“A machine will never do that” — and then it does

There’s a pattern I can’t get out of my head. A smart, respected person looks at a technology, says with total conviction “this will never happen” — and a few years later that exact thing is everyday life.

In 1997, Piet Hut of the Institute for Advanced Study in Princeton told the New York Times about the board game Go: “It may be a hundred years before a computer beats humans at Go — maybe even longer.” In March 2016, AlphaGo beat world-class player Lee Sedol 4–1. Not in a hundred years. In nineteen.

Douglas Hofstadter, one of the most brilliant minds in cognitive science, wrote an essay in the Atlantic in 2018 titled “The Shallowness of Google Translate”. He approvingly quoted the line “No reasonable person thinks that a machine translation can ever achieve elegance and style.” In 2023, in a conversation published in June, that same Hofstadter sounded like this: “I don’t think it’s interesting. I think it’s terrifying. I hate it. I think about it practically all the time, every single day.” And: “It’s a very traumatic experience when some of your most core beliefs about the world start collapsing.”

If this feels familiar, maybe it’s from somewhere else. In 2017, Jamie Dimon called Bitcoin a fraud at a Barclays conference: “It’s worse than tulip bulbs. It won’t end well. Someone is going to get killed.” In January 2018 he said he regretted the remark. Warren Buffett called Bitcoin “probably rat poison squared” at the 2018 Berkshire annual meeting; Charlie Munger called crypto trading “just dementia.” And in 1998 Paul Krugman wrote that by 2005 it would become clear the internet’s impact on the economy was “no greater than the fax machine’s” — reasoning, among other things: “most people have nothing to say to each other!”

I’m not writing this to look smarter than these people. I’m not. I’m writing it because the same pattern is running again right now, and this time we get to watch it happen. This is my timeline of it. It’s deliberately built to keep growing.

How it was before (so the leap lands)

Before ChatGPT arrived in November 2022, “AI that writes text” was a thing for insiders.

GPT-3, the model behind it, was published as a preprint on May 28, 2020 (“Language Models are Few-Shot Learners”), and the API beta launched on June 11, 2020. 175 billion parameters — and a context window of just 2,048 tokens, a few paragraphs, that was all the memory it had. Training was estimated to cost 4.6 million dollars; on a single GPU it would have taken roughly 355 years.

And you couldn’t just get in. GPT-3 sat behind a waitlist. It wasn’t until November 18, 2021 — a year and a half later — that OpenAI removed the waitlist and made the API generally available.

The most famous piece of text from that period says a lot about the real state of things. On September 8, 2020, the Guardian printed an op-ed written by GPT-3: “A robot wrote this entire article. Are you scared yet, human?” It opened with “I am not a human. I am Artificial Intelligence.” and included the line “I have no desire to wipe out humans.” Sounds impressive — until you read the fine print. GPT-3 had produced eight attempts (fed by a Berkeley student), and the editors cut the best parts together. So “entirely written by AI” was a generous rounding, let’s say.

The experts were sober about it, as you’d expect. On August 22, 2020, Gary Marcus and Ernest Davis ran a headline in the MIT Technology Review: “GPT-3, Bloviator: OpenAI’s language generator has no idea what it’s talking about.” In a deliberately unsystematic set of tests — the authors explicitly said it was not a benchmark — GPT-3 came out roughly half clearly right, half clearly wrong on everyday reasoning. In March 2021, Emily Bender and colleagues coined the term that stuck, at the FAccT conference: “stochastic parrots” — machines that parrot word patterns without understanding anything. Marcus’s verdict two years later, in March 2022: “For all its fluency, GPT-3 can neither integrate information from basic web searches nor reason about the most basic everyday phenomena.”

And the images? On January 5, 2021, OpenAI showed DALL·E — a 12-billion-parameter offshoot of GPT-3 that painted images from text. Low resolution, often off. A curiosity, not a tool.

Hold that picture: waitlist, parrot, curated text, pixelated kitten. That was the state of things three and a half years ago.

The timeline

From here it gets dense. I’ll set the key releases with dates — and next to each, what they made practically possible for the first time. This list keeps growing; I’ll add to it as new things come.

May/June 2020 — GPT-3 as preprint and API beta. In OpenAI’s own test, evaluators correctly spotted machine-written short articles only 52% of the time — barely better than a coin flip. First time: text that sounds human at paragraph length.

January 2021 — DALL·E 1. First time: an image from arbitrary text — but crude and small.

June 21, 2022 — GitHub Copilot becomes generally available, 10 dollars a month. First time: an AI that types along as you code, affordable for any developer.

July/August 2022 — the image breakthrough comes in a rush: DALL·E 2 (announced April 6, beta July 20), Midjourney (open beta July 12), Stable Diffusion (August 22, open source, runs locally on consumer hardware). First time: photorealistic, composable images — and with Stable Diffusion, on your own machine, no cloud.

Late August / early September 2022 — Jason Allen’s Midjourney image “Théâtre D’opéra Spatial” wins first place at the Colorado State Fair (digitally manipulated photography category, 300 dollars). The judges didn’t know it was AI. The public debate followed in September.

November 30, 2022 — ChatGPT, a free preview built on GPT-3.5. A million users in five days, around 100 million a month after two months. First time: the parrot gets a chat window — and the whole world can walk in.

