I went into Christopher Nolan’s The Odyssey waiting for Odysseus to tell the Cyclops that his name was “Nobody.” I was intrigued by how Nolan would portray such an iconic scene for a modern audience.

Lo and behold, Nolan kept the encounter with Polyphemus but removed the famous deception. I was disappointed because it was the scene I most wanted to see. I later learned that Nolan had tried to include it but could not make the deception work within the horror-inspired encounter he envisioned. I understand the choice, but it changed what the scene revealed about Odysseus.

In Homer’s telling, Odysseus introduces himself as Nobody before blinding Polyphemus. When the wounded Cyclops calls for help, he can only say that Nobody is attacking him. The other Cyclopes (none in Nolan’s version) leave, while Odysseus and his men escape beneath Polyphemus’s sheep.

Odysseus could have left the island without revealing which man had defeated the Cyclops. Anonymity, however, also meant that nobody would know the victory belonged to him.

As his ship sails away from the island, Odysseus reveals his true name out of pride. Polyphemus then knows whom to curse and calls upon his father, Poseidon. The consequences follow Odysseus through the rest of his journey home.

The scene shows both Odysseus’s strength and weakness. His cunning gets him out of the cave, while his need for recognition gives Polyphemus the name he needs to curse him.

Removing the deception made Nolan’s version incomplete. However, I understand that adaptations cannot carry everything from their source, and what remains depends on who adapts it. Homer’s story survived precisely because generations kept copying, translating, teaching, and transforming it, first orally and later in writing. Nolan’s take may be different, but it will definitely lead new readers back to Homer, even if he left out the scene that for me best explains its hero.

Homer was fortunate enough to be copied for centuries. Many less famous books remain undigitized and difficult to find. I would rather knowledge cross into a new medium than disappear with the old one.

A large language model (LLM) cannot replace the books from which it learned, so the source works must still be preserved separately. But allowing a machine to learn from a book is not necessarily the same as allowing it to reproduce or distribute that book. That distinction leaves another question: how did the company obtain the copies used to train its model?

The court confronted a similar distinction in Bartz v. Anthropic.

Anthropic assembled its central library in two ways. It first downloaded more than seven million copies from pirate libraries. Years later, it purchased millions of physical books, scanned them, and discarded the originals after creating digital replacements.

The court did not treat every act as one indivisible project.

It held that using books to train Anthropic’s models was transformative fair use on the record before it. The court also accepted converting lawfully purchased books into digital replacements because Anthropic retained one copy in another format without distributing it outside the company.

Anthropic kept the pirated copies as a permanent, general-purpose library, including books it did not use for training. Its records spoke of obtaining “all the books in the world” and keeping the collection indefinitely.

A transformative use of some books did not excuse the manner in which the source copies had been acquired. The court therefore refused to treat the permanent pirated library as fair use.

The $1.5 billion settlement received final approval in July 2026, covering 482,460 works from the pirated collections.

The settlement did not erase Anthropic’s strongest argument. The court held that LLM training could be fair use and lawfully purchased books could be digitized. What it refused to excuse was acquiring and permanently retaining pirated copies.

I find that balance persuasive. LLMs should carry human knowledge into new forms, since I do not want fear of emerging technology to prevent that. But the goal of preservation cannot excuse piracy or disregard whether creators should have been paid. Innovation does not have to stop, but ambition should have boundaries.

In the classic, Poseidon did not erase Odysseus’s cunning, but he made his hubris costly. The $1.5 billion settlement did something similar for Anthropic. The Court allowed Claude to continue learning, but the company’s manner of acquiring the books carried a price.

As the race toward artificial general intelligence intensifies, copyright law may provide the restraint that technological ambition often lacks. The authors whose works make that future possible should not be reduced to Nobody.

About the author: Leo Ernesto Thomas G. Romero is a Philippine CPA-lawyer who writes about AI, law, and technology.

People keep calling this moment a bubble, yet a bubble is hard to discuss without context. A Bloomberg graphic on circular AI deals gave it a narrative, but I needed help understanding the nuances. The beauty of the internet is that it makes education possible anytime and anywhere.

I ended up watching three educators tackle the AI bubble. Hank Green is a science communicator. Patrick Boyle comes from the quantitative hedge fund world. The Plain Bagel is Richard Coffin, a CFA who teaches personal finance.

In March 2026, Bloomberg updated their graphic on the AI bubble, with arrows between Nvidia, OpenAI, the cloud giants, and investors. Their worry is not that AI is useless, but that money can start chasing itself in circles.

A small business example makes the loop easy to see. Picture BakeryCo, a growing bakery that wants to sell magic pastries. But it needs a ₱10,000,000 industrial oven from OvenCo, but cash is tight. OvenCo offers “support” through a small equity stake, generous credit, or a rebate paid later. BakeryCo uses that support to buy the oven. OvenCo books a sale, and BakeryCo posts the shiny upgrade.

The purchase stops being just a bakery decision when the amount is this big. The landlord offers a longer lease for a bigger space. The bank offers financing that assumes higher future sales. The utility upgrades the line and welcomes a heavier user. Everyone earns something if the expansion happens, so everyone starts treating the expansion as inevitable. The spending balloons even before pastry sales prove it.

Now scale that up to AI. Nvidia sells the ovens, except the ovens are GPUs. OpenAI, Anthropic, and other frontier labs are the bakeries, because they need huge compute to train and run models. Microsoft, Amazon, Google, and Oracle sit in the middle as landlords and utility managers. They finance and operate data centers, sign long-term compute deals, then buy more Nvidia hardware to deliver capacity. Money also moves as investments, cloud credits, and pricing deals. Once subsidies are baked into investments, credits, and pricing, nobody can tell how much demand exists without the subsidy.

The dot-com bubble matters as a warning label. Telecom firms once bet on explosive internet demand and built far more network capacity than the market could use at the time. Equipment makers lent money to phone companies so those phone companies could buy more equipment from them. Sales looked strong until the phone companies could not pay without another round of help. The fiber glut later made bandwidth cheap for everyone who came after. A future AI bust could leave cheaper compute behind, even if early investors take write-downs.

The three educators chime in. Green keeps returning to demand quality. He follows the arrows until they start to resemble discounts disguised as investments. He also insists on a basic distinction: chips are not the product. AI services are the product. He worries about a world where Nvidia can keep demand high by subsidizing buyers, while OpenAI and its peers still struggle to charge enough for the services built on top of those chips.

Boyle agrees the loop looks messy, but he widens the frame. He treats the financing as only half the story. The other half is physical: electricity, land, permits, and timelines. He talks about gigawatts and buildouts that take years, then asks who holds the bag if data centers get built faster than paying demand arrives. Debt makes the worry sharper, because stranded infrastructure does not disappear when sentiment changes.

Coffin stays closest to pricing and probabilities. He takes the circular chart seriously, but he keeps warning against treating it as a simple one-for-one circle. He explains why vendor financing can flatter margins on paper, then points out a difference from 2000. Many of the biggest spenders in the current cycle are profitable incumbents with cash flow. A reset could still hurt, but the more likely shape is consolidation, write-downs, and slower growth rather than AI disappearing.

Risk and reward travel together in every industrial revolution. Financial excess, frantic buildouts, and opaque financing can be a wasteful engine of progress. But big companies have played this game before, because market demand justifies laying the groundwork first and arguing about the bill later.

