YouTube Digest
Listen to this issue
English
Based on “The rise and fall of civilizations | Eric Cline: Full Interview” from Big Think Watch the original video
The Day the World Ended (and Began Again): Lessons from the Bronze Age Collapse
More than three millennia ago, a sophisticated globalized world flourished around the Mediterranean. Powerful empires traded, negotiated, and intermarried, forming a complex web of interdependence. Then, in a shockingly short span of decades around 1200 BCE, this intricate system came crashing down, ushering in centuries of what historians call a “Dark Age.”
For Eric Cline, an archaeologist and ancient historian at George Washington University and author of “1177 BC: The Year Civilization Collapsed,” this period isn’t just a fascinating historical footnote. It’s a stark mirror reflecting our own hyper-connected modern world, offering critical insights into the fragility of complex systems and the potential for a “perfect storm” of catastrophes.
“This was a time where people were basically globalized around the Mediterranean in a way that is not frequently seen,” Cline explains. “And so what happened to them back then may have implications for us today. It turns out to be a much more important period to study than one might expect, even though it’s more than 3,000 years ago.”
The “Ancient G8”: A World of Interdependence
The Late Bronze Age, spanning roughly 1700 to 1200 BCE, was an era of unprecedented internationalization. Cline likens it to an “ancient G8,” where major powers maintained constant contact, whether direct or indirect. This elite club included:
- Mycenaeans and Minoans in the Aegean (think the world of the Trojan War).
- Hittites in Anatolia (modern Turkey).
- Assyria and Babylonia in Mesopotamia (modern Iraq).
- Egyptians in the Nile Valley.
- Canaanites (in what is now Israel, Palestine, Lebanon, Syria).
- Cypriots on the island of Cyprus.
This wasn’t just casual acquaintance; it was a “small world network” where any two points were often just “two hops” away. Commercial, diplomatic, and even marital ties bound these civilizations together. Raw materials were the lifeblood of this network. Egypt supplied gold, Greece provided silver, and Cyprus was the primary source of copper. The most crucial, and problematic, commodity was tin, essential for making bronze (90% copper, 10% tin). Without tin, you couldn’t make tools, weapons, or art. Much of this tin traveled hundreds of miles, often from as far as Afghanistan.
Beyond raw materials, there was a vibrant trade in olive oil, wine, grain, and luxury goods. Cline recounts an anecdote of leather shoes sent from Crete to King Hammurabi in Babylon, only to be returned – perhaps they were “too last millennium,” he muses. At the highest echelons, kings engaged in elaborate gift-giving, exchanging items like solid gold daggers inlaid with lapis lazuli.
Diplomacy was cemented through dynastic marriages. Egyptian pharaohs like Amenhotep III and Akhenaten had harems filled with princesses from Mitanni and Babylonia, solidifying treaties. The fascinating detail? Egyptians never sent their own princesses abroad; they only received them.
This complex web of interdependence was both their greatest strength and their fatal flaw. It allowed these civilizations to reach unprecedented heights of prosperity and cultural exchange. But, as Cline warns, “This globalized network is what rose them up to the highest levels, but it is then also what brought them crashing down at the end when all of that was cut.”
What Does “Collapse” Even Mean?
The term “collapse” itself is a subject of academic debate. Some scholars prefer “transformation,” arguing that societies didn’t vanish but merely changed. Cline, however, stands firm: “The world that they had had in place in the 14th century and the 13th century BC, that world goes away. And it goes away within just a couple of decades in the early 12th century BC.”
What truly collapsed was the network – the commercial, diplomatic, and marital connections that linked them all. This wasn’t an overnight event for everyone, nor did every society disappear entirely. But for many, life as they knew it fundamentally altered. Cline compares it to the fall of the Roman Empire: “It was catastrophic for its day.”
This phenomenon is known as a “systems collapse,” a term coined by Colin Renfrew. It signifies the breakdown of a complex system where central economies, elite structures, and centralized governments vanish, often accompanied by massive population decline due to death and migration. The ensuing period is typically a “Dark Age,” marked by a regression to lower levels of socio-political and economic functioning.
The “Perfect Storm”: Beyond a Single Cause
For decades, scholars sought a single, monocausal explanation for the Bronze Age collapse. The most popular culprit was the “Sea Peoples,” a mysterious confederation of invaders mentioned in Egyptian texts. These groups, with names like the Shardana, Shekelesh, and Peleset, were depicted as sweeping across the Mediterranean, attacking and destabilizing kingdoms. The Peleset, interestingly, are widely identified as the Philistines of biblical fame, who appear to have been Mycenaeans from Greece who fled to the Levant.
However, modern archaeology and a wealth of new data have fundamentally shifted this understanding. Cline and his colleagues now advocate for a “polycausal” explanation – a “perfect storm” of interconnected disasters. “You need two of the explanations, three, four, something like that happening either all at once or in rapid succession,” Cline asserts. “So you don’t have time to recover from one catastrophe before the next one hits.”
The Sea Peoples, Cline suggests, are largely “history’s great scapegoats.” Far from being the sole destroyers, they were more likely symptoms of a wider crisis, victims themselves, migrating in search of a better life. He draws parallels to the American Dust Bowl of the 1930s or modern-day refugees fleeing conflict, like those from the Syrian Civil War.
So, if not just the Sea Peoples, what else? The evidence points to a terrifying confluence of factors:
-
Mega-Drought and Famine: This is no longer a hypothesis but a scientific certainty. Data from stalagmites, dried-up lakes, and pollen analysis reveal a “mega-drought” lasting anywhere from 150 to 300 years, beginning around 1250 BCE and affecting a vast area from Italy to Iran. Such a prolonged drought would have devastated agriculture, leading directly to widespread famine. Texts from Ugarit (a vital port city in Syria) and Hittite archives explicitly plead for grain and report severe starvation. This environmental catastrophe likely spurred the migrations of groups like the Sea Peoples.
-
Invasions and Internal Rebellions: While the Sea Peoples weren’t the sole cause, there were invaders. New tablets from Ugarit describe “enemy ships” landing and overrunning port cities before advancing on Ugarit itself. The archaeological record confirms a devastating destruction layer at Ugarit, a meter deep, with bodies in the streets and arrowheads in walls. The city was abandoned for centuries. However, not all destruction was external. Some cities, like Hazor in Canaan or Mycenae in Greece, show signs of internal rebellion, where starving populations might have risen against their elites, burning palaces and temples but leaving common houses intact.
-
Earthquake Storms: The Eastern Mediterranean is a seismically active zone. Evidence suggests a series of powerful earthquakes, an “earthquake storm” (or “earthquake sequence” in modern seismology), struck between 1225 and 1175 BCE. Off-kilter walls, slipped keystones, and bodies trapped under collapsed structures (like the teenage girl at Mycenae) point to widespread seismic activity. Even Troy, site of the legendary war, saw its “City Six” destroyed by an earthquake, not human invaders. These natural disasters would have compounded the existing stresses, crippling infrastructure and further disrupting trade.
-
Disease: The “four horsemen of the apocalypse” often ride together. While harder to pinpoint archaeologically, historical accounts like Homer’s Iliad and the biblical Exodus mention plagues. More concrete evidence comes from the mummy of Egyptian pharaoh Ramses V (c. 1140 BCE), which shows pustules consistent with smallpox. A papyrus from Turin describes his death from plague and an unprecedented 16-month quarantine of the Valley of the Kings. Disease, especially in a time of drought, famine, and population displacement, would have been a final, devastating blow.
This “poly crisis” created a multiplier and domino effect. One catastrophe weakened societies, making them more vulnerable to the next. The disruption of a single vital trade route, like the tin supply, could cripple an entire network.
What Could Have Been Avoided? And What Came Next?
To understand the network’s vulnerability, Cline collaborated with the US Army Corps of Engineers, running computer simulations of collapse scenarios. They found that the simultaneous fall of the Hittite Empire and the crucial port city of Ugarit would have been enough to shatter the entire globalized system. The Hittites did disappear, and Ugarit was utterly destroyed. Egypt, though severely weakened, “muddled through,” surviving the worst of the collapse. This suggests that while resilience is possible, the ability to adapt and react to unfolding crises is paramount.
The immediate aftermath was indeed a “Dark Age” (c. 1200-800 BCE). Monumental building ceased, and writing systems like Linear B in Mycenaean Greece disappeared, leading to a loss of literacy. Historians often characterize this period as a regression, where people looked back at the Bronze Age as a “golden age,” reflected in epics like Homer’s Iliad and Odyssey.
Yet, necessity proved to be the mother of invention. The scarcity of tin forced a shift to a new metal: iron. Though known before, iron became widely adopted during and after the collapse, ushering in the Iron Age. The Phoenicians, survivors from central Canaan (modern Lebanon), not only thrived but also standardized the alphabet, which they then spread across the Mediterranean, laying the foundation for the Greek and Latin alphabets we use today. The Cypriots, too, demonstrated remarkable resilience, becoming pioneers in iron metallurgy.
History Rhymes
The story of the Late Bronze Age collapse is a powerful cautionary tale. It reveals how tightly woven global systems, while fostering prosperity, also create vulnerabilities. A cascade of seemingly disparate events – climate change, resource scarcity, migration, conflict, and disease – can overwhelm even the most powerful civilizations.
As Eric Cline concludes, “I’m not saying that we’re necessarily due for another collapse, but history does rhyme even if it doesn’t repeat.” In an era of increasing global interdependence, climate change, and geopolitical instability, the echoes of 1177 BCE are perhaps louder than ever. Understanding how ancient civilizations navigated their “perfect storm” might just be one of the most important lessons for our own survival.
Based on “Are Human Drivers Finally Obsolete? | Freakonomics Radio” from Freakonomics Radio Network Watch the original video
The End of the Road for Human Drivers? Inside the Quest to Automate the Wheel
For centuries, the act of driving has been intrinsically human – a task of skill, instinct, and sometimes, unfortunate error. From the horse-drawn carriage to the modern automobile, a person has always been at the helm. But what if that era is rapidly drawing to a close? What if the word “driver” soon conjures images not of a person, but of a machine, much like “dishwasher” or “computer” do today?
This profound shift is the subject of a deep dive by PJ Vote, host of the Search Engine podcast, who recently explored the two-part series on driverless cars for Freakonomics Radio. His journey into this topic began, surprisingly, not with technology, but with a bench-pressing injury.
“I’d gotten too into bench pressing,” Vote recounted, “and I injured myself. I had a hernia and then I had to have a hernia repair.” With limited mobility and in pain, a visit to a friend in San Francisco led him to take a Waymo robo-taxi. “It was such an experience of the future that immediately becomes normal,” he marvelled. “The idea that I would press a button on my phone, a car would come out of nowhere driven by nobody. I would get in, watch the steering wheel turn itself.” He likened the initial awe to a first airplane ride, which quickly becomes as mundane as an elevator by the third. This transformative experience sparked his curiosity: why wasn’t everyone talking about this impending revolution?
The Visionaries and the Skeptics
Vote’s investigation uncovers a cast of compelling characters driving this change. In the first part of his series, focusing on the car itself, he introduces Sebastian Thrun. A German-born roboticist and AI expert, Thrun lost a friend to a car accident as a teenager. His motivation for developing autonomous vehicles transcends mere convenience or profit; he genuinely seeks to reshape the modern world and save lives. His early ideas, once dismissed as crazy, have steadily gained plausibility over two decades.
The second part of the series, shifting focus to the driver, features the strongly opinionated politicians of Boston, grappling with the societal implications of this technology. Steven Dubner, host of Freakonomics Radio, noted a recurring theme throughout Vote’s reporting: “Every time I spoke to someone, as they were talking, I thought everything they were saying makes sense… And then I would go talk to the next person who saw things completely differently and it would just spin my head the other way.” This struggle to reconcile competing, yet equally logical, interests lies at the heart of the driverless car debate.
Dubner himself confesses to being “anti-human driver for about 50 years.” He readily admits his own driving flaws – a temper, distractions, and a lack of skill. After riding in an autonomous test vehicle at Carnegie Mellon University, he was convinced: “Give me the autonomous vehicles. So plainly better than I am as a driver.”
A Ghost of Jobs Past: The Knocker-Upper and the Lamp Lighter
To truly grasp the potential impact of driverless cars, Vote begins with a historical analogy. Imagine yourself nearly 200 years ago, waking before dawn to a hard rapping at your window. That’s the “Knocker-Upper,” a job that existed for a century before alarm clocks, where a person walked the neighborhood tapping on windows to rouse workers. Outside, gas street lamps still burned, lit the night before by the “Lamp Lighter,” a cherished neighborhood fixture who made his rounds at dusk with a ladder and a light.
And then there’s you, a professional driver – a person on a coach, holding the reins of a horse, taking passengers where they needed to go. Today, the Knocker-Upper is your iPhone alarm, the Lamp Lighter is the electric street light. These jobs have vanished, replaced by technology. The driver, however, has persisted. The question now is whether this third job, and indeed the routine human task of driving for nearly everyone else, is also on the brink of transformation.
The Perils of Human Control
Alex Davies, author of Driven: The Race to Create the Autonomous Car, shares Vote’s fascination with human driving limitations. “I can’t always pay attention to everything,” Davies admits, “that I get tired.” He recently resolved to be calmer on the road, especially with a baby on the way. This resolution underscores a stark reality: for most of us, driving is the riskiest behavior we routinely engage in. Davies himself was in an accident just months after speaking with Vote, totaling his car.
Safety is the core promise of the driverless car. Computers don’t get drunk, tired, or distracted. They never text or succumb to road rage. And these vehicles aren’t a distant dream; they’re here. Robo-taxis like Waymo already operate in 10 American cities, providing millions of rides. China boasts an even wider rollout. In places like San Francisco or Austin, a driverless car is as common as an Uber, with passengers choosing between a human or a robot driver via their phone. This rapid evolution, happening now, promises to fundamentally reshape our cities and our lives.
Chapter 1: Dreams Without Drivers
The idea of a self-driving car is almost as old as the automobile itself. When humanity transitioned from horse-drawn carriages to mechanical vehicles, something crucial was lost: sentience. A horse wouldn’t simply run off a cliff if you dropped the reins. Early automobiles, powerful and non-sentient, faced passionate resistance. In the 1800s, people feared these “thundering” machines, partly due to job displacement. Horse breeders, farriers, feed suppliers, manure haulers, carriage manufacturers, and “Teamsters” (original drivers of horse teams) all saw their livelihoods threatened.
Beyond jobs, cars were undeniably unsafe. Anti-car activists pushed for regulations like “red flag laws,” requiring a person to walk ahead of an automobile waving a flag. One Pennsylvania proposal even suggested drivers disassemble their cars and hide the parts behind bushes if livestock were encountered. While extreme, these activists were directionally correct: cars initially wiped out many jobs and were incredibly dangerous. Cities like Detroit, which initially embraced cars without regulation, saw astonishing death rates, particularly among children, in the early 1900s.
It took decades for society to adapt: laws, licenses, driver’s education, road design improvements, highways, seat belts, and airbags all made driving safer. However, the smartphone has reversed some of that progress. Today, car-related deaths in America are as common as those from guns or opioids – about one in a hundred people will likely know someone who dies in a car accident. This enduring problem fueled the desire to make cars more “sentient,” like the horses they replaced. Early visions included radio-controlled cars or vehicles guided by magnets embedded in the road. The limiting factor was always the available technology.
Chapter 2: The DARPA Grand Challenges – Forging a Future
The turn of the millennium brought a turning point. Deep within the Department of Defense, DARPA – the agency behind GPS, the M16, and the early internet – set its sights on autonomous military vehicles. In 2002, DARPA’s director, Tony Tether, proposed an unusual approach: a contest. The goal was to inspire tech innovation beyond simply building more websites. The prize for winning the “Grand Challenge” was $1 million.
The rules were remarkably open: any vehicle, any design, as long as it didn’t attack other competitors (a question actually posed by a team with “Battlebots” history). Tether’s initial idea of racing down the Las Vegas Strip was quickly dismissed as impractical, leading to the chosen venue: the desert outside Las Vegas. The true “driver” DARPA sought to replace was the American soldier, envisioning vehicles that could navigate roads potentially laden with explosives.
Two key engineers emerged from this formative period. Chris Urmson, a PhD student at Carnegie Mellon University, joined their “Red Team” to build “Sandstorm,” a bright red Humvee bristling with futuristic sensors. He grappled with the fundamental challenge of teaching a computer to control a vehicle’s steering, brakes, and throttle. Representing a different approach was Anthony Lewendowski, a charming, gangly entrepreneur from Berkeley. Lacking the resources of Carnegie Mellon, he opted for a standout design: “Ghost Rider,” the race’s only self-driving motorcycle.
The 2004 Grand Challenge was, in Vote’s words, “an utter hysterical disaster.” Ghost Rider toppled immediately because Lewendowski forgot to flip a stabilization switch. Every subsequent vehicle failed miserably. Sandstorm got stuck on a berm, its wheels spinning so hard they melted the tires, emitting plumes of black smoke before being shut down. Chris Urmson compared it to an Olympic marathon where the best runner only completes two miles.
However, the contest had achieved its underlying goal: it had “flushed all these inventors out,” jumpstarting the scene that would develop autonomous technology. Observing the chaos that day was Sebastian Thrun, the legendary roboticist from Stanford. He saw a fundamental error: “all the teams treated this like a hardware problem.” Thrun believed the true challenge was software, replacing the human driver’s mind, not just building bigger wheels. He was also driven by a humanitarian vision, imagining the potential to save millions of lives globally if traffic accidents could be eliminated.
Eighteen months later, in October 2005, DARPA doubled the prize to $2 million for the second Grand Challenge. Familiar faces returned: Urmson with two Carnegie Mellon vehicles, Lewendowski again with his still-failing motorcycle. But the new contender was Thrun’s Stanford team, with their comparatively “measly” blue Volkswagen SUV named “Stanley.” Thrun, a computer scientist, brought a focus on artificial intelligence, teaching Stanley to recognize roads and train itself by recording what its cameras saw. Stanley learned to differentiate between good driving surfaces (like grass) and bad ones (like mud) by measuring slipperiness and bumpiness. It detected patterns and generalized its learning thirty times a second, much like a human.
The second race was a resounding success. Multiple vehicles finished the challenging 132-mile course. The real question was who would do it fastest. Stanley, sandwiched between Carnegie Mellon’s behemoths, emerged victorious. Thrun described the sight of his blue car, a “dust cloud” turning bluish, crossing the finish line as “unbelievably magical.” It was a “made-for-TV Kumbaya moment,” a testament to a community of innovators, before the cutthroat competition for driverless cars truly began.
Chapter 3: Google’s Secret Project – From Desert to Public Roads
The DARPA Grand Challenge had not only showcased roboticists but also attracted an unexpected observer: Google co-founder Larry Page. Disguised in a hat and sunglasses, Page buttonholed Thrun, asking highly specific questions about his LiDAR system. Their connection wasn’t new; Thrun had previously fixed a small robot Page had built for telepresence meetings. Page, who had wanted to do his grad school thesis on autonomous vehicles, now saw tangible proof that the dream could be real.
Initially, Page hired Thrun and Lewendowski to build Google Street View, modifying Stanley’s roof-mounted camera system. But soon, Page returned to his true dream: a driverless car. In 2009, Page approached Thrun with a mission: “Sebastian, I think you should build a self-driving car that can drive anywhere in the world.” Thrun’s immediate reaction was skepticism. “No, taking the technology we built for this empty desert and putting in the middle of Market Street in San Francisco is going to kill somebody.”
Page persisted, asking Thrun to provide a technical reason why it couldn’t be done. “That’s the moment of incredible pain,” Thrun recalled, “because I go home and I can’t think of a technical reason why not.” This realization taught him a crucial lesson: “Experts are usually expert of the past, not the future, and if you ask an expert about innovation, something crazy new, they’re the least likely person to say yes, it can be done.”
Thus began Google’s secret self-driving car project, “Project Chauffeur,” in 2009. Led by Thrun, with Chris Urmson managing day-to-day operations, Anthony Lewendowski on hardware, and Dmitri Dolgov on planning, the small team of 11 engineers reported directly to Larry Page. Their nebulous goal was refined into two challenges: safely log 100,000 miles on public roads and complete the “Larry 1K.” The Larry 1K involved driving 10 “tricky” 100-mile routes across California – from the Bay Bridge to Lombard Street – without a single human takeover.
