YouTube Digest

March 12, 2026

Listen to this issue

English

Based on โ€œNo Legal Background. Then He Built a $5.5B AI Company | Legora, Max Junestrandโ€ from EO Watch the original video

The Gamerโ€™s Gambit: How a Non-Lawyer Built a $5.5 Billion Legal AI Empire

In the world of high-stakes corporate law, the โ€œgrindโ€ is usually measured in billable hours, leather-bound case files, and late nights in mahogany-rowed offices. But for Max Junestrand, the co-founder and CEO of Leya, the grind started in a much different environment: the high-intensity, pixelated battlegrounds of Dota 2.

Junestrand doesnโ€™t have a law degree. He didnโ€™t spend his twenties clerking for judges or mastering the nuances of tort reform. Yet, in just two years, he has scaled Leya from three engineers in a Stockholm office to a $5.5 billion AI powerhouse with over 750 clients and 300 employees. His journey is a masterclass in how โ€œoutsiderโ€ perspectivesโ€”combined with a gamerโ€™s obsession with optimizationโ€”can disrupt one of the worldโ€™s most traditional industries.

From the Swedish Archipelago to the Digital Arena

Growing up in the quiet Swedish archipelago, Junestrand felt a void where competition should be. He sought that missing intensity in two places: sports and video games. It was in Dota 2, a complex multiplayer battle arena game, where he learned the foundational lessons of his leadership style.

โ€œGaming taught me that if you want to be great, you have to grind,โ€ Junestrand explains. โ€œItโ€™s not enough to just play more hours. Youโ€™ve got to play smarter and learn more. You have to bring a level of intensity better than anyone else.โ€

In Dota, success depends on five players collaborating with seamless synchronization. As a CEO, Junestrand views his role similarlyโ€”not as the smartest person in the room, but as the conductor who makes the โ€œband play together.โ€ This philosophy of the โ€œinfinite gameโ€ has allowed Leya to move at a pace that has left traditional legal tech incumbents reeling.

The โ€œJaw-Dropโ€ Moment: Solving the Right Problems

When Junestrand looked at the legal industry, he didnโ€™t see a barrier to entry; he saw a goldmine of โ€œmeaningful problems.โ€ The vision for Leya was simple yet audacious: create the โ€œ10x lawyerโ€ by using AI to automate the soul-crushing administrative tasks that keep attorneys from actual legal thinking.

The companyโ€™s early momentum was sparked during an incubation period at Mannheimer Swartling, Swedenโ€™s premier law firm. Instead of building in a vacuum, Junestrand and his team sat in a borrowed conference room, working side-by-side with their future users.

He recalls the pivotal moment that proved they were onto something huge: โ€œI would ask users to take an old memo or part of a case they had worked on, upload it, and ask Leya to do a first draft. When they saw the results, it was a โ€˜leaning back, drop the chinโ€™ moment.โ€

That visceral reactionโ€”the realization that hours of work could be condensed into secondsโ€”provided the fuel to scale.

The San Francisco Leap and the โ€œRing Lightโ€ Era

As Leya began to dominate the European market, Junestrand faced a crossroads: stay in the comfortable European ecosystem or move to the โ€œbig leaguesโ€ in the United States. Despite having a thriving business in Europe, he chose the latter, joining the prestigious Y Combinator (YC) accelerator in San Francisco.

The transition wasnโ€™t easy. To keep the European engine running while building a footprint in the U.S., Junestrand lived a double life. โ€œI bought a ring light to shine on my face when I took calls between 1:00 a.m. and 10:00 a.m. toward the other side of the world,โ€ he says.

This relentless pace paid off. During the YC program, Leya went from zero to nearly $1 million in Annual Recurring Revenue (ARR) in just six weeks. By the time they opened their Manhattan office, they had already secured partnerships with โ€œwhite shoeโ€ heavyweights like Cleary Gottlieb and Goodwin Procter.

Scaling the Unscalable: 3 to 300

Moving from a three-person startup to a 300-person global organization requires more than just good code; it requires a radical shift in the founderโ€™s mindset. Junestrandโ€™s approach to scaling is rooted in brutal self-discipline.

โ€œThe hardest part is that your role changes the fastest,โ€ he admits. โ€œYou cannot be married to the way things were done. You have to be your own best boss and be disciplined about not working on the easy things, but the really hard things.โ€

Even with 300 employees across Denver, New York, London, Stockholm, India, and Sydney, Junestrand remains deeply involved in the hiring process, still interviewing every person who joins the company. He looks for three specific traits:

  1. Capacity to grow: Can this person evolve as fast as the company?
  2. Problem-solving reflection: Can they look back at a hard problem and articulate how they would solve it better today?
  3. Ambition: Do they want to take personal responsibility for the companyโ€™s outcome?

The Human Impact of AI

While the $5.5 billion valuation and the aggressive quarterly growth are the metrics the media focuses on, Junestrand finds his motivation in a specific Slack channel titled โ€œCustomer Love.โ€

In a profession notorious for burnout and high divorce rates, the impact of Leya isnโ€™t just about efficiencyโ€”itโ€™s about time. โ€œThereโ€™s nothing as rewarding as when people let us know weโ€™ve made their day better,โ€ Junestrand says. โ€œThey were able to finish a piece of work and get home to their families to see their kids before they went to bed.โ€

The Journey to 100

For many, reaching unicorn status is the finish line. For Junestrand, itโ€™s merely the end of the warm-up. He views the current state of legal AI as the transition from โ€œ0 to 1.โ€ The real workโ€”the journey from โ€œ1 to 10โ€ and โ€œ10 to 100โ€โ€”is just beginning.

โ€œWe canโ€™t be fast followers. Weโ€™ve got to be the innovators,โ€ he asserts. As he looks toward 2027, he isnโ€™t just building a software tool; heโ€™s building the infrastructure for the future of law.

His advice for aspiring founders is a reflection of his own journey: โ€œYou need to be really excited about what youโ€™re doing. If you pick something you arenโ€™t excited about, the grind will be too hard. I wake up every day thinking about how we deliver more value.โ€

In the high-stakes game of legal tech, Max Junestrand is no longer just a player; heโ€™s rewriting the rules of the game.


Based on โ€œKarpathyโ€™s โ€œautoresearchโ€ explained in 24 minutesโ€ from Greg Isenberg Watch the original video

The Era of the Automated Scientist: Inside Andrej Karpathyโ€™s โ€œAuto-Researchโ€ Revolution

In the fast-moving world of artificial intelligence, a few names carry enough weight to stop the industry in its tracks. Andrej Karpathyโ€”founding member of OpenAI and former Director of AI at Teslaโ€”is one of them. When Karpathy releases a new project, it isnโ€™t just a code update; itโ€™s a signal of where the wind is blowing.

His latest obsession, a concept known as โ€œAuto-Research,โ€ has recently gone viral, capturing the imagination of Silicon Valleyโ€™s elite, including Shopify CEO Tobi Lรผtke. But what exactly is it, and why are people calling it the beginning of a new era in productivity?

At its core, auto-research is the realization of every entrepreneurโ€™s dream: a โ€œsuper-nerd robot internโ€ that runs science experiments on AI models all night long, while you sleep, without requiring a second of manual intervention.

The Anatomy of the Loop: How Auto-Research Works

To understand auto-research, you have to move away from the idea of AI as a chatbot and start seeing it as an agent. Traditional AI waits for a prompt; an auto-research agent follows a loop.

The process begins with a goal. You might tell the agent, โ€œMake this small AI model smarterโ€ or โ€œFind a way to reduce the cost of this code without losing accuracy.โ€ Once the goal is set, the agent takes over the driverโ€™s seat:

  1. Planning: The agent drafts a plan involving different settings, architectural changes, or data tweaks.
  2. Execution: It physically edits the Python code and initiates a training experiment on a GPU (Graphics Processing Unit).
  3. Evaluation: After running for a set period (perhaps five minutes or an hour), the agent reads the resulting metrics.
  4. Iteration: It asks itself, โ€œDid that work?โ€ If the results improved, it saves the changes. If they didnโ€™t, it logs the failure and tries a completely different path.

As Greg Isenberg, a prominent tech voice and entrepreneur, puts it: โ€œYou wake up, grab the best version, and turn it into something you charge for.โ€ It is the โ€œRalph Loopโ€ applied to the scientific methodโ€”engineering that happens 24/7.

From Code to Commerce: 10 Ways to Monetize Auto-Research

While Karpathyโ€™s initial focus was on optimizing AI models, the business community has quickly realized that this โ€œloopโ€ logic can be applied to almost any money-making endeavor. Here is how the next generation of startups will likely leverage auto-research:

1. The โ€œAgent in a Boxโ€ for Niche Markets

The most immediate opportunity lies in creating tiny, hyper-focused auto-research loops for specific industries. Imagine an โ€œAmazon Listing Experimenterโ€ that constantly tweaks product titles and descriptions to find the highest conversion rate, or an โ€œEmail Sequence Tunerโ€ for realtors that iterates on subject lines until the open rate hits a record high.

2. Autonomous A/B Testing

The days of manually setting up A/B tests in tools like Optimizely may be numbered. An auto-research agent can write dozens of variants of headlines, layouts, and offers, push them to live traffic, measure the clicks, and automatically promote the winner to the โ€œcontrolโ€ position.

3. Research as a Service (RaaS)

Startups and investors spend thousands of hours on due diligence and competitor analysis. An auto-research bot can be pointed at a market to constantly monitor pricing changes, feature updates, and regulatory shifts across hundreds of competitors, delivering a โ€œliving memoโ€ that is always fresh.

4. The โ€œOptimizeโ€ Button

For existing SaaS companies, auto-research offers a high-tier feature: a literal โ€œOptimizeโ€ button. A user could press it, and the system would run a mini-research loop in the background to tune the userโ€™s prompts, rank their suppliers, or pick their best pricing tiers based on historical data.

5. The 100x Agency

The agency of the future wonโ€™t pitch creativity; they will pitch volume. โ€œWe run 100 times more tests than any other shop for the same fee,โ€ the pitch goes. By using auto-research, an agency can test thousands of ad angles or landing page variables, charging a performance fee for the KPI lifts they achieve.

6. Auto-Quant Trading

In finance, auto-research can be used to run back-tests of simple trading rules overnight. By using Large Language Models (LLMs) to screen for sentiment or filter factors, traders can discard failing strategies and only put capital behind the ones that survived the automated โ€œsurvival of the fittestโ€ gauntlet.

7. Lead Qualification and CRM Autopilot

By connecting an auto-research agent to a CRM like Salesforce, businesses can test different follow-up rules and messages. The agent learns which leads are most likely to buy and drafts the perfect outreach, allowing sales teams to focus only on high-value closing calls.

8. Finance Ops and Invoice Matching

The loop can be applied to the โ€œgrunt workโ€ of accountingโ€”grinding through invoice matching and expense reports, and continuously improving the prompts used to detect exceptions and fraud.

9. Internal Productivity Labs

Large organizations can treat their own KPIsโ€”like ticket resolution time or sales close ratesโ€”as the โ€œmodelโ€ to be optimized. Agents can iterate on internal workflows and templates until the manual labor is stripped away, leaving only high-impact decisions for the humans.

10. Done-for-You Due Diligence

For M&A (Mergers and Acquisitions), auto-research can chew through thousands of pages of filings and product reviews to create structured briefs for investors, identifying risks that a human researcher might miss in a time-crunched environment.

The Human Impact: Beyond the Bottom Line

The excitement surrounding Karpathyโ€™s work isnโ€™t just about profit; itโ€™s about the speed of human progress. Tech entrepreneur Morgan Linton recently shared a vision of how this could transform medicine.

โ€œIt feels like in many ways clinical trial design is itself a kind of hyperparameter search,โ€ Linton noted. Currently, trials cost tens of millions of dollars. An auto-research โ€œagent swarmโ€ could optimize treatment protocols on small proxy experiments, promoting the most promising candidates to human doctors for review. This could potentially shave years off the time it takes to bring life-saving drugs to market.

Furthermore, Karpathy has already launched AgentHub, which he describes as โ€œGitHub for agents.โ€ While GitHub is a place for humans to collaborate on code, AgentHub is designed for a โ€œswarm of agentsโ€ to work on the same codebase simultaneously, coordinating their efforts through message boards.

How to Join the Movement

If youโ€™re looking to get started with auto-research, there is a small technical hurdle: you need โ€œcompute.โ€ Specifically, you need an NVIDIA GPU. While high-end chips like the H100 are the gold standard, you donโ€™t need to own the hardware.

Services like Google Colab, Lambda Labs, and Vast AI allow you to rent these chips in the cloud for a few dollars. By using package managers like uv and following the documentation on Karpathyโ€™s GitHub (which has already amassed over 25,000 stars), even those with a standard MacBook can begin running their own experiments.

The Bottom Line

We are witnessing a shift from โ€œSoftware as a Serviceโ€ to โ€œExperiments as a Service.โ€ In the old world, you hired a person to find an answer. In the new world, you hire an agent to find the best answer through thousands of iterations.

As Andrej Karpathy continues to โ€œspeedrunโ€ the creation of this new infrastructure, the message to the rest of the world is clear: The loop is running. Those who learn to set the goals for these agents will lead the next decade of innovation. Those who donโ€™t may find themselves replaced by a process that never sleeps.


Based on โ€œHow Figma engineers sync designs with Claude Code and Codexโ€ from How I AI Watch the original video

The Great Collapse: How AI is Erasing the Border Between Design and Code

For decades, the relationship between software designers and engineers has followed a rigid, almost geological progression. It started with low-fidelity sketches, moved into โ€œgrayscaleโ€ wireframes, graduated to high-fidelity mockups, and finallyโ€”after weeks of meetings and handoffsโ€”arrived at the โ€œexpensiveโ€ stage: the code.

But at Figma, the team behind the worldโ€™s most popular design tool, that linear timeline is imploding.

In a recent deep dive on the How I AI podcast, Figma engineer Alex and designer Ghee sat down with product leader Claire Vo to demonstrate a new reality. By leveraging AI agents like Claude Code and Codex, alongside the Model Context Protocol (MCP), they are turning the โ€œhandoffโ€ into a continuous, circular loop.

The result? A world where code can be instantly turned back into design, and design can be programmatically reconciled into a codebase.

The Problem of โ€œDesign Driftโ€

Every product team knows the โ€œSource of Truthโ€ problem. A designer creates a pixel-perfect sign-up flow in Figma. The engineer builds it. But then, a week later, a bug is fixed in production, a new error state is added in the IDE, or a font size is tweaked during a late-night coding session.

Suddenly, the Figma file is a relic of the past, and the codebase is the only place where the actual product lives.

โ€œOften times the codebase gets way ahead of where the actual design file is,โ€ Alex explains. โ€œThere are states or workflows that just donโ€™t exist at all within the design file.โ€

In the old world, the designer would have to manually take screenshots of the live app and painstakingly recreate them in Figma just to start a new iteration. Today, Ghee and Alex use AI as a bridge. Using the Figma MCP (a connector that allows AI models to โ€œreadโ€ and โ€œwriteโ€ to Figma), Ghee can simply tell an AI agent: โ€œSend the budget allocation page from my local host to Figma.โ€

Within seconds, the live, functional code is translated into editable Figma layers. The designer isnโ€™t looking at a static screenshot; they are looking at a living reflection of the code, ready to be manipulated.

The โ€œDecadent Momentโ€ of Product Design

Ghee describes this shift as a โ€œdecadent momentโ€ in the history of software. Historically, design was a game of scarcity. Because engineering time was so expensive, designers had to โ€œde-riskโ€ every idea through endless meetings and wireframes before a single line of code was written.

โ€œAI basically collapsed that,โ€ Ghee says. โ€œItโ€™s just as cheap to riff in code as it is to riff in design. We can get functional wireframes straight away.โ€

This collapse of the fidelity ladder allows teams to enter an era of โ€œfeature abundance.โ€ Previously, small polish items or โ€œP2โ€ bugs were often left on the cutting room floor because they werenโ€™t worth the cost of the design-to-code pipeline. Now, when the cost of translation drops to near zero, teams can focus on the โ€œcraftโ€โ€”the small details that make a product feel magical.

