ ๋ณํ๊ฐ ์ผ์ด๋์ง ์์์ ๋ฟโ์ด๋ผ๋ฉฐ, ์์ผ๋ก๋ ๊ธฐ์ ๋ค์ด ์ง์ ์์ด์ ํธ๋ฅผ ๊ตฌ์ถํ๊ฑฐ๋ ์์ด์ ํธ๊ฐ ๋๊ตฌ ์ฌ์ฉ์ ๋ฅ์ํด์ง๋ ๋ ๊ฐ์ง ๋ฐฉ์์ด ๊ฒฐํฉ๋ ๊ฒ์ด๋ผ๊ณ ๋ด๋ค๋ดค์ต๋๋ค.
๊ณต๊ธ๋ง ํ๋ณด์ ์ ์ ์ ๋ณ๋ชฉ ๊ด๋ฆฌ
์๋น๋์์ ํด์์ ๋ํ ๋ ๋ค๋ฅธ ํด์์ ๋ง๋ํ ๊ตฌ๋งค ์ฝ์ (purchase commitments)์ ํตํด ํฌ์ํ ๋ถํ๋ค์ ์ ์ ํ๊ณ ์๋ค๋ ๊ฒ์ ๋๋ค. ์ธ๋ฏธ์ ๋๋ฆฌ์์ค(SemiAnalysis)๋ ์๋น๋์๊ฐ ํ์ด๋๋ฆฌ, ๋ฉ๋ชจ๋ฆฌ, ํจํค์ง ๋ถ์ผ์์ ๊ฑฐ์ 1์ฒ์ต ๋ฌ๋ฌ์ ๋ฌํ๋ ๊ตฌ๋งค ์ฝ์ ์ ์ฒด๊ฒฐํ์ผ๋ฉฐ, ํฅํ 2์ฒ5๋ฐฑ์ต ๋ฌ๋ฌ๊น์ง ๋์ด๋ ์ ์๋ค๊ณ ๋ณด๋ํ์ต๋๋ค. ์ด์ ๋ํด ํฉ CEO๋ โ๋ค๋ฅธ ์ฌ๋์ด ํ๊ธฐ ์ด๋ ค์ด ์ผ ์ค ํ๋โ๋ผ๋ฉฐ ์ด ์ญ์ ์๋น๋์์ ๊ฐ์ ์ค ํ๋์์ ์ธ์ ํฉ๋๋ค.
๊ทธ๋ฌ๋ ๋จ์ํ ๋์ผ๋ก๋ง ๊ฐ๋ฅํ ๊ฒ์ ์๋๋๋ค. ์๋น๋์๋ ๊ณต๊ธ๋ง ์๋ฅ์ CEO๋ค์ ์ง์ ๋ง๋ โ์ด ์ฐ์ ์ด ์ผ๋ง๋ ์ปค์ง์ง, ์ ๊ทธ๋ ๊ฒ ๋ ์ง, ๊ทธ๋ฆฌ๊ณ ๋ด๊ฐ ๋ฌด์์ ๋ณด๊ณ ์๋์งโ๋ฅผ ์ค๋ช ํ๋ฉฐ ์๊ฐ์ ์ฃผ๊ณ ์ค๋ํฉ๋๋ค. ์ด๋ฌํ ๊ณผ์ ๋๋ถ์ ๊ณต๊ธ์ ์ฒด๋ค์ ์๋น๋์๋ฅผ ์ํด ๋ง๋ํ ํฌ์๋ฅผ ๊ธฐ๊บผ์ด ๊ฐํํฉ๋๋ค. ๊ทธ ์ด์ ๋ ์๋น๋์๊ฐ ์ด๋ค์ ๊ณต๊ธ์ ๊ตฌ๋งคํ๊ณ ์์ฌ์ ํ๋ฅ ๊ณต๊ธ๋ง(downstream supply chain)์ ํตํด ํ๋งคํ ๋ฅ๋ ฅ์ด ์์์ ์๊ธฐ ๋๋ฌธ์ ๋๋ค. GTC(GPU Technology Conference)์ ๊ฐ์ ํ์ฌ๋ฅผ ํตํด ์๋น๋์๋ AI ์ํ๊ณ ์ ์ฒด๋ฅผ ํ์๋ฆฌ์ ๋ชจ์ ์๋ฅ์ ํ๋ฅ ํํธ๋๋ค์ด ์๋ก ๋ง๋๊ณ AI์ ๋ฐ์ ์ ์ง์ ๋ชฉ๊ฒฉํ๊ฒ ํจ์ผ๋ก์จ ๊ณต๊ธ๋ง์ ํจ๊ณผ์ ์ผ๋ก ์ฐ๊ฒฐํ๊ณ ๊ฐํํฉ๋๋ค.
ํฉ CEO๋ ๋ณ๋ชฉ ํ์(bottlenecks)์ด ๋ฐ์ํ๋ ์ฆ์ ํด๊ฒฐํ๊ธฐ ์ํด ์ ์ ์ ์ผ๋ก ๋ ธ๋ ฅํ๋ค๊ณ ์ค๋ช ํฉ๋๋ค. ์ผ๋ก๋ก, ํ๋ ์ฌ๊ฐํ ๋ณ๋ชฉ์ด์๋ CoWoS(Chip-on-Wafer-on-Substrate) ํจํค์ง ๊ธฐ์ ์ ์๋น๋์์ 2๋ ๊ฐ์ ์ง์ค์ ์ธ ํฌ์์ ๋ ธ๋ ฅ์ผ๋ก ๊ณต๊ธ๋์ด ํฌ๊ฒ ๋์ด ์ด์ ๋ ๋ฌธ์ ๊ฐ ๋์ง ์๋๋ค๊ณ ๋งํฉ๋๋ค. TSMC๋ CoWoS ๊ณต๊ธ์ด ๋ ผ๋ฆฌ ๋ค์ด ๋ฐ ๋ฉ๋ชจ๋ฆฌ ์์์ ๋ณด์กฐ๋ฅผ ๋ง์ถฐ์ผ ํ๋ค๋ ๊ฒ์ ์ธ์ํ๊ณ ์์ผ๋ฉฐ, ๋ฏธ๋ ํจํค์ง ๊ธฐ์ ๋ ํจ๊ป ํ์ฅํ๊ณ ์์ต๋๋ค.
๊ทธ๋ฌ๋ ์ฅ๊ธฐ์ ์ธ ๊ด์ ์์ ํฉ CEO๊ฐ ์ฐ๋ คํ๋ ์ง์ ํ ๋ณ๋ชฉ์ ๋ฐ๋์ฒด ์ ์กฐ ์ญ๋ ์์ฒด๊ฐ ์๋๋ผ๊ณ ๋งํฉ๋๋ค. EUV(Extreme Ultraviolet) ๋ ธ๊ด ์ฅ๋น๋ ํน(fab) ์ฆ์ค ๋ฑ์ 2~3๋ ๋ด์ ํด๊ฒฐ ๊ฐ๋ฅํ ๋ฌธ์ ์ด๋ฉฐ, ์ถฉ๋ถํ ์์ ์ ํธ๋ง ์๋ค๋ฉด ๋น ๋ฅด๊ฒ ํ์ฅ๋ ์ ์๋ค๋ ๊ฒ์ ๋๋ค. ์คํ๋ ค ๋ ํฐ ๋ฌธ์ ๋ โ๋ฐฐ๊ด๊ณต(plumbers)๊ณผ ์ ๊ธฐ ๊ธฐ์ ์(electricians)โ ๊ฐ์ ์๋ จ๊ณต์ ๋ถ์กฑ, ๊ทธ๋ฆฌ๊ณ ์๋์ง ์ ์ฑ ์ด๋ผ๊ณ ์ง์ ํฉ๋๋ค. ์๋ก์ด ์ฐ์ ์ ๋ง๋ค๊ณ , ์นฉ ์ ์กฐ, ์ปดํจํฐ ์ ์กฐ, AI ํฉํ ๋ฆฌ ๋ฑ์ ์ฌ์ฐ์ ํ(reindustrialize)ํ๋ ค๋ฉด ๋ง๋ํ ์๋์ง๊ฐ ํ์ํ๋ฉฐ, ์ด๋ฌํ ์๋์ง ์ ์ฑ ์ ํจ์ฌ ๋ ์ค๋ ์๊ฐ์ด ๊ฑธ๋ฆฌ๋ ๋ฌธ์ ์ ๋๋ค.