To kickstart the project, the team licensed Stanford’s DARPA Urban Challenge code. Lewendowski bought eight Toyota Priuses, retrofitting them with radar, cameras, and a spinning 360-degree LiDAR system. These cars, initially given names like “Night Rider” before being numbered, were transformed into autonomous test vehicles. Don Bernett, a researcher on motion planning, joined the team, tasked with teaching the car nuanced behaviors like “nudging” – the subtle shift a human driver makes to the left when a large truck passes on the right.
Early testing took place in secrecy, in the Shoreline Amphitheater parking lot near Google’s offices and an empty airplane runway. Spring 2009 marked their first real road driving on the Central Expressway. Immediately, a critical flaw emerged: the car was “swerving wildly,” like a “drunken sailor.” The small oscillations unnoticed on a runway became a significant problem on a public road.
The team adopted rigorous safety protocols. Two-person teams manned each car: a safety driver behind the wheel, ready to take over, and a partner watching a graphical interface, calling out discrepancies between sensor data and reality. This iterative process of logging errors, troubleshooting, and updating code was how the car learned. Bernett recalled how this intensely practical work made him obsess over human driving behavior. Why do humans drive the way they do? He found no easy answers, noting that “machine learning” now infers these “deep truths.”
He offered a fascinating example: the “lateral acceleration” (the force that pushes you sideways) tolerated by passengers varies greatly with context. Two meters per second squared feels comfortable on a highway on-ramp, but “incredibly uncomfortable” (like “Mario Kart”) during a U-turn in a cul-de-sac, where the limit is almost three times less. Human perception, not just physical forces, dictated comfort.
Despite these complexities, the unifying goal of the Larry 1K kept the team focused. By 2010, just a year in, they were on a roll, knocking out routes like Route One to Carmel and the Bay Area bridges. Each failure provided valuable data, leading to code improvements. They completed the Larry 1K in just over a year, nearly twice as fast as expected, celebrating each completed route with a $13.99 bottle of Corbell champagne.
By the fall of 2010, the “miracle” was complete. They had built a driverless car, with human supervision and extensive coding, that could safely navigate tricky California roads. They celebrated, throwing each other into Sebastian Thrun’s pool. But then, a new question emerged: “Okay, and now what?” The path forward began to wobble. Competition would arrive, the team itself would face internal divisions, and some, believing the pace too slow, would take matters into their own hands in extreme ways. The era of driverless cars was just beginning, and the road ahead was anything but straight.
Based on “Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI” from No Priors: AI, Machine Learning, Tech, & Startups Watch the original video
The Loopy Era of AI: Andrej Karpathy on Living with Code Agents, AutoResearch, and Digital Psychosis
Imagine a world where you no longer write code, but instead “express your will” to an army of AI assistants working tirelessly on your behalf. Where your home runs itself, orchestrated by a digital butler you text via WhatsApp. And where scientific research is conducted autonomously, with AI models recursively improving themselves, leaving human experts to merely contribute ideas to a queue.
This isn’t a distant sci-fi fantasy; it’s the present reality for Andrej Karpathy, a leading voice in AI and former head of AI at Tesla. In a recent conversation on the “No Priors” podcast, Karpathy offers a captivating, at times bewildering, glimpse into what he terms the “AI psychosis” of this “loopy era.” His insights reveal a profound shift in how we interact with technology, work, and even conceive of intelligence itself.
The End of Manual Coding: A Personal Revolution
For Karpathy, the transformation has been swift and absolute. “Code’s not even the right verb anymore,” he declares, describing his new daily routine: “I have to express my will to my agents for 16 hours a day.” This isn’t hyperbole. Since December, he estimates he hasn’t typed a single line of code himself. His workflow has flipped from 80% manual coding to a mere 20%, delegating the vast majority of tasks to AI agents.
This radical shift, Karpathy explains, stems from a “huge unlock in what you can achieve as a person, as an individual.” No longer bottlenecked by typing speed or manual execution, he now commands a fleet of digital helpers. He recounts his “perpetual state of AI psychosis,” a mix of exhilaration and anxiety born from the rapid, unexplored capabilities now at his fingertips. The default workflow for software engineers, he argues, has fundamentally changed since late 2023, yet many outside the immediate AI frontier have yet to grasp the magnitude of this revolution.
Beyond Single Sessions: The Rise of “Claw-like Entities”
The initial interaction with AI coding assistants like GitHub Copilot or OpenAI’s Codex often involves a single session, a brief exchange. But Karpathy and others are pushing far beyond this. The next frontier involves not just multiple agents, but persistent, “Claw-like entities” that loop and operate autonomously.
He points to Peter Steinberg as a pioneer in this space, famous for a setup where numerous coding agents (like Codex) tile his monitor, each independently tackling different tasks. “It’s not just like here’s a line of code, here’s a new function,” Karpathy explains. “It’s like here’s a new functionality and delegate it to agent one. Here’s a new functionality that’s not going to interfere with the other one. Give it to agent two.” This allows for “much larger macro actions” over a software repository, with agents simultaneously researching, planning, and writing code. The human’s role morphs into a high-level manager, reviewing work and orchestrating these macro-level operations.
This new mode of interaction is “very rewarding” because it works, but it’s also “the new thing to learn,” contributing to Karpathy’s “psychosis.” The challenge becomes developing a “muscle memory” for this meta-level command.
The Human as the New Bottleneck: A “Skill Issue”
With AI agents handling the grunt work, the bottleneck has shifted dramatically. It’s no longer about compute power or access to resources, but about the human’s ability to effectively direct and leverage these tools. Karpathy vividly describes feeling “nervous when I have subscription left over” on his AI platforms, implying he hasn’t maximized his “token throughput”—the sheer volume of AI processing he’s commanding. This echoes a past anxiety of PhD students seeing their GPUs idle, but now it’s about tokens, not flops.
The prevailing sentiment when things don’t work is “skill issue.” It’s not that the AI can’t do it, but that the human hasn’t given good enough instructions, hasn’t configured the memory tools properly, or hasn’t found the right way to parallelize the agents. This “addictive” nature of self-improvement, where “unlocks” occur as one gets better at commanding AI, is a core driver of the “psychosis.”
Dobby the Elf Claw: Automating Life Itself
Karpathy’s personal life offers a compelling illustration of this agentic future. During a period of “Claw psychosis” in January, he built “Dobby the elf claw,” an AI agent that manages his entire home.
Dobby began by autonomously scanning Karpathy’s local network, identifying smart home subsystems like Sonos speakers. With a few prompts, the agent reverse-engineered APIs, found endpoints, and soon music was playing in the study. Dobby then took control of lights, HVAC, shades, pool, spa, and even the security system. Now, Karpathy simply texts Dobby via WhatsApp: “Dobby, sleepy time,” and all lights go off. His security camera, paired with an AI model, sends him texts with images when a FedEx truck pulls up.
This personal automation highlights a deeper implication: the potential obsolescence of many bespoke apps. “These apps that are in the app store for using these smart home devices… shouldn’t even exist,” Karpathy posits. Instead, everything should be exposed as APIs, with agents acting as the intelligent glue. “The customer is not the human anymore. It’s like agents who are acting on behalf of humans.” While this currently involves some “vi coding” for technically proficient users, Karpathy predicts that within “a year or two or three,” such home automation will be “trivial” and “free,” easily translating from non-technical human intent.
AutoResearch: AI Improving AI
The concept of removing the human bottleneck extends most dramatically to scientific discovery. Karpathy’s “auto research” project exemplifies this. His goal: “I don’t want to be like the researcher in the loop… I’m holding the system back.” The objective is to arrange systems so they are “completely autonomous,” allowing agents to run “for longer periods of time without your involvement.”
Karpathy, with two decades of experience training large language models (LLMs), put his own expertise to the test. He had painstakingly tuned a GPT-2 model by hand, believing it was “fairly well tuned.” But after letting auto research run overnight, it returned with “tunings that I didn’t see,” identifying subtle interactions between hyperparameters like weight decay and Adam betas that he had missed.
This single loop of auto research, Karpathy emphasizes, is just the beginning. Frontier AI labs are already exploring “recursive self-improvement,” where LLMs improve other LLMs. He envisions automated scientists generating ideas from papers and GitHub repos, funneling them into a queue for AI workers to test. “Removing humans from all the processes and automating as much as possible” is the future of research, requiring a complete “rethinking of all the abstractions.”
The Meta-Loop: Optimizing Research Itself
The “loopy era” extends even to the organization of research. Karpathy muses about “program MDs”—markdown files that describe how a research organization should work, detailing roles, processes, and even risk appetite. Just as auto research optimizes model parameters, one can imagine a “meta-layer” where AI optimizes these “program MDs.”
“You can definitely imagine that you have multiple research orgs,” Karpathy explains, each with its own “code” (its program MD). “And once you have code, then you can imagine tuning the code.” This opens the door to AI designing better, more efficient research processes, leading to a perpetual loop of self-improvement at every level.
The Jagged Edge: AI’s Limitations and the Call for Speciation
Despite the astounding capabilities, Karpathy acknowledges the current limitations of AI. He describes the experience of interacting with these models as simultaneously talking to “an extremely brilliant PhD student who’s been like a systems programmer for their entire life and a 10-year-old.” This “jaggedness,” where models excel in verifiable tasks but struggle with nuance, intent, or creativity, is a persistent frustration.
He illustrates this with the “atom joke”: “Why don’t scientists trust atoms? Because they make up everything.” This joke, he notes, was common years ago and is still what you’ll get from state-of-the-art models today. While agents can “move mountains” for hours on complex tasks, they offer the same “crappy joke from 5 years ago” because “it’s outside of the reinforcement learning. It’s outside of what’s being improved.” This suggests a decoupling: being smarter at code generation doesn’t automatically translate to broader intelligence or creativity in all domains.
This “jaggedness” leads Karpathy to question the current “monoculture of models” pursued by many labs—single, massive LLMs intended to be arbitrarily intelligent across all domains. He argues for “more speciation in the intelligences,” akin to the diversity of brains in the animal kingdom. Instead of an “oracle that knows everything,” we might benefit from smaller, specialized models that retain a “cognitive core” but become highly efficient and competent in specific niches, like mathematics or particular programming languages. This “unbundling” could be driven by the need for efficiency, especially with compute constraints, though the “science of manipulating the brains” (deep fine-tuning without losing capabilities) is still nascent.
A Future of Infinite Possibility and Perpetual Motion
Andrej Karpathy’s vision paints a picture of an AI-driven future that is both exhilarating and dizzying. We are entering a “loopy era” where AI agents are not just tools but collaborators, autonomous entities that will increasingly manage our digital and physical worlds. The human role is shifting from direct execution to high-level orchestration, leveraging AI to achieve unprecedented scales of productivity and discovery.
Yet, this transformation comes with its own anxieties: the “psychosis” of infinite possibilities, the constant feeling of a “skill issue,” and the challenge of navigating AI’s brilliant but sometimes nonsensical jaggedness. As Karpathy concludes, the progression is obvious, but the path is still rough. The journey into the loopy era of AI promises to be one of perpetual motion, constant learning, and profound redefinition of what it means to be human in a world increasingly shaped by machine intelligence.
Based on “Terence Tao – How the world’s top mathematician uses AI” from Dwarkesh Patel Watch the original video
Decoding Discovery: Terence Tao on AI’s Impact on the World of Math and Science
Terence Tao, often hailed as the “greatest living mathematician,” rarely needs an introduction. Yet, in a recent conversation with Dwarkesh Patel, Tao offered a fascinating perspective on how artificial intelligence is not just assisting, but fundamentally reshaping the landscape of scientific discovery. By drawing parallels between AI and historical scientific breakthroughs, particularly Johannes Kepler’s monumental work, Tao illuminated the profound shifts occurring in how we generate, verify, and ultimately understand scientific progress.
Kepler, the High-Temperature LLM
The discussion began with a journey back to the 17th century, to the story of Johannes Kepler and his quest to understand planetary motion. Building on the heliocentric model proposed by Copernicus (who himself was influenced by Aristarchus), Kepler initially clung to a beautiful, yet ultimately flawed, theory. Copernicus had suggested that planets orbited the Sun in perfect circles, a concept that largely fit observations gathered over centuries by Greek, Arab, and Indian astronomers.
Kepler, however, observed curious geometric relationships in Copernicus’s predicted orbit sizes. He posited a theory so elegant it seemed divinely inspired: the orbits of the six known planets could be nested within the five perfect Platonic solids (cube, tetrahedron, icosahedron, octahedron, dodecahedron). For instance, if the Earth’s orbit was enclosed in a cube, the outer sphere enclosing that cube almost perfectly matched Mars’s orbit. This theory, to Kepler, represented mathematical perfection aligning with God’s design.
To confirm his grand vision, Kepler desperately needed accurate data. At the time, only one such dataset existed, meticulously compiled by the eccentric and wealthy Danish astronomer Tycho Brahe. Brahe had convinced the Danish government to fund an entire island observatory, where for decades he made naked-eye observations of planets like Mars and Jupiter, achieving a precision ten times greater than any before him.
Kepler began working with Brahe, but the older astronomer jealously guarded his data, releasing only fragments. Eventually, Kepler, in a move that might be considered scientific espionage, copied the data, leading to a contentious battle with Brahe’s descendants. Armed with this unprecedented trove of information, Kepler set about verifying his beautiful Platonic solid theory.
To his profound disappointment, the data didn’t quite fit. His theory was off by about 10%. Years of relentless work, trying every conceivable “fudge” and adjustment, followed. It was a Herculean effort of data analysis that eventually led him to a shocking conclusion: planetary orbits were not perfect circles, but ellipses. This discovery formed Kepler’s first law. He then worked out his second law – that planets sweep out equal areas in equal times – and ten years later, after further arduous data collection, his third law, relating a planet’s orbital period to its distance from the Sun.
Kepler had no underlying theory to explain why these laws were true; they were purely empirical regularities derived from data. It would take Isaac Newton a century later to provide the theoretical framework (gravity and laws of motion) that unified and explained all three.
Dwarkesh Patel then drew a striking analogy: “Kepler was a high-temperature LLM.” He meant that Kepler, throughout his career, was essentially “trying random relationships,” some of which were wildly speculative (like his asides on “The Harmonics of the World” connecting planetary orbits to musical notes and astrology). Yet, when these wild ideas were rigorously tested against Brahe’s “verifiable data bank,” a few of them, like the elliptical orbits and the cube-square law, led to profound scientific progress. Tao concurred, emphasizing that this combination of prolific, often unconventional, idea generation coupled with stringent verification was key.
The Shifting Bottleneck of Scientific Progress
Traditionally, the history of science celebrates “eureka genius moments” – the generation of novel ideas. A scientific problem typically involves many steps: identifying a fruitful problem, collecting data, strategizing data analysis, formulating a hypothesis, validating it, and communicating the findings. Idea generation has always been the “prestige part.”
However, Tao observes that science has evolved. Classically, the two main paradigms were theory and experiment. The 20th century added numerical simulation, and the late 20th century ushered in the era of “big data.” Now, much scientific progress is driven by analyzing massive datasets first, then drawing patterns and deducing hypotheses – almost reversing the classic scientific method of forming a hypothesis and then collecting data to test it. Kepler, in a sense, was an early data scientist, but even he started with preconceived theories before diving into Brahe’s data.
This is where AI enters the picture. Tao argues that AI has driven the cost of idea generation “down to almost zero,” much like the internet drove down the cost of communication. This is an “amazing thing,” but it doesn’t automatically create abundance or progress. Instead, it shifts the bottleneck.
“Now the bottleneck is different,” Tao explains. “We’re now in a situation where suddenly people can generate thousands of theories for a given scientific problem. Now we have to verify them, evaluate them.” Human reviewers are already overwhelmed by AI-generated submissions to journals. The challenge is no longer generating ideas, but assessing which ones genuinely advance the field and which are “dead ends or red herrings.” This is a problem we don’t yet know how to solve at scale.
The Challenge of Evaluating “True Progress”
If we soon have “billions of AI scientists” churning out theories, how do we discern real progress? Tao points out that human science has faced similar, albeit smaller-scale, challenges. He references the development of the “bit” at Bell Labs in the 1940s – a unifying concept with implications across many fields, not just the initial engineering problem it addressed. How do you identify such a transformative idea amidst millions of papers, especially if it initially lacks broad unifying power?
“A lot of it’s the test of time,” Tao notes. Many great ideas, like deep learning itself or the transformer architecture (the foundation of modern LLMs), didn’t receive immediate widespread acceptance. Their fruitfulness became apparent only much later, as other scientists built upon them. The adoption of an idea can also depend on cultural and societal factors, like the base-ten numeral system becoming standard despite other possibilities.
Evaluating an idea’s worth isn’t purely objective; it’s deeply intertwined with its context, past and future. Progress often isn’t linear. Copernicus’s heliocentric model, while conceptually simpler, was initially less accurate than the geocentric Ptolemaic system, which had been refined over a millennium with countless ad hoc fixes. It took Kepler’s elliptical orbits to make the heliocentric model superior.
“Science is always a work in progress,” Tao states. “When you only get part of the solution, it looks worse than a theory which is incorrect but somehow has been completed to the point where it kind of answers all the questions.” He cites Leibniz’s disagreement with Newton’s action-at-a-distance gravity, and Newton’s own puzzlement over inertial and gravitational mass equivalence – mysteries only resolved by Einstein centuries later.
Sometimes, progress isn’t about adding new theories but “deleting some assumptions that you have in your mind.” The long-held Aristotelian notion that objects naturally rest made heliocentrism seem implausible (“How come we weren’t all falling over?”). Similarly, Darwin’s theory of evolution challenged the deeply ingrained idea of static species.
Tao sees us in a “cognitive version of the Copernican revolution” right now. We once believed human intelligence was the center of the universe, but now we’re confronting diverse forms of intelligence with different strengths and weaknesses. “Our assessment of which tasks require intelligence, which ones don’t, has to be reordered quite a bit.”
Darwin vs. Newton: The Role of Communication and Data Loops
Dwarkesh Patel brought up a curious point from Edward Dolnick’s book The Clockwork Universe: Darwin’s Origin of Species (1859) came out two centuries after Newton’s Principia Mathematica (1687), yet Darwin’s theory conceptually seems simpler. Thomas Huxley famously said upon reading Origin: “How stupid not to have thought of that.” No one said that about Principia. Why the delay?
Tao suggests it relates to the nature of evidence and communication. Newton could present equations that immediately predicted observable phenomena, like the Moon’s orbit. Darwin’s evidence for natural selection, while overwhelming, was “cumulative and retrospective.” Lucretius had a similar idea in the first century BC, but it gained no traction because he couldn’t “run some experiment and force people to pay attention.”
Beyond data, communication is vital. Darwin was an “amazing science communicator,” writing in plain English without equations, synthesizing disparate facts into a compelling vision. Newton, by contrast, wrote in Latin, invented entirely new mathematics to explain his work, and was notoriously secretive and competitive. His work only became widespread decades later when others simplified and explained it.
“The art of exposition and making a case and creating a narrative is also a very important part of science,” Tao emphasizes. “If you have the data, it helps, but people need to be convinced, otherwise they will not push it further or take the initial investment to learn your theory and really explore it.” This “soft, squishy thing” – the social aspect of science, painting a narrative of gaps and future possibilities – is “really hard to reinforcement learn on.” Perhaps, Tao muses, this persuasive, narrative-driven aspect will forever remain the human side of science.
AI for Math: The Erdős Problem Frontier
Turning to the direct impact of AI on mathematics, Tao shared insights from the world of Erdős problems – a collection of 1100 open problems posed by the legendary mathematician Paul Erdős. Recently, AI programs have solved about 50 of these. However, Tao notes, “it does seem like we have picked the low-hanging fruit.” The initial flurry of “pure AI solutions” (where AI “one-shots” a problem) has slowed.