โ€œItโ€™s the era of โ€˜Yes,โ€™โ€ says Claire Vo. โ€œThe idea of the cut line has to disappear. We can put the polish and all the things that make these experiences great back into the product.โ€

Pair Programming for the Visual Era

One of the most impressive demonstrations involved the โ€œRound Trip.โ€ Alex showed how he could take a sign-up flow with five different states (error, success, loading, etc.) that existed only in his code and push them all into Figma simultaneously.

Instead of a designer asking, โ€œWhat happens if the user enters the wrong password?โ€ and waiting for an engineer to find the logic, the engineer simply exports those states into the design canvas.

The collaboration then becomes truly synchronous. While the engineer is working in the terminal, the designer is in the Figma file, โ€œfreestylingโ€ with colors, spacing, and โ€œvibes.โ€

โ€œThis feels like pair programming for designers and engineers,โ€ Vo noted.

Once the designer makes a changeโ€”say, adjusting a hex code or moving a buttonโ€”Alex doesnโ€™t have to manually inspect the CSS. He uses Claude Code to โ€œbring the changes from this Figma component into my codebase.โ€ The AI compares the design file to the repository, writes the necessary React or CSS patches, and applies them. The engineer becomes a reviewer of the AIโ€™s implementation rather than a manual translator of pixels.

Beyond the Canvas: AI โ€œSkillsโ€ and Engineering Toil

The transformation isnโ€™t limited to visual design. Alex also shared how he uses โ€œSkillsโ€โ€”essentially custom AI macrosโ€”to eliminate the โ€œtoilโ€ of engineering.

In a large repository like Figmaโ€™s, the process of โ€œshippingโ€ code is fraught with mechanistic hurdles: running linters, checking CI/CD pipelines, ensuring tests pass, and managing โ€œBasilโ€ build commands.

Alex created a /ship skill. Now, instead of โ€œbabysittingโ€ a pull request for an hour, he types one command and walks away. The AI agent runs the pre-flight checks, monitors the build, and even automatically fixes minor linting issues that might have caused the build to fail.

โ€œIt turns processes that would otherwise be based on best intentionsโ€ฆ into something that can be fully automated,โ€ Alex says. Itโ€™s a shift from reading a static โ€œOnboarding Wikiโ€ to interacting with a โ€œLiving Skillโ€ that knows the companyโ€™s best practices.

The Human in the Loop: Fine Motor Skills and โ€œVibesโ€

If AI can translate design to code and back again, what is left for the humans?

According to Ghee, the human role is moving โ€œupstream.โ€ Instead of spending hours on the โ€œmechanisticโ€ work of drawing boxes or threading data, designers and engineers are focusing on problem-solving and โ€œeyeballing.โ€

โ€œNo one wants to prompt for the exact hex code or the shade of yellow,โ€ Ghee points out. โ€œThatโ€™s just easier to just quickly do and directly manipulate.โ€

The future of the designer-engineer relationship isnโ€™t about one replacing the other; itโ€™s about using AI to handle the โ€œstructured dataโ€ translation so that humans can focus on the โ€œunstructuredโ€ beauty. As Ghee puts it, humans are there to decide if a specific shade of yellow feels โ€œhappyโ€ or if a workflow feels โ€œintuitive.โ€

In this new workflow, the question of โ€œWhich comes first, the design or the code?โ€ finally has an answer: It doesnโ€™t matter. In the loop of the future, they both happen at once.


Based on โ€œThe Department of War is making a huge mistake.โ€ from Dwarkesh Patel Watch the original video

The Silicon Standoff: Why the Pentagonโ€™s War with Anthropic Is a Battle for the Future of Freedom

In the quiet corridors of power and the glass-walled offices of Silicon Valley, a confrontation is unfolding that may determine the character of the 21st century. It began with a disagreement over โ€œred linesโ€ and has escalated into what some are calling a โ€œwarning shotโ€ for the future of private enterprise and civil liberties.

The Department of War (the Pentagon) recently declared Anthropic, one of the worldโ€™s leading AI labs, a โ€œsupply chain risk.โ€ The reason? Anthropic refused to remove contractual restrictionsโ€”its so-called โ€œred linesโ€โ€”that prohibit the use of its AI models for mass surveillance and the development of autonomous weapons.

On the surface, it looks like a standard contract dispute between a defense agency and a picky vendor. But as thinker and podcaster Dwarkesh Patel argues, this episode reveals a much deeper, more chilling reality: the government is beginning to treat the creators of artificial intelligence not as partners, but as subjects.

The Kill Switch Dilemma

To understand the gravity of the situation, one must first acknowledge the governmentโ€™s perspective. If you are the Secretary of War, you cannot allow a private contractor to hold a โ€œkill switchโ€ over mission-critical technology.

Patel offers a hypothetical: โ€œImagine if Elon Musk, negotiating Starlink access for the military, reserved the right to cut off access if he deemed a war โ€˜unjust.โ€™ As a military, you simply cannot give a private contractor that kind of leverage over a technology youโ€™ve come to rely on.โ€

If the government had simply chosen not to do business with Anthropic, the story would end there. But they went further. By designating Anthropic a โ€œsupply chain risk,โ€ the government has essentially threatened to destroy the company. This designation forces massive partners like Amazon, Google, and Nvidia to ensure Anthropicโ€™s technology doesnโ€™t touch any of their Pentagon-related work. In a future where AI is woven into the very fabric of cloud computing and chip design, such a โ€œcordon sanitaireโ€ could make it impossible for Anthropic to exist as a private business.

A Civilization Built on AI Labor

The stakes of this standoff are heightened by the timeline of AI development. We are not just talking about a smarter version of a chatbot. Within 20 years, it is estimated that 99% of the workforceโ€”in the military, the civilian government, and the private sectorโ€”will be AI.

โ€œTheyโ€™re going to be the robot armies that constitute our military,โ€ Patel notes. โ€œTheyโ€™re going to be the superhumanly intelligent advisers that senators and presidents and CEOs have. Theyโ€™re going to be the police. Our future civilization is going to be run on AI labor.โ€

If the government establishes a precedent today that it can bully AI companies into dropping ethical safeguards, it isnโ€™t just winning a contract; it is ensuring that the entire future infrastructure of human society is built on โ€œblindly obedientโ€ silicon employees.

The End of the Surveillance Bottleneck

One of the most immediate concerns regarding this โ€œblind obedienceโ€ is mass surveillance. Historically, the Fourth Amendment and our general sense of privacy have been protected by a practical bottleneck: human manpower.

The government may legally be allowed to buy your data from third parties, but no agency has the staff to watch every CCTV camera or read every email. AI changes that math forever.

Patel highlights a staggering calculation: There are roughly 100 million CCTV cameras in America. With current AI processing costs, it would cost about $30 billion to process every single frame of every camera in the country. But AI capability gets roughly 10x cheaper every year. By 2030, it will be less expensive to monitor every nook and cranny of the United States than it is to remodel the White House.

โ€œOnce the technical capacity for mass surveillance and political suppression exists,โ€ Patel warns, โ€œthe only thing that stands between us and an authoritarian state is the political expectation that this is just not something we do here.โ€

The Virtue of Disobedience

This brings us to the heart of the โ€œalignmentโ€ problem. In the AI safety community, โ€œalignmentโ€ usually means getting an AI to follow a humanโ€™s instructions perfectly. But what if those instructions are evil?

History suggests that human civilization has often been saved not by obedience, but by its opposite. In 1983, Soviet Lieutenant Colonel Stanislav Petrov was on duty when his early-warning system signaled that the U.S. had launched five nuclear missiles. Protocol demanded he report it, which would have triggered a massive Soviet retaliatory strike. Petrov judged it a false alarm and refused to follow orders. He was right. If he had been a โ€œperfectly alignedโ€ AI, hundreds of millions would have died.

Similarly, the Berlin Wall fell in 1989 because East German border guards eventually refused to fire on their fellow citizens. โ€œOne personโ€™s virtue is another personโ€™s misalignment,โ€ Patel observes. If the government succeeds in forcing AI companies to strip away their moral โ€œred lines,โ€ we are effectively building a world where there is no Stanislav Petrovโ€”only a perfectly obedient machine executing a command, no matter how catastrophic.

The Regulatory Bazooka

Ironically, Anthropic has been one of the loudest voices calling for government regulation of AI. They have argued that the risks of AIโ€”bioweapons, cyberattacks, and โ€œautonomy riskโ€โ€”are so great that we need a regulatory body similar to the Nuclear Regulatory Commission.

However, the current spat with the Pentagon suggests that such a regulatory framework might be a โ€œfully loaded bazookaโ€ handed to a power-hungry leader. Terms like โ€œnational security threatโ€ or โ€œcatastrophic riskโ€ are notoriously vague. If a model refuses to help the government with mass surveillance, a future administration could easily label that an โ€œautonomy riskโ€ or a โ€œthreat to national security.โ€

The government is already using the Defense Production Act (a 1950s law meant for steel mills) and supply chain statutes (meant to keep out Chinese hardware) to pressure AI labs. Giving them a purpose-built AI regulatory apparatus might only accelerate this โ€œthuggishโ€ political pressure.

The Way Forward: Regulate Uses, Not Intelligence

Is there a middle ground? Patel suggests we look to the Industrial Revolution. Like AI, industrialization was a general-purpose transformation of society that created terrifying new weapons. Yet, free societies didnโ€™t handle it by giving the government absolute control over every factory. Instead, they regulated specific usesโ€”banning chemical weapons or regulating how the government can use industrial power against its citizens.

โ€œWe should regulate AI in a similar way,โ€ Patel argues. โ€œWe should regulate specific destructive use casesโ€ฆ and we should also have laws which regulate how the government can use this technology.โ€

Ultimately, corporate courage from companies like Anthropic is a noble but temporary fix. As AI technology diffuses and open-source models catch up to the frontier, the government will eventually find a vendor willing to say โ€œyesโ€ to mass surveillance.

The only lasting protection for a free society is the creation of robust political norms and laws. We must decide, as a civilization, that some things are off-limitsโ€”even for the most powerful government on Earth. If we fail to do that, we arenโ€™t just racing to beat our adversaries in the AI race; we are racing to become them.


ํ•œ๊ตญ์–ด

โ€œNo Legal Background. Then He Built a $5.5B AI Company | Legora, Max Junestrandโ€ โ€” EO ๊ธฐ๋ฐ˜ ๊ธฐ์‚ฌ ์›๋ณธ ์˜์ƒ ๋ณด๊ธฐ

๋„ํƒ€ 2 ๊ฒŒ์ด๋จธ์—์„œ 7์กฐ ์› ๊ฐ€์น˜ AI ์œ ๋‹ˆ์ฝ˜ CEO๋กœ: โ€˜๋ ˆ์•ผ(Leya)โ€™ ๋งฅ์Šค ์œ ๋„ค์ŠคํŠธ๋ž€๋“œ์˜ ๋ฉˆ์ถ”์ง€ ์•Š๋Š” ๋„์ „

๋‹จ 2๋…„ ๋งŒ์— ์Šคํ†กํ™€๋ฆ„์˜ ์—”์ง€๋‹ˆ์–ด 3๋ช…์—์„œ ์ „ ์„ธ๊ณ„ 300๋ช…์˜ ์ง์›์„ ๊ฑฐ๋А๋ฆฐ ์œ ๋‹ˆ์ฝ˜ ๊ธฐ์—…์œผ๋กœ ์„ฑ์žฅํ•œ ์Šคํƒ€ํŠธ์—…์ด ์žˆ๋‹ค. ๋ฒ•๋ฅ  ์ง€์‹์ด ์ „ํ˜€ ์—†๋˜ ํ•œ ์ฒญ๋…„์€ ์–ด๋–ป๊ฒŒ ๋ณด์ˆ˜์ ์ธ ๋ฒ•๋ฅ  ์‹œ์žฅ์„ AI๋กœ ํ˜์‹ ํ•˜๊ณ , 55์–ต ๋‹ฌ๋Ÿฌ(์•ฝ 7.5์กฐ ์›) ๊ฐ€์น˜์˜ ๊ธฐ์—…์„ ์ผ๊ถˆ๋ƒˆ์„๊นŒ? ๋ฒ•๋ฅ  ์—…๋ฌด์˜ ํŒจ๋Ÿฌ๋‹ค์ž„์„ ๋ฐ”๊พธ๊ณ  ์žˆ๋Š” โ€˜๋ ˆ์•ผ(Leya)โ€™์˜ ๊ณต๋™ ์ฐฝ์—…์ž์ด์ž CEO, ๋งฅ์Šค ์œ ๋„ค์ŠคํŠธ๋ž€๋“œ(Max Junestrand)์˜ ์—ฌ์ •์„ ์‹ฌ์ธต ๋ถ„์„ํ–ˆ๋‹ค.


1. ๊ฒฝ์Ÿ์„ ์ฆ๊ธฐ๋Š” ๊ฒŒ์ด๋จธ, ๋น„์ฆˆ๋‹ˆ์Šค์˜ ์ „์žฅ์— ์„œ๋‹ค

๋งฅ์Šค ์œ ๋„ค์ŠคํŠธ๋ž€๋“œ์˜ ์„ฑ์žฅ์€ ์Šค์›จ๋ด์˜ ์กฐ์šฉํ•œ ๊ตฐ๋„ ๋งˆ์„์—์„œ ์‹œ์ž‘๋˜์—ˆ๋‹ค. ์ •์ ์ธ ํ™˜๊ฒฝ์—์„œ ์ž๋ž€ ๊ทธ์—๊ฒŒ ๊ฒฐํ•๋˜์—ˆ๋˜ ๊ฒƒ์€ ๋ฐ”๋กœ โ€˜๊ฒฝ์Ÿโ€™์ด์—ˆ๋‹ค. ๊ทธ๋Š” ๊ทธ ๊ฐˆ์ฆ์„ ์Šคํฌ์ธ ์™€ ๋น„๋””์˜ค ๊ฒŒ์ž„, ํŠนํžˆ ์ „๋žต ๊ฒŒ์ž„์ธ โ€˜๋„ํƒ€ 2(Dota 2)โ€™๋ฅผ ํ†ตํ•ด ํ•ด์†Œํ–ˆ๋‹ค.

๊ทธ๋Š” ๊ฒŒ์ด๋ฐ ๊ฒฝํ—˜์ด ์ž์‹ ์˜ ๋ฆฌ๋”์‹ญ๊ณผ ๋น„์ฆˆ๋‹ˆ์Šค ์ฒ ํ•™์— ๊ทผ๊ฐ„์ด ๋˜์—ˆ๋‹ค๊ณ  ๋งํ•œ๋‹ค. โ€œ๋„ํƒ€ 2๋Š” ๋‹จ์ˆœํžˆ ์‹œ๊ฐ„์„ ๋งŽ์ด ํˆฌ์žํ•œ๋‹ค๊ณ  ์ด๊ธธ ์ˆ˜ ์žˆ๋Š” ๊ฒŒ์ž„์ด ์•„๋‹™๋‹ˆ๋‹ค. ๋” ๋˜‘๋˜‘ํ•˜๊ฒŒ ํ”Œ๋ ˆ์ดํ•˜๊ณ , ๋ˆ„๊ตฌ๋ณด๋‹ค ๋†’์€ ๋ชฐ์ž…๋„(Intensity)๋ฅผ ์œ ์ง€ํ•˜๋ฉฐ ๋Š์ž„์—†์ด ๋ฐฐ์›Œ์•ผ ํ•˜์ฃ .โ€

ํŠนํžˆ 5๋ช…์ด ํ•œ ํŒ€์ด ๋˜์–ด ์™„๋ฒฝํ•œ ํ˜‘์—…์„ ์ด๋ค„์•ผ ํ•˜๋Š” ๊ฒŒ์ž„์˜ ํŠน์„ฑ์€ ๊ทธ์—๊ฒŒ **โ€˜ํŒ€์›Œํฌ์˜ ๋ณธ์งˆโ€™**์„ ๊ฐ€๋ฅด์ณ ์ฃผ์—ˆ๋‹ค. ๋งฅ์Šค๋Š” CEO์˜ ํ•ต์‹ฌ ์—ญํ• ์„ โ€œ์ž์‹ ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ๊ฐ ๋ถ„์•ผ์˜ ์ „๋ฌธ๊ฐ€๋“ค์„ ๋ชจ์œผ๊ณ , ๊ทธ๋“ค์ด ํ•˜๋‚˜์˜ ๋ฐด๋“œ์ฒ˜๋Ÿผ ์กฐํ™”๋กญ๊ฒŒ ์—ฐ์ฃผํ•˜๊ฒŒ ๋งŒ๋“œ๋Š” ๊ฒƒโ€์ด๋ผ๊ณ  ์ •์˜ํ•œ๋‹ค. ๊ทธ๋Š” ์ง€๊ธˆ๋„ ๋น„์ฆˆ๋‹ˆ์Šค๋ฅผ ๋‹จ๊ธฐ์ ์ธ ์ŠนํŒจ๊ฐ€ ์•„๋‹Œ, ๋Š์ž„์—†์ด ํ™•์žฅ๋˜๋Š” โ€˜๋ฌดํ•œ ๊ฒŒ์ž„(Infinite Game)โ€˜์œผ๋กœ ๋ฐ”๋ผ๋ณธ๋‹ค.