๊ฒฝ์ ์ฐ์: TPU๋ฅผ ๋์ด์ โ๊ฐ์ ์ปดํจํ โ
๊ตฌ๊ธ(Google)์ TPU(Tensor Processing Unit)๊ฐ ํด๋ก๋(Claude)์ ์ ๋ฏธ๋(Gemini) ๊ฐ์ ์ฃผ์ AI ๋ชจ๋ธ ํ๋ จ์ ์ฌ์ฉ๋๋ค๋ ์ ์ ๋ํ ์ง๋ฌธ์ ํฉ CEO๋ ์๋น๋์๊ฐ โํ ์ ์ฒ๋ฆฌ ์ฅ์น๊ฐ ์๋ ๊ฐ์ ์ปดํจํ (accelerated computing)โ์ ๊ตฌ์ถํ๋ค๊ณ ๊ฐ์กฐํฉ๋๋ค. ๊ฐ์ ์ปดํจํ ์ ๋ถ์ ์ญํ, ์์ ์์ญํ, ๋ฐ์ดํฐ ์ฒ๋ฆฌ, ์ ์ฒด ์ญํ, ์ ์ ๋ฌผ๋ฆฌํ ๋ฑ ํจ์ฌ ๋ ๊ด๋ฒ์ํ ๋ถ์ผ์ ์ฌ์ฉ๋๋ฉฐ, AI๋ ๊ทธ ์ค ํ ๋ถ์ผ์ผ ๋ฟ์ ๋๋ค. ์๋น๋์๋ ๋ฒ์ฉ ์ปดํจํ ์์ ๊ฐ์ ์ปดํจํ ์ผ๋ก์ ์ ํ์ ํตํด ์ปดํจํ ๋ฐฉ์์ ์ฌ์ฐฝ์กฐํ์ผ๋ฉฐ, ๊ทธ ์์ฅ ๋๋ฌ ๋ฒ์๋ ์ด๋ค TPU๋ ASIC(Application-Specific Integrated Circuit)๋ ๊ฐ์ง ์ ์๋ ์์ค์ด๋ผ๊ณ ๋งํฉ๋๋ค.
์๋น๋์์ ํต์ฌ ๊ฐ์ ์ โ์ผ๋ฐ์ ์ธ ํ๋ก๊ทธ๋๋ฐ ๊ฐ๋ฅ์ฑ(general programmability)โ์ ์์ต๋๋ค. ์๋ก์ด ์ดํ ์ ๋ฉ์ปค๋์ฆ(attention mechanism)์ ๊ฐ๋ฐํ๊ฑฐ๋, ํ์ฐ(diffusion)๊ณผ ์๊ธฐํ๊ท(autoregressive) ๊ธฐ์ ์ ์ตํฉํ๋ ๋ชจ๋ธ์ ๋ง๋ค ๋, ์ ์ฐํ๊ฒ ํ๋ก๊ทธ๋๋ฐ ๊ฐ๋ฅํ ์ํคํ ์ฒ๊ฐ ํ์์ ์ ๋๋ค. ๋ฌด์ด์ ๋ฒ์น(Mooreโs Law)์ด ์ฐ๊ฐ 25% ์ ๋์ ์ฑ๋ฅ ํฅ์์ ์ ๊ณตํ๋ ๋ฐ๋ฉด, ์๋น๋์๊ฐ ํํผ(Hopper)์์ ๋ธ๋์ฐ(Blackwell)๋ก 50๋ฐฐ์ ์๋์ง ํจ์จ ํฅ์์ ์ด๋ฃฌ ๋น๊ฒฐ์ ์๋ก์ด ๋ชจ๋ธ๊ณผ ์๊ณ ๋ฆฌ์ฆ์ ๋ฐ๋ช ์ ์์ต๋๋ค. MoE(Mixture of Experts)์ ๊ฐ์ด ๋ถ์ฐ๋ ์ปดํจํ ์์คํ ์ ์ต์ ํ๋ ์๋ก์ด ๋ชจ๋ธ์ ๊ฐ๋ฐํ๊ณ , CUDA(Compute Unified Device Architecture)๋ฅผ ํตํด ์๋ก์ด ์ปค๋(kernels)์ ๊ตฌํํ๋ ๊ฒ์ด ๊ฐ๋ฅํ๊ธฐ ๋๋ฌธ์ ๋๋ค.
CUDA๋ ์๋น๋์๊ฐ ํ๋ก์ธ์, ์์คํ , ํจ๋ธ๋ฆญ(fabric), ๋ผ์ด๋ธ๋ฌ๋ฆฌ, ์๊ณ ๋ฆฌ์ฆ ์ ๋ฐ์ ๊ฑธ์ณ ๋ณํ๋ฅผ ์ผ์ผํฌ ์ ์๋ โ๊ทน๋จ์ ๊ณต๋ ์ค๊ณ(extreme co-design)โ๋ฅผ ๊ฐ๋ฅํ๊ฒ ํฉ๋๋ค. ์ด๋ฌํ ์ ์ฐ์ฑ์ AI ์๊ณ ๋ฆฌ์ฆ์ ๋น ๋ฅธ ๋ฐ์ ์ ์ด์งํ๋ฉฐ, ์๋น๋์๊ฐ ๋ชจ๋ ์ข ๋ฅ์ ์์ฉ ํ๋ก๊ทธ๋จ์ ๊ฐ์ํํ ์ ์๋ ์ ์ผํ ๊ธฐ์ ์ผ๋ก ์๋ฆฌ๋งค๊นํ๊ฒ ํ์ต๋๋ค.
ํฉ CEO๋ CUDA ์ํ๊ณ์ ์ธ ๊ฐ์ง ํต์ฌ ๊ฐ์น๋ฅผ ๊ฐ์กฐํฉ๋๋ค. ์ฒซ์งธ, ํ๋ถํจ๊ณผ ํ๋ก๊ทธ๋๋ฐ ์ฉ์ด์ฑ: ๋ชจ๋ ํ๋ ์์ํฌ์ ์๊ณ ๋ฆฌ์ฆ์ ์ง์ํ๋ฉฐ, ๊ฐ๋ฐ์๋ค์ด ์ ๋ขฐํ ์ ์๋ ๊ธฐ๋ฐ ์์์ ํ์ ์ ์ธ ์์ ์ ์ํํ ์ ์๋๋ก ํฉ๋๋ค. ๋์งธ, ๋ฐฉ๋ํ ์ค์น ๊ธฐ๋ฐ(install base): ์์ต ๊ฐ์ GPU๊ฐ ์ ์ธ๊ณ์ ๋ณด๊ธ๋์ด ์์ด, ๊ฐ๋ฐ๋ ์ํํธ์จ์ด๋ ๋ชจ๋ธ์ด ์ด๋์์๋ ์ ์ฉํ๊ฒ ์ฌ์ฉ๋ ์ ์์ต๋๋ค. ์ ์งธ, ํด๋ผ์ฐ๋ ๋ ๋ฆฝ์ฑ(versatility): ๊ตฌ๊ธ, ์๋ง์กด, ์ ์ , OCI ๋ฑ ๋ชจ๋ ํด๋ผ์ฐ๋์์ ์๋น๋์ ์์คํ ์ ์ด์ํ ์ ์์ผ๋ฉฐ, ์จํ๋ ๋ฏธ์ค(on-premise) ํ๊ฒฝ๋ ์ง์ํ์ฌ ๊ณ ๊ฐ์๊ฒ ์ต๊ณ ์ ์ ์ฐ์ฑ์ ์ ๊ณตํฉ๋๋ค.