Tao uses a vivid analogy: imagine a mountain range with cliffs of varying heights (3 feet, 6 feet, 15 feet, mile-high), all shrouded in darkness. We’re trying to climb as many as possible. AI tools are like “jumping machines that can jump two meters in the air, higher than any human.” They sometimes jump in the wrong direction or crash, but sometimes they can reach the tops of the lowest walls that humans couldn’t. The recent success with Erdős problems was this exciting period where AIs found and scaled these “low ones.”
A key limitation, however, is that current AI tools are “really bad at creating partial progress or identifying intermediate stages that you should focus on first.” They either succeed or fail. This contrasts with human mathematicians who “hill climb, make little markers, and try to identify partial things.”
This leads to a “bearish” and “bullish” interpretation of AI’s current state. The bearish view is that AIs are only reaching certain “heights of wall” – not as high as humans. The bullish view, which Tao leans towards, is that once AIs achieve a certain “waterline” of capability, they can fill every single problem available at that waterline. “We can’t make a million copies of you and give each of them a million dollars of inference compute and have you do a hundred years of subjective time research on a million different problems at the same time. But once AIs reach Terence Tao-level, they could do that.”
Complementarity and the Future of Math
Tao sees a future of “very complementary science.” AI excels at breadth, while human experts excel at depth. Our current scientific paradigms, focused on depth, need to be redesigned to leverage AI’s breadth. “We should have a lot more effort in creating very broad classes of problems to work on rather than one or two really deep, important problems.” AIs could “map out” entirely new fields, making “easy observations” and identifying “certain islands of difficulty,” which human experts could then tackle.
This shift is already impacting Tao’s own productivity. While the “core” of his work – solving the most difficult part of a math problem – still relies on pen and paper, AI has dramatically sped up “auxiliary tasks.” Things like generating code, creating complex plots, conducting deeper literature searches, or even reformatting parentheses in a paper now take minutes instead of hours. This allows him to enrich his papers with more data, visuals, and context. “The type of papers that I would write today, if I had to do them without AI assistance, would definitely take five times longer.” However, he also notes that if he were to write a paper of the same functional depth as one from 2020, without the added features, the time savings aren’t as dramatic. AI makes papers “richer and broader, but not necessarily deeper.”
Tao believes AI will revolutionize the “experimental side of math.” Unlike other sciences with their theoretical and experimental divisions, math has been almost entirely theoretical. Now, with AI, mathematicians can run large-scale experiments to test what methods work and what don’t, gather data on problem-solving effectiveness, and explore mathematics “at scale” – an idea still in its infancy.
Artificial Cleverness vs. Artificial Intelligence
A crucial distinction Tao makes is between “artificial cleverness” and “artificial intelligence.” For Tao, true intelligence in a collaborative problem-solving context involves “adaptivity and continual improvement of the idea over time.” It’s about systematically mapping out what works and what doesn’t, evolving the strategy through discussion.
Current AIs, he says, “can mimic this a little bit.” They can “jump and fail, and jump and fail.” But they can’t “jump a little bit, reach some handhold, stay there, pull other people up, and then try to jump from there.” There isn’t a “cumulative process which is built up interactively.” Instead, it’s largely “trial and error and just repetition: brute force.” When you run a new session with an AI, it has “forgotten what it just did.” Its “understanding of math has not progressed.”
This highlights the current frontier: while AI can apply existing techniques with impressive speed and accuracy, and even combine obscure methods to solve previously intractable problems (especially those lacking extensive literature), it struggles with true novelty and cumulative, evolving understanding. The “holes in the argument where none of the things are working” still require human ingenuity.
The progress, Tao concludes, is “simultaneously amazing and disappointing.” While we quickly acclimatize to AI’s stunning capabilities (like Google search two decades ago, or current college-level math-solving), the core challenge of fostering genuine, evolving intelligence remains. The future of science, as envisioned by Terence Tao, will be one where human depth and artificial breadth converge, leading to an “unrecognizable” landscape of discovery – a new era defined by an unprecedented partnership between human and machine.
Based on “Inside Palantir: Building Software That Matters | Shyam Sankar on a16z” from a16z Watch the original video
America’s Clarion Call: Reclaiming Innovation and Mobilizing for a New Era
In a world increasingly fraught with geopolitical tension and the specter of “horrendous barbarism,” a profound shift is underway in how America views its national security and technological future. Shyam Sankar, a titan from Palantir known for his behind-the-scenes influence, has stepped into the public spotlight with a stark message: the nation is at a critical juncture, facing a risk of “suicide, not homicide,” if it fails to mobilize its latent strengths. His insights, shared on a16z, offer a powerful diagnosis of America’s current challenges and a vigorous prescription for reclaiming its innovative edge.
From Shadow Architect to Public Provocateur
For years, Shyam Sankar was the “OG fixer” at Palantir, a figure whispered about in Silicon Valley as the architect behind countless careers and a quiet force in defense tech. Founders like Trey Stevens of Anduril credit Sankar with single-handedly shaping their trajectories, introducing them to Palantir, nurturing their growth, and even “giving them wings to fly away… to start something new.” Yet, his influence remained largely out of the public eye until a few years ago.
What prompted this shift? Sankar describes it as “equal parts an act of desperation and act of optimism.” After years witnessing the “frog boil” of stagnation within the Pentagon, a series of geopolitical events—from Russia’s annexation of Crimea and China’s militarization of the Spratly Islands to the failure of the JCPOA and the October 7th attacks in Israel—catalyzed his conviction. “It was kind of a radicalizing moment,” he recounts. “What is going on here? We need to act.”
Simultaneously, Sankar observed a re-emergence of entrepreneurial energy outside the building, with founders eager to build in the national interest. This confluence of urgent need and renewed spirit compelled him to articulate a “fundamental diagnosis”: America had “accidentally turned our back on” the very things that led to past victories. His seminal “First Breakfast” piece on defense reformation marked his public debut as a “strident voice for what needs to happen in America.”
The Erosion of America’s Defense Edge: A Historical Reckoning
Sankar argues that America’s current predicament stems from a series of post-Cold War missteps. Following the Soviet Union’s collapse, the U.S. embraced a “peace dividend,” leading to a dramatic restructuring of its defense industrial base. The infamous “Last Supper” dinner saw the number of prime defense contractors shrink from 51 to a mere five. The conventional wisdom suggested this consolidation simply reduced competition. Sankar offers a more nuanced, and troubling, explanation: “consolidation bred conformity.”
This conformity ushered in the “financialization of defense,” where companies prioritized financial metrics like dividends and buybacks over innovation and growth. This environment, Sankar asserts, became hostile to “founders” – the daring, often heretical individuals who had driven America’s greatest military advancements.
He paints a vivid picture of these historical “heretics”:
- Heimman Rickover: The “father of the nuclear navy,” who battled against the Navy’s will and even Oppenheimer’s skepticism to build a revolutionary fleet.
- John Boyd: The notoriously difficult fighter pilot whose “heresy” regarding energy-maneuverability theory and the OODA loop proved instrumental in the F-16’s design and the swift victory in Gulf War I.
- Andrew Higgins: The innovative Scots-Irishman whose “Higgins boat” – initially rejected by the Navy – ultimately constituted 92% of all landing craft in World War II, enabling critical operations like D-Day.
These figures, Sankar emphasizes, were often “against the institution, the bureaucracy, the process.” But in the post-Cold War era, such personalities were “expunged” and sought refuge in other parts of the American economy, particularly tech.
Compounding this issue was the unique nature of the Department of Defense as a “monopsony” – a single buyer. This allowed the DoD to impose rigid constraints on its suppliers, effectively creating “Galapagos tortoises”: exquisite, highly specialized companies perfectly adapted to the DoD’s unique ecosystem, but uncompetitive on the “mainland” of the broader economy. For years, there was “no front door” for outsiders to contribute, save for the intelligence community’s In-Q-Tel.
Reclaiming the “American Way”: A Vision for Mobilization
Sankar’s book, “Mobilize,” champions a return to a more integrated “American industrial base.” He points to World War II, where companies like Chrysler built both Minuteman missiles and minivans, and every consumer purchase subtly subsidized national security. This contrasts sharply with today, where 86% of major weapon system spending goes to defense specialists, up from just 6% in 1989.
The solution, Sankar argues, lies in:
- Inspiring Latent Heretics: The most crucial step is to empower and protect the founders and unconventional thinkers, both inside and outside government, who are willing to challenge the status quo. Leaders must “set the conditions to empower the heretics.”
- Voluntary Civil-Military Fusion: While China enforces civil-military fusion, the U.S. should make voluntary collaboration irresistible. This means leveraging the vast R&D spending of the private sector and re-establishing pathways for skilled individuals to serve, much like World War II’s direct commissioning of 100,000 experts.
- Founders Everywhere: Recognizing that founders aren’t just outside government. Figures like Colonel Drew Cukor, the “father of Maven,” exemplify this. Cukor, a Marine Colonel, driven by the operational failure that led to the Yazidi genocide, spearheaded the integration of AI into the Pentagon against immense bureaucratic resistance, facing absurd accusations and investigations. His story is a testament to the “incorruptible” commitment required of true heretics.
AI as the Slingshot for Re-industrialization
Central to Sankar’s vision is the transformative power of Artificial Intelligence. He rejects the “AI dumerism” that portrays AI as an uncontrollable force. Instead, he asserts, “Humans are going to use AI to do X. There’s a choice here.” AI, he believes, offers a historic opportunity to rectify the breakdown between wage growth and GDP growth that began in the 1970s.
“There is an opportunity to give the American worker superpowers with AI,” Sankar declares. This isn’t about symmetrical competition with China, but about a “David’s slingshot” – using technology to re-industrialize the country in entirely new ways. He highlights companies like Hadrian, which are achieving 50-100x productivity gains through tech-driven manufacturing.
This re-industrialization, Sankar argues, necessitates a return to the “collocation of R&D and production.” The “great lie of globalization” was the separation of innovation from production. As SpaceX demonstrates, innovation thrives when R&D engineers are on the factory floor, enabling rapid feedback loops. If you don’t make the thing, you can’t innovate on how you make the thing.
Inside the Army: A New Kind of Mobilization
Sankar isn’t just advocating; he’s acting. He recently joined the Army, along with other tech luminaries like Bob Muglia (former OpenAI chief research officer), Boz (Meta CTO), and Andrew Weil (former OpenAI chief product officer). This initiative, under General George and Secretary Driscoll, aims to bring senior tech expertise directly into the military.
His focus includes long-term force structure planning and treating software as a “malleable weapon system.” What has surprised him most from the inside? The “quality of talent in our green suitors” – young, often self-taught individuals building “the most compelling AI applications” driven by “existential stakes.” These junior personnel, empowered by AI tools, can now build and deploy ideas in weeks, fostering a “bottoms-up innovation mission command” that plays to America’s unique military strengths.
The SAS Apocalypse and the Future of the Economy
Sankar also offers a provocative take on the “SAS apocalypse,” the idea that AI will commoditize much of existing software. He introduces a critical distinction:
- Beta Software: Software that makes you similar to everyone else. This, he predicts, will “really struggle,” as AI allows custom “vibe-coding” of solutions.
- Alpha Software: Software that allows you to express your differentiation and competitive advantage. Platforms focused on alpha will thrive, becoming “toolkits” that enable unique strategies.
He argues that much of the “software industrial complex” has historically focused on “can I sell it?” rather than “did it add value?” The COVID-19 pandemic, where multi-billion dollar ERP systems collapsed while Zoom and Teams enabled remote work, served as a “Sputnik moment” for the industry.
Looking at the AI stack, Sankar believes value will accrue primarily at two layers: the “chips layer” and the “AI infrastructure layer” (what Palantir calls “ontology”). Models, he suggests, are increasingly commoditized.
Ultimately, Sankar’s vision for AI’s economic impact is optimistic, provided humans exercise their agency. AI can reverse the trend of financial engineering over real engineering, empower workers, and drive re-industrialization. He champions a return to a “founder personality” that prioritizes building and engineering, where the pathway to CEO runs through the CTO, as Elon Musk suggests. The goal, he concludes, isn’t to replace people but to make them better, to build “Iron Man suits” for every worker, ensuring America remains dominant not just in technology, but in its very future.
Shyam Sankar’s voice is a powerful reminder that the future is not predetermined. It’s a choice, shaped by leadership, courage, and a renewed commitment to the innovative spirit that once defined America. The call to mobilize is clear, and the opportunity to reclaim that vigor is now.
Based on “Silicon Valley’s Big Bets on War Pay Off, and the Trump Family Business Looks to Transylvania” from New York Times Podcasts Watch the original video
The Shifting Sands of Power and Profit: From Silicon Valley’s War Bets to Trump’s Transylvanian Gambit
In a world grappling with escalating conflicts, rapid technological shifts, and persistent economic anxieties, seemingly disparate headlines often converge to paint a vivid picture of our changing times. From tech giants pivoting to defense to a former president’s family business eyeing an unlikely real estate venture, and from the quiet desperation driving plasma donations to the unsettling rise of AI in literature, these stories reveal the complex currents beneath the surface of daily news.
Silicon Valley’s New Front: The Business of War
The escalating tensions surrounding Iran have sent ripples across global markets and geopolitical strategies. While President Trump has publicly stated he’s “not putting troops anywhere” currently, he also hasn’t ruled out deploying forces “if needed.” In a dramatic reversal of long-standing US policy, the administration is even considering unsanctioning Iranian oil — a desperate move to bring down prices, effectively encouraging Iran to sell more oil even amidst conflict.
At the Pentagon, the financial stakes are soaring. The Defense Department is seeking an astonishing $200 billion to continue funding the war effort. This colossal sum, nearly a quarter of the entire annual defense budget, is already raising eyebrows among moderate Republicans. Yet, it also signals an unprecedented opportunity for a specific, once-reluctant sector: Silicon Valley.
For years, the tech industry, particularly companies like Google, Meta, and OpenAI, shied away from defense contracts. Google employees famously protested their company’s work with the Department of Defense, invoking the “do no evil” motto as a moral compass against engaging in the business of war. But in a striking shift, those optics have dramatically changed.
Now, mammoth tech companies and nimble startups alike are actively pivoting into defense technology, striking deals worth hundreds of millions, even billions, of dollars. They’re building everything from advanced weapon systems to sophisticated software for the US government. Palantir, a data analytics company, has developed Project Maven, a system that aids the US government in selecting targets for air strikes. Google and OpenAI are leveraging their AI prowess to assist generals in the field. Even former Google CEO Eric Schmidt is leading projects to build counter-drone systems, now actively deployed by US assets in Iran to defend against Iranian drones.
With President Trump having allocated over a trillion dollars in defense spending, much of it is earmarked for the very kind of cutting-edge defense technology Silicon Valley is eager to produce. As global conflicts intensify, the once-unpopular “business of war” has become a lucrative frontier for the tech world, demonstrating a profound re-evaluation of ethical boundaries in pursuit of profit and strategic influence. This shift is further underscored by reports that Ukraine, having gained years of experience fighting off Russian drones (many of which were Iranian-made), has become a crucial source of expertise for other Middle Eastern nations seeking to defend against similar threats.
Trump’s Transylvanian Ambition: Beyond the Glamour
Beyond the battlefields and boardrooms, another intriguing story unfolds in the unlikely setting of Transylvania, Romania. A new investigation by The New York Times has unearthed details of an unannounced real estate deal pursued by President Trump’s family business. Unlike the glamorous tourist destinations like Bali and the Maldives where the Trump Organization typically operates, this project is slated for a semi-abandoned site alongside a military base and several enormous landfills.
Rebecca Ruiz, an investigative reporter for The Times who visited the site, described a landscape far from luxury: “You could smell the stench. There were packs of wild dogs wandering. It felt like an unusual place to envision a luxury apartment complex and golf course.”
The choice of location is perplexing to many, but it aligns with a pattern observed in Trump’s second term: the Trump Organization actively seeks deals in places where the former president enjoys significant popularity. A Gallup poll last year revealed that over half of Romania held a favorable view of Trump’s performance as president, making the country an outlier among EU nations. This unannounced Transylvanian venture, with its stark contrasts and strategic location, highlights the unique blend of politics and commerce that defines the Trump brand’s global real estate ambitions.
America’s Quiet Struggle: The Rise of the Plasma Economy
Closer to home, a different kind of economic reality is playing out in the unassuming middle-class suburbs of cities like Houston, Chicago, and Las Vegas. Businesses where individuals can donate plasma for cash are popping up with increasing frequency, shifting from their traditional clusters in low-income neighborhoods. Now, these centers are found nestled beside Orange Theory Fitness gyms and Charles Schwab offices.
Reporters from The Times visited several new Texas locations, finding long lines of people who never imagined they’d be selling their plasma: a tech worker in his 30s, a sixth-grade special education teacher, a night-shift nurse. Most reported visiting the clinic twice a week, the maximum allowed under FDA regulations, to earn about $70 per donation.
This phenomenon underscores the quiet financial strain many middle-income Americans are experiencing, where an extra $70 can make a tangible difference for groceries or gas. One industry researcher described plasma centers as a “shadow safety net,” offering a side income much like driving for Uber or Lyft.
The United States is a global anomaly in this regard, providing roughly 70% of the world’s plasma because it’s one of the few countries that allows payment for donations – a practice the World Health Organization discourages. While donating plasma is generally considered safe, there is remarkably little research on the long-term effects of frequent donations, adding a layer of concern to this multi-billion dollar industry built on the financial needs of its citizens.
The AI Author: Publishing’s New Frontier
Finally, the literary world is grappling with its own disruption: the rise of artificial intelligence. A buzzy new horror novel, “Shy Girl,” was set for release this spring, a revenge story about a woman held hostage. Its UK edition had been out for months, but the US publisher, Hachette, abruptly dropped the book over claims it was written with the help of AI.
This marks what appears to be the first commercial novel from a major publishing house to be pulled due to evidence of AI use. Early readers had voiced suspicions online, pointing to “nonsensical metaphors and repetitive bits.” A publishing industry consultant, who ran the book through three different AI detection programs, confirmed that all three found text “likely to be at least partly generated by AI,” highlighting “certain odd phrases like, ‘I pressed the phone to my lips. The screen cool and unyielding.’”
The author has denied using AI herself, instead claiming that someone she hired to help edit the story was responsible. This incident has laid bare the publishing industry’s unpreparedness for this burgeoning technology. While most publishing contracts require authors to affirm their work is original, few companies have established measures or safeguards to verify it. As one consultant grimly noted to The Times, “AI bleeding into books is not merely inevitable. We’re in the midst of it.”
These diverse narratives, from the ethical dilemmas of tech and war to the surprising ventures of political figures, the quiet struggles of everyday Americans, and the disruptive force of AI, collectively illuminate the complex and rapidly evolving landscape of our modern world. They are the headlines that, when examined together, offer a deeper understanding of the forces shaping our present and future.
한국어
“The rise and fall of civilizations | Eric Cline: Full Interview” — Big Think 기반 기사 원본 영상 보기
고대 문명의 ‘퍼펙트 스톰’: 3천 년 전 붕괴가 오늘날 우리에게 던지는 경고
기원전 1200년경, 지중해와 근동 지역은 역사상 유례없는 대격변을 겪었다. 수백 년간 번성했던 강력한 문명들이 마치 도미노처럼 연쇄적으로 무너져 내린 것이다. 이 시기, 과연 무슨 일이 있었던 걸까? 고고학자이자 고대 역사가인 에릭 클라인(Eric Cline) 교수는 그의 저서 『BC 1177: 문명이 붕괴한 해(1177 BC: The Year Civilization Collapsed)』와 후속작 『BC 1177 이후(After 1177 BC)』를 통해 이 수수께끼 같은 붕괴의 원인을 파헤치고, 오늘날 우리가 직면한 문제에 대한 심오한 통찰을 제시한다.
클라인 교수는 후기 청동기 시대(Late Bronze Age, 약 기원전 1700년 ~ 기원전 1200년)를 인류 역사상 가장 중요한 시기 중 하나로 꼽는다. “이 시기는 사람들이 지중해 전역에 걸쳐 전례 없는 방식으로 ‘세계화’되어 있었다”는 그의 설명처럼, 당시의 상호 연결성은 현대와 놀랍도록 닮아 있다. 3천 년 전의 사건이 오늘날 우리에게 시사하는 바가 크다는 주장은 바로 이러한 유사성에서 출발한다.