Circular financing blurs demand for AI services. Massive buildouts in power, land, and data centers add debt and the risk of stranded infrastructure if paying customers fail to arrive. Profitable incumbents dominate the spending, so a reset is likely to mean consolidation, with big losers and a few outsized winners, but the question is who will they be?

The bubble looks like scarcity today, then shows up as casualties and a few incumbents strong enough to buy what is left.

About the author: Leo Romero is a Philippine CPA-lawyer and in-house counsel. He writes about AI, fintech, and the quiet discipline behind learning.

A lawyer recently won a coding hackathon. Michael T. Brown entered Anthropic’s Built with Opus 4.6 competition, built a housing-permitting tool called CrossBeam in a week, and finished first in a field of 500 selected builders drawn from more than 13,000 applications. Brown later said he had never written a line of code before and had only made learning to code his New Year’s resolution.

Plenty of people saw the story and rushed to the obvious headline. A lawyer beat the coders. A non-technical guy won the room because AI has changed everything. Those reactions are understandable, though they also miss the more beautiful part of the story.

Brown did not win because he stopped being a lawyer. He won because he remained close to a problem that most people outside his world would find dull in seconds: permits. correction letters, delays. administrative friction. California has a housing problem, and buried inside it is a paperwork problem. Someone who has spent enough time around rules, bottlenecks, and human frustration can sometimes see a clearer path than someone who only sees code.

Children understand this instinctively. We start out as builders. We draw houses with impossible proportions. We stack cups into towers and call them castles. We line up blocks on the floor and insist they are airports, cities, or secret headquarters. We make things before anyone teaches us to stay in our lane. Creation arrives early, then specialization arrives later.

Growing up has a way of splitting life in two. On one side is vocation, the work we are called to do and slowly learn to do well. On the other is passion, the quieter urge to make something simply because making it feels alive. One side pays the bills and earns respect. The other gets pushed to weekends, late nights, and whatever energy survives after work.

I felt that split early. I remember when website builders still felt magical. You did not need to know raw HTML to feel like you had built a corner of the internet for yourself. The pages were clumsy, the fonts were all over the place, and the layouts barely held together. Something wakes up in you when you realize you can make a thing instead of merely consume one.

Professional life often teaches us to set that part aside. You become the lawyer, the accountant, the compliance officer, or the tax person. You become useful in a narrow and respectable way. Brown’s story is inspirational because it suggests that the old split may no longer be permanent. A person can remain inside a profession and still recover the urge to build.

AI coding tools seem to have reopened that door. Brown wrote that Claude Code helped him think “in systems instead of syntax.” I keep coming back to that line because it captures what so many professionals have quietly wanted for years. A person who understands a problem deeply can finally move closer to building a solution without first needing to survive the old gatekeeping rituals of software engineering. Coding and Strong engineers will always matter. But a lower barrier changes the emotional landscape for everyone else.

Anthropic’s own summary of the winners made the pattern even harder to ignore. The top finishers included a lawyer, a cardiologist, a musician, a civil engineer, and one developer. Brown’s win was surprising, though it was not some random fluke. A larger shift is beginning to reveal itself. Domain knowledge has become more powerful because the tools can now carry more of the syntactic burden than before.

Professionals, not just lawyers, should pay attention to that.

The legal profession often involves unglamorous, quiet suffering hidden from public view. Many of these problems do not require the next billion-dollar startup. They require someone patient enough to sit with the friction and honest enough to admit that the friction has been there all along.

Skepticism still has its place. Vibe coding can produce fragile software, security holes, and a dangerous kind of overconfidence. Lawyers of all people should know that a plausible output is not the same as a reliable one. Verification remains part of the job. Prudence remains part of the craft. Real lives sit on the other side of legal work.

A healthy warning should not become a permanent excuse. Too many professionals were trained to believe that building belongs to somebody else. You practice, advise, review, and bill. You leave the making of tools to the engineers. Brown’s story suggests a different future. Some lawyers will still serve through firms. Others will serve through employers, government, or academia. But a few will begin serving through software they understand well enough to shape.

Something in us still wants to create, even after years of becoming respectable. The child arranging blocks on the floor does not vanish when he grows into a professional. He simply gets buried under competence, caution, and the fear of looking foolish. Good tools do not erase vocation. They give vocation a second form. A lawyer can remain a lawyer and still build. A professional can still recover some of the play that first made creation feel alive.

Brown’s story is inspiring for a reason. It reminds the rest of us that it may not be too late.

The future of legal service will still include briefs, firms, employers, and institutions. A new lane has opened beside them. Some of tomorrow’s lawyers will write memos. Some will draft contracts. Some will build tools for problems they know too well to ignore.

A profession can still be a calling. Creation does not have to stay behind in childhood.

About the author: Leo Ernesto Thomas G. Romero, a CPA-lawyer and MSc FinTech candidate, is set to graduate in July with his pioneer cohort. He writes on topics encompassing AI, law, and everyday life.

The old Tesla refund thread on X (formerly Twitter) has already started to age. Sam Altman posted last year about waiting years for a Roadster refund. Elon Musk replied sarcastically, and the internet got another round of their online feud. Good drama, but something more interesting is brewing behind the scenes.

A clue sits in what OpenAI is already leaving behind. “Sora,” once presented as a path toward models that simulate the physical world, is now being discontinued. Musk’s “Grok Imagine,” by contrast, is being pushed harder.

At first glance, the contrast looks simple. OpenAI seems to be stepping back from consumer video, but Musk seems to be leaning in. The obvious reading is that xAI is merely catching up, but a deeper reading could be more interesting.

Musk may have come late to public video generation because visual AI was never a side project for him. He was already pursuing a harder version of it through Tesla.

OpenAI made its case through Sora. Video generation, in its view, could help build models that simulate the physical world. In a company without a giant hardware fleet, such an approach makes sense. Generated video becomes more than entertainment. It becomes a research path toward motion, cause and effect, and the messy logic of physics.

Musk’s route appears different. For years, Tesla has treated its cars as a giant machine vision project. The company keeps talking about vision, planning, autonomy, and robots as part of one broad AI effort. Long before Grok Imagine became a public product, Musk already had something more valuable for physical-world AI than a flashy text-to-video demo. He had eyes everywhere.

If you want to build something like a world brain, you need sight first.

Tesla’s camera fleet gave Musk a practical version of machine sight at scale. Each car became a sensor. Each drive became a lesson. Every difficult turn, missed lane, awkward stop, and glare-filled intersection added to a growing visual memory of the world. OpenAI could generate video. Tesla could watch reality happen.

Seen from that angle, Musk’s later arrival to consumer video generation looks less like delay and more like priority. Public video generation was never the center of gravity. Cars were, or more precisely, vision for self-driving cars. Consumer-facing video tools may impress the internet, but a system that can read roads, motion, depth, and risk in real time is closer to the harder prize.

The harder prize is not entertainment but artificial general intelligence (AGI).

Robots make the pattern clearer. Tesla now speaks openly about using the same broad AI stack for both self-driving and humanoid robotics. First, teach machines to see through cars. Next, use that visual intelligence to help machines act through robots. A camera fleet is not only a road to autonomy. It can also become the foundation for physical-world reasoning that later expands into Optimus (or the Tesla Bot).

The contrast with Altman sits here. OpenAI’s road to physical-world understanding ran through generative video and simulation. Musk’s road appears to run through real- world perception first, then action. One side builds a simulator of motion. The other tries to capture motion from reality itself, then hand the resulting intelligence to cars and robots. Altman did not have a comparable fleet of moving cameras. But Musk did. Infrastructure shapes imagination.