์ด๊ฑฐ๋ AI ๊ธฐ์ ๊ณผ์ ๊ด๊ณ์ ์๋น๋์์ ์ ๋ต
์๋น๋์ ๋งค์ถ์ 60%๊ฐ ์์ 5๊ฐ ํ์ดํผ์ค์ผ์ผ๋ฌ(hyperscalers)์์ ๋ฐ์ํ๋ค๋ ์ง์ ์ ๋ํด ํฉ CEO๋ ์ด๋ค ๊ธฐ์ ๋ ์๋น๋์์ ์ ๋ฌธ์ฑ์ด ํ์ํ๋ค๊ณ ์ค๋ช ํฉ๋๋ค. ์๋น๋์์ GPU ์ํคํ ์ฒ๋ ๋ง์น โF1 ๋ ์ด์โ์ ๊ฐ์์, ์ต๊ณ ์ ์ฑ๋ฅ์ ๋์ด๋ด๊ธฐ ์ํด์๋ ์๋นํ ์ ๋ฌธ ์ง์์ด ํ์ํ๋ฉฐ, ์๋น๋์๋ AI ์ฐ๊ตฌ์ ํํธ๋๋ค๊ณผ ๊ธด๋ฐํ ํ๋ ฅํ์ฌ ์คํ(stack)์ ์ต์ ํํ๊ณ 2๋ฐฐ, 3๋ฐฐ, ์ฌ์ง์ด 50%์ ์ฑ๋ฅ ํฅ์์ ์ด๋์ด๋ ๋๋ค. ์ด๋ ๊ณง ๋งค์ถ ์ฆ๊ฐ๋ก ์ง๊ฒฐ๋๋ ๋ง๋ํ ๊ฐ์น๋ฅผ ์ ๊ณตํฉ๋๋ค.
ํฉ CEO๋ ์๋น๋์๊ฐ โ์ธ๊ณ ์ต๊ณ ์ ์ฑ๋ฅ ๋๋น ์ด์์ ๋น์ฉ(Performance per TCO)โ๊ณผ โ์ํธ๋น ํ ํฐ(Tokens per Watt)โ์ ์ ๊ณตํ๋ค๊ณ ๊ฐ์กฐํฉ๋๋ค. ๋ฐ์ดํฐ์ผํฐ๋ฅผ 1๊ธฐ๊ฐ์ํธ(Gigawatt) ๊ท๋ชจ๋ก ๊ตฌ์ถํ ๋, ์๋น๋์ ์ํคํ ์ฒ๋ ๋์ผํ ์ ๋ ฅ์ผ๋ก ์ต๋์ ํ ํฐ ์์ ๋งค์ถ์ ์ฐฝ์ถํ ์ ์๋ค๋ ๊ฒ์ ๋๋ค.
๊ณผ๊ฑฐ ์คํธ๋กํฝ(Anthropic)์ด ๊ตฌ๊ธ์ด๋ AWS(Amazon Web Services)์ ํฌ์๋ฅผ ๋ฐ์ TPU๋ ํธ๋ ์ด๋์(Trainium)์ ์ฌ์ฉํ๊ฒ ๋ ๋ฐฐ๊ฒฝ์ ๋ํด ํฉ CEO๋ โ๋น์ ์๋น๋์๊ฐ ๊ณต๊ธ์ ์ฒด๋ก์ ์์ญ์ต ๋ฌ๋ฌ ๊ท๋ชจ์ ํฌ์๋ฅผ ํ ์ฌ๋ ฅ์ด ์์์ผ๋ฉฐ, ๋ฒค์ฒ์บํผํธ(VC)์ด ๊ทธ๋งํ ํฌ์๋ฅผ ํ ๊ฒ์ด๋ผ๊ณ ์๊ฐํ๋คโ๊ณ ์ธ์ ํฉ๋๋ค. ๊ทธ๋ โ๊ทธ๊ฒ์ด ๋์ ์ค์์๋คโ๋ฉฐ, โ๋ค์๋ ๊ฐ์ ์ค์๋ฅผ ๋ฐ๋ณตํ์ง ์์ ๊ฒโ์ด๋ผ๊ณ ๋จ์ธํ์ต๋๋ค. ํ์ฌ ์๋น๋์๋ ์คํAI(OpenAI), ์คํธ๋กํฝ ๋ฑ ์ ๋ง AI ๋ฉ์ ์ ๊ทน์ ์ผ๋ก ํฌ์ํ๊ณ ์์ผ๋ฉฐ, ์ด๋ ์ด๋ค์ ์ฑ์ฅ์ ๋๊ณ AI ์ฐ์ ๋ฐ์ ์ ๊ธฐ์ฌํ๊ธฐ ์ํจ์ด๋ผ๊ณ ์ค๋ช ํฉ๋๋ค.
์๋น๋์๊ฐ ์ง์ ํด๋ผ์ฐ๋ ์ฌ์ ์๊ฐ ๋์ง ์๋ ์ด์ ์ ๋ํด์๋ ๋ค์ ํ๋ฒ โํ์ํ ๋งํผ, ์ต์ํ์ผ๋กโ๋ผ๋ ์ฒ ํ์ ๊ฐ์กฐํฉ๋๋ค. ์ปดํจํ ํ๋ซํผ ๊ตฌ์ถ๊ณผ CUDA ์ํ๊ณ ์กฐ์ฑ์ ์๋น๋์๊ฐ ์๋๋ฉด ์๋ฌด๋ ํ์ง ์์์ ์ผ์ด์ง๋ง, ํด๋ผ์ฐ๋ ์ฌ์ ์ ์๋น๋์๊ฐ ํ์ง ์์๋ ๋๊ตฐ๊ฐ๋ ํ ๊ฒ์ด๋ผ๋ ํ๋จ์ ๋๋ค. ๋์ ์๋น๋์๋ ์ฝ์ด์๋ธ(CoreWeave), ์์ค์ผ์ผ(Nscale), ๋ค๋น์ฐ์ค(Nebius)์ ๊ฐ์ ์ ์ AI ํด๋ผ์ฐ๋(neoclouds)์ ํฌ์ํ๊ณ ์ง์ํ์ฌ ์ํ๊ณ ์ ์ฒด๊ฐ ๋ฒ์ฑํ๋๋ก ๋๋ ๋ฐ ์ง์คํฉ๋๋ค. ๋ํ โ์น์๋ฅผ ๊ณ ๋ฅด์ง ์๋๋ค(donโt pick winners)โ๋ ์์น์ ๋ฐ๋ผ ๋ชจ๋ ์ ๋ง AI ๊ธฐ์ ์ ํฌ์ํ๊ณ ์ง์ํจ์ผ๋ก์จ ์์ฅ์ ๋ค์์ฑ์ ์กด์คํ๊ณ ํ์ ์ ์ฅ๋ คํฉ๋๋ค.