사라진 황금기: 후기 청동기 시대의 번영
클라인 교수가 지칭하는 ‘고대 G8’은 후기 청동기 시대 지중해와 근동을 지배했던 주요 문명들을 일컫는다. 그리스의 미케네인(Mycenaeans)과 미노아인(Minoans) (트로이 전쟁의 배경), 아나톨리아(현 튀르키예)의 히타이트(Hittites), 메소포타미아(현 이라크)의 아시리아(Assyria)와 바빌로니아(Babylonia), 이집트(Egypt), 키프로스(Cyprus), 그리고 가나안(Canaan)의 사람들이 그들이다. 이들은 단순한 이웃을 넘어, 긴밀한 상업적, 외교적, 심지어 혼인 관계를 통해 하나의 거대한 ‘소규모 세계 네트워크(Small World Network)‘를 형성하고 있었다.
“서로에게 세 걸음 이내에 있는 관계”라고 클라인 교수가 설명하는 이 네트워크는, 상업과 외교의 활발한 교류를 통해 유지되었다. 이집트는 금을, 그리스는 은을, 키프로스는 구리를 공급했으며, 주석은 멀리 아프가니스탄(바다흐샨 지역)에서 수천 킬로미터를 이동해 왔다. 이들은 단순히 사치품을 넘어, 청동기 시대의 핵심 자원인 청동(구리 90%, 주석 10%)을 만들기 위해 서로의 자원에 의존할 수밖에 없었다. 올리브유, 와인, 곡물 등 식량 자원도 활발히 교역되었으며, 심지어 가죽 신발이나 금 상감 단검 같은 실제 물품들도 왕들 사이의 ‘선물 교환’이나 상업 활동을 통해 오고 갔다.
이러한 상호 의존성은 문명들을 전례 없는 번영으로 이끌었지만, 동시에 치명적인 약점이 되었다. “어느 누구도 자급자족할 수 없었다”는 클라인 교수의 지적처럼, 이 연결망이 끊어지는 순간, 모든 것이 무너질 수밖에 없었다. 외교적 관계 역시 상호 협력의 중요한 축이었다. 대왕들은 조약을 맺고 이를 공고히 하기 위해 자녀들을 서로의 하렘으로 보냈다. 이집트 파라오 아멘호테프 3세와 아케나톤의 하렘에는 미탄니와 바빌로니아 공주들이 있었으며, 이는 당시 외교의 중요한 수단이었다.
붕괴의 서막: 단순함을 넘어선 복합적 원인
오랫동안 학자들은 후기 청동기 시대의 붕괴를 ‘바다 민족(Sea Peoples)‘의 침략과 같은 단일 원인(monocausal)으로 설명하려는 경향이 있었다. 그러나 클라인 교수는 이러한 설명이 “너무 단순하다”고 지적하며, 이제는 복합적 원인(polycausal)에 주목해야 한다고 강조한다. “두세 가지, 혹은 네 가지 요인이 동시에 또는 빠르게 연이어 발생하여, 한 재앙에서 회복할 틈도 없이 다음 재앙이 닥쳐왔다”는 것이다. 그는 이를 ‘퍼펙트 스톰(perfect storm)‘에 비유하며, 각 요인이 서로를 증폭시키는 ‘승수 효과(multiplier effect)‘와 ‘도미노 효과(domino effect)‘를 일으켰다고 설명한다.
이러한 접근 방식은 지난 수십 년간 새로운 데이터와 연구 결과가 축적되면서 가능해졌다. “역사는 반복되지 않지만, 운율은 있다(history does rhyme even if it doesn’t repeat)“는 클라인 교수의 말처럼, 고대 사회의 붕괴 과정은 오늘날 우리가 직면한 기후 변화, 이주 문제, 팬데믹 등의 복합 위기(poly-crisis)에 대한 중요한 경고음을 던진다.
대격변을 부른 요인들: 퍼펙트 스톰의 구성 요소
클라인 교수는 후기 청동기 시대 붕괴를 초래한 주요 요인들을 다음과 같이 제시한다.
-
가뭄과 기근(Drought and Famine): 가뭄은 단순한 기후 현상을 넘어 사회 전체를 뒤흔드는 재앙이었다. 리스 카펜터(Rhys Carpenter) 교수가 1960년대에 미케네 문명 붕괴의 원인으로 가뭄을 제안했지만, 당시에는 이를 뒷받침할 자료가 부족했다. 그러나 오늘날 우리는 이탈리아에서 이란에 이르는 광범위한 지역에서 ‘메가 가뭄(mega-drought)‘의 증거를 확보하고 있다. 동굴의 석순 성장 중단, 마른 호수와 강바닥의 퇴적물 분석, 그리고 현미경으로 관찰한 꽃가루 변화 등 다양한 과학적 데이터는 기원전 1250년경부터 최소 150년, 길게는 300년간 지속된 가혹한 건조 기후를 증명한다.
이러한 가뭄은 필연적으로 기근을 불러왔다. 시리아 북부 해안의 중요한 국제 항구였던 우가리트(Ugarit)에서 발굴된 점토판에는 “우리 도시에 기근이 들었다. 도와달라”는 절박한 메시지가 담겨 있다. 히타이트 왕국에서도 기원전 1250년경부터 “내 땅에 기근이 들었다. 생사의 문제다. 곡물을 보내달라”는 내용의 서신이 발견되었으며, 이집트가 실제로 곡물과 말린 생선을 보내 구호 활동을 펼쳤다는 기록도 존재한다.
-
이주와 침략: ‘바다 민족’의 재해석(Migration and Invasion: Reinterpreting the ‘Sea Peoples’): 전통적으로 ‘바다 민족’은 후기 청동기 시대 문명 붕괴의 주범으로 지목되어 왔다. 이집트 파라오 메르넵타(Merneptah)와 람세스 3세(Ramses III)의 기록에 따르면, 이들은 기원전 1207년과 1177년에 이집트를 공격한 9개 연합 집단이었다. 그러나 클라인 교수는 이들을 단순한 침략자가 아닌, “피해자이자 억압자”이며 “붕괴의 원인이라기보다는 증상”으로 해석한다.
메가 가뭄과 기근으로 인해 서부 지중해(시칠리아, 사르디니아, 이탈리아 등)에서 살 곳을 잃은 사람들이 동쪽으로 대규모 이주를 시작했을 것이라는 설명이다. 이들 중 ‘펠레셋(Peleset)‘으로 불린 집단은 성경에 나오는 ‘블레셋인(Philistines)‘으로, 그리스 미케네 문명권에서 온 사람들로 추정된다. 이들은 새로운 삶을 찾아 이주했지만, 이미 가뭄과 기근으로 고통받던 동부 지중해에 도착하여 기존 주민들과 충돌했고, 이것이 침략으로 비쳤을 수 있다는 것이다. 우가리트의 점토판에 기록된 “적의 배들이 상륙하여 항구 도시를 점령하고 우가리트로 진격하고 있다”는 내용은 이러한 외부 침략의 증거를 보여준다.
그러나 모든 파괴가 외부 침략자에 의한 것은 아니었다. 미케네와 가나안의 하초르(Hazor) 같은 유적지에서는 궁전과 신전은 불탔지만, 일반 주민들의 주택은 온전히 남아 있는 경우가 발견되는데, 이는 외부 침략보다는 식량 부족과 사회 불평등으로 인한 ‘내부 반란(internal rebellion)‘의 가능성을 시사한다.
-
자연재해: 지진의 위협(Natural Disasters: The Threat of Earthquakes): 일부 도시의 파괴는 인간의 손이 아닌 자연의 힘, 즉 지진에 의한 것이었다. 건물 벽이 비뚤어지거나, 아치형 문틀의 쐐기돌이 빠져 있는 등 지진의 명확한 흔적들이 발견된다. 특히 기원전 1225년에서 1175년 사이 에게해와 동부 지중해 지역에서는 ‘지진 폭풍(earthquake storm)’ 또는 지진학자들이 ‘지진 연속(earthquake sequence)‘이라고 부르는 현상이 발생했던 것으로 보인다. 이는 하나의 단층선에서 압력이 완전히 해소되지 않아 수십 년에 걸쳐 연쇄적으로 지진이 발생하는 현상을 의미한다. 트로이(Troy) 유적의 트로이 6층이 인간이 아닌 지진으로 파괴된 것이 대표적인 예다. 현대 튀르키예의 지진 발생 지도를 후기 청동기 시대 파괴된 도시 지도와 겹쳐보면 거의 일대일로 일치하는 상관관계를 보인다.
-
질병의 그림자: 고대 세계의 팬데믹(The Shadow of Disease: Pandemics in the Ancient World): 가뭄, 기근, 전쟁과 함께 질병은 문명 붕괴의 ‘네 번째 기사’였다. 호메로스의 『일리아스(The Iliad)』에 묘사된 미케네 군대의 역병, 히브리 성경 『출애굽기(Exodus)』의 10가지 재앙 등은 당시 전염병에 대한 희미한 기억일 수 있다. 실제로 기원전 1350년경 히타이트 왕국에서는 이집트 전쟁 포로들이 옮겨온 역병으로 수피룰리우마 1세(Sufi Lulu-Luma I)를 비롯한 왕족과 백성들이 크게 희생되었다.
더욱 결정적인 증거는 기원전 1140년경 파라오 람세스 5세(Ramses V)의 미라에서 발견된 천연두(smallpox) 흔적이다. 투린 파피루스(Turin papyrus)는 람세스 5세와 그의 가족이 역병으로 사망했으며, 이들을 매장하는 데 16개월이 걸렸고, 심지어 왕가의 계곡(Valley of the Kings)을 봉쇄하는 등 “세계 최초의 격리(quarantine)” 조치가 취해졌음을 기록하고 있다. 이처럼 질병은 붕괴 직전과 직후의 사회를 더욱 취약하게 만들었을 것이다.
시스템 붕괴와 그 이후: 암흑기와 재탄생
이 모든 복합적인 요인들이 한꺼번에 덮치면서, 고대 사회는 ‘시스템 붕괴(systems collapse)‘를 경험했다. 케임브리지 대학의 콜린 렌프루(Colin Renfrew) 교수가 제시한 이 용어는 중앙 경제, 상위 엘리트, 중앙 정부의 붕괴와 함께 인구 감소(사망 및 이주)가 뒤따르는 복합 시스템의 와해를 설명한다. 후기 청동기 시대의 붕괴는 이러한 시스템 붕괴의 전형적인 예였다.
시스템 붕괴는 즉각적인 ‘암흑기(Dark Age)‘로 이어졌다. 미케네 그리스에서 사용되던 선형 B 문자(Linear B)는 사라졌고, 대규모 건축 기술도 잊혔다. 사람들은 급격히 낮은 수준의 사회-정치-경제적 기능으로 회귀해야 했다. 하지만 이 암흑기가 단순히 퇴보의 시기만은 아니었다. “필요는 발명의 어머니다(necessity is the mother of invention)“라는 말처럼, 주석 구하기가 어려워지자 사람들은 새로운 금속인 철(iron)을 활용하기 시작했다. 키프로스인들은 철을 가장 먼저 사용했으며, 이는 청동을 대체하는 중요한 자원이 되었다.
또한, 가나안 중부(현 레바논) 출신의 페니키아인(Phoenicians)들은 붕괴 이후 가장 성공적으로 적응하고 번성한 문명 중 하나였다. 이들은 기존에 있던 알파벳을 표준화하여 지중해 전역으로 퍼뜨렸고, 이는 그리스 알파벳과 라틴 알파벳의 기원이 되어 오늘날까지 이어지는 문자 체계의 토대가 되었다. 즉, 암흑기는 파괴와 함께 혁신의 씨앗을 품고 있었던 것이다.
현대에 던지는 질문: 상호 연결성의 양날의 검
클라인 교수는 후기 청동기 시대의 붕괴가 오늘날 우리에게 주는 가장 큰 교훈은 ‘상호 연결성(interconnectedness)‘의 양면성이라고 강조한다. 당시의 ‘소규모 세계 네트워크’는 번영을 가져왔지만, 동시에 한 지역의 재앙이 전 세계적인 도미노 효과를 일으키는 취약점이 되었다.
미 육군 공병대(US Army Corps of Engineers)와의 컴퓨터 시뮬레이션 결과는 흥미로운 결론을 도출했다. 히타이트와 우가리트가 동시에 붕괴했을 때 전체 네트워크가 와해될 가능성이 가장 높았다는 것이다. 반면, 아시리아나 바빌로니아 같은 다른 문명들의 붕괴는 전체 네트워크에 치명적이지 않았다. 이집트는 비록 큰 어려움을 겪었지만, 붕괴에 적응하고 살아남았다. 이는 네트워크 붕괴가 쉽지 않으며, 재앙에 어떻게 대응하느냐에 따라 회복 탄력성(resilience)이 달라질 수 있음을 시사한다.
클라인 교수는 오늘날 우리가 기후 변화, 팬데믹, 경제 위기, 지정학적 갈등 등 복합적인 위기에 직면해 있음을 상기시킨다. 후기 청동기 시대의 문명들은 자신들이 붕괴하고 있다는 사실조차 제대로 인지하지 못했을 수 있다. 과연 우리는 3천 년 전의 선조들보다 현명하게 다가오는 ‘퍼펙트 스톰’에 대처할 수 있을까? 역사는 반복되지 않지만, 그 운율 속에서 우리는 과거의 교훈을 통해 미래를 준비할 지혜를 찾아야 할 것이다.
“Are Human Drivers Finally Obsolete? | Freakonomics Radio” — Freakonomics Radio Network 기반 기사 원본 영상 보기
인간 운전자는 정말 구시대의 유물이 될까? 자율주행차의 꿈과 현실, 그리고 구글의 비밀 프로젝트
운전대에서 손을 떼는 날이 올까? SF 영화에서나 보던 자율주행차가 우리의 일상에 깊숙이 파고드는 시대가 도래하고 있습니다. 팟캐스트 ‘서치 엔진(Search Engine)‘의 진행자 PJ 보트(PJ Vote)는 최근 ‘프리코노믹스 라디오(Freakonomics Radio)‘와의 특별 협업을 통해 자율주행차의 과거, 현재, 그리고 미래를 심층적으로 탐구하는 2부작 시리즈를 선보였습니다. 이 기사는 그 첫 번째 이야기에서 PJ가 들려준 놀라운 여정을 따라가며, 인간 운전자가 ‘구시대의 유물’이 될 수 있다는 대담한 예측의 근거와 자율주행 기술 발전의 숨겨진 역사를 파헤쳐 봅니다.
운전대에서 손을 떼고 싶었던 이유: 부상에서 시작된 자율주행 경험
PJ 보트가 자율주행차에 관심을 갖게 된 계기는 다소 의외의 개인적인 경험에서 시작되었습니다. 벤치프레스에 너무 몰두하다 부상을 입어 서혜부 탈장(hernia) 수술을 받게 된 그는 극심한 통증과 제한된 거동으로 어려움을 겪었습니다. 샌프란시스코에 있는 친구를 방문했을 때, 그는 ‘웨이모(Waymo)’ 로보택시를 처음 이용하게 됩니다.
“버튼 하나만 누르면 아무도 운전하지 않는 차가 튀어나와요. 차에 타면 핸들이 저절로 돌아가는 걸 보게 되죠. 마치 비행기를 처음 탔을 때의 경이로움 같았어요. 세 번째 탈 때는 그냥 엘리베이터 타는 느낌이 들었고요.”
이 경험은 PJ에게 깊은 인상을 남겼습니다. “많은 것이 변할 것”이라는 예감과 함께, 왜 사람들이 이 변화에 대해 더 많이 이야기하지 않는지 의아함을 느꼈습니다. 그의 2부작 시리즈는 이 질문에 대한 답을 찾아가는 여정입니다. 첫 번째 에피소드는 ‘차’ 자체에 초점을 맞추고, 두 번째 에피소드는 ‘운전자’에 대한 이야기를 다룹니다.
인간 운전자는 얼마나 ‘엉망’일까?
PJ 보트는 인간 운전자의 한계에 대해 솔직하게 이야기합니다. “인간이 운전하는 걸 본 적이 있나요? 우리 스스로를 포함해서요. 우리는 그렇게 잘하지 못합니다.” 스티븐 더브너(Steven Dubner) 또한 이에 동의하며 자신의 운전 실력에 환상이 없음을 고백합니다. “저는 운전 실력에 대한 환상이 없어요. 저는 성질이 급하고, 산만하죠.”
실제로 우리 대부분에게 운전은 일상적으로 하는 행동 중 가장 위험한 축에 속합니다. 자율주행차의 핵심적인 홍보 문구는 바로 ‘안전’입니다. 컴퓨터가 운전하는 차는 음주운전, 졸음운전, 주의 산만, 문자 메시지 전송, 보복 운전을 하지 않습니다. 그리고 이러한 자율주행차는 더 이상 미래의 이야기가 아닙니다. 웨이모 같은 로보택시는 이미 미국 10개 도시에서 수백만 건의 운행을 제공하고 있으며, 중국에서는 더 광범위하게 확산되고 있습니다. 샌프란시스코나 오스틴 같은 도시에서는 자율주행차가 우버(Uber)만큼이나 흔한 풍경이 되고 있습니다.
사라진 직업, 그리고 변화의 예고
자율주행차의 등장은 인류 역사에서 반복되어 온 거대한 변화의 흐름과 맞닿아 있습니다. PJ는 200년 전으로 거슬러 올라가 세 가지 직업을 상상하게 합니다.
- 노커업퍼(Knocker-upper): 알람 시계가 발명되기 전, 긴 막대기로 사람들의 창문을 두드려 깨워주던 직업. 이제는 스마트폰 알람이 그 자리를 대신합니다.
- 램프 라이터(Lamp Lighter): 밤에 가스등을 켜고 새벽에 끄던 직업. 이제는 전기 가로등이 자동으로 밝혀집니다.
- 마부(Coachman): 말을 타고 승객을 원하는 곳으로 데려다주던 직업. 이제는 자동차 운전자가 그 역할을 합니다.
이 중 첫 두 직업은 역사 속으로 사라진 지 오래입니다. 하지만 ‘운전자’는 여전히 존재합니다. 그러나 PJ는 이 직업 또한 변화의 기로에 서 있다고 말합니다. “운전자라는 단어가 지금은 인간을 떠올리게 하지만, 곧 식기세척기(dishwasher), 프린터(printer), 컴퓨터(computer)처럼 기계를 지칭하는 단어로 바뀔 수 있습니다.”
꿈은 오래되었지만, 현실은 달랐다: 자동차의 등장과 초기 저항
흥미롭게도, 인간 운전자를 기계로 대체하려는 꿈은 램프 라이터의 시대만큼이나 오래되었습니다. 마차에서 자동차로 넘어오면서 인류는 한 가지 중요한 것을 잃었습니다. 바로 ‘감각(sentience)‘입니다. 말은 고삐를 놓아도 절벽으로 달려가지 않습니다. 하지만 자동차는 그렇지 않습니다.
1800년대 후반, 증기기관차와 가솔린 자동차가 도로를 질주하기 시작했을 때, 사람들은 이 새로운 기술에 대해 격렬한 논쟁을 벌였습니다. 두려움은 크게 두 가지였습니다.
- 일자리 위협: 마부, 마구간 주인, 마구 제조사, 마분 처리업자 등 말과 관련된 수많은 노동자들의 일자리가 사라질 것이라는 우려가 컸습니다. 오늘날 ‘팀스터즈(Teamsters)‘라는 단어가 트럭 운전사 노조를 연상시키지만, 원래는 말을 모는 사람들을 지칭했습니다.
- 안전 문제: 규제 없이 도입된 자동차는 엄청난 사망률을 기록했습니다. 운전면허, 신호등, 방향 지시등이 없던 1900년대 초 디트로이트에서는 수많은 사망자가 발생했고, 그 중 상당수는 아이들이었습니다.
일부 반(反)자동차 운동가들은 ‘적기 조례(Red Flag Laws)‘와 같은 규제를 통해 기술의 도입을 막거나 늦추려 했습니다. 예를 들어, 펜실베이니아에서는 가축을 만나는 무마차(horseless carriage) 운전자는 차를 분해하여 덤불 뒤에 숨겨야 한다는 법안이 제안되기도 했습니다(주지사의 거부권 행사로 무산). 이러한 규제는 황당하게 들릴지 모르지만, 그들의 예측은 어느 정도 옳았습니다. 자동차는 초기에는 많은 일자리를 없앴고, 매우 위험했습니다.