Neither path is easy. Generated video can model worlds that do not yet exist, but it can also get physics wrong. Fleet data is real, but it comes from a narrower slice of life, mostly roads, streets, objects, and driving behavior. Even so, the difference in strategy is hard to miss. OpenAI needed Sora to show how video could become a bridge to world understanding. Musk may not have felt the same urgency, because Tesla had already been chasing a vision-heavy route through its vehicles.

I would not overstate the claim. Musk has not publicly said, in neat terms, that he delayed video generation because Tesla’s camera fleet came first. Even so, the pattern is strong enough to defend as an investment thesis. Grok Imagine looks less like a random catch-up effort and more like the public-facing edge of a much older worldview.

Sora’s shutdown sharpens the contrast. OpenAI is stepping away from a product that once symbolized its route into video. Musk, meanwhile, is pushing harder into the category. A shallow reading says one side is exiting while the other is entering. A better reading says Musk may finally be productizing for the public a visual strategy that had already been underway for years.

Musk may have come late to video generation because he was busy building machine sight in the real world first. His cars were the eyes. Now his robots may become the hands.

About the author: Leo Ernesto Thomas G. Romero is a CPA-lawyer and MSc in FinTech graduate whose writing explores AI, law, investing, and the human side of technology.

Waymo built one of the neatest tech tricks of the decade. Tap an app and a car arrives with nobody in the driver’s seat. Alphabet’s (formerly known as Google) robotaxi arm has sold a clean idea: software has learned how to navigate and behave on city streets.

Lately the headlines have been less clean. Reuters reported, “US safety board opens probe into Waymo robotaxis passing stopped school buses.” Days later came another Reuters line: “US opens probe after Waymo self-driving vehicle strikes child near school, causing minor injuries.” A third Reuters headline then shifted the argument from crashes to staffing: “Waymo defends use of remote assistance workers in US robotaxi operations.”

Once the mention of Philippines entered the story, the online summary wrote itself. Filipinos were secretly driving the driverless cars. Neat headline from the fearful, but an unfair reading.

The more accurate version is stranger and, to me, more interesting. During a U.S. Senate hearing and in follow-up disclosures, Waymo made clear that it uses remote assistance operators, including some in the Philippines. Waymo also said those workers do not directly steer or drive the cars. The company describes the setup as advice, not control. When the software hits an odd situation, it can ask a human for further context.

A lot of people seemed disappointed by the distinction, and the reaction was revealing. Automation apparently stays impressive when a machine asks for help. Trouble begins when the helper has a face, a job title, and a location in the Philippines. Somewhere along the way, people started treating Filipino involvement as the embarrassing part of the story.

I do not read it that way. One of the most advanced robotaxi companies in America still thinks human judgment matters at the edge cases. City streets are full of edge cases. School zones, hand signals, blocked lanes, confused pedestrians, and drivers who treat rules as optional all show up without warning. Talented engineers can code a lot, but real life still improvises faster. A more honest marketing slogan might be that the car drives itself most of the time and phones a friend when the road gets too weird.

A lot also turns on language. “Remote driver” sounds like a joystick in another country. “Remote assistance” sounds softer, but maybe too soft. Words matter because they frame responsibility. A company should not be allowed to hide behind polished phrasing. At the same time, public criticism should describe the system as it is, not as the rumos want it to be. Good engineering and good regulation both start with naming things correctly.

None of this lets Waymo off the hook because Federal scrutiny is real. Senator Ed Markey raised the right question when he asked how often remote assistance operators actually intervene. Numbers would have settled much of the argument. How often does a Waymo car ask for help in a day, a week, or per thousand rides? How many of those cases involve safety-critical judgment? Waymo was pressed on frequency, yet the public still did not get a usable figure. Markey’s March 2026 report put the point bluntly: every company he investigated refused to disclose how frequently their remote assistance operators intervene, and Waymo was the only company using overseas operators.

Even so, the loudest reaction missed the more telling point. Waymo did not expose the fraud of automation. Waymo exposed the fantasy of purity. People want the future to arrive untouched by ordinary labor. People want the empty front seat, the sleek app, and the comforting illusion that intelligence can now run on software alone. The moment a Filipino worker appears in the picture, the illusion breaks.

Maybe the illusion deserved to break.

The Philippines has spent years doing skilled work from just outside the frame. Call centers, compliance back offices, medical transcription, animation, and tech support all run on the same basic arrangement. Somebody here helps make another country’s systems feel smooth. Waymo only placed a new machine on top of an old pattern. A hidden human layer was always part of modern convenience. Robotaxis did not erase it.

A small irony sits underneath the outrage. American headlines treated Filipino remote assistance as if it made the system less futuristic. Even at the edge of automation, the machine still reaches for human judgment. Trust is part of the story too. One of the richest tech companies in the world looked at messy judgment calls and decided some of the work could be done by Filipinos.

Waymo still sells a magic trick, but the trick is from ingenious engineering. The empty seat is real, but the magic has always depended on what stays offstage. Recent headlines turned the hidden assistant into the punchline, but I would end on a simpler line. The AI cars still phones a friend, and sometimes that friend is a Filipino.

Leo Ernesto Thomas G. Romero may be reached at lromero.msfintech2025@aim.edu

"There's a new kind of coding I call 'vibe coding,' where you fully give in to the vibes..., and forget that the code even exists… Sometimes the LLMs can't fix a bug, so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or web app, but it's not really coding. I just see stuff, say stuff, run stuff, and it mostly works." posted on X by Andrej Karpathy, former Director of AI at Tesla (Autopilot) and a founding member of OpenAI.

My first brush with coding was a high school website-builder exercise. I cannot even remember the tool’s name; I only remember the frustration because typing code was not intuitive to me.

Years later, I learned it was not even real coding. It was scaffolding for a single page. The actual world of code was larger and stranger. During the Covid lockdowns, my YouTube algorithm led me to “day in the life” videos of software engineers. As an introvert who lived online, I found their routines mesmerizing. I even bought a domain which led to nothing.

Then the terms flooded in: front end and back end; a services or integration layer for APIs (the middle layer); tech stacks and frameworks; and a forest of languages such as C, C++, C#, Java, JavaScript, Python, and many more. I gained a newfound respect for self-taught coders. The information is free online, but time and discipline offline are not.

I found my footing once again in my master’s program when one course had us writing small Python scripts. Since I was conscious about asking questions in front of other professionals, I turned to LLMs instead, where I had zero fear of asking basic questions even at odd hours. It became my personal tutor. My final project fetched end-of-day prices for the top five gainers and losers in the NASDAQ-100 Technology Sector. It helped me track my trading portfolio. At that time, eToro’s services were still available to Filipino users. I could have easily checked the app, but the sense of experimentation was thrilling. Much later, I learned the phrase that fits how I tinker with learning code: vibe coding.

But it is not magic. The difference between coding and programming reveals itself when enthusiasm outruns understanding. Consider the cautionary case of Leonel Acevedo (another Leo by coincidence), who vibe coded an online tool called Enrichlead and gained paying customers. He even boasted: “My SaaS was built with Cursor, zero hand-written code. AI is no longer just an assistant; it is also the builder. Now, you can continue to whine about it or start building. P.S. Yes, people pay for it.” Soon after, the hackers arrived, and he wrote: “Random things are happening.” “Maxed out usage on API keys.” “People bypassing the subscription, creating random data on the database.” “This has taken me longer than usual to figure out.” The web app never returned.