GPU ํ ๋น๊ณผ ์์ฅ ์ ๋ต
GPU ๋ถ์กฑ ์ํฉ์์ ์๋น๋์๊ฐ โ์ต๊ณ ๊ฐ ์ ์ฐฐ์(highest bidder)โ๊ฐ ์๋ ์ ์ ํด๋ผ์ฐ๋์ GPU๋ฅผ ์ฐ์ ๋ฐฐ์ ํ๋ค๋ ์๋ฌธ์ ๋ํด ํฉ CEO๋ ์ด๋ฅผ โ์ ํ ์ฌ์ค์ด ์๋๋คโ๋ผ๊ณ ๋ถ์ธํฉ๋๋ค. GPU ํ ๋น์ ์์น์ ๋ช ํํฉ๋๋ค. ์ฒซ์งธ, ์์ ์์ธก(forecasting)์ ํตํด ํํธ๋๋ค๊ณผ ๊ณต๊ธ๋ง์ ์กฐ์จํฉ๋๋ค. ๋์งธ, ๊ตฌ๋งค ์ฃผ๋ฌธ์(purchase order)๋ฅผ ์ ์ถํด์ผ ํฉ๋๋ค. ์ ์งธ, ๊ธฐ๋ณธ์ ์ผ๋ก โ์ ์ฐฉ์(first in, first out)โ ์์น์ ๋ฐ๋ฆ ๋๋ค. ๋ค๋ง, ๋ฐ์ดํฐ์ผํฐ ์ค๋น ์ํ ๋ฑ ๊ณ ๊ฐ์ฌ์ ์ค๋น ์ํฉ์ ๋ฐ๋ผ ๊ณต์ฅ ์ฒ๋ฆฌ๋(throughput)์ ๊ทน๋ํํ๊ธฐ ์ํ ์ผ๋ถ ์กฐ์ ์ ์์ ์ ์๋ค๊ณ ์ค๋ช ํฉ๋๋ค.
ํฉ CEO๋ โ์ฐ๋ฆฌ๋ ๊ฒฐ์ฝ ์ต๊ณ ๊ฐ ์ ์ฐฐ ๋ฐฉ์์ ์ฌ์ฉํ์ง ์๋๋คโ๊ณ ๊ฐ์กฐํ๋ฉฐ, ์ด๋ โ๋์ ์ฌ์ ๊ดํโ์ด๋ผ๊ณ ์ผ์ถํฉ๋๋ค. ๋์ ์๋น๋์๋ ๊ฐ๊ฒฉ์ ์ค์ ํ๊ณ , ๊ณ ๊ฐ์ ๊ทธ ๊ฐ๊ฒฉ์ ๊ตฌ๋งค ์ฌ๋ถ๋ฅผ ๊ฒฐ์ ํ๋ ๋ฐฉ์์ผ๋ก ์ฌ์ ์ ์ด์ํฉ๋๋ค.
๊ฒฐ๋ก
์ ์จ ํฉ CEO์ ํต์ฐฐ์ ์๋น๋์์ ํด์๊ฐ ๋จ์ํ ๊ธฐ์ ์ ์ฐ์๋ ์์ฅ ์ ์ ์จ์ ๋์ด์ ๋ณตํฉ์ ์ธ ์์๋ค๋ก ์ด๋ฃจ์ด์ ธ ์์์ ๋ณด์ฌ์ค๋๋ค. โ์ ์์์ ํ ํฐ์ผ๋กโ์ ๋ณํ์ด๋ผ๋ AI ์๋์ ๊ทผ๋ณธ์ ์ธ ๊ณผ์ ๋ฅผ ํด๊ฒฐํ๋ ๋ฐ ์ง์คํ๋ ๊ธฐ์ ์ ๋ฆฌ๋์ญ, ๊ด๋ฒ์ํ๊ณ ์ ์ฐํ CUDA ์ํ๊ณ, ๊ณต๊ธ๋ง ์ ๋ฐ์ ๋ํ ์ ๋ต์ ์ํฅ๋ ฅ, ๊ทธ๋ฆฌ๊ณ โํ์ํ ๋งํผ, ์ต์ํ์ผ๋กโ๋ผ๋ ์ฒ ํ ์๋ ํํธ๋์ญ์ ํตํด AI ์ฐ์ ์ ํ์ ์ ์ด์งํ๋ ๋น์ ์ด ๋ฐ๋ก ๊ทธ๊ฒ์ ๋๋ค. ์ด๋ฌํ ๋ค์ธต์ ์ธ ํด์๋ ์๋น๋์๊ฐ AI ์๋์ ์ง์์ ์ผ๋ก ์ฑ์ฅํ๊ณ ๊ทธ ์ํฅ๋ ฅ์ ํ๋ํ ์ ์๋ ํต์ฌ ๋๋ ฅ์ด ๋ ๊ฒ์ ๋๋ค.
โTrumpโs Risky Strategy to Blockade Iranโs Blockadeโ โ New York Times Podcasts ๊ธฐ๋ฐ ๊ธฐ์ฌ ์๋ณธ ์์ ๋ณด๊ธฐ
ํธ๋ผํ์ ์ํํ ์ ๋ฒ: ํธ๋ฅด๋ฌด์ฆ ํดํ ๋ด์, ์ด๋์ ๊ตด๋ณต์ํฌ๊น?
๋ด์ํ์์ฆ์ ๋ณด๋์ ๋ฐ๋ฅด๋ฉด, ๋ฏธ๊ตญ์ ํ์ฌ ํธ๋ฅด๋ฌด์ฆ ํดํ(Strait of Hormuz)์์ ์ด๋์ ๋ํ ๊ณ ๋์ ์ํ์ ์๋ฐํ๋ โ๋ด์ ์นํจ ๊ฒ์โ์ ๋ฒ์ด๊ณ ์์ต๋๋ค. ๋ฏธ๊ตญ์ ์ด ์ ๋ต์ ํตํด ์ด๋๊ณผ์ ์ ์์ ๋ฏธ๊ตญ์ด ์ํ๋ ๋ฐฉ์์ผ๋ก ์ข ์์ํค๊ณ , ์ด๋ ๊ฒฝ์ ๋ฅผ ๋ง๋น์์ผ ํ์ ํ ์ด๋ธ๋ก ๋์ด๋ด๋ ค๋ ์๋๋ฅผ ๊ฐ์ง๊ณ ์์ต๋๋ค. ๊ณผ์ฐ ์ด ๋ด์ ์ ๋ต์ ์ฑ๊ณตํ ์ ์์๊น์? ๊ทธ๋ฆฌ๊ณ ์ด ์ ๋ต์ด ๊ฐ์ ธ์ฌ ์ํ๊ณผ ๊ตญ์ ์ฌํ์ ๋ฏธ๋๋ ์ด๋ ํ ๊น์?
์ด๋ฒ ๋ผ์ด๋ํ ์ด๋ธ ํ ๋ก ์๋ ๋ฐฑ์ ๊ด ํนํ์ ๋ฐ์ด๋น๋ ์๊ฑฐ(David Sanger), ์๋์ง ์ ๋ฌธ ๊ธฐ์ ๋ ๋ฒ ์นด ์๋ฆฌ์(Rebecca Elliot), ๊ตฐ์ฌ ์ ๋ฌธ ๊ธฐ์ ์๋ฆญ ์๋ฏธํธ(Eric Schmidt)๊ฐ ์ฐธ์ฌํ์ฌ ํธ๋ผํ ํ์ ๋ถ์ ์ด๋ ๋ด์ ์ ๋ต์ ๋ฐฐ๊ฒฝ, ์ํ, ๊ทธ๋ฆฌ๊ณ ์ด๊ธฐ ์ฑ๊ณผ์ ๋ํด ์ฌ์ธต์ ์ผ๋ก ๋ถ์ํ์ต๋๋ค.