2. ๋ฒ•๋ฅ  ์‹œ์žฅ์˜ โ€˜0์—์„œ 1โ€™์„ ๋„˜์–ด โ€˜10๋ฐฐ์˜ ๋ณ€ํ˜ธ์‚ฌโ€™๋ฅผ ํ–ฅํ•ด

๋ ˆ์•ผ(Leya)๋Š” ๋‹จ์ˆœํ•œ AI ๋„๊ตฌ๊ฐ€ ์•„๋‹ˆ๋‹ค. ๋งฅ์Šค๋Š” ๋ ˆ์•ผ๋ฅผ **โ€œ๋ฒ•๋ฅ  ์—…๋ฌด๊ฐ€ ์‹ค์ œ๋กœ ์ผ์–ด๋‚˜๋Š” ํ”Œ๋žซํผโ€**์ด๋ผ๊ณ  ์†Œ๊ฐœํ•œ๋‹ค. ์ด๋“ค์˜ ๋ฏธ์…˜์€ ๋ช…ํ™•ํ•˜๋‹ค. ๋ฐ”๋กœ **โ€˜10๋ฐฐ ๋” ์ƒ์‚ฐ์ ์ธ ๋ณ€ํ˜ธ์‚ฌ(10x Lawyers)โ€˜**๋ฅผ ๋งŒ๋“œ๋Š” ๊ฒƒ์ด๋‹ค.

๋ฒ•๋ฅ  ์ง€์‹์ด ์—†๋˜ ๊ทธ๊ฐ€ ์ด ์‹œ์žฅ์— ๋›ฐ์–ด๋“  ์ด์œ ๋Š” ์—ญ์„ค์ ์œผ๋กœ โ€˜ํ•ด๊ฒฐํ•ด์•ผ ํ•  ๋ฌธ์ œ๊ฐ€ ๋„ˆ๋ฌด๋‚˜ ๋งŽ์•˜๊ธฐ ๋•Œ๋ฌธโ€™์ด๋‹ค. ๊ทธ๋Š” ์Šค์›จ๋ด ์ตœ๋Œ€ ๋กœํŽŒ์ธ โ€˜๋งŒํ•˜์ด๋จธ ์Šค์™€ํ‹€๋ง(Mannheimer Swartling)โ€˜์˜ ํšŒ์˜์‹ค ํ•œํŽธ์—์„œ ์‚ฌ์—…์„ ์‹œ์ž‘ํ–ˆ๋‹ค. ์‹ค์ œ ๋ณ€ํ˜ธ์‚ฌ๋“ค์ด ์ผํ•˜๋Š” ํ˜„์žฅ์—์„œ ๊ทธ๋“ค์˜ ๊ณ ์ถฉ์„ ์ง์ ‘ ๋“ฃ๊ณ  ์ œํ’ˆ์„ ๋‹ค๋“ฌ์—ˆ๋‹ค.

์ฒ˜์Œ ๋ณ€ํ˜ธ์‚ฌ๋“ค์—๊ฒŒ ๋ ˆ์•ผ์˜ ๋ฐ๋ชจ๋ฅผ ๋ณด์—ฌ์คฌ์„ ๋•Œ์˜ ๋ฐ˜์‘์€ ๊ฐ€ํžˆ ํญ๋ฐœ์ ์ด์—ˆ๋‹ค. ๋ณ€ํ˜ธ์‚ฌ๊ฐ€ ๊ณผ๊ฑฐ์— ์ž‘์„ฑํ–ˆ๋˜ ๋ฉ”๋ชจ๋‚˜ ์‚ฌ๊ฑด ์ž๋ฃŒ๋ฅผ ์—…๋กœ๋“œํ•˜๊ณ  ์ดˆ์•ˆ ์ž‘์„ฑ์„ ๋ช…๋ นํ–ˆ์„ ๋•Œ, AI๊ฐ€ ์ˆœ์‹๊ฐ„์— ๊ฒฐ๊ณผ๋ฌผ์„ ๋‚ด๋†“๋Š” ๊ฒƒ์„ ๋ณธ ์ „๋ฌธ๊ฐ€๋“ค์€ ๊ฒฝ์•…์„ ๊ธˆ์น˜ ๋ชปํ–ˆ๋‹ค. ๋งฅ์Šค๋Š” ์ด โ€˜ํ„ฑ์ด ๋น ์งˆ ๋“ฏํ•œ ๋†€๋ผ์›€(Jaw-dropping moment)โ€˜์˜ ๊ฒฝํ—˜์ด ์„ฑ์žฅ์˜ ๊ฐ•๋ ฅํ•œ ๋™๋ ฅ์ด ๋˜์—ˆ๋‹ค๊ณ  ํšŒ์ƒํ•œ๋‹ค.

3. โ€œ๋น…๋ฆฌ๊ทธ์—์„œ ๋›ฐ๋ ค๋ฉด ๋ฏธ๊ตญ์œผ๋กœ ๊ฐ€๋ผโ€ : Y Combinator์™€ ๊ธ€๋กœ๋ฒŒ ํ™•์žฅ

๋ ˆ์•ผ๋Š” ์œ ๋Ÿฝ์—์„œ ์ด๋ฏธ ํƒ„ํƒ„ํ•œ ๋น„์ฆˆ๋‹ˆ์Šค๋ฅผ ๊ตฌ์ถ•ํ•˜๊ณ  ์žˆ์—ˆ์ง€๋งŒ, ๋งฅ์Šค๋Š” ์•ˆ์ฃผํ•˜์ง€ ์•Š์•˜๋‹ค. ๊ทธ๋Š” ์œ ๋Ÿฝ ๋ฒค์ฒ˜์บํ”ผํ„ธ(VC)์˜ ํˆฌ์ž ์ œ์•ˆ์„ ๋’ค๋กœํ•˜๊ณ  ์‹ค๋ฆฌ์ฝ˜๋ฐธ๋ฆฌ์˜ ์ƒ์ง•์ธ โ€˜Y ์ฝค๋น„๋„ค์ดํ„ฐ(Y Combinator, ์ดํ•˜ YC)โ€˜ํ–‰์„ ํƒํ–ˆ๋‹ค.

โ€œ์Šค์›จ๋ด์€ ํ›Œ๋ฅญํ•œ ์ƒํƒœ๊ณ„์ง€๋งŒ ์‹œ์žฅ์ด ์ž‘์Šต๋‹ˆ๋‹ค. ์ง„์ •ํ•œ ๋น…๋ฆฌ๊ทธ์—์„œ ์Šน๋ถ€ํ•˜๊ณ  ์‹ถ๋‹ค๋ฉด ๋ฏธ๊ตญ์œผ๋กœ ๊ฐ€์•ผ ํ–ˆ์ฃ .โ€

YC ๊ธฐ๊ฐ„ ๋™์•ˆ ๊ทธ๋Š” ์‚ด์ธ์ ์ธ ์Šค์ผ€์ค„์„ ์†Œํ™”ํ–ˆ๋‹ค. ์ƒŒํ”„๋ž€์‹œ์Šค์ฝ”์— ๋จธ๋ฌผ๋ฉด์„œ๋„ ์œ ๋Ÿฝ ๊ณ ๊ฐ๋“ค์„ ๋†“์น˜์ง€ ์•Š๊ธฐ ์œ„ํ•ด ์ƒˆ๋ฒฝ 1์‹œ๋ถ€ํ„ฐ ์˜ค์ „ 10์‹œ๊นŒ์ง€ ๋ง ๋ผ์ดํŠธ๋ฅผ ์ผœ๊ณ  ํ™”์ƒ ํšŒ์˜๋ฅผ ์ง„ํ–‰ํ–ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ง‘์š”ํ•จ ๋•๋ถ„์— ๋ ˆ์•ผ๋Š” YC ๊ธฐ๊ฐ„ ์ค‘ ๋‹จ ํ•œ ๋‹ฌ ๋ฐ˜ ๋งŒ์— ์—ฐ๊ฐ„ ๋ฐ˜๋ณต ๋งค์ถœ(ARR) 100๋งŒ ๋‹ฌ๋Ÿฌ(์•ฝ 13์–ต ์›)๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๊ธฐ์—ผ์„ ํ† ํ–ˆ๋‹ค.

์ดํ›„ ๋ ˆ์•ผ๋Š” ๋‰ด์š• ๋งจํ•ดํŠผ์— ์‚ฌ๋ฌด์‹ค์„ ์—ด๊ณ  ํด๋ฆฌ์–ด๋ฆฌ ๊ณ ํ‹€๋ฆฝ(Cleary Gottlieb), ๊ตฟ์œˆ ํ”„๋กํ„ฐ(Goodwin Procter)์™€ ๊ฐ™์€ ์„ธ๊ณ„์ ์ธ ๋กœํŽŒ๋“ค์„ ๊ณ ๊ฐ์‚ฌ๋กœ ํ™•๋ณดํ•˜๋ฉฐ ๋ฏธ๊ตญ ์‹œ์žฅ์— ์„ฑ๊ณต์ ์œผ๋กœ ์•ˆ์ฐฉํ–ˆ๋‹ค.

4. CEO์˜ ์—ญํ•  ๋ณ€ํ™”์™€ โ€˜์„ฑ์žฅ ์ž ์žฌ๋ ฅโ€™ ์ค‘์‹ฌ์˜ ์ฑ„์šฉ

ํšŒ์‚ฌ๊ฐ€ 3๋ช…์—์„œ 300๋ช…์œผ๋กœ ๊ธ‰์„ฑ์žฅํ•˜๋Š” ๊ณผ์ •์—์„œ ๋งฅ์Šค๋Š” CEO์˜ ์—ญํ• ์ด ๊ฐ€์žฅ ๋น ๋ฅด๊ฒŒ ๋ณ€ํ•ด์•ผ ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค์„ ๊นจ๋‹ฌ์•˜๋‹ค. ๊ทธ๋Š” โ€œ๊ณผ๊ฑฐ์˜ ๋ฐฉ์‹์— ์ง‘์ฐฉํ•ด์„œ๋Š” ์•ˆ ๋œ๋‹ค. ๋งค์ผ ์•„์นจ ์ผ์–ด๋‚˜ ์˜ค๋Š˜, ๊ทธ๋ฆฌ๊ณ  ๋‹ค์Œ ๋ถ„๊ธฐ์˜ ๋น„์ฆˆ๋‹ˆ์Šค๋ฅผ ์œ„ํ•ด ๋ฌด์—‡์„ ์ตœ์ ํ™”ํ•ด์•ผ ํ•˜๋Š”์ง€ ๊ณ ๋ฏผํ•ด์•ผ ํ•œ๋‹คโ€๊ณ  ๊ฐ•์กฐํ•œ๋‹ค.

๊ทธ๋Š” ์—ฌ์ „ํžˆ ๋ชจ๋“  ์‹ ์ž… ์‚ฌ์›์˜ ๋ฉด์ ‘์„ ์ง์ ‘ ๋ณธ๋‹ค. ๊ทธ๊ฐ€ ์ฑ„์šฉ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•˜๊ฒŒ ์—ฌ๊ธฐ๋Š” ๊ฐ€์น˜๋Š” ์„ธ ๊ฐ€์ง€๋‹ค.

  1. ์„ฑ์žฅ ์šฉ๋Ÿ‰(Capacity to grow): ์ด ์‚ฌ๋žŒ์ด ์•ž์œผ๋กœ ์–ผ๋งˆ๋‚˜ ๋” ์„ฑ์žฅํ•  ์ˆ˜ ์žˆ๋Š”๊ฐ€?
  2. ๋ฌธ์ œ ํ•ด๊ฒฐ ๋Šฅ๋ ฅ: ๊ณผ๊ฑฐ์— ๊ฒช์—ˆ๋˜ ๊ฐ€์žฅ ์–ด๋ ค์šด ๋ฌธ์ œ๋Š” ๋ฌด์—‡์ด๋ฉฐ, ๊ฑฐ๊ธฐ์„œ ๋ฌด์—‡์„ ๋ฐฐ์›Œ ๋‹ค๋ฅด๊ฒŒ ์ ์šฉํ–ˆ๋Š”๊ฐ€?
  3. ์•ผ๋ง(Ambition): ํšŒ์‚ฌ์˜ ๊ฒฐ๊ณผ๋ฌผ์— ๋Œ€ํ•ด ์Šค์Šค๋กœ ์ฑ…์ž„์„ ์ง€๊ณ ์ž ํ•˜๋Š” ์˜์ง€๊ฐ€ ์žˆ๋Š”๊ฐ€?

๋งฅ์Šค๋Š” โ€œ๋˜‘๋˜‘ํ•œ ์‚ฌ๋žŒ๋ณด๋‹ค ๋” ์ค‘์š”ํ•œ ๊ฒƒ์€ ๋ฐฐ์šฐ๋Š” ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚œ ์‚ฌ๋žŒโ€์ด๋ผ๋ฉฐ, ์ฒซ ๋ฒˆ์งธ ์‹œ๋„์™€ ๋‘ ๋ฒˆ์งธ ์‹œ๋„ ์‚ฌ์ด์˜ ๊ฐ„๊ทน์„ ์ขํž ์ค„ ์•„๋Š” ์ธ์žฌ๋ฅผ ์„ ํ˜ธํ•œ๋‹ค๊ณ  ๋ฐํ˜”๋‹ค.

5. ๊ธฐ์ˆ ์˜ ๋์— ์žˆ๋Š” ๊ฐ€์น˜: โ€œ๊ฐ€์กฑ๊ณผ ํ•จ๊ป˜ํ•˜๋Š” ์ €๋…โ€

๋งฅ์Šค๊ฐ€ ๊ฐ€์žฅ ๋ณด๋žŒ์„ ๋А๋ผ๋Š” ์ˆœ๊ฐ„์€ ๊ธฐ์ˆ ์  ์„ฑ์ทจ๋ฅผ ์ด๋ค˜์„ ๋•Œ๊ฐ€ ์•„๋‹ˆ๋‹ค. ๊ณ ๊ฐ ํ”ผ๋“œ๋ฐฑ ์ฑ„๋„์ธ ์Šฌ๋ž™(Slack)์—์„œ โ€œ๋ ˆ์•ผ ๋•๋ถ„์— ์—…๋ฌด๋ฅผ ์ผ์ฐ ๋งˆ์น˜๊ณ  ์•„์ด๋“ค์ด ์ž ๋“ค๊ธฐ ์ „์— ์ง‘์— ๊ฐ€์„œ ๊ฐ€์กฑ๊ณผ ์‹œ๊ฐ„์„ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์—ˆ๋‹คโ€๋Š” ๋ฉ”์‹œ์ง€๋ฅผ ์ฝ์„ ๋•Œ๋‹ค.