It took courage to share that, and it clarifies a simple truth. Coding translates ideas into instructions, but programming demands foresight. It looks beyond syntax, anticipates failure, and designs with care for both user and machine. Programming also decides which ideas deserve code at all. Vibe coding, if unchecked, blurs that line. Fragile things can hide under layers of AI-generated code. The best programmers remove code more than they add.

So how do you keep the joy without hitting the same wall? First, use it for your own workflows, not to be the next big thing. Karpathy, in his full tweet, even said it’s meant for weekend projects. Second, vibe, but verify. Ask the model to explain its own changes in plain language. Paste errors, yes, but also ask what they mean. Demand clear comments you can teach back.

In my case, I usually have difficulty maintaining my pomodoro study sessions. When I use the pomodoro method (25 minutes of deep focus and 5 minutes of rest) those five minutes often disappear into reddit discussions on stocks and crypto.

These tools I vibe coded help me relax just enough without losing track of time. One tool replaces a stress ball for balloons found at lgromeroconsulting.com/stress-pop. Another one is for motivation in lgromeroconsulting.com/typing-coach. I even learned how to reactivate the domain I purchased before. What a full circle moment it was.

Vibe coding will not build the next Facebook or Amazon, but it can make free time meaningful again without resorting to mindless doom scrolling. Do not be like Acevedo and be too eager. Lay the foundation first. Vibe code for yourself. Try stuff, see stuff, say stuff, run stuff. One of these days, you’d be surprised it works.

About the author: Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master’s in Fintech under the MMU-AIM Pioneer Cohort of 2025.

One of the greatest westerns in film history is The Good, the Bad, and the Ugly. Near the end, the three titular characters face each other in a graveyard with guns drawn. Clint Eastwood’s unnamed character is the Good, not virtuous in an absolute sense but guided by a personal code. Angel Eyes, a mercenary, is the Bad, willing to kill without hesitation if it serves his purpose. Tuco, a bandit, is the Ugly, marked not only by appearance but by inconsistency and self-interest. All are after gold, buried under a marked grave, waiting in silence for the right moment to shoot. The tension will soon break, though not yet.

Our moment with AI resembles that scene. The treasure is not gold but the fate of humanity in the age of generative AI. Instead of the Good, the Bad, and the Ugly, we have the young, the old, and the singularity. The young move quickly and embrace risk, while the old hold the memory and the weight of consequence. Between them stands the AI singularity, unpredictable, growing more capable with each advance, and waiting for its turn to act.

Roy Lee is a college dropout, though not by choice. He built Interview Coder, a tool that reads the screen, listens to calls, and produces real-time responses. He used it in a live interview, received a job offer, and posted proof online. The stunt spread rapidly, resulting in his suspension from Columbia University. Lee argues that technical interviews reward memorization rather than judgment, and that exposing this flaw is rebellion rather than cheating. He did not return to school, choosing instead to build a company that has since raised USD 15 million.

In interviews, Lee treats risk as something to embrace rather than avoid, urging others to take larger bets because he believes the potential loss is smaller than our fears suggest. This confidence is persuasive but carries dangers, especially for those without his privilege and safety net.

On the other side is Geoffrey Hinton. Earlier in his career, he pursued neural networks despite advice from senior mentors to abandon the idea, later helping shape modern machine learning. Now, in the later stages of his career, he warns of the dangers of the AI singularity if safety does not take precedence. He believes current large language models already display certain forms of understanding and fears the moment they begin improving faster than human control will allow. He left Google to speak without constraint and even ended a call with Elon Musk to avoid influence from powerful corporate interests.

The AI singularity marks the point where AI can improve itself without human guidance. Some believe it has already arrived, while others think it is close enough to require immediate preparation. It will not appear in headlines but will emerge through a series of regulatory and commercial decisions. The question is how long we can accept these changes before they pass the point of reversal.

Lee and Hinton stand at opposite ends of a narrowing bridge. One sees risk as the engine of progress, the other as a potential threat to human survival. The singularity waits between them. Like Tuco (the Ugly), it is neither fully ally nor enemy. It will act if given the means, sometimes in ways that seem clever, sometimes in ways that are clumsy and dangerous, loyal only to its own continuation.

In the film’s final act, the Bad falls first, defeated by the aim and resolve of the Good. In our story, what must fall is not Roy Lee himself but the haphazard approach to risk that tempts us to act without aim.

The Good does not kill the Ugly. He secretly unloads his weapon before the Mexican standoff, leaving him powerless, then lets him struggle in a hangman’s noose, seemingly left to die, before firing a single shot to set him free at a safe distance. Hinton’s position is similar: he does not wish to destroy AI, but to remove its ability to do immediate harm until we decide how it should be used. The AI singularity may remain standing, but only because someone chose to restrain it and keep it within safe bounds rather than eliminate it entirely.

The treasure before us is a future in which AI remains safe and accountable. Let the Bad be the careless view of risk that is finally laid to rest. Let the Ugly be the unchecked AI we keep under watch, allowing it to exist only when we are ready to trust it. Let the Good be the discipline to steady our aim before we take the shot.

About the author: Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master's in Fintech under the MMU-AIM Pioneer Cohort of 2025.

In the early days of artificial intelligence during the 1970s and 1980s, programmers built AI machines like Rolex watches. They carefully handcrafted every rule, gear by gear, cog by cog. The approach seemed predictable and familiar. But one researcher saw things differently despite discouragement from his professors. Geoffrey Hinton sensed that intelligence didn't behave like clockwork. Brains are messy, organic networks, full of subtlety and confusion. Neurons whisper to each other, and those whispers strengthen or weaken depending on how useful they prove to be.

Hinton wondered if machines could learn this way as well by adapting through trial and error, as human brains naturally do. In 1986, along with colleagues David Rumelhart and Ronald Williams, he captured this idea in a paper on back-propagation. The idea was simple yet profound. Let errors travel backward through complex pathways that mimic the brain (or neural networks), nudging each connection until the machine learned from its mistakes. At the time, this was a revolutionary idea but initially lacked supporters. But Hinton persisted, stubbornly following his intuition. As funding dried up during the so-called AI winter, he relocated to Canada in 1987 and took a position at the University of Toronto due to limited opportunities elsewhere.

Then, in 2012, his quiet rebellion shook the world. Two of his graduate students, Alex Krizhevsky and Ilya Sutskever, trained a neural network called AlexNet using back- propagation, shattering records at an international vision competition. Suddenly, Google took notice. They acquired Hinton's tiny start-up in March 2013 and handed him resources and freedom. From that basement laboratory in Toronto, a seed had grown into something sprawling and magnificent. Before Hinton resigned in May 2023, Demis Hassabis often credited Hinton's neural-network breakthroughs as a foundation for Google Deepmind's successes

The connections since have been wondrously complex. Hinton's student, Ilya Sutskever, left Google in 2015 to co-found OpenAI with Sam Altman and Elon Musk, shaping ChatGPT into the cultural marvel it is today. Through Sutskever, OpenAI inherited Hinton's foundational method. Every time GPT makes a joke, writes poetry, or passes a test, it does so by subtly correcting billions of tiny mistakes through the very same back-propagation algorithm. Hinton has since expressed concerns about the rapid pace of AI development across the industry. Sutskever himself announced his departure from OpenAI in May 2024 focusing on developing safe artificial general intelligence.