๋ฐฐ๊ฒฝ: ๋ด์์ ์์๊ณผ ๋ฏธ๊ตญ์ ๋ชฉํ
์ด๋ฒ ๋ด์ ์์ด๋์ด๋ ์ด๋์ด ๋ฏธ๊ตญ์ ๋ถํต๋ น JD ๋ฐด์ค(JD Vance)๋ฅผ ํํค์คํ ํ์์์ ์๋ฌด๋ฐ ์ฑ๊ณผ ์์ด ๋๋ ค๋ณด๋ธ ์งํ ๋ถ๊ฑฐ์ก์ต๋๋ค. ํด์ ์๋ ๋ถ๊ตฌํ๊ณ ์ด๋์ด ์ธ๊ณ ์์ ์ด์ก์ ํต์ฌ ํต๋ก์ธ ํธ๋ฅด๋ฌด์ฆ ํดํ์ ํต์ ๊ถ์ ๋์ง ์๊ณ ์คํ๋ ค ํตํ๋ฃ๋ฅผ ๋ถ๊ณผํ๊ธฐ ์์ํ์, ํธ๋ผํ ํ์ ๋ถ๋ ํผ๋์ค๋ฌ์ด ์ํฉ์ ์ง๋ฉดํ์ต๋๋ค.
1979๋ ์ด์ฌ๋ ํ๋ช ์ ๋ถ๊ฐ ์๋ฆฝ๋ ์ด๋ 47๋ ๊ฐ ์ด๋์ ํธ๋ฅด๋ฌด์ฆ ํดํ์ ํตํด ๋ชจ๋ ์ ๋ฐ์ ์์ ๋ก์ด ํตํ์ ํ์ฉํด์์ต๋๋ค. ๊ทธ๋ฌ๋ ์ต๊ทผ ์ด๋์ ๊ฐ์๊ธฐ ํดํ์ ํตํ์ ๋ง๊ณ ํตํ๋ฃ๋ฅผ ๋ถ๊ณผํ๊ธฐ ์์ํ์ต๋๋ค. ๋ง์น ๊ณ ์๋๋ก ํจ๊ฒ์ดํธ์ฒ๋ผ, ์ ๋ฐ ํ ์ฒ๋น 2๋ฐฑ๋ง ๋ฌ๋ฌ์ ๋ฌํ๋ ํตํ๋ฃ๋ฅผ ์๊ตฌํ ๊ฒ์ ๋๋ค. ๋ฏธ๊ตญ์ผ๋ก์๋ ์ด๋ฌํ ์ํฉ์ ์ฉ๋ฉํ ์ ์์๊ณ , ๋ฏธ ํด๊ตฐ์ด ํดํ์ ํต์ ๊ถ์ ์ด๋์ผ๋ก๋ถํฐ ๋์ฐพ์์ผ ํ ํ์์ฑ์ ๋๊ผ์ต๋๋ค.
๊ตฐ์ฌ์ ์ผ๋ก โํด์ ๋ด์(naval blockade)โ๋ ๋ช ๋ฐฑํ ์ ๋ ํ์์ด์ ์ ์ ํ์๋ก ๊ฐ์ฃผ๋ฉ๋๋ค. ์ด๋ ํ ๊ตญ๊ฐ๊ฐ ๊ตฐ์ฌ๋ ฅ์ ๋์ํ์ฌ ๋ค๋ฅธ ๊ตญ๊ฐ์ ์ ๋ฐ ํตํ์ ์ํํ๊ฑฐ๋ ์ค์ ๊ฐ์ ๋ก ์ ๋ฐ์ ๋ํฌํ๋ ๊ฒ์ ์๋ฏธํฉ๋๋ค. ํ์ฌ ๋ฏธ๊ตญ์ ํญ๊ณต๋ชจํจ๋ถํฐ ๊ตฌ์ถํจ, ํด๋ณ๋ ์์ก์ ์ ์ด๋ฅด๊ธฐ๊น์ง 10,000๋ช ์ด์์ ํด๊ตฐ ๋ณ๋ ฅ๊ณผ 12์ฒ ์ด์์ ์ ํจ์ ํธ๋ฅด๋ฌด์ฆ ํดํ ์ธ๊ณฝ์ ๋ฐฐ์นํ์ฌ ์ด๋ ํญ๊ตฌ๋ก ํฅํ๋ ์ ๋ฐ๋ค์ ๊ฐ์ํ๊ณ ์์ต๋๋ค. ํธ๋ผํ ๋ํต๋ น์ ์ด๋ฅผ โ๋ด์โ๋ผ๊ณ ๋ถ๋ฅด์ง๋ง, ์ค์ ๋ก๋ โ๊ฒ์ญ(quarantine)โ์ ๊ฐ๊น์ด ํํ๋ก ์ด์๋๊ณ ์์ต๋๋ค. ๋๋ก ๊ณผ ๊ณต๊ฐ ์ ๋ณด๋ฅผ ํ์ฉํด ์์ฌ์ค๋ฌ์ด ์ ๋ฐ์ ์๋ณํ๊ณ , ๋ฌด์ ์ ํตํด ์ ์ง ๋ช ๋ น์ ๋ด๋ฆฌ๊ฑฐ๋ ํ์์ ํด๋ณ๋๋ ๋ค์ด๋น์ค(Navy SEALs)์ ํฌ์ ํด ์ ๋ฐ์ ์น์ ํ์ฌ ํ๋ฌผ์ ๊ฒ์ฌํ ์ ์์ต๋๋ค.
๊ถ๊ทน์ ์ธ ๋ชฉํ๋ ์ด๋ ๊ฒฝ์ , ํนํ ์์ ์์ถ์ ์ง์์ํค๋ ๊ฒ์ ๋๋ค. ์ด๋์ ์ค๋ซ๋์ ํธ๋ฅด๋ฌด์ฆ ํดํ์ ํตํด ์์ ๋ฅผ ์์ถํ๋ฉฐ ๊ฒฝ์ ๋ฅผ ์ ์งํด์๊ณ , ํนํ ์ด๋ ํ๋ช ์๋น๋(IRGC)๋ ์์ ์์ถ ์์ต์ ๊ฑฐ์ ์ ๋ถ๋ฅผ ์ ์ ์๊ธ์ผ๋ก ํ์ฉํ๊ณ ์์์ต๋๋ค. ๋ฐ๋ผ์ ๋ฏธ๊ตญ์ ๋ด์๋ ์ด๋์ ์ฃผ์ ์์ ์์ ์ฐจ๋จํ์ฌ ์ด๋์ ํ์ ํ ์ด๋ธ๋ก ๋ณต๊ท์ํค๊ณ , ํต ๋น์ถ๋ ํฌ๊ธฐ ๋ฐ ์ฐ๋ผ๋ ๋์ถ ์ค๋จ๊ณผ ๊ฐ์ ๋ฏธ๊ตญ์ ์๊ตฌ๋ฅผ ๊ด์ฒ ์ํค๋ ค๋ ์ด์ค์ ์ธ ๋ชฉ์ ์ ๊ฐ์ง๊ณ ์์ต๋๋ค.
ํธ๋ผํ ํ์ ๋ถ์ ์ํํ ๋๋ฐ: ์ ์ฌ์ ์ํ๋ค
์ด๋ฌํ ํด์ ๋ด์๋ ํธ๋ผํ ํ์ ๋ถ๊ฐ ๊ฐ์ํด์ผ ํ ๋ง๋ํ ์ํ์ ๋ดํฌํ๊ณ ์์ต๋๋ค.