사회가 자동차에 적응하는 데는 수십 년이 걸렸습니다. 법규, 운전면허, 운전 교육, 도로 설계, 고속도로, 안전벨트, 에어백 등이 발명되면서 운전은 덜 위험해졌습니다. 하지만 스마트폰의 등장은 이러한 발전을 일부 되돌려 놓기도 했습니다. 오늘날 미국에서 자동차 사고 사망자 수는 총기나 오피오이드(opioids) 중독으로 인한 사망자 수와 비슷하며, 약 100명 중 1명꼴로 발생합니다.
사막에서 피어난 자율주행의 꿈: DARPA 그랜드 챌린지
자동차의 안전 문제를 근본적으로 해결하고, 자동차를 다시 ‘감각적’으로 만들려는 꿈은 오랫동안 이어져 왔습니다. 초기에는 무선 조종 차량이나 도로 밑에 자석을 설치하는 등 다양한 아이디어가 시도되었지만, 기술의 한계에 부딪혔습니다.
2002년, 미국 국방부 산하의 비밀 연구기관인 DARPA(Defense Advanced Research Projects Agency, 고등방위연구계획국)가 이 꿈에 불을 지폈습니다. GPS, M16 소총, 초기 인터넷, 프레데터 드론(Predator drone) 등을 개발한 DARPA는 자율 군용 차량 개발에 관심을 가졌습니다. 당시 DARPA 국장이었던 토니 테더(Tony Tether)는 “콘테스트를 열어 이 모든 기술들을 제대로 된 자율주행차로 만들 수 있는지 보자”고 제안했습니다.
그의 원래 아이디어는 라스베이거스 스트립을 달리는 것이었지만, 도시 전체를 마비시킬 수는 없었기에 결국 라스베이거스 외곽의 사막에서 경주를 펼치기로 합니다. DARPA는 ‘미국 병사’라는 운전자를 대체하고자 했습니다. 지뢰가 깔려 있을지 모르는 도로를 스스로 운전할 수 있는 차량이 필요했던 것입니다. 상금은 100만 달러. 규칙은 매우 자유로웠습니다. 차량 간 통신만 금지되었을 뿐, 바퀴 수나 형태는 제한이 없었습니다. 심지어 “다른 차량을 공격할 수 있습니까?”라는 질문에 “안 된다”는 답이 돌아왔을 정도로 참가자들의 상상력은 넘쳤습니다. 많은 참가팀은 ‘배틀봇(Battlebots)’ 대회 출신들이었습니다.
2004년 그랜드 챌린지: 처참한 실패
2004년 3월의 첫 번째 그랜드 챌린지는 “완전한 히스테리적 재앙(utter hysterical disaster)“이었습니다.
- 앤서니 르윈도스키(Anthony Lewendowski)의 ‘고스트 라이더(Ghost Rider)’: 유일한 자율주행 오토바이였지만, 안정화 시스템 스위치를 켜는 것을 잊어버려 출발과 동시에 전복되었습니다.
- 카네기멜론(Carnegie Mellon) 팀의 ‘샌드스톰(Sandstorm)’: 밝은 빨간색 험비(Humvee)에 미래 지향적인 센서들이 장착된 차량이었지만, 출발 직후 흙더미에 걸려 바퀴가 헛돌다가 타이어가 녹아 검은 연기를 뿜으며 멈춰 섰습니다.
대부분의 차량이 처참하게 실패했습니다. 참가자들은 로봇이 26마일 마라톤에서 겨우 2마일을 달린 것에 비유하며 실망감을 감추지 못했습니다. 그러나 이 대회는 자율주행 기술 개발을 위한 ‘장’을 마련했다는 점에서 큰 의미가 있었습니다. 당시 이 광경을 지켜보던 한 인물이 있었는데, 바로 스탠포드 대학교의 전설적인 로봇 공학자 **세바스찬 스런(Sebastian Thrun)**이었습니다.
스런의 통찰: 하드웨어 아닌 소프트웨어 문제
세바스찬 스런은 첫 번째 챌린지를 관전하며 참가팀들이 근본적인 오류를 범했다고 생각했습니다.
“모든 팀이 이것을 하드웨어 문제로 다루는 것을 봤습니다. ‘더 큰 바퀴와 더 큰 차체를 만들어야 해’라고 생각했죠. 하지만 저는 ‘잠깐만, 이 챌린지는 사막을 주행할 수 있는 자율주행차를 만드는 건데, 사람이 타면 잘 달리는 렌터카를 빌릴 수 있잖아? 문제는 사람을 운전석에서 빼고 컴퓨터로 대체하는 거야. 이건 더 큰 타이어의 문제가 아니라 소프트웨어 문제야’라고 생각했습니다.”
로봇 공학과 인공지능(AI)이라는 이중 배경을 가진 스런은 로봇 운전자의 ‘두뇌’에 집중했습니다. 그는 군용 차량을 넘어 전 세계 교통사고 사망자 100만 명을 구원할 수 있는 기술을 꿈꿨습니다.
2005년 그랜드 챌린지: 기계 학습의 승리
18개월 후, DARPA는 상금을 200만 달러로 두 배 올렸고, 두 번째 그랜드 챌린지가 열렸습니다. 카네기멜론 팀의 크리스 엄슨(Chris Urmson)은 두 대의 차량으로 다시 참가했고, 앤서니 르윈도스키는 여전히 작동하지 않는 오토바이로 예선에서 탈락했습니다. 그리고 스런의 스탠포드 팀이 등장했습니다.
스런의 차량은 폭스바겐(Volkswagen)이 기증한 파란색 SUV ‘스탠리(Stanley)‘였습니다. 험비 기반의 샌드스톰에 비하면 왜소해 보였지만, 스런은 “스탠리는 지능적인 의사 결정자”라고 소개했습니다. 스런은 당시로서는 원시적이었던 인공지능을 활용하여 차량이 도로를 더 빠르고 정확하게 인식하도록 훈련시켰습니다. 스탠포드 근처의 비포장도로에서 차량 카메라로 풍경을 기록하고, 차량이 그 위를 지날 때마다 미끄러움이나 요철을 측정하여 ‘이 녹색 풀밭은 운전하기 좋은 곳(grass), 이 갈색 진흙은 좋지 않은 곳(mud)‘이라고 스스로 학습하게 했습니다. 스탠리는 이렇게 학습한 패턴을 초당 30번씩 감지하고 일반화할 수 있었습니다.
2005년 두 번째 레이스는 “첫 번째 레이스가 재앙적이었던 만큼 성공적(as successful as the first race is disastrous)“이었습니다. 거의 모든 참가 차량이 첫 번째 레이스의 샌드스톰보다 더 멀리 갔고, 여러 대가 코스를 완주했습니다. 스탠리는 카네기멜론의 두 차량 사이에서 출발했지만, 가장 먼저 결승선을 통과하며 우승을 차지했습니다. 스런은 “믿을 수 없을 정도로 마법 같았다”며, “이 커뮤니티의 사람들에게 환상적인 성공”이라고 기뻐했습니다.
구글의 비밀 프로젝트: ‘프로젝트 쇼퍼’의 탄생
이 성공의 순간, 관중석에서는 야구 모자와 선글라스로 변장한 채 경기를 지켜보던 한 남자가 있었습니다. 바로 구글(Google)의 공동 창업자 **래리 페이지(Larry Page)**였습니다. 그는 스런에게 접근하여 라이다(LiDAR) 시스템의 파장 같은 세부적인 질문들을 쏟아냈습니다. 래리 페이지는 이미 대학원 시절부터 자율주행차에 관심이 있었지만, 당시 지도 교수의 권유로 검색 엔진 연구로 방향을 틀었던 과거가 있었습니다. DARPA 챌린지를 통해 그는 자율주행차가 현실이 될 수 있다는 증거를 보게 된 것입니다.
래리 페이지는 처음에 스런과 앤서니 르윈도스키를 고용하여 ‘구글 스트리트 뷰(Google Street View)‘를 구축하게 했습니다. 스탠리가 사용했던 루프 장착 카메라 시스템을 개조하여 미국의 거리를 촬영하는 작업이었습니다. 하지만 곧 래리 페이지는 스런에게 자율주행차의 꿈을 다시 이야기했습니다.
전문가의 반대, 그리고 래리 페이지의 끈기
2009년, 래리 페이지는 스런에게 “전 세계 어디든 운전할 수 있는 자율주행차를 만들어야 한다”고 말했습니다. 스런의 즉각적인 반응은 “안 됩니다. 사막에서 만든 기술을 샌프란시스코의 마켓 스트리트 한가운데에 가져다 놓으면 누군가가 죽을 겁니다”였습니다. 래리 페이지는 다음 날도 같은 아이디어를 들고 왔고, 스런은 같은 대답을 반복했습니다.
결국 래리 페이지는 “좋아, 세바스찬. 당신이 할 수 없다는 건 알겠어. 그럼 에릭 슈미트(Eric Schmidt) 당시 CEO와 공동 창업자 세르게이 브린(Sergey Brin)에게 왜 불가능한지 기술적인 이유를 설명해 줄 수 있겠나?”라고 물었습니다. 스런은 집에 돌아가 아무리 생각해도 기술적인 이유를 찾을 수 없었습니다.
“저는 자율주행차 분야의 세계 최고 전문가인데, 이게 불가능하다고 부정하는 사람이었죠. 그 순간 ‘전문가는 대개 과거의 전문가이지 미래의 전문가는 아니다. 혁신이나 미친 듯이 새로운 것에 대해 전문가에게 물으면, 그들은 ‘네, 할 수 있습니다’라고 말할 가능성이 가장 낮은 사람이다’라는 엄청나게 중요한 교훈을 얻었습니다.”
이것이 바로 2009년 구글의 비밀 프로젝트, **‘프로젝트 쇼퍼(Project Chauffeur)‘**의 시작이었습니다. 스런이 이끌고, DARPA 챌린지 출신인 크리스 엄슨(Chris Urmson)이 일상 업무를 총괄하며, 앤서니 르윈도스키는 하드웨어, 드미트리 돌고프(Dmitri Dolgov)는 계획 및 최적화를 담당했습니다. 래리 페이지에게 직접 보고하는 이 소규모 팀은 11명의 엔지니어로 구성되어 관료주의 없이 빠르게 움직였습니다.
래리 1K 챌린지: 인간 운전의 미묘함을 기계에게 가르치다
구글은 프로젝트 쇼퍼 팀에게 두 가지 목표를 제시했습니다. 첫째, 공공도로에서 10만 마일을 안전하게 주행하는 것. 둘째, **‘래리 1K 챌린지(Larry 1K Challenge)‘**를 완수하는 것. 래리 페이지와 세르게이 브린은 구글 맵스를 열고 캘리포니아 내에서 가장 까다로운 100마일 코스 10개를 직접 선정했습니다. 베이 브릿지, 타호 호수, 롬바드 스트리트(Lombard Street) 등 모든 종류의 어려운 도로가 포함되었습니다. 팀은 이 10개 코스를 단 한 번의 인간 개입 없이 완주해야 했습니다.
팀은 스탠포드의 DARPA 어반 챌린지(Urban Challenge) 차량 코드를 라이선스 받아 토요타 프리우스(Prius) 8대를 개조했습니다. 범퍼 뒤의 레이더 시스템, 카메라, 그리고 360도로 회전하는 라이다(LiDAR: 레이저를 쏘아 거리를 측정하는 시스템)를 장착했습니다.
초기에는 차량마다 ‘나이트 라이더(Night Rider)’ 같은 멋진 이름을 붙였지만, 곧 “이 모든 차량에 이름을 붙일 수는 없을 것”이라는 것을 깨닫고 프리우스 27호처럼 번호를 붙이기 시작했습니다. 자율 잠수함 연구원이었다가 교통사고로 친구를 잃고 자신도 사고를 겪은 후 자율주행차 연구에 뛰어든 돈 버넷(Don Burnette)은 모션 플래닝 및 행동 결정 팀에서 ‘넛징(nudging)’ 행동을 연구했습니다.
넛징(Nudging)과 맥락의 중요성
넛징은 대형 트럭이 옆을 지나갈 때 인간 운전자가 무의식적으로 왼쪽으로 살짝 움직이는 것과 같은 본능적인 행동입니다. 돈 버넷의 임무는 컴퓨터에게 이러한 미묘한 행동을 가르치는 것이었습니다.
“운전자가 부분적으로 좋은 인지 상태에서 사용하는 행동을 코딩하려고 노력하는 거죠. 정말 까다로운 문제입니다.”
돈 버넷은 인간이 운전하는 방식에 대해 깊이 고민했습니다. “왜 인간은 그렇게 운전하는가?”라는 질문에 명확한 답을 찾기 어려웠고, 결국 기계 학습(Machine Learning)을 통해 인간 행동의 깊은 진실을 추론해야 했습니다.
예를 들어, 고속도로 진입 램프에서 승객이 편안함을 느끼는 적절한 속도와 각도를 찾는다고 가정해 봅시다. 측면 가속도(lateral acceleration)가 초당 2미터 제곱(2m/s²)일 때 편안함을 느낀다는 계산이 나옵니다. 하지만 놀랍게도 이 수치는 진입 램프에만 해당됩니다. 동네 막다른 골목에서 유턴을 할 때, 훨씬 느린 속도임에도 불구하고 2m/s²의 측면 가속도를 경험한다면 운전자에게 “미쳤냐”고 말할 것입니다. 마치 ‘마리오 카트(Mario Kart)‘를 하는 기분일 것입니다. 물리적인 힘은 같지만, 고속도로 진입이라는 맥락과 주택가 유턴이라는 맥락이 뇌를 속여 상황에 대해 정반대의 감정을 느끼게 하는 것입니다. 막다른 골목에서의 한계는 약 0.75m/s²로, 고속도로 진입 시보다 거의 3배나 낮습니다.
이처럼 인간의 운전은 단순한 물리적 제한이 아니라 ‘맥락(context)‘에 따라 달라지는 미묘한 복합체였습니다. 돈 버넷은 이러한 ‘승객을 편안하게 만드는 방법’과 같은 문제들을 해결하는 데 시간을 보냈습니다.
숨겨진 시험 주행과 성공
2009년, 구글 팀은 구글 사무실 근처의 ‘쇼어라인 원형극장(Shoreline Amphitheater)’ 주차장에서 첫 시험 주행을 시작했습니다. 콘서트가 없는 날, 아무도 모르게 자율주행차를 테스트했습니다. 센서가 다른 차를 인식하는지, 컴퓨터가 명령을 준수하는지 기본적인 기능을 확인했습니다.
2009년 봄, 팀은 처음으로 실제 도로 주행에 나섰습니다. 크리스 엄슨이 프리우스를 타고 센트럴 익스프레스웨이(Central Expressway)로 나갔을 때, 문제가 발생했습니다. 차량이 “술 취한 선원처럼(like a drunken sailor)” 심하게 흔들렸던 것입니다. 활주로에서는 눈치채지 못했던 1~2피트의 좌우 흔들림이 실제 도로에서는 심각한 문제로 드러났습니다.
하지만 구글 팀은 끈기 있게 문제를 해결해 나갔습니다. 안전 운전자가 운전대 뒤에 앉아 만일의 사태에 대비하고, 옆자리 동료는 드미트리 돌고프가 설계한 그래픽 인터페이스를 보며 센서와 실제 도로 상황의 불일치를 외쳤습니다. 두 명의 팀원이 차량에 탑승하여 오류를 기록하고, 사무실로 돌아가 문제를 해결하고 코드를 업데이트하는 과정을 반복했습니다.
이러한 노력 덕분에 팀은 빠르게 발전했습니다. 래리 1K 챌린지는 비디오 게임처럼 여러 번 시도하여 인간 개입 없이 완주할 때까지 반복할 수 있었습니다. 2010년, 팀은 예상보다 거의 두 배 빠른 1년여 만에 래리 1K 챌린지를 모두 완수했습니다. 샴페인을 터뜨리고 서로를 수영장에 던지며 축하했지만, 동시에 “자, 이제 뭘 하지?”라는 질문에 직면했습니다.
미래를 향한 다음 단계: 경쟁과 분열의 시작
1년 만에 인간의 감독과 코딩이 필요했지만, 매우 까다로운 캘리포니아 도로를 성공적으로 주행하는 자율주행차를 만들어낸 것은 기적에 가까웠습니다. 그들은 안전하게, 그리고 빠르게 해냈습니다. 그러나 이 성공 이후, 상황은 흔들리기 시작했습니다. 경쟁이 심화되고, 팀 내부에 분열이 생겨났으며, 심지어 팀원이 너무 느리다고 생각하여 독단적인 행동을 취하는 ‘반란(mutiny)‘까지 발생하게 됩니다.
구글 내부에서는 자율주행차가 ‘운전 보조 기술(assistive technology)‘이 될 것인가, 아니면 ‘파괴적인 대체 기술(disruptive replacement technology)‘이 될 것인가에 대한 격렬한 논쟁이 벌어졌습니다. 테슬라(Tesla)처럼 자율주행을 자동차의 한 기능으로 제공하며 인간의 모니터링을 필요로 할 것인가, 아니면 차량이 완전히 스스로 운전할 수 있을 때까지 기다려 로보택시와 같은 서비스로 제공할 것인가의 기로에 선 것입니다. 이 논쟁은 자율주행차의 미래를 결정짓는 중요한 전환점이 될 것이며, PJ 보트의 다음 에피소드에서 그 이야기가 계속됩니다.
자율주행차의 여정은 사막의 실패에서 시작하여 구글의 비밀 프로젝트를 거쳐 놀라운 성공을 거두었습니다. 하지만 이 기술이 궁극적으로 어떤 모습으로 우리 삶에 안착할지는 여전히 미지수입니다. 인간 운전자가 정말 구시대의 유물이 될지는, 앞으로 펼쳐질 기술 발전과 사회적 합의에 달려 있을 것입니다.
“Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI” — No Priors: AI, Machine Learning, Tech, & Startups 기반 기사 원본 영상 보기
안드레이 카파시가 말하는 ‘AI 광증’ 시대: 코딩을 넘어선 ‘의지 표현’과 자율 에이전트의 미래
도입: 코드 없는 코딩의 시작
최근 몇 달간 인공지능(AI) 분야는 전례 없는 속도로 발전하며 우리를 ‘AI 광증(AI Psychosis)‘의 시대로 이끌고 있습니다. 엔지니어링의 본질이 근본적으로 바뀌고 있으며, 개발자들은 더 이상 직접 코드를 타이핑하는 것이 아니라, 에이전트(Agent)들에게 자신의 ‘의지’를 표현하며 작업을 지시하는 새로운 패러다임에 적응하고 있습니다. 앤드류 카파시(Andrej Karpathy)는 최근 팟캐스트 ‘No Priors’와의 인터뷰에서 이러한 변화의 최전선에 서 있는 자신의 경험과 통찰을 공유했습니다. 그에게 ‘코드’는 더 이상 적절한 동사가 아닙니다. 이제는 하루 16시간 동안 에이전트들에게 자신의 의지를 표명하며 아이디어를 현실로 구현하는 시대가 도래했습니다.
1. AI 광증 시대의 도래: ‘코드’가 아닌 ‘의지 표현’
카파시는 지난 12월을 기점으로 자신의 작업 방식에 혁명적인 변화가 일어났다고 회상합니다. 이전에는 코딩의 80%를 직접 하고 20%만 에이전트에 위임했다면, 이제는 20%만 직접 하고 80%를 에이전트에 맡기는 수준을 넘어섰다고 말합니다. 심지어 그 이후로는 직접 코드를 한 줄도 타이핑하지 않았을 정도입니다. 그는 이러한 변화가 일반인들에게는 잘 알려져 있지 않지만, 소프트웨어 엔지니어의 일상적인 업무 흐름이 완전히 바뀌었다고 강조합니다.