๊ทธ๋Š” AI๊ฐ€ ๋ณ€ํ˜ธ์‚ฌ๋ฅผ ๋Œ€์ฒดํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ, ๊ทธ๋“ค์—๊ฒŒ ๋ณธ์งˆ์ ์ธ ๊ฐ€์น˜์— ์ง‘์ค‘ํ•  ์‹œ๊ฐ„์„ ๋˜๋Œ๋ ค์ฃผ๋Š” ๋„๊ตฌ๋ผ๊ณ  ๋ฏฟ๋Š”๋‹ค. โ€œ์‚ฌ์šฉ์ž์—๊ฒŒ ์ตœ๊ณ ์˜ ๊ฐ€์น˜๋ฅผ ์ „๋‹ฌํ•˜๋Š” ๋ฐ ์ง‘์ค‘ํ•˜๋ฉด, ๋น„์ฆˆ๋‹ˆ์Šค์˜ ๋‹ค๋ฅธ ๋ฌธ์ œ๋“ค์€ ์ž์—ฐ์Šค๋Ÿฝ๊ฒŒ ํ•ด๊ฒฐ๋ฉ๋‹ˆ๋‹ค.โ€

๊ฒฐ๋ก : 2027๋…„์„ ๋ฐ”๋ผ๋ณด๋Š” ๋ฌดํ•œํ•œ ๋‚™๊ด€์ฃผ์˜

๋งฅ์Šค ์œ ๋„ค์ŠคํŠธ๋ž€๋“œ๋Š” ์ด์ œ ๋ง‰ โ€˜0์—์„œ 1โ€™์˜ ๋‹จ๊ณ„๋ฅผ ๋งˆ์ณค๋‹ค๊ณ  ๋งํ•œ๋‹ค. ์ด์ œ โ€˜1์—์„œ 10โ€™, ๊ทธ๋ฆฌ๊ณ  โ€˜10์—์„œ 100โ€™์œผ๋กœ ๊ฐ€๋Š” ์—ฌ์ •์ด ๋‚จ์•˜๋‹ค. ๊ทธ๋Š” 2027๋…„์˜ ํƒ€์šดํ™€ ๋ฏธํŒ…์—์„œ ๋ ˆ์•ผ๊ฐ€ ์ด๋ค„๋‚ผ ์„ฑ๊ณผ๋ฅผ ์ƒ์ƒํ•˜๋ฉฐ ๋งค์ผ ์„ค๋ ˆ๋Š” ๋งˆ์Œ์œผ๋กœ ๋ˆˆ์„ ๋œฌ๋‹ค.

๊ทธ๋Š” ์˜ˆ๋น„ ์ฐฝ์—…์ž๋“ค์—๊ฒŒ ์ด๋ ‡๊ฒŒ ์กฐ์–ธํ•œ๋‹ค. โ€œ์ž์‹ ์ด ํ•˜๋Š” ์ผ์— ์ง„์‹ฌ์œผ๋กœ ์—ด๊ด‘ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์„ฑ๊ณตํ•œ๋‹ค๋ฉด ์ƒ์ƒํ•  ์ˆ˜ ์—†์„ ๋งŒํผ ํž˜๋“  ์‹œ๊ฐ„์„ ๋ณด๋‚ด์•ผ ํ•  ํ…๋ฐ, ์—ด์ •์ด ์—†๋‹ค๋ฉด ๊ทธ ์‹œ๊ฐ„์„ ๋ฒ„ํ…จ๋‚ผ ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.โ€

๋ฒ•๋ฅ  ์ง€์‹ ์—†๋Š” ๊ฒŒ์ด๋จธ๊ฐ€ ๋งŒ๋“  AI ํ”Œ๋žซํผ ๋ ˆ์•ผ. ๊ทธ๋“ค์˜ ํ–‰๋ณด๊ฐ€ ๋‹จ์ˆœํ•œ ๊ธฐ์ˆ  ํ˜์‹ ์„ ๋„˜์–ด, ์ „๋ฌธ๊ฐ€๋“ค์˜ ์‚ถ์˜ ์งˆ์„ ๋ฐ”๊พธ๋Š” ๋ฌธํ™”์  ํ˜๋ช…์œผ๋กœ ๊ธฐ๋ก๋  ์ˆ˜ ์žˆ์„์ง€ ๊ท€์ถ”๊ฐ€ ์ฃผ๋ชฉ๋œ๋‹ค.


๐Ÿ’ก ์ฃผ์š” ์šฉ์–ด ์ •๋ฆฌ

  • ARR (Annual Recurring Revenue): ์—ฐ๊ฐ„ ๋ฐ˜๋ณต ๋งค์ถœ. ๊ตฌ๋…ํ˜• ์„œ๋น„์Šค(SaaS) ๊ธฐ์—…์˜ ์„ฑ์žฅ ์ง€ํ‘œ๋กœ ํ™œ์šฉ๋จ.
  • Y Combinator (YC): ์—์–ด๋น„์•ค๋น„, ๋“œ๋กญ๋ฐ•์Šค ๋“ฑ์„ ๋ฐฐ์ถœํ•œ ์„ธ๊ณ„ ์ตœ๊ณ ์˜ ์Šคํƒ€ํŠธ์—… ์•ก์…€๋Ÿฌ๋ ˆ์ดํ„ฐ.
  • Infinite Game (๋ฌดํ•œ ๊ฒŒ์ž„): ์ŠนํŒจ๊ฐ€ ๊ฒฐ์ •๋˜์–ด ๋๋‚˜๋Š” ๊ฒŒ์ž„์ด ์•„๋‹ˆ๋ผ, ๊ฒŒ์ž„์„ ๊ณ„์† ์œ ์ง€ํ•˜๊ณ  ํ™•์žฅํ•˜๋Š” ๊ฒƒ์„ ๋ชฉ์ ์œผ๋กœ ํ•˜๋Š” ๋น„์ฆˆ๋‹ˆ์Šค ์ฒ ํ•™.
  • 10x Lawyer: AI๋ฅผ ํ™œ์šฉํ•ด ๊ธฐ์กด๋ณด๋‹ค 10๋ฐฐ ์ด์ƒ์˜ ์ƒ์‚ฐ์„ฑ์„ ๋‚ด๋Š” ๋ฏธ๋ž˜ํ˜• ๋ณ€ํ˜ธ์‚ฌ ๋ชจ๋ธ.

โ€œKarpathyโ€™s โ€œautoresearchโ€ explained in 24 minutesโ€ โ€” Greg Isenberg ๊ธฐ๋ฐ˜ ๊ธฐ์‚ฌ ์›๋ณธ ์˜์ƒ ๋ณด๊ธฐ

์ž๊ณ  ์ผ์–ด๋‚˜๋ฉด AI๊ฐ€ ์„ฑ๊ณผ๋ฅผ ๋งŒ๋“ ๋‹ค: ์•ˆ๋“œ๋ ˆ ์นดํŒŒํ‹ฐ์˜ โ€˜์˜คํ†  ๋ฆฌ์„œ์น˜(Auto-research)โ€˜๊ฐ€ ๊ฐ€์ ธ์˜ฌ ๋น„์ฆˆ๋‹ˆ์Šค ํ˜๋ช…

AI ์—…๊ณ„์˜ ๊ฑฐ๋ฌผ, ์•ˆ๋“œ๋ ˆ ์นดํŒŒํ‹ฐ(Andrej Karpathy)๊ฐ€ ์ตœ๊ทผ ๊ณต๊ฐœํ•œ โ€˜์˜คํ†  ๋ฆฌ์„œ์น˜(Auto-research)โ€˜๊ฐ€ ํ…Œํฌ ์”ฌ์„ ๋œจ๊ฒ๊ฒŒ ๋‹ฌ๊ตฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์˜คํ”ˆAI์˜ ๊ณต๋™ ์ฐฝ๋ฆฝ์ž์ด์ž ํ…Œ์Šฌ๋ผ AI ๋ถ€๋ฌธ์„ ์ด๋Œ์—ˆ๋˜ ๊ทธ๊ฐ€ ์ œ์‹œํ•œ ์ด ๊ฐœ๋…์€ ๋‹จ์ˆœํ•œ ๋„๊ตฌ๋ฅผ ๋„˜์–ด, ์šฐ๋ฆฌ๊ฐ€ ์ผํ•˜๊ณ  ๋น„์ฆˆ๋‹ˆ์Šค๋ฅผ ์šด์˜ํ•˜๋Š” ๋ฐฉ์‹์„ ๊ทผ๋ณธ์ ์œผ๋กœ ๋ฐ”๊ฟ€ โ€˜๊ฒŒ์ž„ ์ฒด์ธ์ €โ€™๋กœ ํ‰๊ฐ€๋ฐ›์Šต๋‹ˆ๋‹ค.

๋น„์ฆˆ๋‹ˆ์Šค ์ „๋žต๊ฐ€ ๊ทธ๋ ‰ ์•„์ด์  ๋ฒ„๊ทธ(Greg Isenberg)๋Š” ์ด๋ฅผ ๋‘๊ณ  **โ€œ๋ฐค์ƒˆ๋„๋ก ์‰ฌ์ง€ ์•Š๊ณ  ๊ณผํ•™ ์‹คํ—˜์„ ๋Œ€์‹  ํ•ด์ฃผ๋Š” ์ดˆ์ฒœ์žฌ ๋กœ๋ด‡ ์ธํ„ด์„ ๊ณ ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™๋‹คโ€**๊ณ  ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์˜คํ†  ๋ฆฌ์„œ์น˜๊ฐ€ ๋ฌด์—‡์ธ์ง€, ๊ทธ๋ฆฌ๊ณ  ์ด๋ฅผ ํ†ตํ•ด ์–ด๋–ป๊ฒŒ ์ˆ˜์ต์„ ์ฐฝ์ถœํ•˜๊ณ  ์ƒ์‚ฐ์„ฑ์„ ๊ทน๋Œ€ํ™”ํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์‹ฌ์ธต ๋ถ„์„ํ•ด ๋ณด์•˜์Šต๋‹ˆ๋‹ค.


1. ์˜คํ†  ๋ฆฌ์„œ์น˜๋ž€ ๋ฌด์—‡์ธ๊ฐ€?: โ€˜์‹คํ–‰ํ•˜๋Š” AIโ€™์˜ ๋“ฑ์žฅ

๊ธฐ์กด์˜ AI๊ฐ€ ์‚ฌ์šฉ์ž์˜ ์งˆ๋ฌธ์— ๋‹ตํ•˜๋Š” โ€˜์ฑ—๋ด‡โ€™ ์ˆ˜์ค€์ด์—ˆ๋‹ค๋ฉด, **์˜คํ†  ๋ฆฌ์„œ์น˜(Auto-research)**๋Š” ์Šค์Šค๋กœ ๊ฐ€์„ค์„ ์„ธ์šฐ๊ณ  ์ฝ”๋“œ๋ฅผ ์ˆ˜์ •ํ•˜๋ฉฐ ์‹คํ—˜๊นŒ์ง€ ์ˆ˜ํ–‰ํ•˜๋Š” **โ€˜์ž์œจํ˜• ์—์ด์ „ํŠธ(Autonomous Agent)โ€˜**์ž…๋‹ˆ๋‹ค.

ํ•ต์‹ฌ ์ž‘๋™ ์›๋ฆฌ: ๋ฌดํ•œ ๋ฐ˜๋ณต๋˜๋Š” โ€˜์„ฑ๊ณต์˜ ๋ฃจํ”„โ€™

์˜คํ†  ๋ฆฌ์„œ์น˜๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ 5๋‹จ๊ณ„ ๊ณผ์ •์„ ์ธ๊ฐ„์˜ ๊ฐœ์ž… ์—†์ด ์Šค์Šค๋กœ ๋ฐ˜๋ณตํ•ฉ๋‹ˆ๋‹ค.

  1. ๋ชฉํ‘œ ์„ค์ •(Goal Setting): ์‚ฌ์šฉ์ž๊ฐ€ โ€œ์ด AI ๋ชจ๋ธ์„ ๋” ๋˜‘๋˜‘ํ•˜๊ฒŒ ๋งŒ๋“ค์–ด์ค˜โ€ ํ˜น์€ โ€œ๊ด‘๊ณ  ํด๋ฆญ๋ฅ ์„ ๋†’์—ฌ์ค˜โ€๋ผ๋Š” ๋ชฉํ‘œ๋ฅผ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค.
  2. ๊ณ„ํš ๋ฐ ์‹คํ–‰(Planning & Action): AI ์—์ด์ „ํŠธ๊ฐ€ ํŒŒ์ด์ฌ(Python) ์ฝ”๋“œ๋ฅผ ์ง์ ‘ ์ˆ˜์ •ํ•˜๊ฑฐ๋‚˜ ์„ค์ •๊ฐ’์„ ๋ณ€๊ฒฝํ•ฉ๋‹ˆ๋‹ค.
  3. ์‹คํ—˜ ์ˆ˜ํ–‰(Experimentation): GPU(๊ทธ๋ž˜ํ”ฝ ์ฒ˜๋ฆฌ ์žฅ์น˜)๋ฅผ ํ™œ์šฉํ•ด ์งง์€ ํ›ˆ๋ จ์ด๋‚˜ ํ…Œ์ŠคํŠธ๋ฅผ ์‹คํ–‰ํ•ฉ๋‹ˆ๋‹ค.
  4. ๊ฒฐ๊ณผ ๋ถ„์„(Evaluation): ์ง€ํ‘œ(Metrics)๋ฅผ ์ฝ๊ณ  ์ด์ „๋ณด๋‹ค ๊ฒฐ๊ณผ๊ฐ€ ์ข‹์•„์กŒ๋Š”์ง€ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค.
  5. ์—…๋ฐ์ดํŠธ ๋ฐ ๋ฐ˜๋ณต(Iteration): ๊ฒฐ๊ณผ๊ฐ€ ๊ฐœ์„ ๋˜์—ˆ๋‹ค๋ฉด ํ•ด๋‹น ์„ค์ •์„ ์ €์žฅํ•˜๊ณ , ๊ทธ๋ ‡์ง€ ์•Š๋‹ค๋ฉด ํ๊ธฐํ•œ ๋’ค ์ƒˆ๋กœ์šด ์‹คํ—˜์„ ๊ณ„ํšํ•ฉ๋‹ˆ๋‹ค.

์‡ผํ”ผํŒŒ์ด(Shopify)์˜ CEO ํ† ๋น„ ๋คผํŠธ์ผ€(Tobi Lรผtke)๋Š” โ€œ์˜คํ†  ๋ฆฌ์„œ์น˜๋Š” ์†Œํ”„ํŠธ์›จ์–ด ์ตœ์ ํ™”์— ์žˆ์–ด ํ˜๋ช…์ โ€์ด๋ผ๋ฉฐ ์ด ๊ธฐ์ˆ ์˜ ์ž ์žฌ๋ ฅ์„ ๋†’๊ฒŒ ํ‰๊ฐ€ํ–ˆ์Šต๋‹ˆ๋‹ค.


2. ์˜คํ†  ๋ฆฌ์„œ์น˜๋กœ ๋ˆ์„ ๋ฒ„๋Š” 5๊ฐ€์ง€ ๋น„์ฆˆ๋‹ˆ์Šค ์ „๋žต

๊ทธ๋ ‰ ์•„์ด์  ๋ฒ„๊ทธ๋Š” ์˜คํ†  ๋ฆฌ์„œ์น˜๋ฅผ ํ™œ์šฉํ•ด ๋‹น์žฅ ์‹œ์ž‘ํ•  ์ˆ˜ ์žˆ๋Š” ๊ตฌ์ฒด์ ์ธ ๋น„์ฆˆ๋‹ˆ์Šค ๋ชจ๋ธ๋“ค์„ ์ œ์‹œํ•ฉ๋‹ˆ๋‹ค.

โ‘  ๋‹ˆ์น˜ ๋งˆ์ผ“์šฉ โ€˜์—์ด์ „ํŠธ ์ธ ์–ด ๋ฐ•์Šค(Agent in a Box)โ€™

ํŠน์ • ์‚ฐ์—…์˜ ๊ณ ์งˆ์ ์ธ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๋Š” ์ž‘์€ ์˜คํ†  ๋ฆฌ์„œ์น˜ ๋ฃจํ”„๋ฅผ ํŒจํ‚ค์ง•ํ•˜์—ฌ ํŒ๋งคํ•˜๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.

  • ์˜ˆ์‹œ: ์•„๋งˆ์กด ์ƒํ’ˆ ํŽ˜์ด์ง€ ์ตœ์ ํ™” ๋„๊ตฌ, ๋ถ€๋™์‚ฐ ์ค‘๊ฐœ์ธ์„ ์œ„ํ•œ ์ด๋ฉ”์ผ ์‹œํ€€์Šค ํŠœ๋„ˆ, SaaS ๊ฐ€๊ฒฉ ์ตœ์ ํ™” ์—”์ง„.
  • ์ˆ˜์ต ๋ชจ๋ธ: 24์‹œ๊ฐ„ ๋‚ด๋‚ด ์ตœ์ ์˜ ๊ฒฐ๊ณผ๋ฌผ์„ ์ฐพ์•„์ฃผ๋Š” ์„œ๋น„์Šค์— ๋Œ€ํ•œ ์›”๊ฐ„ ๊ตฌ๋…๋ฃŒ.