-
์ด๋์ ๋ณด๋ณต: ๊ฐ์ฅ ํฐ ์ํ์ ์ด๋ ํ๋ช ์๋น๋(IRGC)๊ฐ ๋ด์์ ๋งน๋ ฌํ ๋ฐ๊ฒฉํ ์ ์๋ค๋ ์ ์ ๋๋ค. ์ด๋์ ์ด๋ฏธ ๋ฏธ ํด๊ตฐ ํจ์ ์ ๊ณต๊ฒฉํ๊ฒ ๋ค๊ณ ์ํํ์ผ๋ฉฐ, ์ด๋ ํดํ ์ํ์์ ๋๊ท๋ชจ ๊ตฐ์ฌ์ ์ถฉ๋๋ก ์ด์ด์ง ์ ์์ต๋๋ค.
-
์ค๊ตญ์ ๋ฐ๋ฐ: ์ด๋์ด ์์ถํ๋ ์์ ์ ์ฝ 90%๋ ์ค๊ตญ์ผ๋ก ํฅํ๋ฉฐ, ๋๋ถ๋ถ ์ค๊ตญ ์ ๋ฐ์ ํตํด ์ด์ก๋ฉ๋๋ค. ํธ๋ผํ ๋ํต๋ น์ 4์ฃผ ์์ ๋ฒ ์ด์ง์ ๋ฐฉ๋ฌธํ์ฌ ๋ฏธ-์ค ๊ด๊ณ ๊ฐ์ ๊ณผ ๋ฌด์ญ ํ์์ ๋ ผ์ํ ์์ ์ด์์ผ๋, ์ด๋์ฐ ์์ ๋ฅผ ์ฃ๊ณ ๊ฐ๋ ์ค๊ตญ ์ ๋ฐ์ด ๋ฏธ ํด๊ตฐ์ ์ํด ํํญ๋นํ๋ค๋ฉด, ์ด๋ ๋ฏธ-์ค ๊ด๊ณ์ ์ฌ๊ฐํ ์ ์ํฅ์ ๋ฏธ์น ์ ์์ต๋๋ค. ์ด๊ฐ๋๊ตญ ์ ๋๊ตญ์ ์๊ทนํ๋ ๋งค์ฐ ํจ๊ณผ์ ์ธ ๋ฐฉ๋ฒ์ด๊ธฐ ๋๋ฌธ์ ๋๋ค. ๊ฒ๋ค๊ฐ ์ค๊ตญ์ด ์ด๋์ ๋ฌด๊ธฐ๋ฅผ ์ง์ํ๋ ๊ฒ์ ๊ณ ๋ คํ๊ณ ์๋ค๋ ๋ณด๋๋ ๋์ ๊ธด์ฅ์ ๋์ฑ ๊ณ ์กฐ๋๊ณ ์์ต๋๋ค.
-
๊ธ๋ก๋ฒ ์๋์ง ์์คํ ์ ํผ๋: ์ด๋์ด ํ๋ฅด์์๋ง ์ ์ญ์ ์๋์ง ์ธํ๋ผ์ ๋ํ ๊ณต๊ฒฉ์ ์ฌ๊ฐํ๋ค๋ฉด, ์ด๋ ๊ธ๋ก๋ฒ ์๋์ง ์์คํ ๊ณผ ์ธ๊ณ ๊ฒฝ์ ์ ์ฅ๊ธฐ์ ์ธ ์ํ์ ์ด๋ํฉ๋๋ค. ์ ์ ์์ค๊ณผ ์์ฐ ์์ค์ด ํ๊ดด๋๋ฉด ์๋์ง ๊ณต๊ธ์ด ์ฅ๊ธฐ๊ฐ ์ค๋จ๋ ์ ์์ต๋๋ค. ๊ตญ์ ์๋์ง๊ธฐ๊ตฌ(IEA)๋ ์ด๋ฏธ ์ด ์ง์ญ์ 80๊ฐ ์ด์์ ์๋์ง ์์ค์ด ํผํด๋ฅผ ์ ์์ผ๋ฉฐ, ์ ์ ์ด์ ์์ค์ผ๋ก ์์ฐ๋์ ํ๋ณตํ๋ ๋ฐ ์ต๋ 2๋ ์ด ๊ฑธ๋ฆด ์ ์๋ค๊ณ ์ถ์ ํ๊ณ ์์ต๋๋ค.
๋ด์์ ์ด๊ธฐ ์ฑ๊ณผ์ ๋ฏธ์์ ๊ณผ์
์ ๋ต์ ์ด๋ก ์ ๋ถ๊ณผํ๋ ๋ด์๋ ์ด์ ํ์ค์ด ๋์์ต๋๋ค. ์ด๊ธฐ 48์๊ฐ ๋์์ ์ํฉ์ ์ดํด๋ณด๋ฉด, ๋ฏธ๊ตญ ์ค์์ฌ๋ น๋ถ(US Central Command)๋ ๋ด์ ๊ฐ์ 24์๊ฐ ๋์ ๋จ ํ ๊ฑด์ ๋ด์ ์๋ฐ์ด๋ ์ ์ง๋ ์์์ผ๋ฉฐ, ์ด๋ ํญ๊ตฌ๋ฅผ ๋ ๋ฌ๋ 6์ฒ์ ์ ๋ฐ์ด ๋ฏธ ํด๊ตฐ์ ์ฐ๋ฝ์ ๋ฐ๊ณ ํํญํ๋ค๊ณ ๋ฐํํ์ต๋๋ค. ์ด๋ ์ด๋์ ์์ ์์ถ์ ๋ง๋ ๋ฐ๋ ์ฑ๊ณตํ๊ณ ์๋ ๊ฒ์ผ๋ก ๋ณด์ ๋๋ค.
ํ์ง๋ง ์ด๋ โ์ ๋ฐ์ ์ฑ๊ณตโ์ ๋ถ๊ณผํฉ๋๋ค. ๋ฏธ๊ตญ์ ์ด๋์ ์์ ์์ถ์ ๋ง์์ง๋ง, ์๋์๋ฏธ๋ฆฌํธ(UAE)๋ ๋ค๋ฅธ ์๋ ๊ตญ๊ฐ๋ค์ ์ ๋ฐ๋ค์ด ์ด๋์ ์ํ์ ๋ซ๊ณ ์์ ํ๊ฒ ํดํ์ ํต๊ณผํ ์ ์๋๋ก ๋ณด์ฅํ๋ ๊ฒ์ด ๋จ์์์ต๋๋ค. ํ์ฌ์ ์ํฉ์ ๋ฏธ๊ตญ์ด ์ด๋์ โ๋ด์โ์ ๋ํด โ๋ด์โ๋ฅผ ๊ฐํ๋, ๋งค์ฐ ๋ณต์กํ ์์์ผ๋ก ์ ๊ฐ๋๊ณ ์์ต๋๋ค. ์ ์ธ๊ณ ๊ฒฝ์ ๋ ์ด๋์ฐ ์์ ๊ฐ ์์ ํ ์ฐจ๋จ๋ ๊ฒ์ธ์ง, ์๋๋ฉด ๋ค๋ฅธ ๊ตญ๊ฐ์ ์ ๋ฐ๋ค์ด ๋ฏธ ํด๊ตฐ์ ๋ณดํธ ์๋ ํดํ์ ํต๊ณผํ ์ ์๊ฒ ๋ ๊ฒ์ธ์ง์ ์ด๊ฐ์ ๊ณค๋์ธ์ฐ๊ณ ์์ต๋๋ค.