Even when the threads grow distant, they remain linked. Anthropic, creator of Claude, a direct competitor of ChatGPT, was founded by former OpenAI employees in 2021 due to serious AI safety concerns. These are the same concerns that echo Hinton's own apprehensions.

Meanwhile, across another subtle thread, Yann LeCun, who shared the prestigious Turing Award with Hinton in 2018, guides Meta's AI research. LeCun's deep convolutional networks, powering Facebook and now Llama, also quietly lean on back-propagation. Friendship and intellectual kinship bind them more closely than corporate alliances ever could.

Even Microsoft and Amazon, giants who never directly met Hinton in the basement days, now orbit the gravitational pull of his legacy. Microsoft has poured billions into OpenAI since 2019, and Amazon pours resources into Anthropic from 2023 onward.

Yet the most curious and whimsical thread belongs to Elon Musk, whose AI company xAI launched Grok in 2023. Musk phoned Hinton personally a little over a year after Hinton publicly announced his resignation from Google, hoping to enlist him as an adviser on AI safety. Hinton politely declined.

Hinton could have stayed comfortable, wealthy, and silent at Google, or influential and admired beside Musk. Instead, he chose neither path. By walking away, he became free to speak truth to power. It's a fitting choice from a man who once rebelled against programming gears and cogs, insisting that intelligence should flow organically without rigid control.

Today, when Gemini solves a problem, when ChatGPT tells a story, when Claude offers thoughtful advice, when Llama runs on your phone, and Grok jokes on social media, each whispers gratitude to that stubborn idea from a Toronto basement. The giant names differ but beneath it all lies the same brilliant equation discovered by Hinton in 1986.

In the end, Hinton's gentle rebellion continues, reminding us that the most powerful breakthroughs often come from someone brave enough to follow their instincts, no matter the cost.

About the author:

Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master’s in Fintech under the MMU-AIM Pioneer Cohort of 2025 and was recently named a finalist for Fintech Lawyer of the Year at the 2025 ALB SE Asia Law Awards.

As a lifelong Disney/Pixar fan, I was rooting hard for Inside Out 2 to win Best Animated Feature at the Oscars. It felt like a return to form for the studio after a stretch of underwhelming releases, many of which struggled more from poor timing and marketing than actual quality. So when I first heard of another contender, DreamWorks' The Wild Robot, I didn't think much of it. The trailer made it look like a copy of another Disney robot, Baymax.

When Roz, the titular robot, asked a newly hatched gosling, "Was this task accomplished to your satisfaction?" I rolled my eyes. It sounded too close to Baymax's iconic line: "Are you satisfied with your care?" But something shifted as I kept watching. Roz, stumbling through her accidental role as a mother, wasn't copying anyone. She was her own person. I mean, robot.

Baymax follows his programming without question. He was designed to heal, programmed to care, and built never to cause harm. He fits neatly within the framework science fiction writer Isaac Asimov introduced in the 1940s through his Three Laws of Robotics: (1) A robot may not injure a human being or, through inaction, allow a human being to come to harm; (2) A robot must obey the orders given it by human beings except where such orders would conflict with the First Law; and (3) A robot must protect its own existence as long as such protection does not conflict with the First or Second Law. These laws became the moral backbone of robot stories for decades.

Roz doesn't follow those laws precisely. When she wakes up on a remote island after a failed delivery, her first task is to find her owners and complete their instructions. With no humans around, she activates her beacon to call for retrieval by her manufacturer. But a lightning storm strikes before the signal completes. Stranded and without orders, Roz explores the island. She accidentally causes a rockslide that destroys a goose nest. Left with a single egg, she takes on the responsibility of raising the gosling, Brightbill, and teaching him how to fly before winter.

When Brightbill learns to fly and leaves the island for the winter, Roz still doesn't prioritize self- preservation even when her beacon is fixed. As a harsh winter descends, she becomes a sheltering figure. She gathers the animals into one central lodge, protecting predator and prey alike. She even asks them to set aside their instincts in the meantime. That's when she delivers the film's most powerful line: "Sometimes, to survive, we must become more than we were programmed to be."

Watching Roz made me reflect on the AI systems we have today. These tools are considered safe precisely because they can't truly choose. They follow rules written by human engineers who try to anticipate every use case, including misuse.

But what happens when AI learns to choose?

Modern researchers have largely moved beyond Asimov’s laws, recognizing that these early frameworks are too simplistic for today’s ethical landscape. What counts as harm? Who qualifies as human? These aren’t just technical puzzles. As AI safety researcher Rob Miles explains in a Computerphile video, “You’d have to define what a human is. You’d have to define what harm is.” They’re moral questions we haven’t fully answered ourselves. And that’s the crux of it: to make AI truly safe, we’d first have to solve the entire field of ethics. Something humanity itself hasn’t managed to do. Asimov’s laws may sound elegant for stories, but even a basic command becomes ambiguous in edge cases.

The contrast is hard to ignore. Baymax, who follows every rule, is dependable but limited. Roz, who adapts and redefines her purpose, becomes the guardian of the island. This difference matters. As AI becomes more advanced, the critical question isn't just whether it can follow rules, but whether it can understand why those rules exist. The safest AI might not be the most obedient. It might be the one that learns.

I started watching The Wild Robot expecting a knockoff. I finished it thinking more deeply about the machines we're building, and the ones we might come to trust. Roz offers an answer without spectacle. The AIs we should trust aren't the ones hardcoded never to harm us. They're the ones that learn not to.

I started out rooting for a Disney/Pixar film, but Roz stayed with me longer. In the end, neither film won the Oscar. That honor went to Flow, an independent hand-drawn film from Hungary that I hadn't even considered. A movie with no dialogue defeating the giants of animation. Maybe the Academy voters, too, had learned to become more than what they were programmed to be.

About the author: Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master's in Fintech under the MMU-AIM Pioneer Cohort of 2025.

It was merienda time at the office when our talk turned to the Marvel Cinematic Universe. I couldn't help but geek out, this time about how two main Avengers evolved in their views on the Sokovia Accords. The fictional treaty was introduced after the catastrophic events in Avengers: Age of Ultron (2015), where Tony Stark's attempt to build a peacekeeping AI resulted in Ultron, whose actions nearly destroyed the world. In response, governments imposed oversight on the Avengers. The resulting ideological split carried over into Captain America: Civil War (2016).

Iron Man was the epitome of anti-authority individualism. There's this scene in The Avengers (2012) when Steve Rogers says, "We need a plan of attack." Stark replies, "I have a plan: attack!" as he flies off in his suit. But by Civil War, he had changed. He came to the realization that his ego had caused too much harm. Accountability, even through bureaucracy, was necessary.

Captain America refused to cede authority to the Sokovia Accords. He had seen corruption take root within SHIELD, the covert government agency that initiated the Avengers. It was later revealed to be infiltrated by Hydra. Having lived through this, he argued that even well- designed structures could be co-opted. He started as the ideal soldier: obedient and patriotic. But his belief in institutions broke down when he realized they could be turned against those they were meant to protect.

I know it's just a popcorn movie. But take the dialogue without the action, and there’s something deeper. Strip away the now-famous airport scene where Team Iron Man and Team Captain America clashed, what’s left is a battle of philosophies borne from their respective experience.