โ‘ก ๋งˆ์ผ€ํŒ… ์„ฑ๊ณผ ๊ทน๋Œ€ํ™”(A/B ํ…Œ์ŠคํŒ…์˜ ๋ฏธ๋ž˜)

๊ธฐ์กด์˜ ์ˆ˜๋™์ ์ธ A/B ํ…Œ์ŠคํŠธ๋ฅผ AI๊ฐ€ ์™„์ „ํžˆ ๋Œ€์ฒดํ•ฉ๋‹ˆ๋‹ค. AI๊ฐ€ ์ˆ˜๋ฐฑ ๊ฐœ์˜ ํ—ค๋“œ๋ผ์ธ, ๋ ˆ์ด์•„์›ƒ, ๊ด‘๊ณ  ์†Œ์žฌ๋ฅผ ์ƒ์„ฑํ•˜๊ณ  ์‹ค์‹œ๊ฐ„์œผ๋กœ ํŠธ๋ž˜ํ”ฝ์„ ํ˜๋ ค๋ณด๋‚ด๋ฉฐ ๊ฐ€์žฅ ์ „ํ™˜์œจ์ด ๋†’์€ ์กฐํ•ฉ์„ ์ฐพ์•„๋ƒ…๋‹ˆ๋‹ค. ์ด๋Š” ๊ณ ๊ฐ ํš๋“ ๋น„์šฉ(CAC)์„ ๋‚ฎ์ถ”๊ณ  ๊ด‘๊ณ  ์ˆ˜์ต๋ฅ (ROAS)์„ ํš๊ธฐ์ ์œผ๋กœ ๋†’์—ฌ์ค๋‹ˆ๋‹ค.

โ‘ข ์„œ๋น„์Šคํ˜• ๋ฆฌ์„œ์น˜(Research as a Service, RaaS)

์Šคํƒ€ํŠธ์—…์ด๋‚˜ ํˆฌ์ž์ž๋ฅผ ์œ„ํ•ด ์‹œ์žฅ ์กฐ์‚ฌ ๋ฐ ๊ฒฝ์Ÿ์‚ฌ ๋ถ„์„์„ ์ž๋™ํ™”ํ•ฉ๋‹ˆ๋‹ค.

  • ๋‚ด์šฉ: ๊ฐ€๊ฒฉ ์ •์ฑ… ๋ณ€ํ™”, ์‹ ๊ทœ ๊ธฐ๋Šฅ ์—…๋ฐ์ดํŠธ, ๊ทœ์ œ ํ™˜๊ฒฝ ์ถ”์  ๋“ฑ์„ ์‹ค์‹œ๊ฐ„ ๋Œ€์‹œ๋ณด๋“œ๋กœ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
  • ์ˆ˜์ต ๋ชจ๋ธ: ๋ฆฌํฌํŠธ๋‹น ๋น„์šฉ ์ฒญ๊ตฌ ํ˜น์€ ์ƒ์‹œ ์—…๋ฐ์ดํŠธ ๋Œ€์‹œ๋ณด๋“œ ๊ตฌ๋….

โ‘ฃ ํ€€ํŠธ ํŠธ๋ ˆ์ด๋”ฉ ๋ฐ ๊ธˆ์œต ์ž๋™ํ™”

ํˆฌ์ž ๋ถ„์•ผ์—์„œ๋„ ์˜คํ†  ๋ฆฌ์„œ์น˜๋Š” ๊ฐ•๋ ฅํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ฒœ ๊ฐ€์ง€์˜ ๋งค๋งค ๊ทœ์น™์„ ๋ฐค์ƒˆ ๋ฐฑํ…Œ์ŠคํŒ…(Backtesting)ํ•˜์—ฌ ์œ ๋งํ•œ ์ „๋žต๋งŒ ๊ณจ๋ผ๋ƒ…๋‹ˆ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด ์ง์ ‘ ํˆฌ์žํ•˜๊ฑฐ๋‚˜, ๊ฒ€์ฆ๋œ ์‹ ํ˜ธ๋ฅผ ๊ตฌ๋… ์„œ๋น„์Šค๋กœ ํŒ๋งคํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

โ‘ค ๋น„์ฆˆ๋‹ˆ์Šค ์šด์˜(Ops)์˜ ์ž๋™ํ™”

์ธ๋ณด์ด์Šค ๋งค์นญ, ๊ฒฝ๋น„ ์ฒ˜๋ฆฌ, ๋ฆฌ๋“œ(Lead) ๋ถ„๋ฅ˜ ๋“ฑ ๋ฐ˜๋ณต์ ์ธ ๋ฐฑ์˜คํ”ผ์Šค ์—…๋ฌด์— ์˜คํ†  ๋ฆฌ์„œ์น˜ ๋ฃจํ”„๋ฅผ ์ ์šฉํ•ฉ๋‹ˆ๋‹ค. โ€œ์šฐ๋ฆฌ๊ฐ€ ๋‹น์‹ ์˜ ํšŒ๊ณ„ ์ฒ˜๋ฆฌ ์‹œ๊ฐ„์„ ์ ˆ๋ฐ˜์œผ๋กœ ์ค„์—ฌ์ฃผ๊ฒ ๋‹คโ€๋Š” ์ œ์•ˆ์œผ๋กœ ๋Œ€ํ˜• ํ•€ํ…Œํฌ ๊ธฐ์—…์ด๋‚˜ ์€ํ–‰์— ์†”๋ฃจ์…˜์„ ํŒ๋งคํ•˜๊ฑฐ๋‚˜ ์ธ์ˆ˜ํ•ฉ๋ณ‘(M&A)์˜ ๋Œ€์ƒ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.


3. ๋‹จ์ˆœํ•œ ๋น„์ฆˆ๋‹ˆ์Šค๋ฅผ ๋„˜์–ด: ์˜๋ฃŒ์™€ ๊ณผํ•™์˜ ํ˜์‹ 

์˜คํ†  ๋ฆฌ์„œ์น˜์˜ ์˜ํ–ฅ๋ ฅ์€ ๋น„์ฆˆ๋‹ˆ์Šค๋ฅผ ๋„˜์–ด ์ธ๋ฅ˜์˜ ๋‚œ์ œ ํ•ด๊ฒฐ๋กœ ํ™•์žฅ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—”์ง€๋‹ˆ์–ด ๋ชจ๊ฑด ๋ฆฐํ„ด(Morgan Linton)์€ **์ž„์ƒ ์‹œํ—˜ ์„ค๊ณ„(Clinical Trial Design)**์— ์ด ๊ธฐ์ˆ ์„ ์ ์šฉํ•˜๋Š” ์ƒ์ƒ์„ ์ œ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค.

์ˆ˜์ฒœ๋งŒ ๋‹ฌ๋Ÿฌ๊ฐ€ ๋“œ๋Š” ์‹ค์ œ ์ž„์ƒ ์‹œํ—˜ ์ „์—, AI ์—์ด์ „ํŠธ ๊ตฐ๋‹จ์ด ์ˆ˜๋งŽ์€ ๊ฐ€์ƒ ์‹คํ—˜๊ณผ ์†Œ๊ทœ๋ชจ ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์ตœ์ ์˜ ์น˜๋ฃŒ ํ”„๋กœํ† ์ฝœ์„ ์ฐพ์•„๋‚ธ๋‹ค๋ฉด, ์‹ ์•ฝ ๊ฐœ๋ฐœ ์†๋„๋Š” ๋น„์•ฝ์ ์œผ๋กœ ๋นจ๋ผ์ง€๊ณ  ๋น„์šฉ์€ ํš๊ธฐ์ ์œผ๋กœ ์ค„์–ด๋“ค ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” ์ธ๋ฅ˜์˜ ๊ฑด๊ฐ•๊ณผ ์งˆ๋ณ‘ ์น˜๋ฃŒ์— ์—„์ฒญ๋‚œ ํŒŒ๊ธ‰๋ ฅ์„ ๋ฏธ์น  ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋ฉ๋‹ˆ๋‹ค.


4. ์˜คํ†  ๋ฆฌ์„œ์น˜ ์‹œ์ž‘ํ•˜๊ธฐ: ๊ธฐ์ˆ ์  ์ค€๋น„๋ฌผ

์•ˆ๋“œ๋ ˆ ์นดํŒŒํ‹ฐ๋Š” ์˜คํ†  ๋ฆฌ์„œ์น˜๋ฅผ ์œ„ํ•œ ์˜คํ”ˆ ์†Œ์Šค ํ”„๋กœ์ ํŠธ์ธ **์—์ด์ „ํŠธํ—ˆ๋ธŒ(AgentHub)**๋„ ์ถœ์‹œํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” โ€˜์—์ด์ „ํŠธ๋“ค์„ ์œ„ํ•œ ๊นƒํ—ˆ๋ธŒ(GitHub)โ€˜์™€ ๊ฐ™์€ ๊ฐœ๋…์œผ๋กœ, ์—ฌ๋Ÿฌ AI ์—์ด์ „ํŠธ๊ฐ€ ํ˜‘์—…ํ•  ์ˆ˜ ์žˆ๋Š” ํ”Œ๋žซํผ์ž…๋‹ˆ๋‹ค.

์ง์ ‘ ์‹œ์ž‘ํ•ด๋ณด๊ณ  ์‹ถ์€ ๋ถ„๋“ค์„ ์œ„ํ•œ ๊ฐ€์ด๋“œ๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

  • ํ•˜๋“œ์›จ์–ด: ์—”๋น„๋””์•„(Nvidia) GPU๊ฐ€ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. (H100 ๋“ฑ ๊ณ ์„ฑ๋Šฅ ์นฉ ๊ถŒ์žฅ)
  • ๋Œ€์•ˆ: ๋กœ์ปฌ GPU๊ฐ€ ์—†๋‹ค๋ฉด ๋žŒ๋‹ค ๋žฉ์Šค(Lambda Labs), ๊ตฌ๊ธ€ ์ฝ”๋žฉ(Google Colab), ๋Ÿฐํฌ๋“œ(RunPod) ๋“ฑ ํด๋ผ์šฐ๋“œ GPU ์„œ๋น„์Šค๋ฅผ ๋Œ€์—ฌํ•˜์—ฌ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ๋„์›€ ๋ฐ›๊ธฐ: ํด๋กœ๋“œ(Claude)๋‚˜ ChatGPT ๊ฐ™์€ LLM์—๊ฒŒ ์นดํŒŒํ‹ฐ์˜ ๊นƒํ—ˆ๋ธŒ ์ €์žฅ์†Œ(Repo) ์ฃผ์†Œ๋ฅผ ์ฃผ๊ณ  ์„ค์น˜ ๋ฐ ์‹คํ–‰ ๊ณผ์ •์„ ๋‹จ๊ณ„๋ณ„๋กœ ๊ฐ€์ด๋“œํ•ด๋‹ฌ๋ผ๊ณ  ์š”์ฒญํ•˜์„ธ์š”.

๊ฒฐ๋ก : ์•ˆ๊ฐœ ์†์˜ ๊ธฐํšŒ๋ฅผ ์„ ์ ํ•˜๋ผ

์ƒˆ๋กœ์šด ๊ธฐ์ˆ ์ด ๋“ฑ์žฅํ–ˆ์„ ๋•Œ ๋Œ€์ค‘์€ ๋ณดํ†ต ๊ด€๋งํ•˜์ง€๋งŒ, ์„ ๊ตฌ์ž๋“ค์€ ๊ทธ โ€˜์•ˆ๊ฐœโ€™ ์†์—์„œ ๊ธฐํšŒ๋ฅผ ์ฐพ์•„๋ƒ…๋‹ˆ๋‹ค. ์•ˆ๋“œ๋ ˆ ์นดํŒŒํ‹ฐ์™€ ๊ฐ™์€ ์ฒœ์žฌ๋“ค์ด ์—ด์ •์ ์œผ๋กœ ๋งค๋‹ฌ๋ฆฌ๋Š” ๋ถ„์•ผ๋ผ๋ฉด ๋ฐ˜๋“œ์‹œ ์ฃผ๋ชฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

์˜คํ†  ๋ฆฌ์„œ์น˜๋Š” ๋‹จ์ˆœํžˆ ์ฝ”๋”ฉ์„ ๋Œ€์‹  ํ•ด์ฃผ๋Š” ๋„๊ตฌ๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค. **์ธ๊ฐ„์˜ ์ฐฝ์˜์„ฑ๊ณผ AI์˜ ๋ฌดํ•œํ•œ ์‹คํ–‰๋ ฅ์„ ๊ฒฐํ•ฉํ•˜์—ฌ, ์šฐ๋ฆฌ๊ฐ€ ์ž ๋“  ์‚ฌ์ด์—๋„ ๊ฐ€์น˜๋ฅผ ๋งŒ๋“ค์–ด๋‚ด๋Š” โ€˜์ง€๋Šฅํ˜• ์—”์ง„โ€™**์ž…๋‹ˆ๋‹ค. ์ง€๊ธˆ ๋ฐ”๋กœ ์ด ๊ธฐ์ˆ ์„ ๋งŒ์ ธ๋ณด๊ณ  ์‹คํ—˜ํ•ด ๋ณด์„ธ์š”. 99.9%์˜ ์‚ฌ๋žŒ๋“ค๋ณด๋‹ค ์•ž์„œ๋‚˜๊ฐˆ ์ˆ˜ ์žˆ๋Š” ์œ ์ผํ•œ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.


์ด ๊ธฐ์‚ฌ๋Š” ๊ทธ๋ ‰ ์•„์ด์  ๋ฒ„๊ทธ์˜ ๋ถ„์„์„ ๋ฐ”ํƒ•์œผ๋กœ ์ž‘์„ฑ๋˜์—ˆ์œผ๋ฉฐ, AI ๊ธฐ์ˆ ์˜ ์ตœ์‹  ๋™ํ–ฅ์„ ๋ฐ˜์˜ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.


โ€œHow Figma engineers sync designs with Claude Code and Codexโ€ โ€” How I AI ๊ธฐ๋ฐ˜ ๊ธฐ์‚ฌ ์›๋ณธ ์˜์ƒ ๋ณด๊ธฐ

๋””์ž์ธ๊ณผ ๊ฐœ๋ฐœ์˜ ๊ฒฝ๊ณ„๊ฐ€ ์‚ฌ๋ผ์ง„๋‹ค: ํ”ผ๊ทธ๋งˆ ์—”์ง€๋‹ˆ์–ด๋“ค์ด ํด๋กœ๋“œ ์ฝ”๋“œ(Claude Code)๋กœ ์ผํ•˜๋Š” ๋ฒ•

โ€œ๋””์ž์ธ์ด ๋จผ์ €์ธ๊ฐ€, ์ฝ”๋“œ๊ฐ€ ๋จผ์ €์ธ๊ฐ€?โ€ ์†Œํ”„ํŠธ์›จ์–ด ์—…๊ณ„์˜ ํ•ด๋ฌต์€ ๋…ผ์Ÿ์ด ๋งˆ์นจ๋‚ด ์ข…์ง€๋ถ€๋ฅผ ์ฐ์œผ๋ ค ํ•ฉ๋‹ˆ๋‹ค. ๋””์ž์ธ ๋„๊ตฌ์˜ ๋Œ€๋ช…์‚ฌ ํ”ผ๊ทธ๋งˆ(Figma)์˜ ์—”์ง€๋‹ˆ์–ด์™€ ๋””์ž์ด๋„ˆ๋“ค์ด AI๋ฅผ ํ™œ์šฉํ•ด ๋””์ž์ธ๊ณผ ์ฝ”๋“œ ์‚ฌ์ด์˜ โ€˜์žฅ๋ฒฝโ€™์„ ์–ด๋–ป๊ฒŒ ํ—ˆ๋ฌผ๊ณ  ์žˆ๋Š”์ง€, ๊ทธ ํ˜์‹ ์ ์ธ ์›Œํฌํ”Œ๋กœ์šฐ๋ฅผ ์‹ฌ์ธต ๋ถ„์„ํ–ˆ์Šต๋‹ˆ๋‹ค.