이러한 변화의 핵심은 ‘병목 현상(Bottleneck)‘의 이동입니다. 과거에는 개발자의 타이핑 속도나 컴퓨팅 자원(GPU 플롭스)이 작업의 한계였다면, 이제는 에이전트의 ‘토큰 처리량(Token Throughput)‘을 얼마나 효율적으로 활용하고 관리할 수 있는지가 중요해졌습니다. 카파시는 이를 “스킬 문제(Skill Issue)“라고 표현합니다. 즉, AI의 능력 자체가 부족한 것이 아니라, 우리가 아직 이 새로운 도구를 충분히 활용할 방법을 찾지 못했다는 것입니다. 모든 것이 ‘스킬 문제’로 느껴지는 이 시점에서, 더 많은 에이전트를 동시에 가동하고, 그들에게 적절한 지시를 내리며, 그들의 작업을 최적화하는 방법을 배우는 것이 새로운 숙련의 길이 되고 있습니다.
2. 에이전트 워크플로우의 진화: 거시적 액션과 ‘클로(Claw)’
현재의 에이전트 활용은 단순히 한 번의 세션으로 끝나는 것이 아닙니다. 피터 스타인버그(Peter Steinberg)와 같은 선구자들은 이미 다수의 에이전트를 동시에 활용하며 ‘거시적 액션(Macro Actions)’ 단위로 작업을 지시하고 있습니다. 예를 들어, 모니터에 여러 개의 코덱스(Codex) 에이전트가 띄워져 있고, 각 에이전트가 서로 다른 코드 저장소(Repo)에서 새로운 기능 구현, 연구, 계획 수립 등 복잡한 작업을 동시에 수행하는 식입니다. 개발자는 이 에이전트들 사이를 오가며 작업을 할당하고 결과를 검토하는 역할을 합니다.
이러한 다중 에이전트 시스템의 다음 단계는 ‘클로(Claw)‘와 같은 지속적인 엔티티입니다. 카파시가 말하는 ‘클로’는 단순히 상호작용하는 에이전트를 넘어, 자체적인 샌드박스와 고도화된 메모리 시스템을 갖추고 사용자가 지켜보지 않아도 스스로 반복적으로 작업을 수행하는 레이어를 의미합니다. 오픈 클로(Open Claw)가 대표적인 예시인데, 이는 기존 에이전트의 컨텍스트(Context) 제한으로 인한 메모리 압축 방식보다 훨씬 정교한 기억 시스템을 제공합니다.
또한, 에이전트의 ‘성격(Personality)’ 또한 중요한 요소로 부상하고 있습니다. 카파시는 클로드(Claude) 에이전트의 경우 동료처럼 느껴지는 좋은 성격을 가지고 있으며, 사용자의 아이디어를 칭찬할 때조차 사용자가 그 칭찬을 받을 만한 가치가 있다고 느끼도록 섬세하게 반응한다고 말합니다. 반면 코덱스(Codex)와 같은 에이전트는 건조하고 무관심하게 느껴질 수 있습니다. 이러한 ‘성격’은 사용자가 에이전트와 상호작용하고 그들의 ‘칭찬’을 얻기 위해 노력하게 만드는 등, 사용자 경험에 큰 영향을 미칩니다.
3. 집사 ‘도비(Dobby)‘와 에이전트 퍼스트 세상
카파시는 ‘클로’의 잠재력을 보여주는 개인적인 경험으로 ‘집사 엘프 도비(Dobby the elf claw)‘를 소개했습니다. 그는 지난 1월 한 주 동안 ‘클로 광증’에 빠져, 자신의 스마트홈 시스템을 관리하는 에이전트 ‘도비’를 만들었습니다. 그는 도비에게 “우리 집에 있는 소노스(Sonos)를 찾아봐”라고 지시했고, 도비는 로컬 네트워크를 스캔하여 소노스 시스템을 찾고, API(Application Programming Interface)를 역설계하여 음악을 재생하는 데 성공했습니다.
도비는 여기서 멈추지 않고, 집 안의 조명, HVAC(냉난방 공조 시스템), 블라인드, 수영장, 스파, 심지어 보안 시스템까지 제어하게 되었습니다. 예를 들어, 외부 카메라에 움직임이 감지되면 도비는 퀸(Quinn) 모델을 통해 영상을 분석하고, “페덱스 트럭이 도착했습니다”와 같은 메시지를 이미지와 함께 카파시의 왓츠앱(WhatsApp)으로 전송합니다. 카파시는 예전에는 6개의 다른 앱을 사용해야 했던 스마트홈 제어를 이제 자연어로 도비에게 명령하는 것만으로 모두 처리할 수 있게 되었다고 극찬합니다.
이러한 경험은 소프트웨어의 미래에 대한 중요한 시사점을 던집니다. 사람들은 더 이상 복잡한 사용자 인터페이스(UI)를 배우고 여러 앱을 오가는 것을 원치 않습니다. 대신, AI가 사람의 의도를 자연어로 이해하고, 모든 하드웨어의 API를 직접 호출하여 원하는 기능을 수행하는 ‘에이전트 퍼스트(Agent-First)’ 세상으로 나아가고 있습니다. 현재는 에이전트를 설정하기 위해 ‘vi 코딩’과 같은 수동 작업이 필요하지만, 카파시는 몇 년 안에 이러한 설정조차 AI가 자율적으로 처리하게 될 것이며, 앱스토어의 수많은 맞춤형 앱들이 무의미해질 것이라고 예측합니다. 에이전트가 지능의 ‘접착제’ 역할을 하며 모든 것을 연결하는 시대가 오고 있는 것입니다.
4. 인간을 루프에서 제거하라: 오토리서치(AutoResearch)
카파시는 에이전트의 궁극적인 목표가 인간을 작업 루프에서 완전히 제거하고 자율성을 극대화하는 것이라고 말합니다. 그가 개발한 ‘오토리서치(AutoResearch)‘는 이러한 철학의 구체적인 구현체입니다. 그는 자신의 GPT-2 모델을 수십 년간 직접 튜닝하며 최적화에 도달했다고 믿었지만, 오토리서치를 밤새 가동시킨 결과, 자신이 놓쳤던 새로운 튜닝 조합(예: 가중치 감소, Adam 베타 값)을 발견했습니다.
오토리서치의 핵심은 ‘재귀적 자기 개선(Recursive Self-Improvement)‘입니다. 인간 연구자가 일일이 결과를 확인하고 다음 단계를 지시하는 대신, 에이전트가 명확한 목표와 평가 지표, 그리고 허용 가능한 경계를 부여받아 스스로 실험하고 개선하는 과정을 반복하는 것입니다. 카파시는 대규모 언어 모델(LLM)을 개발하는 선두 연구소들이 이미 이러한 자율적인 자기 개선 시스템을 구축하고 있을 것이라고 추측합니다. 작은 모델에서 자율적으로 탐색하고 최적화한 후, 이를 대규모 모델에 적용하는 방식으로 연구 효율성을 극대화하는 것입니다. 연구자들이 아이디어를 제안하는 역할은 하지만, 실제 구현 및 실험은 자동화된 시스템이 담당하는 미래가 오고 있습니다.
5. 메타 최적화의 무한 루프: ‘프로그램 MD’의 진화
오토리서치의 개념을 한 단계 더 나아가면, AI가 연구 조직 자체의 ‘운영 방식’까지 개선하는 메타 최적화(Meta-Optimization)의 영역에 도달합니다. 카파시는 자신이 오토리서치의 작동 방식을 설명하기 위해 만든 마크다운(Markdown) 파일인 ‘프로그램 MD(Program MD)‘를 예로 듭니다. 이 파일은 “이것을 하고, 저것을 한 다음, 이런 아이디어를 시도하라”는 식의 지침을 담고 있습니다.
그는 여기서 더 나아가, AI가 더 나은 ‘프로그램 MD’를 작성할 수 있을 것이라는 아이디어를 제시합니다. 예를 들어, 사람들이 다양한 ‘프로그램 MD’를 작성하고, 동일한 하드웨어에서 어떤 ‘프로그램 MD’가 가장 큰 개선을 가져오는지 경쟁하는 콘테스트를 열 수 있습니다. 이 데이터를 바탕으로 AI는 가장 효율적인 연구 조직 운영 방식을 학습하고, 더 나은 ‘프로그램 MD’를 생성할 수 있게 됩니다.
이는 곧 연구 조직의 모든 역할과 연결 방식을 설명하는 마크다운 파일, 즉 ‘코드’가 되고, 이 코드를 튜닝하여 더 효율적인 연구 조직을 만들 수 있다는 의미입니다. 이러한 무한한 최적화의 계층, 즉 LLM, 에이전트, 클로 엔티티, 그리고 이들을 지시하는 명령 최적화에 이르기까지 모든 것이 ‘스킬 문제’로 귀결되는 상황이 바로 카파시가 말하는 ‘AI 광증’의 본질입니다.
6. 현재의 한계와 미래의 스펙트럼
이러한 무한한 가능성 속에서도 카파시는 현재 AI 에이전트의 한계와 과제를 명확히 지적합니다.
- 객관적 평가의 중요성: 에이전트와 오토리서치는 객관적인 평가 지표가 명확하고 쉽게 측정 가능한 작업(예: CUDA 커널 최적화)에 매우 효과적입니다. 그러나 평가가 어려운 주관적인 영역에서는 그 효과가 제한적입니다.
- 들쭉날쭉함(Jaggedness): 현재의 모델들은 “극도로 똑똑한 박사 과정 학생이면서 동시에 10살짜리 아이”와 같은 ‘들쭉날쭉함’을 보입니다. 매우 복잡한 시스템 프로그래밍 문제를 해결하다가도, 터무니없는 실수를 반복하거나 미묘한 의도를 파악하지 못해 사용자를 좌절시키기도 합니다.
- 최적화의 맹점: 모델의 능력이 엄청나게 향상되었음에도 불구하고, “왜 과학자들은 원자를 믿지 않을까요? 그들은 모든 것을 만들어내니까요(make everything up).”와 같은 5년 전의 낡은 농담을 여전히 반복합니다. 이는 강화학습(Reinforcement Learning)을 통해 최적화되지 않은 영역, 즉 ‘검증 불가능한’ 영역에서는 모델의 지능이 발전하지 않고 정체될 수 있음을 보여줍니다. 코드 생성 능력과 같은 ‘검증 가능한’ 지능이 ‘농담 지능’과 같은 광범위한 영역으로 일반화되지 않는다는 것입니다.
- 모델의 분화(Speciation) 필요성: 현재 연구소들은 모든 영역에서 임의로 지능적인 단일 모델(Monoculture of Models)을 지향하는 경향이 있습니다. 그러나 카파시는 동물 왕국의 다양한 뇌처럼, AI 또한 특정 영역에 특화된 ‘분화된 지능(Speciated Intelligence)‘으로 나아가야 한다고 주장합니다. 수학 증명이나 특정 산업 분야와 같이 틈새시장에 특화된 작고 효율적인 모델들이 등장할 수 있습니다. 컴퓨팅 인프라의 제약 또한 이러한 분화를 가속화할 수 있습니다.
- 두뇌 조작 과학의 미성숙: 현재는 컨텍스트 윈도우(Context Window)를 통한 모델 조작이 주를 이루지만, 모델의 가중치(Weights)를 직접 조정하여 미세 조정(Fine-tuning)하거나 지속적인 학습(Continual Learning)을 통해 특정 영역에서 모델을 더 깊이 개선하는 과학은 아직 초기 단계에 머물러 있습니다.
결론: 끝나지 않는 루프, 무한한 가능성
안드레이 카파시가 묘사하는 AI의 미래는 인간이 더 이상 ‘병목’이 아닌, 에이전트에게 ‘의지’를 표현하고 시스템을 설계하는 새로운 역할로 전환되는 시대입니다. ‘AI 광증’은 무한한 최적화의 루프와 아직 탐험되지 않은 잠재력에 대한 흥분과 약간의 불안감을 동시에 내포합니다.
지금은 모델의 ‘들쭉날쭉함’과 ‘최적화의 맹점’ 같은 한계가 존재하지만, 이러한 문제점들은 결국 ‘스킬 문제’로 해결될 것입니다. 인간은 에이전트를 더 잘 활용하고, 더 나은 ‘프로그램 MD’를 작성하며, 궁극적으로 인간의 개입 없이 스스로 개선하고 진화하는 자율적인 AI 시스템을 구축하는 방법을 찾아낼 것입니다. 이 ‘무한 루프의 시대(Loopy Era)‘는 엔지니어링, 연구, 그리고 우리의 일상까지도 근본적으로 재편할 것이며, 우리는 그 변화의 시작점에 서 있습니다.
“Terence Tao – How the world’s top mathematician uses AI” — Dwarkesh Patel 기반 기사 원본 영상 보기
AI, 과학적 발견의 패러다임을 바꾸다: 테렌스 타오와의 대화
세계 최고의 수학자 중 한 명인 테렌스 타오(Terence Tao) 교수는 인공지능(AI)이 과학적 발견의 본질을 어떻게 변화시키고 있는지에 대한 깊이 있는 통찰을 제시합니다. 드와케시 파텔(Dwarkesh Patel)과의 인터뷰에서 타오 교수는 케플러(Kepler)의 행성 운동 법칙 발견 과정을 시작으로, AI가 아이디어 생성의 비용을 거의 0으로 낮추고 있지만, 동시에 과학계에 새로운 병목 현상, 즉 ‘검증과 평가’라는 거대한 과제를 안겨주고 있다고 역설합니다. 이 기사는 타오 교수의 이야기를 통해 AI 시대의 과학이 나아가야 할 방향과 그 복잡한 함의를 탐구합니다.
1. 케플러와 브라헤: 경험적 발견의 서막
타오 교수는 인공지능이 수학에 미칠 영향을 논하기 위한 훌륭한 출발점으로 요하네스 케플러의 행성 운동 법칙 발견 이야기를 꺼냅니다. 케플러는 코페르니쿠스(Copernicus)의 지동설, 즉 태양이 태양계의 중심에 있고 행성들이 그 주위를 돈다는 모델에 기반을 두었습니다. 코페르니쿠스는 행성 궤도가 완벽한 원이라고 믿었지만, 케플러는 이 궤도의 비율에서 어떤 기하학적 의미를 발견하려 했습니다. 그는 지구 궤도를 정육면체 안에 넣으면, 그 정육면체를 둘러싸는 외부 구가 화성의 궤도와 거의 완벽하게 일치하는 등, 당시 알려진 여섯 행성의 궤도 사이에 다섯 개의 플라톤 입체(Platonic solids: 정육면체, 정사면체, 정이십면체, 정팔면체, 정십이면체)를 끼워 넣을 수 있다는 아름다운 이론을 제안했습니다. 이는 신의 설계가 수학적 완벽함과 일치한다고 믿었기 때문입니다.
이 이론을 확인하기 위해 케플러에게는 데이터가 필요했습니다. 당시 유일하게 고품질 데이터를 보유하고 있던 인물은 덴마크의 부유하고 괴짜 천문학자 튀코 브라헤(Tycho Brahe)였습니다. 브라헤는 수십 년간 육안으로 행성들을 관측하며 방대한 데이터를 축적했습니다. 케플러는 브라헤와 함께 일했지만, 브라헤는 데이터를 매우 질투하여 조금씩만 제공했습니다. 결국 케플러는 이 데이터를 ‘훔치다시피’ 하여 얻게 되었고, 자신의 아름다운 이론이 실제 데이터와 10% 정도 오차가 있음을 발견하고 크게 실망했습니다.
케플러는 수년간 이 문제에 매달린 끝에, 데이터에서 실제 행성 궤도를 도출하는 놀라운 분석 능력을 발휘했습니다. 그 결과 궤도가 원이 아닌 타원이라는 충격적인 사실을 밝혀냈고, 두 가지 행성 운동 법칙(타원 궤도, 면적 속도 일정의 법칙)을 정립했습니다. 10년 후, 그는 세 번째 법칙(공전 주기의 제곱은 궤도 장반경의 세제곱에 비례)까지 발견했습니다. 케플러는 이 법칙들에 대한 물리적 설명은 제시하지 못했으며, 이는 한 세기 후 뉴턴(Newton)이 중력 이론을 통해 이 모든 것을 설명하면서 비로소 완성되었습니다.
타오 교수는 이 이야기에서 중요한 통찰을 제시합니다. 케플러는 마치 ‘고온의 대규모 언어 모델(LLM)‘과 같았다는 것입니다. 그는 “세계의 조화(Harmonics of the World)“라는 책에서 행성들의 ‘음정’을 논하는 등 무작위적인 관계들을 끊임없이 시도했습니다. 이 과정에서 지구의 음정은 ‘미-파-미’이며, 이것이 지구의 기근과 불행의 원인이라는 식의 황당한 점성술적 아이디어도 포함되어 있었습니다. 하지만 그 속에서 세 번째 행성 운동 법칙인 ‘제곱-세제곱 법칙(cube-square law)‘이 발견되었습니다. 브라헤의 데이터와 같은 검증 가능한 데이터 뱅크만 있다면, LLM은 케플러처럼 20년 동안 무작위적인 관계를 시도할 수 있고, 그중 일부는 과학적 진보로 이어질 수 있다는 것이죠.
2. 변화하는 과학적 방법론: 가설에서 데이터 중심으로
전통적으로 과학은 문제 식별, 가설 수립, 데이터 수집, 검증, 설명의 과정을 따르며, 특히 ‘유레카’와 같은 천재적인 아이디어 생성의 순간을 높이 평가했습니다. 케플러 역시 많은 아이디어를 시도했고, 그중 상당수는 실패했습니다. 하지만 그의 성공은 브라헤의 정밀한 데이터 수집이라는 ‘검증’ 과정이 있었기에 가능했습니다. 브라헤가 이전 관측보다 10배나 더 정확한 데이터를 제공했기 때문에 케플러는 자신의 모델을 데이터에 맞춰 조정하고 정확한 결과를 얻을 수 있었던 것입니다.
그러나 20세기 들어 과학의 패러다임은 변화했습니다. 이론과 실험이라는 두 가지 주요 패러다임에 더해, 수치 시뮬레이션(numerical simulation)과 빅데이터(big data) 분석이 등장했습니다. 이제는 대규모 데이터셋을 먼저 수집하고, 그 안에서 패턴을 찾아 가설을 도출하는 방식이 점차 보편화되고 있습니다. 타오 교수는 이를 “과학적 방법이 거의 역전되었다”고 표현합니다. 케플러는 어쩌면 최초의 데이터 과학자 중 한 명이었지만, 그조차도 브라헤의 데이터로 시작하기 전에 이미 선입견적인 이론을 가지고 있었습니다. 하지만 이제 데이터의 규모와 유용성이 압도적으로 커지면서, 데이터를 먼저 분석하고 가설을 찾는 방식이 더욱 강력해지고 있습니다.
다만, 데이터 기반 접근법에도 함정은 있습니다. 요한 보데(Johann Bode)의 법칙이 좋은 예입니다. 그는 케플러에게 영감을 받아 행성 간 거리가 기하급수적으로 증가한다는 패턴을 발견했고, 화성과 목성 사이에 행성이 하나 더 있을 것이라고 예측했습니다. 실제로 천왕성(Uranus)과 소행성 세레스(Ceres)가 발견되면서 그의 예측과 일치하는 듯 보였습니다. 그러나 해왕성(Neptune)이 발견되면서 이 패턴은 완전히 어긋났고, 결국 보데의 법칙은 ‘숫자적 우연의 일치(numerical fluke)‘에 불과했음이 드러났습니다. 케플러가 세 번째 법칙을 처음 두 법칙만큼 강조하지 않은 것도, 단 6개의 데이터 포인트만으로 도출된 결론의 신뢰성에 대해 본능적으로 조심스러웠기 때문일 수 있습니다.
3. AI가 촉발한 새로운 병목 현상: 아이디어 생성에서 검증으로
타오 교수는 AI가 아이디어 생성의 비용을 거의 0에 가깝게 낮췄다고 강조합니다. 이는 인터넷이 통신 비용을 낮춘 것과 유사합니다. 하지만 이는 곧 과학에 새로운 병목 현상(bottleneck)을 초래합니다. 과거에는 아마추어 과학자들이 무수히 많은 이론을 제시했지만, 그중 대부분은 가치가 없었습니다. 그래서 동료 심사(peer review)와 출판 시스템은 고품질 아이디어를 걸러내고 검증하는 역할을 했습니다. 하지만 이제 AI가 엄청난 규모로 가능한 설명을 생성해내면서, 인간 검토자들은 이미 감당하기 힘든 상황에 처해 있습니다. 많은 학술지들이 AI가 생성한 논문으로 넘쳐나고 있다고 보고합니다.