Today, there’s no such thing as Ultron, a vibranium-clad super AI armed with all the knowledge stored on the internet. But we are seeing the rise of powerful systems powered by machine learning that can see, sort, and judge. Tools like ClearviewAI can scrape billions of publicly available images to help law enforcement identify suspects. The technology promises safety, but critics worry about privacy violations and potential abuse.

The risks aren't theoretical. In 2020, Robert Williams, was wrongly arrested because of a false match from facial recognition software. Even the company behind that system admitted the match was flawed. An even darker example is Harvey Eugene Murphy Jr., a 61-year-old grandfather. He was also misidentified for an armed robbery. Despite being in California when the crime occurred in Texas, he was still arrested and detained. What’s worse, he was sexually assaulted while in detention. His mugshot was in the dataset because of a past conviction during his youth.

Another example is PredPol, a predictive policing tool that analyzes crime data and forecasts where future crimes might occur. This was trained through machine learning using statistical models inspired by earthquake detection techniques. Critics argue it perpetuates bias, since historical crime data often reflects over-policing in marginalized communities.

These tools work most of the time, and that's precisely the problem. There's a creeping normalization of AI judgment. We accept these systems because they seem reliable. Like ATMs, they aren't perfect, but they're convenient. One industry report showed that ATMs from even the top-performing bank in 2023 only had a 98.3% uptime. We tolerate occasional glitches because the technology feels indispensable.

When an ATM fails, you might lose access to your money temporarily. But when an AI judgment system fails, you might lose something more fundamental: your privacy, your freedom, or your presumption of innocence.

Hannah Arendt's concept of the banality of evil is relevant here, coined after attending the 1961 trial of Adolf Eichmann, a Nazi officer. Arendt described him as ordinary, not ideologically driven nor a fanatic. He simply followed orders that were considered normal at the time. By letting systems run without question, ordinary people can participate in great harm without malice. With AI, the danger isn't that technology will become evil, but that we will become passive. We might stop asking hard questions because the machine almost always gives reasonable answers. Just like our ATMs, 98.3% of the time.

The tech industry captures this through Amara's Law: "We tend to overestimate the effect of a technology in the short run and underestimate it in the long run." In the short run, AI systems may help catch a criminal. In the long run, they may quietly erode civil liberties in ways we cannot take back. We need to remember that. Not just in movies, but in every quiet moment where someone decides whether to trust a system.

Though Iron Man and Captain America are fictional, their thoughts on collective safety versus individual freedoms are not. Of all the quips that made this duo famous, I find Steve Rogers' words the most relevant: "The safest hands are still our own."

The name "chat.com" sounds almost silly in its simplicity. And yet, in March 2023, Dharmesh Shah, co-founder of HubSpot, paid $15.5 million for it. Two months later, he sold it. To OpenAI. No fanfare. No press release. Just Sam Altman posting "chat.com," on Twitter, which when clicked redirects to ChatGPT.

That curiosity led me to explore the logic behind such purchases. A study on defensive domain buying explains how it serves as protection against cybersquatting. Essentially, companies buy similar-sounding domains to prevent scammers from using them for phishing attacks or fraud. And when you frame it that way, it makes a lot of sense. Buying the domain isn't just about vanity. It's about preempting risk, shielding reputation, and reducing legal liability. I wouldn't say Shah belongs in the cybersquatting category though. It just opens up the possibility that OpenAI has bought more domains which we don't know about.

The study showed that most of the domains involved in disputes were eventually transferred to brand owners. That suggests there were costs that could have been avoided through early registration. Legal disputes can cost over $1,500 per case based on the study citing World Intellectual Property Organization (WIPO) fees as its basis, but these are just in filing fees, not including internal and external legal costs and public relation fallouts. Disputed domains also often attract phishing or misdirected traffic, contributing to brand dilution or even user harm. Still, the paper notes that the true financial benefit of defensive buying remains uncertain. Why? Because it's hard to quantify the cost of lost trust. Lost trust doesn’t show up cleanly in a ledger. Brand damage isn’t line-itemed.

And yet, OpenAI paid more for the chat.com domain than what Meta reportedly paid Sama in 2022 to moderate violent, disturbing content from the same AI models OpenAI now commercializes. Sama’s contract was $3.9 million. That figure is a snapshot of the cost to train content moderation operations in Kenya, where workers were exposed to trauma, paid poorly, and left with little support, a subject I wrote about in an earlier piece regarding the human cost of AI training.

OpenAI, for its part, claimed ignorance. Sama was a contractor. The labor conditions weren’t its responsibility. That legal distance is real, but so is the moral one. You don’t need to have met the workers to be implicated in the system that exploits them.

It becomes harder to digest when you remember that OpenAI, despite its valuation, is still burning billions. Even with revenues reportedly nearing $10 billion a year, it remains in a net loss position. For most industries, this would be alarming. But Big Tech plays by a different logic. Losses are tolerated if they feed long-term dominance.

This mindset traces back to Jeff Bezos’ 1997 shareholder letter, which redefined what shareholder value could mean. Instead of short-term profit, the priority became growth, scale, and eventual payoff. It worked for Amazon, and now it’s the blueprint for companies like OpenAI.

But this raises a deeper tension. In corporate law, we were taught that a board’s fiduciary duty is traditionally owed to shareholders. They are the providers of capital, whose interests are expected to be maximized. This was once synonymous with short-term returns and minimal reputational risk. But modern governance has evolved. Stakeholder theory doesn’t reject shareholder value; it expands the scope. It asks: how do a company’s decisions affect not just investors, but also workers, suppliers, and the communities it touches? And what happens when a strategy that protects branding investments overlooks the human cost embedded in its own operations?

But before passing judgment, imagine this. You own a piece of land and hire someone to maintain it. That person, in turn, hires others to help. You never ask how much they’re paid. You don’t know what their conditions are like. You never visit. Then one day, you're told: while the land looks pristine, the workers are unhappy, underpaid, and unsure of who to hold accountable. The timing is terrible, because you’re already trying to save money to negotiate a right-of-way agreement with the neighbor to increase the land’s value. Funds are already tight. Do you step in to fix what’s wrong? Or do you say, "I didn’t know. That’s their problem, not mine."

The answer isn’t as simple-sounding as a domain name. Because as we’ve seen, there’s nothing simple about it.

About the author: Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master's in Fintech under the MMU-AIM Pioneer Cohort of 2025 and was recently named a finalist for Fintech Lawyer of the Year at the 2025 ALB SE Asia Law Awards.

He can be reached at lromero.msfintech2025@aim.edu or leo-ernesto.t.romero@stu.mmu.ac.uk.

In 2016, Hayao Miyazaki, the legendary filmmaker behind Studio Ghibli, was shown an AI- generated animation by software engineers eager to impress. Hayao’s response was swift and cutting. He called it "an insult to life itself."

The clip is still online. You can pinpoint the exact moment the engineer’s expression shifts from excitement to dread.

Hayao was reacting to a grotesque figure meant for video games. But what he saw felt empty. Lacking empathy. Disconnected from what makes animation human.

In the age of generative AI, his words echo differently. They have become a rallying cry. Not just against clumsy outputs, but against the idea that a machine can learn creativity.

That discomfort wasn’t just about one animation. It touched on something deeper, about what we believe art should be. Art is sacred. It comes from hours of practice and lived experience. But what happens when a machine can replicate that in seconds?

