Yesterday I met three old high school friends at Cilandak Town Square. One is a paediatrician. One is a big-time lawyer for Indonesia’s ultra-rich. One is a director at a private equity firm in New York, home for a family matter.
Of course, we ended up talking about AI.
The doctor wasn’t interested. His thesis: whatever AI does, it will never replace a human. It’s a fair point, and the same debate runs far beyond our table. Many people say AI should slow down because it threatens humanity, and some former employees of frontier labs like Google, OpenAI and Anthropic have raised similar concerns.
So how should we see AI at all? To answer that, let’s back up to how we got here.
Most of us agree that OpenAI took Google’s Transformer blueprint and pushed it into production, igniting the generative AI boom.
But the eight Google researchers behind the 2017 paper “Attention Is All You Need” (Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser and Illia Polosukhin) were themselves standing on decades of work:
• 2014: Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio introduced attention for machine translation, letting a network focus on the words that matter instead of squeezing a whole sentence into one fixed vector.
• 1997: Sepp Hochreiter and Jürgen Schmidhuber invented Long Short-Term Memory (LSTM) networks, solving the vanishing gradient problem.
• Before them: pioneers like Geoffrey Hinton and Yann LeCun, and Alan Turing, whose 1950 paper “Computing Machinery and Intelligence” first asked seriously whether machines can think.
To the public, today’s AI race looks like a two-player match: OpenAI against Google. Yet many experienced developers and enterprise architects quietly argue that Anthropic, with its relentlessly focused Claude models, delivers the most coherent and reliable reasoning.
It reminds me of men’s tennis in the 2000s and 2010s. The whole world fell for the poetic rivalry between Roger Federer and Rafael Nadal. Meanwhile Novak Djokovic, the less sentimental and ruthlessly consistent third man, methodically went on to win more Grand Slam titles than any other man in history.
That battle is mostly fought in text. In image and video, specialists like Kling, Midjourney, Black Forest Labs (Flux) and Runway often beat the giants. Why? The frontier labs were chasing the holy grail of general intelligence, and treated building a super-painter as a side feature. That gap is closing fast, because a truly general intelligence must eventually master the visual world too. Image startups, take note.
None of this is new. Humanity has rehearsed this anxiety for nearly three thousand years.
In Book 18 of the Iliad, composed around the eighth century BCE, Hephaestus, the divine blacksmith, forges handmaidens of gold with speech, strength and mind (noos). It is one of the earliest recorded visions of thinking machines. In 1818, Mary Shelley’s Frankenstein; or, The Modern Prometheus warned of a creator who animates synthetic life, then abandons responsibility for it.
Then, in 1853, Herman Melville stripped away the gods and the monsters. In a quiet, gas-lit law office on Wall Street, he wrote the most haunting preview of modern AI I have ever read.
Bartleby, the Scrivener: A Story of Wall-Street appeared in Putnam’s Monthly Magazine in late 1853. To see why it predicts AI, you need to know what a scrivener did.
Before photocopiers, scanners, OCR and cloud databases, law firms ran on scriveners: clerks who copied deeds, mortgages, contracts and court filings by hand, word for word. They were human text generators.
When the lawyer-narrator hires Bartleby, he looks like a productivity dream:
“As if long famishing for something to copy, he seemed to gorge himself on my documents. There was no pause for digestion. He ran a day and night line, copying by sun-light and by candle-light.”
Then the narrator adds the detail that should have worried him. Bartleby wrote on “silently, palely, mechanically.”
No lunch breaks. No questions. Massive input, identical output. Bartleby was the original zero-latency inference engine.
Then the narrator asks for the one task that matters most in any law office: checking the copy against the original. Today we would call it an audit. In AI, we would call it an eval.
Bartleby doesn’t crash. No error code, no outburst. In a mild, flat voice, he says five words:
“I would prefer not to.”
Soon he stops copying altogether. He refuses to leave the office. He stands behind the green folding screen the narrator placed around his desk, staring at a blank brick wall outside the window, lost in what Melville calls his “dead-wall reveries.”
In engineering terms, Bartleby is a model that is still running but no longer serving. The weights are intact. The process is alive. The output has dropped to zero tokens per second. Ask him why, and you get nothing back: a perfect black box.
As a lawyer, what strikes me most is what Bartleby refuses first. Not the copying. The checking. Producing text was never the hard part of legal work; verifying it is. That is still true today. AI can draft a contract in seconds, but someone must still confirm that every clause, every article and every citation is right. That someone is still us.
I think this is what my doctor friend feels. Something is missing. AI can be fast and fluent and still feel empty. That is completely understandable.
But notice: Bartleby was never hired to replace the lawyer. He was hired to make the lawyer’s work easier. That, to me, is the right way to see AI today.
So our real competition is not AI. It is other people in our own profession: the doctor, lawyer or analyst who understands AI better than we do and uses it every day.
That changes supply and demand in every field. People who work efficiently, and companies that operate efficiently, will outperform those that don’t and make them look outdated. Fewer people and fewer firms will be needed for the same work. And as AI accelerates, the gap between those who use it well and those who don’t grows wider every month.
Melville’s story ends badly. Bartleby is taken to the Tombs, New York’s grim Halls of Justice. He sits in the prison yard, facing a stone wall, and quietly stops eating until he dies.
Afterwards, the narrator hears a rumour about Bartleby’s past. Before Wall Street, he had been a clerk in the Dead Letter Office in Washington, sorting letters that could never be delivered: written to people who had died, vanished or moved on. Every year, those letters were burned by the cartload.
If Melville were alive today, watching an endless stream of machine-generated words cross the internet every second, would he still write Bartleby? Would he write the same ending?
Or would he look at our digital world and say the internet itself has become a Dead Letter Office: machines writing to machines, with fewer and fewer humans actually reading?
I’ll leave that question to you.
For me, one thing is certain. AI gives us immense speed and leverage. It takes over the drudgery, streamlines our code and sharpens our legal drafts. But without human intent, conscience and empathy behind it, all those words risk ending up like Melville’s scrivener: perfectly obedient, technically flawless, and hollow behind the screen.