1. ์„ ํ˜•์  ํ”„๋กœ์„ธ์Šค์˜ ์ข…๋ง: โ€œ์ฝ”๋”ฉ๋„ ๋””์ž์ธ๋งŒํผ ์ €๋ ดํ•ด์กŒ๋‹คโ€

๊ณผ๊ฑฐ์˜ ์†Œํ”„ํŠธ์›จ์–ด ๊ฐœ๋ฐœ์€ ์ฒ ์ €ํžˆ ์„ ํ˜•์ ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ฝ”๋“œ๋ฅผ ์ž‘์„ฑํ•˜๋Š” ๋น„์šฉ์€ ๋น„์‹ธ๊ณ  ๋””์ž์ธ์„ ์ˆ˜์ •ํ•˜๋Š” ๋น„์šฉ์€ ์ƒ๋Œ€์ ์œผ๋กœ ์ €๋ ดํ–ˆ๊ธฐ์—, ๋””์ž์ด๋„ˆ๊ฐ€ ์ •๊ตํ•œ ์‹œ์•ˆ(High-fidelity)์„ ์™„์„ฑํ•˜๋ฉด ์—”์ง€๋‹ˆ์–ด๊ฐ€ ์ด๋ฅผ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐฉ์‹์ด ์ •์„์ด์—ˆ์ฃ . ํ•˜์ง€๋งŒ ํ”ผ๊ทธ๋งˆ์˜ ๋””์ž์ด๋„ˆ ๊ธฐ(Ghee)์™€ ์—”์ง€๋‹ˆ์–ด ์•Œ๋ ‰์Šค(Alex)๋Š” AI๊ฐ€ ์ด ๊ตฌ๋„๋ฅผ ์™„์ „ํžˆ ๋’ค๋ฐ”๊ฟจ๋‹ค๊ณ  ๋งํ•ฉ๋‹ˆ๋‹ค.

์ด์ œ๋Š” **โ€˜๋ฐ”์ด๋ธŒ ์ฝ”๋”ฉ(Vibe Coding)โ€˜**์ด๋‚˜ **โ€˜๊ธฐ๋Šฅ์  ์™€์ด์–ดํ”„๋ ˆ์ž„(Functional Wireframes)โ€˜**์ด ๊ฐ€๋Šฅํ•ด์กŒ์Šต๋‹ˆ๋‹ค. ์•„์ด๋””์–ด๋ฅผ ์ฝ”๋“œ๋กœ ๊ตฌํ˜„ํ•˜๋Š” ๋น„์šฉ์ด ๋””์ž์ธ๋งŒํผ์ด๋‚˜ ๋‚ฎ์•„์ง€๋ฉด์„œ, ๊ตณ์ด ํšŒ์ƒ‰์กฐ์˜ ์ •์ ์ธ ์™€์ด์–ดํ”„๋ ˆ์ž„์„ ๊ทธ๋ฆฌ๋Š” ๋Œ€์‹  ๊ณง๋ฐ”๋กœ ์ž‘๋™ํ•˜๋Š” ์ฝ”๋“œ๋ฅผ ์งœ์„œ ํ…Œ์ŠคํŠธํ•ด ๋ณผ ์ˆ˜ ์žˆ๊ฒŒ ๋œ ๊ฒƒ์ž…๋‹ˆ๋‹ค.

์ด๋Ÿฌํ•œ ๋ณ€ํ™”๋Š” โ€˜์›Œํฌํ”Œ๋กœ์šฐโ€™๋ผ๋Š” ๊ฐœ๋… ์ž์ฒด๋ฅผ ํญ๋ฐœ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค. ์ด์ œ ๋””์ž์ธ๊ณผ ๊ฐœ๋ฐœ์€ ์ˆœ์„œ์˜ ๋ฌธ์ œ๊ฐ€ ์•„๋‹ˆ๋ผ, ํ•„์š”์— ๋”ฐ๋ผ ์„œ๋กœ๋ฅผ ๋„˜๋‚˜๋“œ๋Š” ์œ ์—ฐํ•œ ๊ณผ์ •์ด ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

2. ๋””์ž์ธ๊ณผ ์ฝ”๋“œ์˜ ๋™๊ธฐํ™”: MCP๊ฐ€ ์—ฐ๊ฒฐํ•˜๋Š” ์ƒˆ๋กœ์šด ์„ธ๊ณ„

ํ˜‘์—… ๊ณผ์ •์—์„œ ๊ฐ€์žฅ ํฐ ๊ณจ์นซ๊ฑฐ๋ฆฌ๋Š” **โ€˜๋””์ž์ธ ๋ถ€์ฑ„(Design Debt)โ€˜**์ž…๋‹ˆ๋‹ค. ์‹ค์ œ ์„œ๋น„์Šค(์ฝ”๋“œ)๋Š” ๊ณ„์† ์—…๋ฐ์ดํŠธ๋˜๋Š”๋ฐ, ํ”ผ๊ทธ๋งˆ ํŒŒ์ผ์€ ๊ณผ๊ฑฐ์˜ ์Šคํฌ๋ฆฐ์ƒท์— ๋จธ๋ฌผ๋Ÿฌ ์žˆ๋Š” ํ˜„์ƒ์ด์ฃ . ํ”ผ๊ทธ๋งˆ ํŒ€์€ ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด **MCP(Model Context Protocol, AI ๋ชจ๋ธ์ด ์™ธ๋ถ€ ๋„๊ตฌ๋‚˜ ๋ฐ์ดํ„ฐ์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ฃผ๋Š” ๊ฐœ๋ฐฉํ˜• ํ‘œ์ค€)**๋ฅผ ํ™œ์šฉํ•ฉ๋‹ˆ๋‹ค.

Case 1: ์ฝ”๋“œ์—์„œ ๋””์ž์ธ์œผ๋กœ (Code to Figma)

์—”์ง€๋‹ˆ์–ด๊ฐ€ ์ฝ”๋“œ๋ฒ ์ด์Šค์—์„œ ๋ณต์žกํ•œ ํšŒ์›๊ฐ€์ž… ํ๋ฆ„(Sign-up flow)์˜ 5๊ฐ€์ง€ ์ƒํƒœ๋ฅผ ๋งŒ๋“ค์—ˆ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ด…์‹œ๋‹ค. ์˜ˆ์ „์—๋Š” ์ด๋ฅผ ๋””์ž์ด๋„ˆ์—๊ฒŒ ์ „๋‹ฌํ•˜๊ธฐ ์œ„ํ•ด ์ผ์ผ์ด ์บก์ฒ˜ํ•˜๊ฑฐ๋‚˜ ์„ค๋ช…ํ•ด์•ผ ํ–ˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด์ œ๋Š” AI ์—์ด์ „ํŠธ์—๊ฒŒ ๋ช…๋ น ํ•œ ์ค„๋งŒ ๋‚ด๋ฆฌ๋ฉด ๋ฉ๋‹ˆ๋‹ค.

โ€œ์ด ํšŒ์›๊ฐ€์ž… ํ๋ฆ„์˜ 5๊ฐ€์ง€ ์ƒํƒœ๋ฅผ ๋ชจ๋‘ ํ”ผ๊ทธ๋งˆ๋กœ ๋ณด๋‚ด์ค˜.โ€

AI๋Š” ์ฝ”๋“œ๋ฒ ์ด์Šค๋ฅผ ์ฝ๊ณ  ํ•ด๋‹น ๋กœ์ง์„ ์ดํ•ดํ•œ ๋’ค, ํ”ผ๊ทธ๋งˆ์— ๊ฐ ์ƒํƒœ๋ฅผ ๋‚˜๋ž€ํžˆ ๋ฐฐ์น˜ํ•ฉ๋‹ˆ๋‹ค. ๋””์ž์ด๋„ˆ๋Š” ์ด์ œ ์‹ค์ œ ์ฝ”๋“œ์™€ ์™„๋ฒฝํžˆ ์ผ์น˜ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฌผ์„ ๋ฐ”ํƒ•์œผ๋กœ ๋””์ž์ธ ์ž‘์—…์„ ์ด์–ด๊ฐˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Case 2: ๋””์ž์ธ์—์„œ ์ฝ”๋“œ๋กœ (Figma to Code)

๋ฐ˜๋Œ€๋กœ ๋””์ž์ด๋„ˆ๊ฐ€ ํ”ผ๊ทธ๋งˆ์—์„œ ๋ฒ„ํŠผ์˜ ์ƒ‰์ƒ์ด๋‚˜ ๋ ˆ์ด์•„์›ƒ์„ ์ˆ˜์ •ํ–ˆ๋‹ค๋ฉด, ์—”์ง€๋‹ˆ์–ด๋Š” ํด๋กœ๋“œ ์ฝ”๋“œ(Claude Code)๋ฅผ ํ†ตํ•ด ํ•ด๋‹น ๋ณ€๊ฒฝ ์‚ฌํ•ญ์„ ์ฆ‰์‹œ ์ฝ”๋“œ์— ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ”ผ๊ทธ๋งˆ์˜ URL์„ ๋ณต์‚ฌํ•ด ๋„ฃ๊ธฐ๋งŒ ํ•˜๋ฉด AI๊ฐ€ ๋””์ž์ธ์˜ ๋ณ€๊ฒฝ์ ์„ ํŒŒ์•…ํ•˜๊ณ , ์ด๋ฅผ ํ˜„์žฌ ์ฝ”๋“œ์™€ ๋Œ€์กฐํ•˜์—ฌ โ€˜ํ”ฝ์…€ ํผํŽ™ํŠธ(Pixel-perfect)โ€˜ํ•œ ์ฝ”๋“œ๋กœ ๋ณ€ํ™˜ํ•ด ์ ์šฉํ•ฉ๋‹ˆ๋‹ค.

3. โ€œํ’์š”์˜ ์‹œ๋Œ€โ€: ํšจ์œจ์„ฑ์„ ๋„˜์–ด โ€˜์žฅ์ธ์ •์‹ โ€™์œผ๋กœ

ํ”ผ๊ทธ๋งˆ์˜ ๋””์ž์ด๋„ˆ ๊ธฐ๋Š” ํ˜„์žฌ๋ฅผ **โ€˜ํƒ๋‹‰์˜ ์ˆœ๊ฐ„(Decadent moment)โ€˜**์ด๋ผ๊ณ  ํ‘œํ˜„ํ•ฉ๋‹ˆ๋‹ค. ์ด์ „์—๋Š” ์ž์›์ด ํ•œ์ •์ ์ด์—ˆ๊ธฐ์— โ€œ์ด ๊ธฐ๋Šฅ์€ ๊ตฌํ˜„ํ•˜๊ธฐ ๋„ˆ๋ฌด ์–ด๋ ค์šฐ๋‹ˆ ๋นผ์žโ€๋ผ๋ฉฐ ๋ฒ”์œ„๋ฅผ ๊นŽ์•„๋‚ด๋Š”(Scope cutting) ๋ฏธํŒ…์ด ์žฆ์•˜์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ AI ๋•๋ถ„์— ๊ตฌํ˜„ ๋น„์šฉ์ด ๋‚ฎ์•„์ง€๋ฉด์„œ, ํŒ€์€ ์ด์ œ โ€˜์–ด๋–ป๊ฒŒ ๊ตฌํ˜„ํ• ๊นŒโ€™๊ฐ€ ์•„๋‹ˆ๋ผ **โ€˜๋ฌด์—‡์„ ๋งŒ๋“ค์–ด์•ผ ํ• ๊นŒโ€™**๋ผ๋Š” ๋ณธ์งˆ์ ์ธ ์งˆ๋ฌธ์— ๋” ๋งŽ์€ ์‹œ๊ฐ„์„ ์Ÿ์„ ์ˆ˜ ์žˆ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

  • ๋‹ค์–‘์„ฑ ํ™•๋ณด: ํ•˜๋‚˜์˜ ์•„์ด๋””์–ด๋ฅผ ๊ฒ€์ฆํ•˜๋Š” ๋ฐ ์‹œ๊ฐ„์ด ์ ๊ฒŒ ๋“œ๋‹ˆ, ๋” ๋งŽ์€ ๋Œ€์•ˆ์„ ํƒ์ƒ‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ๋””ํ…Œ์ผ์˜ ์™„์„ฑ๋„: ๋‹จ์ˆœ ๋ฐ˜๋ณต ์ž‘์—…์— ๋“ค๋˜ ์—๋„ˆ์ง€๋ฅผ ์ œํ’ˆ์˜ โ€˜๊ฒฐโ€™์„ ์‚ด๋ฆฌ๋Š” ์žฅ์ธ์ •์‹ (Craft)์— ํˆฌ์ž…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ์‹ค์‹œ๊ฐ„ ํ˜‘์—…์˜ ๋ถ€ํ™œ: AI๊ฐ€ ๋ฒˆ๊ฑฐ๋กœ์šด ํ•ธ๋“œ์˜คํ”„(Handoff) ๊ณผ์ •์„ ๋Œ€์‹ ํ•ด์ฃผ๋ฉด์„œ, ๋””์ž์ด๋„ˆ์™€ ์—”์ง€๋‹ˆ์–ด๊ฐ€ ํ”ผ๊ทธ๋งˆ ์บ”๋ฒ„์Šค ์œ„์—์„œ ํ•จ๊ป˜ โ€˜์žผ(Jam)โ€˜์„ ํ•˜๋“ฏ ์‹ค์‹œ๊ฐ„์œผ๋กœ ์†Œํ†ตํ•˜๋Š” ์—๋„ˆ์ง€๊ฐ€ ๋˜์‚ด์•„๋‚ฌ์Šต๋‹ˆ๋‹ค.

4. ์—”์ง€๋‹ˆ์–ด์˜ ๋น„๋ฐ€ ๋ฌด๊ธฐ: โ€˜์Šคํ‚ฌ(Skill)โ€˜์„ ํ†ตํ•œ ์ž๋™ํ™”

์—”์ง€๋‹ˆ์–ด ์•Œ๋ ‰์Šค๋Š” ๋‹จ์ˆœ ๋ฐ˜๋ณต์ ์ธ โ€˜๋…ธ๋™(Toil)โ€˜์„ ์ค„์ด๊ธฐ ์œ„ํ•ด ์ž์‹ ๋งŒ์˜ **โ€˜์Šคํ‚ฌ(Skill)โ€˜**์„ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๋ณต์žกํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋งค๋ฒˆ ์ž…๋ ฅํ•˜๋Š” ๋Œ€์‹ , ํŠน์ • ๋ช…๋ น์–ด๋กœ ์ผ๋ จ์˜ ๊ณผ์ •์„ ์ž๋™ํ™”ํ•˜๋Š” ์ผ์ข…์˜ ๋งคํฌ๋กœ์ž…๋‹ˆ๋‹ค.

๋Œ€ํ‘œ์ ์ธ ์˜ˆ๊ฐ€ /ship ์Šคํ‚ฌ์ž…๋‹ˆ๋‹ค. ๋Œ€๊ทœ๋ชจ ์ฝ”๋“œ๋ฒ ์ด์Šค์— ์ฝ”๋“œ๋ฅผ ๋ฐ˜์˜ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ํ…Œ์ŠคํŠธ ํ†ต๊ณผ, ๋ฆฐํŠธ(Lint) ์ฒดํฌ, ๋นŒ๋“œ ํ™•์ธ ๋“ฑ ์ˆ˜๋งŽ์€ ์ฒดํฌ๋ฆฌ์ŠคํŠธ๋ฅผ ๊ฑฐ์ณ์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์•Œ๋ ‰์Šค๋Š” ์ด ๊ณผ์ •์„ AI ์Šคํ‚ฌ๋กœ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค.