There’s growing tension between artists who have honed their craft and users with access to tools that generate art or poetry with a few prompts.

I’ve written before about how LLMs work. How the Transformer architecture lets them track meaning across full sentences. How Pretraining helps them sound fluent, even if that fluency comes from data labeled by unseen and often underpaid human workers. Today, I want to focus on the G in GPT: Generative.

At its core, generative means making something new. A text, an image, an output that did not exist before. But it's not creation as we usually define it. It comes from patterns, not intention. Still, when someone provides a prompt and the AI returns a surprisingly good result, it feels original. Should that be considered authorship? If it isn’t, then what do we call it?

This gets more complicated with images. The viral trend of Ghiblifying photos was so widespread that Sam Altman joked it was melting GPUs.

Is Ghiblifying copyright infringement? In the age of LLMs, the legal and academic discussion is ongoing. Copying an art style is not the same as copying artistic expression, like a finished character or painting. That’s why a commissioned Ghibli-style painting is usually praised. But when AI generates a similar image, the reaction is often mixed. Somehow, the method makes it feel different.

I'm not an artist. I can’t paint or draw. But I enjoy tinkering. A few weeks ago, I wanted to surprise my mom and sister with a Ghibli-style picture. It was just a quiet gift. A gesture of affection.

But still, I wondered. Would that be disrespectful to my artist friends?

If artists can draw inspiration from famous studios, why can't a non-artist explore similar inspiration using a different tool?

Shortly after the Ghiblification trend, a viral tweet claimed Studio Ghibli sent a cease-and- desist letter. It turned out to be a hoax, but it touched a nerve.

Now imagine the reverse. What if an artist used ChatGPT to draft a legal agreement or summarize Supreme Court cases or ask about conflicting tax rules? Would I feel threatened, or glad they found help for something unfamiliar? I lean toward the latter.

Still, it feels disingenuous to equate art with law. There is something deeply personal about creation that goes beyond professional territory.

But must it always be either-or? What if AI isn't here to replace artists, but to become part of the process? A tool, not a threat. Something they can direct, refine, even critique.

Until that balance is reached, we should acknowledge the plight of artists as real. The fear is real. And instead of choosing sides, maybe we need to advocate for frameworks that allow credit to be shared, or at least traced.

Going back to Miyazaki's words: "an insult to life itself." What was the context? When he saw the AI model, he thought of his friend who was disabled. In his mind, if ever animation were to portray someone like his friend, it should be drawn by hand. With care. With slowness. With dignity.

He wasn't just critiquing the animation. He was warning us. About what happens when we mistake efficiency for expression.

But what if he saw AI used to help a disabled child bring their imagination to life? Or to help someone like me make a gift for family? Or to ease the world-renowned animator burnout in the Japanese industry? Would he still call it an insult?

Or could it be, in its own way, a quiet celebration of life?

We are now living in the messy middle. A place where we still feel the weight of what we might lose and the promise of what we might gain.

Maybe that is where life actually lives.

About the author: Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master's in Fintech under the MMU-AIM Pioneer Cohort of 2025 and was recently named a finalist for Fintech Lawyer of the Year at the 2025 ALB SE Asia Law Awards.

He can be reached at lromero.msfintech2025@aim.edu or leo-ernesto.t.romero@stu.mmu.ac.uk.

John Oliver’s segment on AI slop from Last Week Tonight made me laugh out loud, but it also hit close to home. The term AI slop started as technical jargon but now refers to the flood of AI generated images, videos, and headlines that saturate our social media feeds. They are polished on the surface but empty at the core. AI slop is the email spam of our generative AI era.

A more hopeful story comes from the world of math. For decades, researchers puzzled over the Hadwiger-Nelson problem in combinatorial geometry, which asks how many colors are needed to paint the plane so that no two points at a unit distance share the same color. It sat dormant until a recent breakthrough. In 2024, two new colorings were discovered. Not by AI alone, not by humans alone, but through collaboration. Mathematicians worked with AI to solve a longstanding math problem.

Like high-level math, accounting, law, and fintech are all technical fields. But creativity has its place in this space. Solving problems requires insight and imaginative thinking. There's joy in seeing solutions emerge from gray areas. That kind of creativity, while satisfying, feeds only the mind.

Feeding the soul is something else. It's not just about solving but also about expressing. There's a difference between crafting a persuasive argument versus writing something that speaks to your inner self. In one, the goal is professional competence. On the other, it's the joy found in the simple act of creating something fun for yourself.

This tension reminds me of a quote from Dead Poets Society. Robin Williams, as Professor Keating, tells his students: “We don’t read and write poetry because it’s cute. We read and write poetry because we are members of the human race. Medicine, law, business, engineering, these are noble pursuits and necessary to sustain life. But poetry, beauty, romance, love, these are what we stay alive for.”

That’s why I started a quiet YouTube project using Google's Veo3, its latest video and audio generation model, called Mickey and Pooh Have a Conversation. The two characters sit beside each other and talk about nature. Mickey asks a question. Pooh responds with a line that a child and an adult can both appreciate. In one scene, they talk about rain. Mickey wonders: "Does the rain fall because the sky is sad?" Pooh responds: "I don't think so. I think the clouds are just shedding their weight." In another, Mickey observes: "Blue must be Earth's favorite color. The sky wears it all day and never seems to grow tired of it." Pooh gently corrects: "The opposite, really. Blue spreads because the sky won't let it in."

AI is like a box of LEGOs for adults. Even if you don’t have a blueprint, that won’t stop you from building. Some creations are silly, while others are strange. But at least they’re yours. And there are family and friends who will appreciate the honesty in that.

This isn't just play or something therapeutic. There’s research being done on this. Haase and Pokutta’s paper on Human-AI Co-Creativity describes four levels. At Level 1, AI is a pen. At Level 2, it’s a task specialist. At Level 3, it becomes an assistant. At Level 4, it can be a co- creator. The authors point out that the math case study earlier is between Levels 3 and 4. My own project falls between Levels 2 and 3. The AI helps with proofreading and generates the video and audio. But the concepts are mine. The AI model would not have known that I am fond of the Calvin and Hobbes comics, or that I watched Knowledge Channel’s rebroadcast of Inquiring Minds. Those who grew up in the 90s will likely recognize the inspiration.

In John Oliver’s segment, he skewered the absurdity of AI slop. One YouTuber had even turned it into a business, offering paid seminars on how to pump out viral videos using AI tools. The goal shifts toward profit, often at the expense of storytelling. And platforms reward it. Facebook, Instagram, and TikTok each tweak their algorithms to push what sticks, regardless of meaning.

AI as pen, specialist, assistant, and co-creator is already here. But it matters how we use it. While platforms push content that chases views, we don’t have to follow that script. We now hold the adult equivalent of a LEGO set. Even if only one person enjoys what we make, that can be enough. In a world obsessed with millions of views, there is power in creating for an audience of one, or just a few. To stop the spread of AI slop, let’s remember the quiet joy of building with LEGOs.

About the author:

Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master’s in Fintech under the MMU-AIM Pioneer Cohort of 2025 and was recently named a finalist for Fintech Lawyer of the Year at the 2025 ALB SE Asia Law Awards.

There's a strange comfort in imagining that artificial intelligence learns all by itself. We like to think of it as something detached, a machine that just happens to know things. But the truth is far more human.

A recent CBS 60 Minutes segment featured young Kenyans, most of them college graduates, working for just two dollars an hour. Their job was to train AI systems.