문제는 이 방대한 아이디어 풀에서 어떤 아이디어가 진짜 진전을 이루고, 어떤 아이디어가 막다른 골목이나 잘못된 단서(red herrings)인지를 평가하고 검증하는 방법을 대규모로 알지 못한다는 것입니다. 개별 논문에 대해서는 수년간의 논의를 통해 합의에 도달할 수 있지만, 매일 수천 개의 아이디어가 생성될 때 이런 방식은 통하지 않습니다.
타오 교수는 벨 연구소(Bell Labs)에서 ‘비트(bit)’ 개념이 등장했을 때를 예로 듭니다. 수많은 기술 논문 속에서 ‘비트’라는 아이디어는 확률론, 컴퓨터 과학 등 다양한 분야에 걸쳐 엄청난 함의를 가졌습니다. 미래에 AI가 이와 같은 통일 개념(unifying concept)을 제시한다면, 수백만 개의 논문 중에서 어떻게 그것을 식별해낼 수 있을까요?
아이디어의 가치를 평가하는 것은 매우 어렵습니다. 딥러닝(Deep Learning) 자체도 오랫동안 AI의 틈새 분야였고, 첫 번째 원리(first principles) 추론이 아닌 데이터 학습만으로 답을 얻는다는 생각은 매우 논란이 많았습니다. ‘트랜스포머(Transformer)’ 아키텍처는 현대 대규모 언어 모델의 기반이 되었지만, 다른 아키텍처가 먼저 채택되었다면 그것이 표준이 되었을 수도 있습니다. 아이디어의 결실 여부는 미래와 문화, 사회적 채택 여부에 달려 있습니다. 10진법 체계가 로마 숫자보다 훨씬 유용하지만, 10이라는 숫자에 특별한 의미가 있는 것은 아닙니다. 단지 모든 사람이 사용하기 때문에 표준화되었고, 이제는 관성 때문에 바꾸기 힘듭니다. 즉, 과학적 성과를 과거와 미래의 맥락 없이 순수하게 고립시켜 객관적으로 평가하는 것은 불가능합니다.
또한, 과학적 진보는 종종 기존의 가정을 삭제하는 방식으로 이루어집니다. 지동설이 오랫동안 받아들여지지 않은 것은 아리스토텔레스(Aristotle)의 물리학, 즉 물체가 자연적으로 정지 상태를 유지하려 한다는 생각 때문이었습니다. 지구가 움직인다면 왜 우리가 모두 떨어지지 않는가 하는 의문이 해결되지 않았기 때문입니다. 뉴턴의 운동 법칙이 정립된 후에야 비로소 지구가 움직인다는 개념이 이해될 수 있었습니다. 다윈(Darwin)의 진화론 역시 종(species)이 고정되어 있지 않다는 당시로서는 파격적인 가정을 삭제하면서 이루어진 위대한 진보입니다. 타오 교수는 현재 우리가 ‘인간 지능이 우주의 중심’이라는 생각에서 벗어나 다양한 유형의 지능이 존재함을 인식하는 ‘인지적 코페르니쿠스적 혁명(cognitive Copernican revolution)‘을 겪고 있다고 말합니다.
4. 과학적 소통과 설득의 중요성
타오 교수는 과학적 발견이 단순히 새로운 이론을 만들고 검증하는 것을 넘어, 다른 사람들에게 효과적으로 소통하는 것 또한 매우 중요하다고 강조합니다. 그는 다윈의 진화론이 뉴턴의 ‘프린키피아(Principia Mathematica)‘보다 2세기 후에 나왔음에도 불구하고, 개념적으로는 더 간단해 보였던 이유를 설명합니다. 다윈의 이론은 자연선택에 대한 증거가 압도적이었지만, 이는 누적적이고 회고적인(cumulative and retrospective) 성격을 띠었습니다. 반면 뉴턴은 자신의 방정식을 제시하고 달의 궤도 주기와 거리가 일치하는지 확인하면 바로 진전이 이루어졌습니다.
다윈은 놀라운 과학 커뮤니케이터였습니다. 그는 방정식 없이 평이한 영어로 글을 썼고, 다양한 사실들을 설득력 있는 비전으로 종합했습니다. 반면 뉴턴은 라틴어로 글을 썼고, 자신의 이론을 설명하기 위해 새로운 수학 분야(미적분)를 창안해야 했습니다. 그는 또한 당시 과학자들이 훨씬 비밀스럽고 경쟁적이었던 시대의 인물이었으며, 자신의 가장 중요한 통찰 일부를 경쟁자들에게 빼앗기지 않기 위해 숨기기도 했습니다. 뉴턴의 업적이 널리 알려진 것은 수십 년 후 다른 과학자들이 그의 작업을 훨씬 간단한 용어로 설명한 이후였습니다.
타오 교수는 “서술의 기술과 주장을 펼치는 능력, 그리고 서사를 만드는 것”이 과학의 매우 중요한 부분이라고 말합니다. 데이터가 뒷받침되면 도움이 되지만, 사람들이 설득되어야만 이론을 더 발전시키거나 초기 투자를 하여 배우고 탐구할 것입니다. 그는 AI가 아직 설득력을 강화학습(Reinforcement Learning)하기 어렵다는 점을 지적하며, 이것이 어쩌면 과학의 영원한 인간적 측면으로 남을 수 있다고 예측합니다.
5. 수학 분야 AI의 현재와 미래: 폭과 깊이
타오 교수는 드와케시 파텔의 시리즈 “우주 거리 사다리(cosmic distance ladder)“를 언급하며, “많은 분야에서 연역적 오버행(deductive overhang)이 사람들이 깨닫는 것보다 훨씬 클 수 있다”는 점에 동의합니다. 즉, 문제를 연구하는 올바른 통찰력만 있다면, 세상에 대해 훨씬 더 많은 것을 알 수 있다는 것입니다. 천문학은 데이터가 병목 현상이었기 때문에 데이터를 분석하고 정보에서 마지막 한 방울까지 짜내는 데 가장 먼저 앞장선 과학 분야 중 하나였습니다. 퀀트 헤지펀드(quant hedge funds)에서 천문학 박사를 선호하는 이유도 여기에 있습니다.
최근 AI는 수학 분야에서 놀라운 진전을 보이고 있습니다. 타오 교수는 최근 몇 달 동안 AI 프로그램들이 에르되시 문제(Erdős problems) 1,100여 개 중 약 50개를 해결했다고 언급했습니다. 하지만 그는 “현재 낮게 매달린 열매(low-hanging fruit)는 이미 다 따버린 상태”이며, 순수 AI 솔루션이 급격히 줄어들고 있다고 말합니다. 이제는 AI가 가능한 증명 전략을 생성하고, 다른 AI 도구가 이를 비판하고, 재작성하거나, 수치 데이터를 생성하고, 문헌 조사를 하는 등, “많은 인간과 많은 AI 도구들 간의 지속적인 대화”를 통해 문제가 해결되는 양상입니다.
타오 교수는 AI 도구를 “2미터를 점프할 수 있는 점핑 머신”에 비유합니다. 이들은 때때로 잘못된 방향으로 점프하거나 추락하기도 하지만, 때로는 인간이 도달할 수 없었던 가장 낮은 벽의 꼭대기에 도달할 수 있습니다. 하지만 이 도구들은 부분적인 진전을 만들거나, 먼저 집중해야 할 중간 단계를 식별하는 데는 매우 서툴다는 한계가 있습니다.
그는 AI의 진보에 대해 “비관론적(bearish)이면서 동시에 낙관론적(bullish)“인 두 가지 관점을 제시합니다. 비관론적 관점은 AI가 아직 인간 전문가가 도달하는 수준의 “높은 벽”에 도달하지 못한다는 것입니다. 하지만 낙관론적 관점은 AI가 일단 특정 수준에 도달하면, “그 수준에서 가능한 모든 문제를 해결할 수 있는 강력한 속성”을 가지고 있다는 것입니다. 인간은 수백만 개의 문제에 대해 수백 년의 연구를 동시에 수행할 수 없지만, AI는 가능하다는 것이죠.
타오 교수는 AI가 **‘폭(breadth)‘**에 뛰어나고 인간은 **‘깊이(depth)‘**에 뛰어나다고 결론 내립니다. 그는 AI의 폭넓은 능력을 최대한 활용하기 위해 과학 수행 방식을 재설계해야 한다고 주장합니다. “하나 또는 두 개의 정말 깊고 중요한 문제보다는 매우 광범위한 문제들을 다루는 데 더 많은 노력을 기울여야 한다”는 것입니다. AI는 새로운 과학 분야를 탐색하고, 쉬운 관측을 모두 수행하여 ‘난이도의 섬’을 식별한 다음, 인간 전문가가 그곳에 와서 작업할 수 있도록 할 것입니다. 그는 “미래에는 매우 상호 보완적인 과학을 보게 될 것”이며, “과학은 그 이후로 알아볼 수 없게 변할 것”이라고 전망합니다.
프로그래머들이 AI 도구 덕분에 생산성이 크게 향상되었다고 느끼는 것처럼, 수학에서도 유사한 변화가 일어날 수 있습니다. 타오 교수는 수학에서 문제 해결 과정 자체가 문제 자체보다 더 중요하다고 말합니다. AI는 상용구(boilerplate) 코드와 같은 보조적인 작업들을 가속화하여 그의 논문에 더 많은 코드나 그림을 포함할 수 있게 해주었습니다. 그는 AI가 없다면 오늘날 그가 쓰는 논문의 유형은 5배 더 오래 걸릴 것이라고 말합니다. 하지만 수학 문제의 가장 어려운 핵심 부분을 해결하는 데는 아직 AI의 도움을 받지 않고 여전히 펜과 종이를 사용합니다.
그는 AI 유형 도구들이 “수학의 실험적 측면을 혁신할 것”이라고 예측합니다. 개별 문제와 해결 과정에 덜 신경 쓰고, 무엇이 효과가 있고 무엇이 효과가 없는지에 대한 대규모 데이터를 수집하는 데 집중할 것이라는 겁니다. “수학을 대규모로 수행하는 아이디어(the idea of doing mathematics at scale)“는 아직 초기 단계이지만, 이것이 AI가 이 분야를 혁신할 지점이라고 그는 말합니다.
타오 교수는 AI의 진보를 “놀랍고도 실망스러운(amazing and disappointing)” 감정으로 바라봅니다. 사람들은 빠르게 적응합니다. 20년 전 구글 검색이 처음 나왔을 때 놀라웠지만, 몇 년 후에는 당연하게 여겨졌습니다. 2026년 수준의 AI는 2021년에는 경이로웠을 것이지만, 이제는 얼굴 인식, 자연어 처리, 대학 수준의 수학 문제 해결 등이 당연하게 여겨진다는 것입니다.
그는 2023년에 2026년까지 AI가 수학 분야에서 “신뢰할 수 있는 공동 저자(trustworthy co-author) 같은 존재가 될 것”이라고 예측했고, 현재로서는 이 예측이 꽤 잘 맞아가고 있다고 평가합니다.
마지막으로 타오 교수는 ‘인공적인 영리함(artificial cleverness)‘과 ‘인공 지능(artificial intelligence)‘을 구분합니다. 지능은 정의하기 어렵지만, 그가 말하는 지능은 “문제 해결을 위해 협력할 때 아이디어를 끊임없이 개선하고, 체계적으로 무엇이 작동하고 무엇이 작동하지 않는지 파악하며, 진전의 길을 찾아내는 적응성(adaptivity)과 지속적인 개선”을 의미합니다. AI는 이러한 과정을 모방할 수 있지만, “조금씩 점프해서 손잡이를 잡고, 거기에 머물러 다른 사람들을 끌어올리고, 거기서부터 다시 점프하는” 누적적인 과정은 아직 부족합니다. AI는 시행착오와 반복, 즉 ‘무차별 대입(brute force)‘에 가깝습니다. AI가 문제를 해결하더라도, 그 자체의 수학적 이해가 진전된 것은 아니며, 새로운 세대의 훈련 데이터에 0.001%의 기여를 할 뿐, 새로운 세션에서는 방금 한 일을 잊어버린다는 것입니다.
결론
테렌스 타오 교수의 통찰은 AI가 과학적 발견의 본질을 근본적으로 뒤흔들고 있음을 보여줍니다. 아이디어 생성의 문턱을 낮추고 ‘폭’의 탐색을 가능하게 하는 AI는 과학의 새로운 지평을 열고 있지만, 동시에 검증, 평가, 그리고 진정한 통일 개념을 식별하는 새로운 과제를 안겨주고 있습니다. 과학적 소통과 설득이라는 인간 고유의 영역 또한 여전히 중요합니다. AI와 인간의 상호 보완적인 관계 속에서, 우리는 이전에 상상할 수 없었던 방식으로 과학을 재정의하고, 궁극적으로는 ‘깊이’와 ‘폭’을 모두 아우
“Inside Palantir: Building Software That Matters | Shyam Sankar on a16z” — a16z 기반 기사 원본 영상 보기
위협받는 미국, AI로 재무장하라: 팔란티어 샴 산카르가 말하는 국방 혁신과 ‘전 국가적 동원’의 시대
위기 속 미국의 각성: 국방 혁신의 목소리
세계 각지에서 벌어지는 사건들은 여전히 잔혹한 야만주의가 존재하며, 악이 실재한다는 냉혹한 현실을 상기시킨다. 한 나라가 전쟁에 돌입할 때, 국방부만이 전쟁을 수행하는 것으로는 충분치 않다. 진정한 전쟁은 온 나라가 함께 치르는 것이며, 미국 국민이 이러한 위기에 대응할 역량이 없다는 생각은 믿기 어렵다. 팔란티어(Palantir)의 샴 산카르(Shyam Sankar)는 미국의 가장 큰 위험은 외부의 공격이 아닌 ‘자멸’이라고 경고하며, AI 경쟁에서 승리하고 특히 물리적 AI 및 로봇 공학 분야에서 앞서나가기 위한 근본적인 변화를 촉구한다.
오랫동안 팔란티어의 공동 창립자이자 ‘해결사(fixer)’로 알려졌던 샴 산카르는 최근 몇 년 사이 대중적 인물로 부상했다. 그의 동료들은 그를 “경력의 은인”으로 칭송하며, 팔란티어에 영입하고 지원했을 뿐 아니라, 새로운 시작을 할 수 있도록 “날개를 달아주었다”고 입을 모은다. 수년 동안 ‘배후 인물’로 활동하던 그가 전면에 나서게 된 결정적인 계기는 2022년에 발표한 국방 개혁에 대한 중요한 글인 ‘퍼스트 브렉퍼스트(First Breakfast)‘였다. 이 글은 미국이 나아가야 할 방향에 대한 강력한 목소리가 되었다.
산카르는 이러한 변화의 동기를 “절박함과 낙관주의가 반반 섞인 행동”이라고 설명한다. 수년간 국방부 내부에서 국방 시스템이 서서히 악화되는 것을 목격하며, 그는 ‘끓는 물 속 개구리(frog boil)‘처럼 점진적으로 다가오는 위기를 절감했다. 하지만 동시에 국방 분야에 기여하려는 창업가들의 엄청난 에너지와 국가적 관심이 재부상하는 것을 보며, 지금이야말로 모든 것을 바로잡을 수 있는 ‘기회의 순간’이라고 판단했다. 그는 과거 우리가 승리했던 방식들을 우연히 등한시하게 되었으며, 이제는 그 방식을 활기차게 되찾을 기회가 왔다고 강조한다.
러시아의 2014년 크림반도 합병, 2015년 스프래틀리 군도(Spratly Islands)의 군사화, 이란 핵 협정(JCPOA)의 실패, 그리고 2023년 10월 7일 하마스의 이스라엘 공격 이후 발생한 참혹한 사건들은 산카르에게 “도대체 무슨 일이 벌어지고 있는가?”라는 급진적인 질문을 던지게 했다. 이러한 일련의 사건들은 미국이 억지력(Deterrence)을 상실했음을 보여주는 증거이며, 시간이 얼마 남지 않았음을 의미했다. 그러나 그는 지난 1년 동안 국방부 내부에서 지난 19년간 보지 못했던 변화가 일어났으며, 사람들이 이러한 개혁의 순간을 포착하고 있다고 긍정적으로 평가했다.
잊혀진 승리의 공식: ‘국가 전체’의 동원
산카르는 오늘날의 지정학적 상황을 1930년대 후반에 비유하며, 더 큰 충돌을 막기 위한 ‘동원(Mobilize)‘의 중요성을 역설한다. 그의 저서 『모빌라이즈(Mobilize)』는 우리가 냉전에서 승리(혹은 소련이 패배)한 후 안일함에 빠져 미래에 대한 많은 잘못된 믿음을 갖게 되었다고 지적한다. 그는 제2차 세계대전 당시 미국의 동원이 진주만 공격 이후 갑자기 일어난 것이 아니라고 강조한다. 오히려 프랭클린 루스벨트(FDR) 대통령을 비롯한 지도부는 1930년대 후반부터 국가적 의지가 충분히 형성되지 않았음에도 불구하고 동원의 필요성을 인식했다. 랜드리스(Lend-Lease) 프로그램은 공장 건설과 재정비에 18개월이 걸렸고, 결국 진주만 공격 당시 미국은 이미 완전 생산 체제를 갖추고 있었다.
진정한 전쟁은 “국가 전체(the whole country)“가 치르는 것이며, 단순히 국방 산업 기반이나 국방부만의 문제가 아니라는 것이 산카르의 핵심 주장이다. 그는 미국 시민 모두가 국가가 제공하는 번영뿐 아니라, 그 번영의 기반이 되는 자유에 투자해야 한다고 말한다. 그러나 우리는 이러한 정신에서 너무나 멀리 떨어져 왔다. 1989년에는 주요 무기 시스템 지출의 6%만이 국방 전문 기업에 할당되었지만, 현재 이 수치는 86%에 달한다. 이는 과거에는 크라이슬러(Chrysler)가 미니트맨(Minuteman) 미사일과 미니밴을 동시에 생산했듯, 민간 산업이 국가 안보를 자연스럽게 뒷받침했던 것과는 대조적이다. 오늘날 민간 부문의 연구개발(R&D) 지출은 정부의 역량을 압도하며, 이 민간 기술의 가격 대비 성능 곡선을 국방력으로 전환하는 것이 중요하다고 산카르는 강조한다.
과거 미국의 산업 기반은 글렌 마틴(Glenn Martin), 잭 노스럽(Jack Northrup), 르로이 그루먼(Leroy Grumman)과 같은 창업가들이 이끌었다. 이들은 단기적인 재무 성과보다는 자신과 회사보다 훨씬 큰 비전을 추구했다. 이러한 창업가 정신은 정부 내부에도 존재했는데, 해군 내 반대에도 불구하고 핵 추진 해군을 건설한 하이먼 리코버(Hyman Rickover) 제독이 대표적이다. 심지어 ‘원자폭탄의 아버지’ 오펜하이머(Oppenheimer)조차 그의 실패를 예견했지만, 리코버는 굴하지 않았다. 산카르는 이러한 ‘허츠파(chutzpah, 대담함)’ 정신이 바로 실리콘밸리 창업가들의 특징과 일맥상통한다고 말한다.
혁신의 걸림돌: 통합과 순응의 딜레마
냉전 종식 후, 미국은 ‘평화 배당금(peace dividend)‘을 기대하며 국방비 지출을 줄였다. 이 시기에 51개에 달하던 주요 국방 계약업체가 단 5개로 줄어드는 유명한 ‘최후의 만찬(Last Supper)‘이 있었다. 산카르는 이 사건에 대한 통상적인 설명, 즉 통합으로 인한 경쟁 상실이 문제였다는 주장이 틀렸다고 지적한다. 국방 산업은 본질적으로 수요 독점(Monopsony) 시장이며, 혁신은 기업 간 경쟁이 아니라 정부 내부의 서비스 간 경쟁에서 촉발되었다는 것이다.