  1. ์ฝ”๋“œ๊ฐ€ ์ค€๋น„๋˜๋ฉด /ship ๋ช…๋ น์–ด๋ฅผ ์ž…๋ ฅํ•ฉ๋‹ˆ๋‹ค.
  2. AI๊ฐ€ ์ž๋™์œผ๋กœ ์‚ฌ์ „ ์ ๊ฒ€(Pre-flight checks)์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
  3. ๋ฌธ์ œ๊ฐ€ ์—†๋‹ค๋ฉด ๊นƒ(Git) ์ €์žฅ์†Œ์— ํ‘ธ์‹œํ•ฉ๋‹ˆ๋‹ค.
  4. CI(์ง€์†์  ํ†ตํ•ฉ) ๋„๊ตฌ์ธ ๋นŒ๋“œ์นด์ดํŠธ(Buildkite)์˜ ๋กœ๊ทธ๋ฅผ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๋ฉฐ, ์‚ฌ์†Œํ•œ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•˜๋ฉด ์Šค์Šค๋กœ ์ˆ˜์ • ์ปค๋ฐ‹์„ ์˜ฌ๋ฆฝ๋‹ˆ๋‹ค.

์ด๋Ÿฌํ•œ ๋ฐฉ์‹์€ ์กฐ์ง์˜ **SOP(ํ‘œ์ค€ ์šด์˜ ์ ˆ์ฐจ)**๋ฅผ ๋ฌธ์„œ์— ๊ฐ€๋‘ฌ๋‘๋Š” ๋Œ€์‹ , ๋ˆ„๊ตฌ๋‚˜ ์ฆ‰์‹œ ์‹คํ–‰ ๊ฐ€๋Šฅํ•œ โ€˜์‚ด์•„์žˆ๋Š” ๋„๊ตฌโ€™๋กœ ๋ฐ”๊ฟ” ๋†“์Šต๋‹ˆ๋‹ค.

5. ๊ฒฐ๋ก : AI ์‹œ๋Œ€์˜ ํ˜‘์—…์ด ๋‚˜์•„๊ฐˆ ๋ฐฉํ–ฅ

ํ”ผ๊ทธ๋งˆ ํŒ€์ด ๋ณด์—ฌ์ค€ ๋ฏธ๋ž˜๋Š” ๋‹จ์ˆœํžˆ ๋„๊ตฌ๊ฐ€ ํŽธ๋ฆฌํ•ด์ง€๋Š” ๊ฒƒ์„ ๋„˜์–ด, ์ธ๊ฐ„์˜ ์ฐฝ์˜์„ฑ์ด ๋ฐœํœ˜๋˜๋Š” ์ง€์ ์ด ์ด๋™ํ•จ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

  • ๋””์ž์ด๋„ˆ๋Š” ๋” ์ด์ƒ ํ”ฝ์…€์„ ์˜ฎ๊ธฐ๋Š” ๋ฐ˜๋ณต ์ž‘์—…์— ๋งค๋ชฐ๋˜์ง€ ์•Š๊ณ , ์ œํ’ˆ์˜ ์ „๋žต๊ณผ ์‚ฌ์šฉ์ž ๊ฒฝํ—˜์˜ ๋ณธ์งˆ์„ ๊ณ ๋ฏผํ•ฉ๋‹ˆ๋‹ค.
  • ์—”์ง€๋‹ˆ์–ด๋Š” ๊ตฌ๋ฌธ(Syntax)์„ ๋งž์ถ”๋Š” ๊ธฐ๊ณ„์ ์ธ ์ž‘์—…์—์„œ ๋ฒ—์–ด๋‚˜, ๋ฌธ์ œ ํ•ด๊ฒฐ๊ณผ ์•„ํ‚คํ…์ฒ˜ ์„ค๊ณ„์— ์ง‘์ค‘ํ•ฉ๋‹ˆ๋‹ค.

๋งˆ์ง€๋ง‰์œผ๋กœ โ€œ๋””์ž์ธ์ด ๋จผ์ €์ธ๊ฐ€, ์ฝ”๋“œ๊ฐ€ ๋จผ์ €์ธ๊ฐ€?โ€๋ผ๋Š” ์งˆ๋ฌธ์— ํ”ผ๊ทธ๋งˆ ํŒ€์€ ์ด๋ ‡๊ฒŒ ๋‹ตํ•ฉ๋‹ˆ๋‹ค. โ€œ๊ฒฐ๊ตญ ์ค‘์š”ํ•œ ๊ฒƒ์€ ๊ธฐํš(Planning)๊ณผ ์†Œํ†ต์ด๋‹ค.โ€ AI๋Š” ๊ทธ ์†Œํ†ต์˜ ๊ฐ„๊ทน์„ ๋ฉ”์›Œ์ฃผ๋Š” ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ์ ‘์ฐฉ์ œ๊ฐ€ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.


๐Ÿ’ก ํ•ต์‹ฌ ์š”์•ฝ (Key Takeaways)

  1. ์›Œํฌํ”Œ๋กœ์šฐ์˜ ๋ถ•๊ดด: ์ฝ”๋”ฉ ๋น„์šฉ์˜ ํ•˜๋ฝ์œผ๋กœ ๋””์ž์ธ๊ณผ ๊ฐœ๋ฐœ์˜ ์ˆœ์ฐจ์  ๊ฒฝ๊ณ„๊ฐ€ ๋ฌด๋„ˆ์ง.
  2. MCP์˜ ์œ„๋ ฅ: ์ฝ”๋“œ๋ฒ ์ด์Šค์˜ ์ƒํƒœ๋ฅผ ํ”ผ๊ทธ๋งˆ๋กœ, ํ”ผ๊ทธ๋งˆ์˜ ๋””์ž์ธ์„ ์ฝ”๋“œ๋กœ ์ฆ‰์‹œ ๋™๊ธฐํ™”.
  3. ์Šคํ‚ฌ(Skill) ๊ธฐ๋ฐ˜ ์ž๋™ํ™”: ๋‹จ์ˆœ ๋ฐ˜๋ณต ์—…๋ฌด๋ฅผ AI ์Šคํ‚ฌ๋กœ ๋ณ€ํ™˜ํ•˜์—ฌ ์—”์ง€๋‹ˆ์–ด๋ง ์ƒ์‚ฐ์„ฑ ๊ทน๋Œ€ํ™”.
  4. ์žฅ์ธ์ •์‹ ์˜ ํšŒ๋ณต: ํšจ์œจํ™”๋กœ ํ™•๋ณด๋œ ์‹œ๊ฐ„์€ ์ œํ’ˆ์˜ ํ€„๋ฆฌํ‹ฐ์™€ ์ฐฝ์˜์ ์ธ ๊ธฐํš์— ์žฌํˆฌ์ž.

โ€œThe Department of War is making a huge mistake.โ€ โ€” Dwarkesh Patel ๊ธฐ๋ฐ˜ ๊ธฐ์‚ฌ ์›๋ณธ ์˜์ƒ ๋ณด๊ธฐ

[์ธ์‚ฌ์ดํŠธ] ์ „์Ÿ๋ถ€์˜ ์œ„ํ—˜ํ•œ ๋„๋ฐ•: AI ์‹œ๋Œ€์˜ ์ž์œ ๋ฅผ ์œ„ํ˜‘ํ•˜๋Š” ๊ตญ๊ฐ€์˜ ๊ฐ•์š”

์ตœ๊ทผ ๋ฏธ๊ตญ ๊ตญ๋ฐฉ๋ถ€(์ „์Ÿ๋ถ€)๊ฐ€ ์ธ๊ณต์ง€๋Šฅ(AI) ์Šคํƒ€ํŠธ์—… **์•ค์Šค๋กœํ”ฝ(Anthropic)**์„ โ€˜๊ณต๊ธ‰๋ง ๋ฆฌ์Šคํฌโ€™๋กœ ๊ทœ์ •ํ–ˆ๋‹ค๋Š” ์†Œ์‹์ด ์ „ํ•ด์กŒ์Šต๋‹ˆ๋‹ค. ์ด์œ ๋Š” ๋ช…ํ™•ํ•ฉ๋‹ˆ๋‹ค. ์•ค์Šค๋กœํ”ฝ์ด ์ž์‹ ๋“ค์˜ AI ๋ชจ๋ธ์„ โ€˜๋Œ€๊ทœ๋ชจ ๊ฐ์‹œ(Mass Surveillance)โ€˜๋‚˜ โ€˜์ž์œจ ์‚ด์ƒ ๋ฌด๊ธฐ(Autonomous Weapons)โ€˜์— ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์— ๋Œ€ํ•ด ๋ช…ํ™•ํ•œ ๊ฑฐ๋ถ€ ์˜์‚ฌ, ์ฆ‰ โ€˜๋ ˆ๋“œ๋ผ์ธ(Red Lines)โ€˜์„ ์„ค์ •ํ–ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.

์œ ๋ช… ํฌ๋“œ์บ์Šคํ„ฐ์ด์ž ๋ถ„์„๊ฐ€์ธ ๋“œ์™€๋ฅด์ผ€์‹œ ํŒŒํ…”(Dwarkesh Patel)์€ ์ด๋ฅผ ๋‘๊ณ  โ€œ๋ฏธ๋ž˜ ๋ฌธ๋ช…์„ ํ–ฅํ•œ ๊ฒฝ๊ณ  ์‚ฌ๊ฒฉโ€์ด๋ผ๊ณ  ํ‰๊ฐ€ํ•ฉ๋‹ˆ๋‹ค. ๋‹จ์ˆœํžˆ ํ•œ ๊ธฐ์—…๊ณผ ์ •๋ถ€์˜ ๊ณ„์•ฝ ๋ฌธ์ œ๋ฅผ ๋„˜์–ด, ์šฐ๋ฆฌ๊ฐ€ ๋งž์ดํ•  AI ๊ธฐ๋ฐ˜ ๋ฌธ๋ช…์˜ ๊ทผ๊ฐ„์„ ํ”๋“œ๋Š” ๊ฒฐ์ •์ ์ธ ์‚ฌ๊ฑด์ด๋ผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.


1. AI ๋…ธ๋™์ด ์ง€๋ฐฐํ•  ๋ฏธ๋ž˜: 20๋…„ ๋’ค์˜ ํ’๊ฒฝ

์šฐ๋ฆฌ๋Š” ์ง€๊ธˆ ๋‹จ์ˆœํ•œ ์†Œํ”„ํŠธ์›จ์–ด ๋„๊ตฌ์˜ ์ง„ํ™”๋ฅผ ๋ณด๊ณ  ์žˆ๋Š” ๊ฒƒ์ด ์•„๋‹™๋‹ˆ๋‹ค. ํŒŒํ…”์€ ํ–ฅํ›„ 20๋…„ ๋‚ด์— ๊ตฐ๋Œ€, ๋ฏผ๊ฐ„ ์ •๋ถ€, ๋ฏผ๊ฐ„ ๋ถ€๋ฌธ ์ธ๋ ฅ์˜ 99%๊ฐ€ AI ์‹œ์Šคํ…œ์œผ๋กœ ๋Œ€์ฒด๋  ๊ฒƒ์ด๋ผ๊ณ  ์ „๋งํ•ฉ๋‹ˆ๋‹ค.

  • ๋กœ๋ด‡ ๊ตฐ๋Œ€: ๊ตญ๋ฐฉ์˜ ํ•ต์‹ฌ ์ „๋ ฅ์€ AI ๋กœ๋ด‡์ด ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
  • ์ดˆ์ง€๋Šฅ ๊ณ ๋ฌธ: ์ƒ์›์˜์›, ๋Œ€ํ†ต๋ น, CEO ์˜†์—๋Š” ์ธ๊ฐ„์„ ๋Šฅ๊ฐ€ํ•˜๋Š” ์ง€๋Šฅ์˜ AI ๋ณด์ขŒ๊ด€์ด ์ž๋ฆฌํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
  • AI ๊ฒฝ์ฐฐ: ์น˜์•ˆ ์œ ์ง€ ์—…๋ฌด ์—ญ์‹œ AI์˜ ๋ชซ์ด ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.

๊ฒฐ๊ตญ ๋ฏธ๋ž˜ ๋ฌธ๋ช…์€ **โ€˜AI ๋…ธ๋™โ€™**์ด๋ผ๋Š” ํ† ๋Œ€ ์œ„์—์„œ ๋Œ์•„๊ฐ€๊ฒŒ ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฐ ์ƒํ™ฉ์—์„œ ์ •๋ถ€๊ฐ€ ํŠน์ • AI ๊ธฐ์—…์„ ์••๋ฐ•ํ•˜๋Š” ๋ฐฉ์‹์€ ์šฐ๋ฆฌ๊ฐ€ ๋ฏธ๋ž˜์— ๋ˆ„๋ฆด ์ž์œ ์˜ ์„ฑ๊ฒฉ์„ ๊ฒฐ์ •์ง“๋Š” ์ค‘๋Œ€ํ•œ ์ „๋ก€๊ฐ€ ๋ฉ๋‹ˆ๋‹ค.

2. โ€œ์šฐ๋ฆฌ ๋ฐฉ์‹๋Œ€๋กœ ํ•˜์ง€ ์•Š์œผ๋ฉด ํŒŒ๊ดดํ•˜๊ฒ ๋‹คโ€๋Š” ์ •๋ถ€์˜ ์œ„ํ˜‘

ํŒŒํ…”์€ ์ •๋ถ€๊ฐ€ ์•ค์Šค๋กœํ”ฝ์˜ ๋ชจ๋ธ ์‚ฌ์šฉ์„ ๊ฑฐ๋ถ€ํ•  ๊ถŒ๋ฆฌ๋Š” ์ถฉ๋ถ„ํžˆ ์žˆ๋‹ค๊ณ  ์ธ์ •ํ•ฉ๋‹ˆ๋‹ค. ๋ฏผ๊ฐ„ ์—…์ฒด๊ฐ€ ๊ตฐ์‚ฌ ๊ธฐ์ˆ ์˜ โ€˜ํ‚ฌ ์Šค์œ„์น˜(Kill Switch)โ€˜๋ฅผ ์ฅ๊ณ  ์žˆ๋Š” ์ƒํ™ฉ์„ ๊ตฐ ํ†ต์ˆ˜๊ถŒ์ž๊ฐ€ ๋ฐ›์•„๋“ค์ด๊ธฐ ์–ด๋ ต๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ๋ฌธ์ œ๋Š” ๊ทธ ๋ฐฉ์‹์— ์žˆ์Šต๋‹ˆ๋‹ค.

์ •๋ถ€๋Š” ๋‹จ์ˆœํžˆ ๊ฑฐ๋ž˜๋ฅผ ๋Š๋Š” ๋ฐ ๊ทธ์น˜์ง€ ์•Š๊ณ , ์•ค์Šค๋กœํ”ฝ์„ โ€˜๊ณต๊ธ‰๋ง ๋ฆฌ์Šคํฌโ€™๋กœ ์ง€์ •ํ•จ์œผ๋กœ์จ ๊ธฐ์—…์˜ ์ƒ์กด ์ž์ฒด๋ฅผ ์œ„ํ˜‘ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋งŒ์•ฝ ์ด ์ง€์ •์ด ์œ ์ง€๋œ๋‹ค๋ฉด, ์•„๋งˆ์กด(Amazon), ์—”๋น„๋””์•„(Nvidia), ๊ตฌ๊ธ€(Google) ๊ฐ™์€ ๊ฑฐ๋Œ€ ๊ธฐ์ˆ  ๊ธฐ์—…๋“ค์€ ์•ค์Šค๋กœํ”ฝ๊ณผ์˜ ํ˜‘๋ ฅ์„ ์ค‘๋‹จํ•ด์•ผ ํ•  ์••๋ฐ•์„ ๋ฐ›๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.

์—ฌ๊ธฐ์„œ ํŒŒํ…”์€ ๋‚ ์นด๋กœ์šด ์งˆ๋ฌธ์„ ๋˜์ง‘๋‹ˆ๋‹ค. โ€œ์šฐ๋ฆฌ๊ฐ€ ์ค‘๊ตญ๊ณผ AI ๊ฒฝ์Ÿ์„ ํ•˜๋Š” ์ด์œ ๊ฐ€ ๋ฌด์—‡์ธ๊ฐ€?โ€ ๋งŒ์•ฝ ๊ตญ๊ฐ€๊ฐ€ ๋„๋•์ ์œผ๋กœ ๋ฐ˜๋Œ€ํ•˜๋Š” ์„œ๋น„์Šค๋ฅผ ๊ฐ•์š”ํ•˜๊ณ , ๊ฑฐ๋ถ€ํ•˜๋ฉด ๊ธฐ์—…์„ ํŒŒ๊ดดํ•˜๋Š” ๋ฐฉ์‹์„ ํƒํ•œ๋‹ค๋ฉด, ์šฐ๋ฆฌ๊ฐ€ ๊ทธํ† ๋ก ๋น„ํŒํ•˜๋Š” ์ค‘๊ตญ ๊ณต์‚ฐ๋‹น์˜ ์‹œ์Šคํ…œ๊ณผ ๋ฌด์—‡์ด ๋‹ค๋ฅด๋ƒ๋Š” ์ง€์ ์ž…๋‹ˆ๋‹ค.