๊ตญ์ ์ฌํ์ ๋ฐ์๊ณผ ํธ๋ฅด๋ฌด์ฆ ํดํ์ ๋ฏธ๋
๋ด์๊ฐ ์์๋ ์ด๊ธฐ 24์๊ฐ ๋์ ์ฝ 20์ฒ์ ์์ ์ด ํธ๋ฅด๋ฌด์ฆ ํดํ์ ํต๊ณผํ์ต๋๋ค. ์ด๋ ์ ์กฐ์ , ํ๋ฌผ์ ๋ฑ์ด ํฌํจ๋ ๊ฒ์ผ๋ก ์ถ์ ๋์ง๋ง, ์ด ์ซ์๊ฐ ์ฅ๊ธฐ์ ์ธ ์ถ์ธ์ ์์์ธ์ง, ์๋๋ฉด ์ผ์์ ์ธ ํ์์ธ์ง๋ ์์ง ๋ถํ์คํฉ๋๋ค.
์ ์ฌ๋ค๊ณผ ์ ์ฅ๋ค์ ๋ถ์๊ฐ์ ๋งค์ฐ ๋์ต๋๋ค. ์ ์ ์ด์ ์๋ ์์ ๋ก์ด ํตํ๊ณผ ์ ์ ์ํ์ด ๋ณด์ฅ๋์์ง๋ง, ์ด์ ๋ ๋์ ๋ณดํ๋ฃ๋ฅผ ์ง๋ถํด์ผ ํ๋ฉฐ ๊ณต๊ฒฉ ์ํ์ ๋ ธ์ถ๋์ด ์์ต๋๋ค. ๋ง์ ์ ๋ฌธ๊ฐ๋ค์ ํธ๋ฅด๋ฌด์ฆ ํดํ์ด ์ ์ ์ด์ ์ ์์ ๋ก์ด ์ํ๋ก ๋์๊ฐ๊ธฐ๋ ์ด๋ ค์ธ ๊ฒ์ด๋ผ๊ณ ์ ๋งํฉ๋๋ค. ์ด๋์ด ํดํ ํต์ ๋ผ๋ โ์ด๊ฐ๋๊ตญ์ ํโ์ ๋ฐ๊ฒฌํ๊ธฐ ๋๋ฌธ์ ๋๋ค. ์์์ ๊ธฐ๋ขฐ๋ฅผ ๋ถ์คํ๊ฑฐ๋ ์ด๊นจ ๋ฐ์ฌ ๋ฏธ์ฌ์ผ๋ก ์ ๋ฐ์ ์ํํ๋ ๊ฒ๋ง์ผ๋ก๋ ํ๋ฅด์์๋ง ์ ์ฒด์ ์ด์ก์ ๋ง๋น์ํฌ ์ ์๋ค๋ ์ฌ์ค์ ๊นจ๋ฌ์ ๊ฒ์ ๋๋ค.
ํธ๋ฅด๋ฌด์ฆ ํดํ์ ์ฅ๊ธฐ์ ์ธ ๋ฏธ๋์ ๋ํ ๋ช ๊ฐ์ง ์ ์์ด ๋ ผ์๋๊ณ ์์ต๋๋ค. ํธ๋ผํ ๋ํต๋ น์ ํ๋ ์์ ๊ณผ ์ด๋ ์ต๊ณ ์ง๋์๊ฐ ํดํ์ ๊ณต๋์ผ๋ก ์ด์ํ๋ ๋ฐฉ์์ ์ธ๊ธํ๊ธฐ๋ ํ์ต๋๋ค. ๋ณด๋ค ํ์ค์ ์ธ ์ ์์ผ๋ก๋ ๊ตญ์ ์ปจ์์์(international consortium)์ ๊ตฌ์ฑํ์ฌ ํดํ์ ํตํ์ ๊ฐ์ํ๊ณ ํตํ๋ฃ๋ฅผ ์ง์ํ๋ฉฐ, ์ด๋ฅผ ์ด๋, ์ค๋ง, ๋ฏธ๊ตญ, ์ค๊ตญ, ์ธ๋, ์ ๋ฝ ๋ฑ ํดํ ํตํ์ ์์กดํ๋ ๊ตญ๊ฐ๋ค์ด ๊ณต์ ํ๋ ๋ฐฉ์์ด ์์ต๋๋ค. ๊ทธ๋ฌ๋ ํธ๋ผํ ํ์ ๋ถ๊ฐ ๊ตญ์ ํ๋ ฅ์ ํฐ ๊ด์ฌ์ ๋ณด์ด์ง ์์ ์ด๋ฌํ ๊ตฌ์์ด ์คํ๋ ์ง๋ ๋ฏธ์ง์์ ๋๋ค. ๋ฐ์ด๋น๋ ์๊ฑฐ์ ๋ ๋ฒ ์นด ์๋ฆฌ์์ ํธ๋ฅด๋ฌด์ฆ ํดํ์ด ๋ ์ด์ ๋์์์ด๋ ํํ์์ฒ๋ผ ๋ฐฉํด๋ฐ์ง ์๋ ์์ ๋ก์ด ๊ตญ์ ์๋ก๊ฐ ๋๊ธฐ๋ ์ด๋ ค์ธ ๊ฒ์ด๋ผ๊ณ ์ ์ ๋ชจ์์ต๋๋ค.
์๋์ง ์์ฅ์ ๋ณํ์ โ๊ณ ํต ๊ฐ๋ดโ์ ๊ฒ์
ํธ๋ฅด๋ฌด์ฆ ํดํ์ ์ํฉ ๋ณํ๋ ์ ์ธ๊ณ ์๋์ง ์์ฅ์ ๊ด๋ฒ์ํ ์ํฅ์ ๋ฏธ์น ๊ฒ์ ๋๋ค. ์ ๋ฌธ๊ฐ๋ค์ ์ธ ๊ฐ์ง ์ฃผ์ ๋ณํ๋ฅผ ์์ํฉ๋๋ค.
- ๋์์ ์ด์ก ๊ฒฝ๋ก ๊ฐ๋ฐ: ๊ฑธํ๋ง ๊ตญ๊ฐ๋ค์ด ํธ๋ฅด๋ฌด์ฆ ํดํ์ ์ฐํํ์ฌ ์ธ๊ณ ์์ฅ์ผ๋ก ์์ ๋ฅผ ์ด์กํ ์ ์๋ ํ์ดํ๋ผ์ธ๊ณผ ๊ฐ์ ๋์์ ๊ฒฝ๋ก๋ฅผ ๋ ๋ง์ด ๊ฑด์คํ ๊ฐ๋ฅ์ฑ์ด ์์ต๋๋ค. ์ฌ์ฐ๋์๋ผ๋น์์ UAE๋ ์ด๋ฏธ ์ด๋ฌํ ์ต์ ์ ์ผ๋ถ ๋ณด์ ํ๊ณ ์์ต๋๋ค.
- ๋ค๋ฅธ ์ง์ญ ์์ ์์ ์ฆ๊ฐ: ํธ๋ฅด๋ฌด์ฆ ํดํ ํต๊ณผ ์ํ์ด ์๋ ๋ค๋ฅธ ์ง์ญ์ ์์ ์ ๋ํ ์์๊ฐ ์ฆ๊ฐํ ์ ์์ต๋๋ค.
- ๋์ฒด ์๋์ง ๋ถ์: ์์ ๊ฐ๊ฒฉ์ด ๋น์ธ์ง๋ฉด ์์๋ ฅ, ํ์๊ด, ๋ฐฐํฐ๋ฆฌ์ ๊ฐ์ ๋์ฒด ์๋์ง์์ ๋งค๋ ฅ์ด ์ปค์ง๊ณ , ์ด์ ๋ํ ํฌ์๊ฐ ์ฆ๊ฐํ ๊ฒ์ ๋๋ค.