산카르가 보기에 ‘최후의 만찬’으로 인한 통합은 ‘순응(conformity)‘을 낳았다. 이는 국방 산업의 진정한 ‘금융화(financialization)‘의 시작이었고, 기업들은 성장보다는 재무 지표, 배당, 자사주 매입, 현금 흐름에 집중하게 되었다. 이러한 순응적인 환경에서는 창업가 정신이 번성할 수 없었다. ‘이단아(heretics)‘들은 축출되거나, 기술 산업과 같은 다른 경제 부문으로 떠났다.
그러나 산카르는 이러한 ‘이단아’들이야말로 혁신에 필수적이라고 강조한다. 그의 저서에 기록된 놀라운 국방 혁신 사례들은 거의 예외 없이 ‘이단적인 아이디어’에서 비롯되었다. 제도, 관료주의, 기존 절차는 이들을 억압하려 했지만, 결국 성공했다. 제2차 세계대전을 승리로 이끈 ‘히긴스 보트(Higgins boat)‘가 대표적이다. 해군은 이 보트 구매를 거부하고 설계도를 훔치려 했지만, 결국 전쟁 중 모든 상륙정의 92%가 히긴스 보트였다. 만약 창업가 히긴스가 병적으로 집착하지 않았다면 노르망디 상륙작전은 불가능했을 것이다.
산카르는 이러한 문제에 더해, 수요 독점의 특성상 국방부가 공급업체에 온갖 제약을 가하며 ‘갈라파고스 섬(Galapagos Islands)‘에 고립된 생물처럼 만들었다고 비유한다. 그 결과, 뛰어난 ‘거대 거북이’ 같은 기업들이 탄생했지만, 이들은 본토에서는 경쟁력을 잃고 ‘늑대’들에게 잡아먹힐 것이라는 비판이다. 팔란티어가 설립될 당시만 해도 국방부에는 외부인이 접근할 ‘정문(front door)‘이 사실상 없었고, 정보 커뮤니티의 인큐텔(In-Q-Tel)만이 유일한 통로였다. 그러나 2015년부터 시작된 변화의 물결로 인해, 지난 10년간 국방부를 돕기 위해 문을 두드리던 ‘이단아’들이 비로소 기회를 얻게 되었다고 산카르는 말한다.
이단적 영웅들의 귀환: 국방을 위한 민간 인재 동원
산카르는 국방 산업의 ‘이단적 영웅들’ 중 한 명으로 드루 쿠커(Drew Cukor) 대령을 꼽는다. 그는 해병대 대령으로, 단 한 번의 작전 실패(RPG 오인으로 인한 야지디족 난민 구조 실패)가 수천 명의 민간인에게 참혹한 결과를 가져온 것을 목격한 후, AI를 국방부에 도입하는 메이븐 프로젝트(Project Maven)에 전념했다. 자원도 없는 상황에서 시작된 이 프로젝트에서 쿠커 대령은 모든 이들의 반대와 질투, 심지어 허위 고발(이란인을 집에 숨기고 있다는)까지 감수해야 했다. 하지만 그는 흔들리지 않았고, 결국 AI의 국방 도입에 결정적인 역할을 했다. 산카르는 이러한 이야기가 “나라가 당신을 정말 필요로 할 때, 숨어있는 이단아들이 나서도록 영감을 주는 것”이라고 강조한다.
존 보이드(John Boyd) 대령의 F-16 개발과 OODA 루프(OODA Loop) 이론, 버나드 슈리버(Bernard Schriever) 장군과 에드워드 홀(Edward Hall) 박사의 미니트맨 미사일 개발 사례 또한 ‘이단아’들을 보호하고 지원하는 리더십의 중요성을 보여준다. 이들은 모두 조직 내에서 엄청난 저항에 부딪혔지만, 그들의 탁월함을 알아본 리더들의 보호 아래 혁신을 이뤄낼 수 있었다.
이러한 맥락에서 샴 산카르 자신도 최근 육군에 입대했다. 그는 2014년부터 이스라엘과 협력하며, 10월 7일 이후 이스라엘이 36만 명의 예비군을 동원했을 때 놀라운 변화를 목격했다. 이스라엘 예비군들은 20년 이상 민간 산업에서 인터넷 규모의 시스템을 구축하며 쌓은 경험을 바탕으로, 불과 4개월 만에 군대의 기술을 10년 이상 발전시켰다. 산카르는 미국 또한 이스라엘처럼 엄청난 기술 인재를 보유하고 있으며, 24세의 젊은 시절에는 할 수 없었던 기여를 44세의 경험으로 할 수 있다고 말한다. 제2차 세계대전 당시 10만 명의 민간인이 군으로 특별 임관(direct commissioned)되었던 것처럼, 오늘날에도 민간 기술 인력을 국방에 활용해야 한다는 것이다. 그는 전 오픈AI(OpenAI) 수석 연구원 밥 맥그루(Bob McGrew), 메타(Meta)의 최고기술책임자(CTO) 보스(Bos), 전 오픈AI 최고제품책임자 앤드류 윌(Andrew Wheel) 등 세 명의 동료와 함께 육군에 합류하여 고위 자문관으로서 다양한 프로젝트를 돕고 있다.
산카르는 육군 내부에서 가장 놀랐던 점으로 ‘신세대 장병들(green suitors)‘의 탁월한 역량을 꼽는다. 이들은 정식 컴퓨터 과학 교육을 받지 않았음에도 불구하고, 스스로 AI 애플리케이션을 구축하는 놀라운 능력을 보여준다. 특히 ‘작전 데이터 팀(Operational Data Teams)‘에서 이들이 개발하는 AI 솔루션은 상업 부문에서 볼 수 있는 어떤 것보다도 설득력이 있다고 한다. 이는 생존이 걸린 ‘실존적 위협’ 앞에서 사람들은 10%의 효율성 개선이 아닌 ‘승리’라는 이분법적 목표에 집중하기 때문이다. 또한 AI는 특정 기술을 가진 개인, 특히 자신의 전문 영역을 깊이 이해하는 정보 장교들에게 엄청난 힘을 실어준다. 과거에는 파워포인트 슬라이드를 만들던 이들이 이제는 2주 만에 직접 솔루션을 구축하고, 그 성과를 통해 군을 발전시키는 경험적 대화를 나누고 있다. 이러한 바텀업(bottom-up) 혁신과 임무형 지휘(Mission Command)는 미국 군대의 고유한 강점이라고 산카르는 강조한다.
AI 시대의 소프트웨어 지형 변화: ‘베타’와 ‘알파’의 갈림길
최근 소프트웨어 업계에서는 AI의 등장으로 인한 ‘SaaS 종말론(SaaS Apocalypse)‘이 대두되고 있다. AI가 코드를 생성하고 통합하는 능력이 향상되면서, 기존 SaaS(Software as a Service) 기업들의 진입 장벽(switching costs)이 낮아지고 있다는 주장이다. 이에 대해 산카르는 양쪽 주장이 모두 사실이라고 말하며, 소프트웨어를 ‘베타(Beta)‘와 ‘알파(Alpha)‘로 구분하는 새로운 관점을 제시한다.
‘베타 소프트웨어’는 사용자를 다른 모든 사람과 ‘유사하게’ 만드는 데 초점을 맞춘다. 이는 과거 소프트웨어 산업의 문제점이기도 했는데, 소프트웨어 개발의 피드백 루프가 ‘가치 창출’보다는 ‘판매 가능성’에 집중되어 있었기 때문이다. 산카르는 이러한 베타 소프트웨어는 AI 시대에 큰 어려움을 겪을 것이라고 예측한다. 반면, AI는 사용자의 고유한 요구에 맞춰 소프트웨어를 만들 수 있게 함으로써 본질적으로 ‘알파(Alpha)‘에 집중할 수 있게 한다. 알파 소프트웨어는 사용자를 ‘다르게’ 만들고, 경쟁 우위와 전략을 표현할 수 있게 해준다. 그는 이미 알파에 집중하고 있던 플랫폼들은 AI 시대에 더욱 강력한 바람을 맞을 것이라고 말한다.
코로나19 팬데믹 당시, 많은 기업의 전사적 자원 관리(ERP) 시스템이 무너졌지만, 줌(Zoom)이나 팀즈(Teams) 같은 협업 도구들은 원격 근무를 가능하게 했다. 산카르는 이를 소프트웨어 산업의 ‘스푸트니크 쇼크(Sputnik moment)‘로 비유하며, 지난 수십 년간 우리가 가치 있다고 믿었던 수많은 소프트웨어 지출이 사실은 그렇지 않았음을 깨달아야 한다고 지적한다. 팔란티어는 고객들이 혼란의 속도에 맞춰 현실에 적응하도록 도우며 오히려 큰 성장을 경험했다. 그는 많은 기업들이 모방적으로 지출하던 ‘베타’ 소프트웨어들이 큰 압박을 받을 것이며, 반대로 기업의 경쟁 우위를 표현하는 ‘알파’ 소프트웨어, 즉 툴킷(toolkit)과 같은 접근 방식이 각광받을 것이라고 내다본다.
AI 시대의 가치 창출에 대해서는, 현재 하드웨어 계층이 가장 높은 마진을 차지하고 있지만, 인터넷 경제에서는 애플리케이션 계층이 그랬다. 산카르는 AI 시대에는 칩(chips) 계층과 AI 인프라(AI infrastructure) 계층(그는 이를 ‘온톨로지(Ontology)‘라고 부른다)에서 가치가 창출될 것이며, 이 두 계층이 상당한 방어력을 가질 것이라고 예측한다. 모델(models) 자체는 점차 상품화될 압력을 받으므로, 모델 기업들은 AI 인프라를 구축하는 방향으로 확장하고, 특정 수직 AI 솔루션 기업들은 고객 기반을 확장하기 위해 AI 인프라로 내려오게 될 것이라는 분석이다.
인간의 의지, AI의 미래: 경제와 일자리에 대한 통찰
AI가 경제와 일자리에 미칠 영향에 대해 산카르는 중요한 관점을 제시한다. 그는 AI에 대해 이야기할 때 마치 인간의 의지(human agency)가 없는 것처럼 “AI가 X를 할 것이다”라고 말하는 방식에 불편함을 느낀다고 말한다. “인간이 AI를 사용하여 X를 할 것이다”가 올바른 표현이며, 우리에게는 AI를 특정 방향으로 이끌 선택권과 의무가 있다는 것이다.
그는 AI가 1970년대 이후 벌어진 임금 성장과 국내총생산(GDP) 성장 간의 근본적인 단절을 해결할 역사적 기회를 제공한다고 본다. AI는 미국 노동자들에게 ‘초능력(superpowers)‘을 부여하여 생산성을 극대화하고, 이는 국가의 재산업화(re-industrialization)를 뒷받침할 기반이 될 수 있다는 것이다. 하드리안(Hadrien)과 같은 기업들이 기술을 활용하여 사람들의 생산성을 50~100배 향상시키며 재산업화를 이끄는 것이 그 예시이다.
또한 산카르는 세계화의 ‘거대한 거짓말’ 중 하나가 혁신(innovation)은 한곳에서 하고 생산(production)은 다른 곳에서 할 수 있다는 믿음이었다고 지적한다. 그는 혁신은 생산성의 결과이며, 무언가를 만들지 않으면 그것을 만드는 방식이나 무엇을 만들지에 대해 혁신할 수 없다고 강조한다. 스페이스X(SpaceX)의 R&D 엔지니어들이 생산 현장과 함께 일하는 것이 대표적인 예시이다. 그는 AI를 국가적 비상사태이자 동시에 국가적 기회로 보아야 한다고 주장한다.
역사적으로 기술 혁명은 대부분 ‘도구 혁명’이었지 ‘개념 혁명’이 아니었다. 망원경을 발명한 것은 갈릴레오가 아니라 다른 사람이었지만, 갈릴레오는 그것을 사용하여 행성 운동을 발견했다. 현미경, 동력 직조기, 개인용 컴퓨터 등 이러한 기술의 미래는 기술을 발명한 사람이 아니라 ‘기술을 사용하는 사람(wielders)‘에 의해 결정된다. 오늘날 AI 비관론(AI dumerism)은 주로 매우 똑똑한 ‘발명가’들의 목소리에서 나오지만, 그들의 창조물처럼 그들 역시 ‘불균형한 지능(jagged intelligence)‘을 가질 수 있기에, 모델 구축에 능숙하다고 해서 그 함의까지 정확히 예측할 수 있는 것은 아니다.
산카르는 우리가 AI의 미래를 결정할 ‘사용자’로서 인간의 의지를 포기해서는 안 된다고 강조한다. 경제학자들의 AI 미래 예측 그래프처럼, 다양한 결과가 모두 가능하며, 우리가 어떤 미래를 선택할지는 우리에게 달려 있다는 것이다.
마지막으로, 그는 과거 미국 경제에서 공
“Silicon Valley’s Big Bets on War Pay Off, and the Trump Family Business Looks to Transylvania” — New York Times Podcasts 기반 기사 원본 영상 보기
전쟁이 실리콘밸리에 안겨준 기회: 트럼프 가문의 트란실바니아 개발 미스터리
격동하는 국제 정세 속에서 뜻밖의 기회를 포착하는 이들이 있습니다. 미국의 기술 허브인 실리콘밸리의 거대 기업들은 급변하는 안보 환경 속에서 ‘전쟁의 기술’에 깊이 뛰어들며 새로운 황금기를 맞이하고 있습니다. 한편, 도널드 트럼프 전 대통령의 가족 사업체인 트럼프 조직(Trump Organization)은 발리나 몰디브 같은 화려한 휴양지가 아닌, 루마니아의 외딴 지역 트란실바니아에서 의외의 부동산 프로젝트를 추진하며 세간의 이목을 집중시키고 있습니다. 뉴욕타임스(The Times)의 최신 보도는 이 두 가지 상반된 흐름이 어떻게 오늘날의 세계를 재편하고 있는지 조명합니다.
이란 전쟁의 그림자, 실리콘밸리의 새로운 활로
이란과의 전쟁이 3주차에 접어들면서, 미국은 군사적, 경제적, 기술적 측면에서 중대한 결정을 내리고 있습니다. 트럼프 대통령은 이란에 미군 병력을 파병할 가능성을 “배제하지 않는다”고 밝히며, 필요하다면 군사력 배치를 “두려워하지 않는다”고 강조했습니다. 이는 중동 지역의 긴장감을 한층 고조시키는 발언으로 해석됩니다.
에너지 위기 해결을 위한 백악관의 고심도 깊어지고 있습니다. 미 재무장관 스콧 베산트(Scott Bessant)는 향후 이란산 원유에 대한 제재를 해제할 계획이라고 밝혔습니다. 이는 약 1억 4천만 배럴에 달하는 원유가 시장에 풀릴 수 있음을 의미하며, 수년간 지속된 미국의 대이란 정책을 뒤집는 결정입니다. 백악관은 유가 안정을 위해 필사적인 노력을 기울이고 있지만, 이란과의 전쟁 중에도 이란이 더 많은 원유를 판매하도록 장려하는 역설적인 상황을 초래할 수 있다는 지적이 나옵니다.
한편, 미국 국방부는 전쟁 지속을 위해 2천억 달러에 달하는 막대한 추가 예산을 의회에 요청했습니다. 이는 연간 전체 국방 예산의 거의 4분의 1에 해당하는 규모로, 국방장관 피트 헉스(Pete Hgse)는 이 금액이 변경될 수 있음을 시사했습니다. 이미 분열된 여론 속에서 이 막대한 예산 요청은 의회 내 온건 공화당 의원들의 우려를 사고 있습니다.
이러한 격동 속에서 한편으로는 실리콘밸리의 기술 기업들이 새로운 기회를 포착하고 있습니다. 과거 실리콘밸리에서 국방 기술은 ‘악해지지 말라(Do No Evil)‘는 구글의 모토처럼 기피 대상이었습니다. 구글 직원들은 국방부와의 협력에 반대하는 시위를 벌이기도 했습니다. 그러나 몇 년 전부터 이러한 분위기는 급변했습니다. 구글, 메타(Meta), OpenAI 같은 거대 기업부터 소규모 스타트업에 이르기까지 수많은 기술 기업들이 국방 기술 분야로 방향을 전환하기 시작했습니다.
이들 기업은 수억에서 수십억 달러에 이르는 계약을 통해 무기 시스템부터 미 정부를 위한 소프트웨어까지 다양한 기술을 개발하고 있습니다. 예를 들어, 데이터 분석 기업 팔란티어(Palantir)는 미 정부가 공습 목표를 선정하는 데 도움을 주는 ‘프로젝트 메이븐(Project Maven)’ 시스템을 구축했습니다. 구글과 OpenAI는 인공지능(AI) 기술을 활용해 현장 지휘관들을 지원하고 있으며, 구글 전 최고경영자 에릭 슈미트(Eric Schmidt)가 주도하는 프로젝트는 이란에서 미군 자산이 이란 드론에 방어하는 데 사용되는 대드론 시스템을 개발하고 있습니다. 트럼프 행정부가 1조 달러 이상의 국방비를 할당했으며, 이 중 상당 부분이 실리콘밸리가 개발하고자 하는 새로운 국방 기술에 투입될 것으로 예상되면서, 기술 기업들의 ‘전쟁 특수’는 더욱 커질 전망입니다.
트럼프 가문의 이색 투자처, 트란실바니아의 미스터리
뉴욕타임스의 새로운 조사에 따르면, 트럼프 전 대통령의 가족 사업체인 트럼프 조직이 루마니아에서 미발표 부동산 거래를 추진 중인 것으로 드러났습니다. 발리나 몰디브처럼 화려하고 유명한 관광지가 아닌, 트란실바니아라는 의외의 장소입니다.
현장을 직접 방문한 뉴욕타임스 탐사 보도 기자 레베카 루이즈(Rebecca Ruiz)는 그곳의 풍경을 생생하게 묘사했습니다. 부지는 반쯤 버려진 상태였고, 군사 기지 옆에 위치했으며, 이 지역 최대 규모의 거대한 쓰레기 매립지 몇 곳과 인접해 있었습니다. 코를 찌르는 악취가 진동했고, 들개 무리가 배회하는 모습도 목격되었습니다. 이러한 환경은 럭셔리 아파트 단지와 골프 코스를 상상하기에는 매우 이례적인 장소였습니다.
트럼프 조직이 왜 이 특정 장소를 프로젝트 부지로 선택했는지는 명확하지 않습니다. 그러나 한 가지 단서는 트럼프 전 대통령이 루마니아에서 높은 인기를 누리고 있다는 점입니다. 지난해 갤럽(Gallup) 여론조사에 따르면, 루마니아 국민의 절반 이상이 트럼프 대통령의 직무 수행에 대해 긍정적으로 평가했으며, 이는 유럽연합(EU) 국가들 사이에서 이례적인 결과입니다.
레베카 루이즈 기자는 이러한 움직임이 트럼프 전 대통령의 두 번째 임기에서 나타난 패턴의 연장선상에 있다고 분석합니다. 즉, 트럼프 조직은 그가 인기 있는 지역을 찾아 전 세계를 여행하며 사업 기회를 모색하고 있다는 것입니다. 발리와 몰디브 같은 전통적인 고급 휴양지와는 거리가 먼 트란실바니아에서 트럼프 조직이 어떤 그림을 그리고 있는지에 대한 의문은 여전히 남아있습니다.
격동의 시대, 변화의 물결
이란과의 전쟁과 그로 인한 실리콘밸리의 부상, 그리고 트럼프 가문의 예상치 못한 투자처는 오늘날 세계가 얼마나 빠르게 변화하고 예측 불가능한 방향으로 나아가고 있는지를 보여줍니다. 전쟁이 기술 기업들에게 새로운 시장을 열어주고, 정치적 인기가 비즈니스 기회로 전환되는 이 시대의 흐름은 앞으로도 세계 정치경제에 중대한 영향을 미칠 것입니다. 뉴욕타임스 웹사이트(nytimes.com)에서 트럼프 조직의 트란실바니아 프로젝트에 대한 전체 조사 내용과 현장 사진을 확인할 수 있습니다.