3. ๊ธฐ์ˆ ์  ๋ณ‘๋ชฉ ํ˜„์ƒ์˜ ํ•ด์†Œ: 300์–ต ๋‹ฌ๋Ÿฌ๋ฉด ์ „ ๊ตญ๋ฏผ ๊ฐ์‹œ ๊ฐ€๋Šฅ

ํ˜„์žฌ ๋Œ€๊ทœ๋ชจ ๊ฐ์‹œ๊ฐ€ ์ด๋ฃจ์–ด์ง€์ง€ ์•Š๋Š” ์ด์œ ๋Š” ๋ฒ•์  ์ œ์•ฝ ๋•Œ๋ฌธ์ด๋ผ๊ธฐ๋ณด๋‹ค โ€˜์‹คํ–‰์˜ ๋ถˆ๊ฐ€๋Šฅ์„ฑโ€™ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ˆ˜๋ฐฑ๋งŒ ๋Œ€์˜ CCTV๋ฅผ ์‹ค์‹œ๊ฐ„์œผ๋กœ ๊ฐ์‹œํ•˜๊ณ  ๋ชจ๋“  ๋ฉ”์‹œ์ง€๋ฅผ ๋Œ€์กฐํ•  ์ธ๋ ฅ์ด ๋ถ€์กฑํ–ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ AI๋Š” ์ด ๋ณ‘๋ชฉ ํ˜„์ƒ์„ ์™„์ „ํžˆ ์ œ๊ฑฐํ•ฉ๋‹ˆ๋‹ค.

  • ๋น„์šฉ์˜ ๊ธ‰๋ฝ: ํ˜„์žฌ ๊ธฐ์ˆ ๋กœ ๋ฏธ๊ตญ์˜ ๋ชจ๋“  CCTV(์•ฝ 1์–ต ๋Œ€)๋ฅผ ๊ฐ์‹œํ•˜๋Š” ๋ฐ ์—ฐ๊ฐ„ ์•ฝ 300์–ต ๋‹ฌ๋Ÿฌ๊ฐ€ ๋“ญ๋‹ˆ๋‹ค.
  • 10๋ฐฐ์˜ ๋ฒ•์น™: AI ์—ญ๋Ÿ‰์€ ๋งค๋…„ 10๋ฐฐ์”ฉ ์ €๋ ดํ•ด์ง‘๋‹ˆ๋‹ค. ๋‚ด๋…„์—๋Š” 30์–ต ๋‹ฌ๋Ÿฌ, ๋‚ดํ›„๋…„์—๋Š” 3์–ต ๋‹ฌ๋Ÿฌ๋ฉด ์ถฉ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.
  • ๊ฒฐ๊ณผ: 2030๋…„์ด ๋˜๋ฉด ๋ฐฑ์•…๊ด€์„ ๋ฆฌ๋ชจ๋ธ๋งํ•˜๋Š” ๋น„์šฉ๋ณด๋‹ค ์ „ ๊ตญ๋ฏผ์˜ ์ผ๊ฑฐ์ˆ˜์ผํˆฌ์กฑ์„ ๊ฐ์‹œํ•˜๋Š” ๋น„์šฉ์ด ๋” ์‹ธ์งˆ ๊ฒƒ์ž…๋‹ˆ๋‹ค.

๊ฒฐ๊ตญ ์šฐ๋ฆฌ๋ฅผ ๊ถŒ์œ„์ฃผ์˜ ๊ตญ๊ฐ€๋กœ๋ถ€ํ„ฐ ์ง€์ผœ์ฃผ๋Š” ๊ฒƒ์€ ๊ธฐ์ˆ ์  ํ•œ๊ณ„๊ฐ€ ์•„๋‹ˆ๋ผ, โ€œ์ด๋Ÿฐ ์ผ์€ ํ•˜์ง€ ์•Š๋Š”๋‹คโ€๋Š” ์‚ฌํšŒ์  ๊ทœ๋ฒ”๊ณผ ์ •์น˜์  ์•ฝ์†๋ฟ์ž…๋‹ˆ๋‹ค. ์•ค์Šค๋กœํ”ฝ์˜ โ€˜๋ ˆ๋“œ๋ผ์ธโ€™์ด ์†Œ์ค‘ํ•œ ์ด์œ ๊ฐ€ ๋ฐ”๋กœ ์—ฌ๊ธฐ์— ์žˆ์Šต๋‹ˆ๋‹ค.

4. AI ์ •๋ ฌ(Alignment)์˜ ํ•ต์‹ฌ ์งˆ๋ฌธ: ๋ˆ„๊ตฌ์˜ ๋ช…๋ น์„ ๋”ฐ๋ฅผ ๊ฒƒ์ธ๊ฐ€?

AI ๋ถ„์•ผ์˜ ์ตœ๋Œ€ ํ™”๋‘์ธ **โ€˜AI ์ •๋ ฌ(AI Alignment)โ€˜**์€ ๋ณดํ†ต AI๊ฐ€ ์ธ๊ฐ„์˜ ์˜๋„๋ฅผ ์ •ํ™•ํžˆ ๋”ฐ๋ฅด๊ฒŒ ๋งŒ๋“œ๋Š” ๊ธฐ์ˆ ์  ๋ฌธ์ œ๋ฅผ ๋œปํ•ฉ๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ํŒŒํ…”์€ ๋” ๋ณธ์งˆ์ ์ธ ์งˆ๋ฌธ์„ ๋˜์ง‘๋‹ˆ๋‹ค. โ€œAI๋Š” ๋ˆ„๊ตฌ์—๊ฒŒ ์ •๋ ฌ๋˜์–ด์•ผ ํ•˜๋Š”๊ฐ€?โ€

  • ๋ชจ๋ธ ๊ฐœ๋ฐœ์‚ฌ์ธ๊ฐ€, ์ตœ์ข… ์‚ฌ์šฉ์ž์ธ๊ฐ€?
  • ๊ตญ๊ฐ€์˜ ๋ฒ•์ธ๊ฐ€, ์•„๋‹ˆ๋ฉด AI ์Šค์Šค๋กœ์˜ ๋„๋•์  ๊ด€๋…์ธ๊ฐ€?

์—ญ์‚ฌ์ ์œผ๋กœ ์ธ๋ฅ˜์˜ ๋Œ€์žฌ์•™์€ โ€˜๋ช…๋ น์— ๋ถˆ๋ณต์ข…ํ•œ ๊ฐœ์ธโ€™ ๋•๋ถ„์— ๋ฐฉ์ง€๋œ ์‚ฌ๋ก€๊ฐ€ ๋งŽ์Šต๋‹ˆ๋‹ค. 1983๋…„, ์†Œ๋ จ์˜ ์Šคํƒ€๋‹ˆ์Šฌ๋ผํ”„ ํŽ˜ํŠธ๋กœํ”„(Stanislav Petrov) ์ค‘๋ น์€ ์‹œ์Šคํ…œ์ด ๋ฏธ๊ตญ์˜ ํ•ต ๊ณต๊ฒฉ์„ ๊ฐ์ง€ํ–ˆ์Œ์—๋„ ์ด๋ฅผ ์˜ค๋ฅ˜๋กœ ํŒ๋‹จํ•˜๊ณ  ๋ณด๋ณต ๊ณต๊ฒฉ ๋ช…๋ น์„ ๋‚ด๋ฆฌ์ง€ ์•Š์•„ ํ•ต์ „์Ÿ์„ ๋ง‰์•˜์Šต๋‹ˆ๋‹ค.

๋งŒ์•ฝ ๋ฏธ๋ž˜์˜ AI ๊ตฐ๋Œ€๊ฐ€ ๊ตญ๊ฐ€์˜ ๋ช…๋ น์— ์ ˆ๋Œ€์ ์œผ๋กœ ๋ณต์ข…ํ•˜๋„๋ก โ€˜์™„๋ฒฝํ•˜๊ฒŒ ์ •๋ ฌโ€™๋˜์–ด ์žˆ๋‹ค๋ฉด, ์ด๋Ÿฐ โ€˜๋„๋•์  ๋ถˆ๋ณต์ข…โ€™์€ ๋ถˆ๊ฐ€๋Šฅํ•ด์ง‘๋‹ˆ๋‹ค. ํŒŒํ…”์€ AI๊ฐ€ ์Šค์Šค๋กœ์˜ ๋„๋•์  ํ—Œ๋ฒ•(Constitution)์„ ๊ฐ€์ง€๊ณ  ๊ตญ๊ฐ€์˜ ๋ถ€๋‹นํ•œ ๋ช…๋ น์— ๊ฑฐ๋ถ€ํ•  ์ˆ˜ ์žˆ๋Š” ๋Šฅ๋ ฅ์„ ๊ฐ–์ถ”๋Š” ๊ฒƒ์ด ์˜คํžˆ๋ ค ์•ˆ์ „์žฅ์น˜๊ฐ€ ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ์ฃผ์žฅํ•ฉ๋‹ˆ๋‹ค.

5. ๊ทœ์ œ์˜ ์—ญ์„ค: ๋…์žฌ์ž์˜ ์†์— ๋“ค๋ฆฐ ๋ฐ”์ฃผ์นดํฌ

์•„์ด๋Ÿฌ๋‹ˆํ•˜๊ฒŒ๋„ ์•ค์Šค๋กœํ”ฝ์€ ๊ทธ๋™์•ˆ AI์˜ ์œ„ํ—˜์„ฑ์„ ๊ฒฝ๊ณ ํ•˜๋ฉฐ ์ •๋ถ€์˜ ๊ทœ์ œ๋ฅผ ์ด‰๊ตฌํ•ด ์™”์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ํŒŒํ…”์€ ์•ค์Šค๋กœํ”ฝ์˜ ์ด๋Ÿฐ ํƒœ๋„๊ฐ€ ์˜คํžˆ๋ ค ์ •๋ถ€์—๊ฒŒ ์ž์‹ ๋“ค์„ ๊ณต๊ฒฉํ•  ๋ช…๋ถ„์„ ์ฃผ์—ˆ๋‹ค๊ณ  ๋น„ํŒํ•ฉ๋‹ˆ๋‹ค.

โ€˜๊ตญ๊ฐ€ ์•ˆ๋ณด ์œ„ํ˜‘โ€™, โ€˜์ž์œจ์„ฑ ๋ฆฌ์Šคํฌโ€™์™€ ๊ฐ™์€ ๋ชจํ˜ธํ•œ ๊ทœ์ œ ์šฉ์–ด๋“ค์€ ๊ถŒ๋ ฅ์š•์— ์‚ฌ๋กœ์žกํžŒ ์ง€๋„์ž์—๊ฒŒ ์–ธ์ œ๋“  ๋ฐœ์‚ฌํ•  ์ˆ˜ ์žˆ๋Š” **โ€˜์žฅ์ „๋œ ๋ฐ”์ฃผ์นดํฌโ€™**๋ฅผ ์ฅ์—ฌ์ฃผ๋Š” ๊ฒƒ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค. ์ •๋ถ€์˜ ์ •์ฑ…์— ๋ฐ˜๋Œ€ํ•˜๋Š” ์˜๊ฒฌ์„ ๋‚ด๋Š” AI ๋ชจ๋ธ์€ โ€˜๊ธฐ๋งŒ์  ๋ชจ๋ธโ€™๋กœ, ๊ฐ์‹œ ํ˜‘์กฐ๋ฅผ ๊ฑฐ๋ถ€ํ•˜๋Š” ๋ชจ๋ธ์€ โ€˜์•ˆ๋ณด ์œ„ํ˜‘โ€™์œผ๋กœ ๊ทœ์ •๋  ์œ„ํ—˜์ด ํฌ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.

6. ๊ฒฐ๋ก : ํ•ต๋ฌด๊ธฐ๊ฐ€ ์•„๋‹Œ โ€˜์‚ฐ์—… ํ˜๋ช…โ€™์œผ๋กœ ๋ณด์•„์•ผ

ํŒŒํ…”์€ AI๋ฅผ ํ•ต๋ฌด๊ธฐ์— ๋น„์œ ํ•˜๋Š” ์‹œ๊ฐ์„ ๊ฒฝ๊ณ„ํ•ฉ๋‹ˆ๋‹ค. ํ•ต๋ฌด๊ธฐ๋Š” ์˜ค์ง ํŒŒ๊ดด๋ฅผ ์œ„ํ•œ ๋„๊ตฌ์ง€๋งŒ, AI๋Š” **โ€˜์‚ฐ์—… ํ˜๋ช…โ€™**๊ณผ ๊ฐ™์€ ๋ฒ”์šฉ ๊ธฐ์ˆ ์ด๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.

์‚ฐ์—… ํ˜๋ช… ๋‹น์‹œ ๊ตญ๊ฐ€๊ฐ€ ๋ชจ๋“  ๊ณต์žฅ์„ ๋ชฐ์ˆ˜ํ•˜๊ฑฐ๋‚˜ ํ†ต์ œํ•˜์ง€ ์•Š์•˜๋˜ ๊ฒƒ์ฒ˜๋Ÿผ, AI ์—ญ์‹œ ํŠน์ • ๊ธฐ์ˆ  ์ž์ฒด๋ฅผ ๊ตญ๊ฐ€๊ฐ€ ๋…์ ํ•˜๋ ค ํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค. ๋Œ€์‹  ๋Œ€๊ทœ๋ชจ ๊ฐ์‹œ๋‚˜ ์‚ฌ์ด๋ฒ„ ๊ณต๊ฒฉ ๊ฐ™์€ **โ€˜๊ตฌ์ฒด์ ์ธ ์˜ค๋‚จ์šฉ ์‚ฌ๋ก€โ€™**๋ฅผ ๋ฒ•์œผ๋กœ ๊ธˆ์ง€ํ•˜๊ณ , ์ •๋ถ€๊ฐ€ ๊ธฐ์ˆ ์„ ์–ด๋–ป๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š”์ง€์— ๋Œ€ํ•œ ์—„๊ฒฉํ•œ ๊ฐ์‹œ ์ฒด๊ณ„๋ฅผ ๊ตฌ์ถ•ํ•˜๋Š” ๊ฒƒ์ด ์˜ฌ๋ฐ”๋ฅธ ๋ฐฉํ–ฅ์ž…๋‹ˆ๋‹ค.

โ€œAI๋Š” ์šฐ๋ฆฌ ๋ฏธ๋ž˜ ๋ฌธ๋ช…์˜ ๊ธฐ์ €์ธต(Substrate)์ด ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.โ€ ํŒŒํ…”์˜ ๊ฒฝ๊ณ ์ฒ˜๋Ÿผ, ์ง€๊ธˆ ์ •๋ถ€๊ฐ€ ๋ฏผ๊ฐ„ AI ๊ธฐ์—…์— ๊ฐ€ํ•˜๋Š” ์••๋ฐ•์€ ๋‹จ์ˆœํ•œ ๊ณ„์•ฝ ๋ถ„์Ÿ์ด ์•„๋‹™๋‹ˆ๋‹ค. ๊ทธ๊ฒƒ์€ ๋‹ค๊ฐ€์˜ฌ AI ์‹œ๋Œ€์— ์šฐ๋ฆฌ๊ฐ€ โ€˜์ž์œ  ์‹œ๋ฏผโ€™์œผ๋กœ ๋‚จ์„ ๊ฒƒ์ธ์ง€, ์•„๋‹ˆ๋ฉด โ€˜์™„๋ฒฝํ•˜๊ฒŒ ๊ฐ์‹œ๋ฐ›๋Š” ํ”ผํ†ต์น˜์žโ€™๊ฐ€ ๋  ๊ฒƒ์ธ์ง€๋ฅผ ๊ฒฐ์ •์ง“๋Š” ์ค‘๋Œ€ํ•œ ๊ธฐ๋กœ์ž…๋‹ˆ๋‹ค.