๊ฒฐ๊ตญ ์ด ์ ์๊ณผ ๋ด์๋ ์ฅ๊ธฐ์ ์ผ๋ก ์ ์ธ๊ณ ์๋์ง ์ธํ๋ผ์ ๊ทผ๋ณธ์ ์ธ ๋ณํ๋ฅผ ์ด๋ํ ์ ์์ต๋๋ค. ๋จ๊ธฐ์ ์ผ๋ก๋ ๋ฏธ๊ตญ๊ณผ ์ด๋ ์ค ์ด๋ ์ชฝ์ด ์ด ๋ด์๋ก ์ธํ ๊ณ ํต์ ๋จผ์ ๊ฒฌ๋์ง ๋ชปํ ๊ฒ์ธ์ง์ ๋ํ ์น์ดํ โ๊ณ ํต ๊ฐ๋ดโ ๊ฒ์์ด ๋ ๊ฒ์ ๋๋ค.
์ด๋์ ์ ๊ฐ ์์น์ด ๋ฏธ๊ตญ ๊ตญ๋ด ์ ์น์ ๋ฏธ์น ์ํฅ(ํนํ ์ค๊ฐ์ ๊ฑฐ) ๋๋ฌธ์ ํธ๋ผํ ๋ํต๋ น์ด ๊ฒฐ๊ตญ ๋ฌผ๋ฌ์ค ๊ฒ์ด๋ผ๊ณ ๋ฒ ํ ํ๊ณ ์์ต๋๋ค. ์ค์ ๋ก ์ด๋ ์ง๋๋ถ๋ ์์ ๋ฏธ๋์ด๋ฅผ ํตํด ํธ๋ผํ์๊ฒ ํ๋ฐ์ ๊ฐ๊ฒฉ์ ๋ํด ์กฐ๋กฑํ๊ธฐ๋ ํ์ต๋๋ค. ๋ฐ๋ฉด ๋ฏธ๊ตญ์ ์ด๋์ ์ฃผ์ ์์ ์์ ์ฐจ๋จํ์ฌ ์ด๋ ํ๋ช ์๋น๋(IRGC)์ ์๊ธ์ค์ ๋๊ณ , ์ด๋์ด ๊ฒฝ์ ์ ์ผ๋ก ๋ฌด๋ฆ ๊ฟ๊ณ ํ์ ํ ์ด๋ธ๋ก ๋์์ฌ ๊ฒ์ด๋ผ๊ณ ๋ฏฟ๊ณ ์์ต๋๋ค.
๋ด์์ ์ง์ ๊ฐ๋ฅ์ฑ๊ณผ ์ ์ ์ฌ๋ฐ ๊ฐ๋ฅ์ฑ
๋ฏธ ๊ตญ๋ฐฉ๋ถ๋ ์ด ๋ด์ ์์ ์ ๋ํต๋ น์ด ์ํ๋ ๋งํผ ์ง์ํ ์ ์๋ค๊ณ ๋ฐํ์ต๋๋ค. ๊ทธ๋ฌ๋ ์ด๋ ๋ค๋ฅธ ์ง์ญ์์์ ๊ตฐ์ฌ์ ์๋ฌด ์ํ์ ๋ง๋ํ ๋น์ฉ์ ์น๋ฌ์ผ ํ ๊ฒ์ ๋๋ค. ์ธ๋-ํํ์ ์ง์ญ์์ ์ค๊ตญ์ ๋์ํ๊ณ ๋ถํ ๋ฌธ์ ์ ๋์ฒํ๊ธฐ ์ํด ํ์ํ ํจ์ ๊ณผ ํ์ฝ์ด ์ ์ฉ๋๊ณ ์์ผ๋ฉฐ, ์ ๋ฝ ์ฌ๋ น๋ถ์์ ์ฐํฌ๋ผ์ด๋๋ก ๋ณด๋ผ ์ ์์๋ ์๊ฒฉ ๋ฏธ์ฌ์ผ๊ณผ ํญํ๋ ์ฐจ์ถ๋๊ณ ์์ต๋๋ค. ํ์ฌ ๋ด์ ์์ ์๋ ์ฝ 10,000๋ช ์ ๋ฏธ ํด๊ตฐ, ํด๋ณ๋ ๋ฐ ๊ธฐํ ๋ณ๋ ฅ์ด ํฌ์ ๋์ด ์์ผ๋ฉฐ, ๋ง์ฝ ์ด ์ง์ญ์ ๊ณ์ ์ง์คํด์ผ ํ๋ค๋ฉด ์ ์ฒด ์์ ๊ท๋ชจ๋ 50,000๋ช ์ด์์ผ๋ก ๋์ด๋ ์ ์์ต๋๋ค.
๊ทธ๋ ๋ค๋ฉด ์ ๋ฉด์ ์ด ์ฌ๊ฐ๋ ๊ฐ๋ฅ์ฑ์ ์ผ๋ง๋ ๋ ๊น์? ๋ฐ์ด๋น๋ ์๊ฑฐ๋ ํธ๋ผํ ๋ํต๋ น์ด ๊ณผ๊ฑฐ 38์ผ๊ฐ ์งํํ๋ ๊ฒ๊ณผ ๊ฐ์ ๊ตฐ์ฌ ์์ ์ ๋ค์ ์ํํ๊ธฐ๋ ๋งค์ฐ ์ด๋ ค์ธ ๊ฒ์ด๋ผ๊ณ ์ ๋งํฉ๋๋ค. ๊ทธ์ ์ง์ง์ธต์ด ๋ถ์ด๋์๊ณ , ์ํ๋ ์ ์ ์ ์ธ ์์ด ๊ตฐ์ฌ ์์ ์ ์ํํ๋ ๊ฒ์ ๋ํด ๋ถ๋ง์ ํ์ถํ์ผ๋ฉฐ, ๋๋งน๊ตญ๋ค๋ ๋ฏธ๊ตญ์ ์ง์์ ๋์์ง ์์๊ธฐ ๋๋ฌธ์ ๋๋ค.
์ด๋ ๋ํ ์ ์ ์ข ์์ ์ํ๊ณ ์์ต๋๋ค. ์ธ์์ด ์์๋๊ธฐ ์ ๋ถํฐ ์ทจ์ฝํ๋ ์ด๋ ๊ฒฝ์ ์ ์ ์์ด ๋ฏธ์น๋ ์ํฅ์ ์์ฒญ๋ฌ์ ๊ฒ์ ๋๋ค. ๋ฐ๋ผ์ ์ด์ ๋ฌธ์ ๋ ์ ์์ด ๋๋ ์ง ์ฌ๋ถ๊ฐ ์๋๋ผ, ๋๊ตฌ์ ์กฐ๊ฑด์ผ๋ก ์ ์์ด ๋๋ ๊ฒ์ธ๊ฐ์ ๋ฌ๋ ค์์ต๋๋ค.
ํ์ฌ์ ๋ด์๋ ์ผ์์ ์ธ ํด์ ์ํ๋ฅผ ์ ์งํ๊ณ ์์ง๋ง, ํธ๋ฅด๋ฌด์ฆ ํดํ์ ๋ฏธ๋์ ๊ตญ์ ์๋์ง ์์ฅ์ ํ๋๋ ์์ธก ๋ถ๊ฐ๋ฅํ ๋ณํ์ ๊ธธ๋ชฉ์ ์ ์์ต๋๋ค. ํํธ, ํ๋์ค์ ์๊ตญ์ ๋ฏธ๊ตญ์ ๋ฐฐ์ ํ ์ฑ ๋ ์์ ์ผ๋ก ํธ๋ฅด๋ฌด์ฆ ํดํ ์ฌ๊ฐ๋ฐฉ ๊ณํ์ ์๋ฆฝํ๊ณ ์๋ค๊ณ ๋ฐํ, ๊ตญ์ ์ฌํ์ ๋ค์ธต์ ์ธ ์์ง์์ ์๊ณ ํ๊ณ ์์ต๋๋ค.