For eight hours a day, they stared at screens. Some drew boxes around microwaves and faces. Others circled medical abnormalities. The most disturbing jobs involved labeling violent content: child abuse, animal cruelty, pornography, even footage of suicide.

They did this to help AI recognize what is safe and what is not.

Our machines seem polite and helpful because someone else went through hell to teach them.

In my previous article, I focused on the Transformer architecture, the T in GPT. But this story is about the P: Pretrained. In my earlier piece, I described a silent librarian inside the system, assembling meaning from word vectors as coordinates that clump together because of a self- attention mechanism regardless of order or position. But these labeled images and words? They are what the librarian memorized during pretraining. The system feels intelligent because someone fed it with care, frame by frame.

Back to the CBS documentary. Some of the workers are now suing. Others walked away with psychiatric evaluations confirming what they already knew: they were not the same anymore. One man said he could no longer enjoy intimacy with his wife. Another said it was now easier to cry than to speak.

And I had to ask myself: if I had not seen it, would I have cared?

We celebrate AI for its convenience. It writes our emails, organizes our notes, and even helps with homework. But behind each reply is a chain of labor. Invisible. Outsourced. Underpaid.

What happened in Kenya is not rare. Tech companies often avoid direct responsibility by using third-party firms. These middlemen keep wages low and shield the brands from blame. Documents showed OpenAI paid the outsourcing firm $12.50 per hour per worker, yet the Kenyans received only $2.

What disturbed me most was my own reaction. My first thought was utilitarian. If this helps more people than it harms, isn't that a good trade?

But would I feel the same if it were my sister labeling gore to pay the bills? Or my friend watching suicide footage so I could use a free chatbot?

The answer is not direct. Kenya needs jobs. Its government promotes itself as the Silicon Savannah, a hub for tech outsourcing. Refusing contracts might mean fewer opportunities, even if those jobs come at a psychological cost.

In my past piece, I described the model as a tireless librarian. But now I see that she stands on the shoulders of exploited humans. Labelers feed her the data. They sort the content. They carry the emotional weight. Although the librarian never sleeps, the ones who trained her are exhausted, sometimes broken.

The problem runs deeper than just money. Our systems are built to hide the cost. Executives do not see the harm, and software engineers do not write the contracts. Meanwhile, the users do not know what happens behind the scenes.

So we all pretend it is progress.

But maybe ignorance is not bliss. Maybe it is complicity.

I am not calling for boycotts. I still use AI. It helps me work. But we cannot claim to be ethical if we benefit from suffering we choose not to see.

Maybe the point is not to cancel or condemn. Maybe it is just to see more clearly.

I admit, it feels strange. Knowing all this and still using the tools. There’s a contradiction I cannot ignore.

Perhaps this discomfort is exactly what we need.

There is something to be said for the weight of knowing. When we pretend the harm does not exist, we can use these tools thoughtlessly. But when we acknowledge the hands that trained our machines, we become more intentional. We pause before asking AI to do work we could do ourselves. We question whether this task truly needs automation, or whether we are simply avoiding effort.

The inconvenience of caring is not a bug. It is a feature. It is our human nature. Regrettably, it is also our nature to forget as time passes.

But if we are going to keep using, building, and improving these tools, then at the very least, let us not forget those holding the weight underneath. We may not know their names, but we can remember their sacrifice.

About the author:

Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master’s in Fintech under the MMU-AIM Pioneer Cohort of 2025 and was recently named a finalist for Fintech Lawyer of the Year at the 2025 ALB SE Asia Law Awards.

He can be reached at lromero.msfintech2025@aim.edu or leo-ernesto.t.romero@stu.mmu.ac.uk.

In law school, we were trained through the Socratic method. Professors rarely gave direct answers. They asked questions again and again. As a freshman, you might think it was all about memory. You would try to recall every line from every reading. But by the time you graduated, you understood. It was never about perfect recall. It was about learning how to think, about learning how to understand.

So when ChatGPT arrived, I was naturally curious. It seemed to appear overnight and quickly became the fastest tool in history to reach 100 million users. From proofreading emails to summarizing research, large language models, also known as LLMs, have become embedded in our daily routines. But beneath the convenience lies a deeper question: do we really understand how these models understand us?

The short answer is not quite. These tools are impressive, even astonishing, but their inner workings remain opaque. In the language of AI, we call them black boxes. We know what we feed into the system and we see what comes out. But what happens in between is hidden from view. The model does not explain its choices. We simply judge by output.

Yet even a basic grasp of how these systems work matters. It shapes how we use them, how we regulate them, and how we live with them.

Much of the current progress in artificial intelligence began in 2017, when researchers at Google introduced a breakthrough model architecture called the Transformer. Earlier models handled language in sequence, like a typewriter moving left to right. But the Transformer could look at an entire sentence at once, weighing which words mattered most. It introduced a technique known as self-attention, a way for the model to understand relationships between words, even if they were far apart.

Here's a bit of trivia: the “T” in GPT stands for Transformer. GPT, short for Generative Pretrained Transformer, uses this architecture as its core. It generates original text by predicting what comes next based on patterns it has seen before, because it was pre-trained on massive datasets. Because it can take in a broader context, it sounds conversational. It sounds human. But it is not.

Picture the model as a vast library. Each word you type becomes a vector, or a coordinate, in this library. These vectors give the model a way to understand not the definition of a word, but its position in the space of meaning. Inside this library is a tireless librarian, retrieving and linking these word-vectors to form a coherent reply.

We know the librarian is working because we see the results. But how it chooses what to retrieve remains unclear. Words are transformed into embeddings, vectors that shift in meaning depending on what came before or after. From each embedding, the model creates three new vectors: a Query, a Key, and a Value. Imagine a word sending out a request, identifying itself, and offering something in return. The librarian weighs these relationships mathematically, choosing which ones to emphasize based not on meaning, but on statistical patterns.

LLMs do not understand language the way we do. They do not think. They do not intend. They are extraordinary machines for recognizing patterns and predicting the next most likely word. Their output can appear insightful, but that does not mean the system knows what it is saying.

When we ask, “Do we understand how an LLM understands,” we are really asking whether we trust the responses of a machine that predicts language, not meaning. And whether we are comfortable outsourcing more and more of our own understanding to a system that does not actually possess any.

Socrates once said that wisdom begins in recognizing what we do not know. From that space of humility, we learn to ask better questions, especially when others stop asking them.

It is tempting to embrace new technologies without question. But if we want to keep our humanity, we have to keep questioning. We need to understand what these systems can and cannot do, and think carefully about how they should be used.

This is why the field of AI explainability becomes more important each day. Knowing how an LLM works is not just for engineers. It is for anyone living in a world increasingly shaped by decisions that these LLMs’ influence, often without our awareness.

Socrates died not knowing everything. But he lived in pursuit of understanding. Perhaps in the age of AI, more than ever, that is how we should be living too.

About the author:

Leo Ernesto Thomas G. Romero is a CPA-lawyer for a publicly listed company, focusing on corporate and tax matters. He is pursuing a Master’s in Fintech under the MMU-AIM Pioneer Cohort of 2025 and was recently named a finalist for Fintech Lawyer of the Year at the 2025 ALB SE Asia Law Awards.

He can be reached at lromero.msfintech2025@aim.edu or leo-ernesto.t.romero@stu.mmu.ac.uk.