March 14, 2023 — GPT-4, multimodal (it understands images too), context up to 32,000 tokens instead of 2,048. That same month Runway’s Gen-2 launches as one of the first usable text-to-video betas. First time: AI that “sees” an uploaded image — and the first moving seconds from text.

October 2023 — DALL·E 3, right inside ChatGPT. First time: images that actually hit what you type, including legible text.

February 15, 2024 — OpenAI announces Sora: up to 60 seconds of highly realistic 1080p video from a single prompt. First time: video you might take for real at first glance.

March 12, 2024 — Cognition shows Devin, the “first AI software engineer” with its own shell, editor, and browser. First time: an AI that works a coding task end to end on its own.

March/April 2024 — Suno V3 (March 21, complete songs with vocals up to four minutes, free too) and Udio (April 10, public beta). First time: a finished song with a voice, from a single sentence.

May 13, 2024 — GPT-4o, an “omni” model for text, audio, and image in one. First time: fluid, real-time talking with the machine.

February 24, 2025 — Claude Code launches as a research preview, generally available in May 2025. First time: coding becomes a dialogue in the terminal — for me too; this page is built that way.

March 25, 2025 — OpenAI builds image generation natively into GPT-4o. The viral “Studio Ghibli” trend overran the service so badly the free version had to be delayed. First time: editing images in conversation — “make the sky warmer,” and it happens.

May 20, 2025 — Google Veo 3: for the first time, sound generated in sync — dialogue, sound effects, lip-synced. Veo 3.1 follows in mid-October 2025. First time: video that doesn’t just look right, it sounds right.

August 7, 2025 — GPT-5, described by OpenAI as a “team of PhD-level experts in your pocket.”

September 30, 2025 — Sora 2: synced sound, better physics, “Cameo” features — plus its own app in TikTok format. First time: AI video as a social medium, not a lab demo.

Take these five years in one sitting, and the parrot of 2020 feels like it’s from another century.

What almost nobody has on their radar

Everyone talks about ChatGPT. Far too few talk about this — and it may be the bigger part of the story.

In November 2020, AlphaFold 2 won the CASP14 competition with a median score around 92 GDT_TS, solving a problem biology had failed at for fifty years: predicting a protein’s three-dimensional fold from the sequence of its amino acids. This is not a gimmick. Protein folding is the foundation for understanding how diseases work and how drugs dock. From 2022 the AlphaFold database held over 200 million structures, more than 214 million by 2024, used by over two million researchers in 190 countries. On May 8, 2024 came AlphaFold 3, which also predicts complexes of proteins, DNA, RNA, and drug molecules.

The recognition came at the highest level: the 2024 Nobel Prize in Chemistry went one half to David Baker for computational protein design, the other half jointly to Demis Hassabis and John Jumper for predicting protein structures. An AI system is at the heart of a Nobel Prize in Chemistry.

And it doesn’t stop at proteins:

  • Medicine: Rentosertib, an AI-discovered drug for the lung disease IPF, showed dose-dependent improvement in lung function after twelve weeks in a Phase IIa study of 71 patients across 22 centers — published on June 3, 2025 in Nature Medicine. In 2023, a deep-learning model (a collaboration between McMaster University and MIT) identified the drug abaucin against a dangerous hospital pathogen, out of roughly 7,500 molecules screened.
  • Materials: DeepMind’s GNoME predicted the stability of 2.2 million new crystals in late 2023. On the same day, the A-Lab in Berkeley showed a robot using ML to synthesize new inorganic materials in 17 days. (In fairness: both papers have since drawn criticism — many consider GNoME’s count of truly new structures inflated, and Nature had to issue a correction on the A-Lab. The breakthrough is real, the scorecard disputed.)
  • Weather: GraphCast predicts up to ten days of weather in under a minute on a single computer — and was more accurate than the ECMWF’s established model on more than 90% of 1,380 verified targets. GenCast followed in late 2024 with better cyclone tracks.
  • Pure mathematics: AlphaProof and AlphaGeometry 2 solved four of six problems at the 2024 International Mathematical Olympiad, 28 of 42 points — silver level, one point short of gold. One of the problems they solved was fully cracked by only five of more than 600 human contestants. And in late 2023, FunSearch found a provably new solution to an open mathematical problem — the first time a language model contributed genuine, verifiable new knowledge.

So the parrot that supposedly only repeats has folded proteins, found drugs, and done mathematics that hardly any human can do.

The numbers

When I want to know whether a technology is for real, I follow the money — and the jobs.

Global corporate investment in AI reached around 252 billion dollars in 2024, according to Stanford HAI. In 2025 it doubled to 581.7 billion. Private investment alone jumped 127.5% to 344.7 billion. Generative AI — the part that makes images, text, and video — grew about 200% in 2025 to roughly 171 billion dollars, nearly half of all private AI investment. The US leads with 285.9 billion in private investment, 23 times China’s.

Usage is keeping pace: 78% of organizations reported using AI in 2024 (up from 55% the year before); McKinsey put it at 88% with regular use in at least one business function in 2025. The catch — and I’ll name it deliberately: only about a third of companies scale beyond pilot projects, and only a minority can show a measurable effect on profit. Lots of experimenting, little following through.

And the work? That’s the sore spot. The IMF estimates that roughly 40% of global employment is AI-exposed, about 60% in advanced economies. Half of those jobs may benefit, the other half may lose core tasks. IMF chief Kristalina Georgieva put it this way: “We are on the brink of a technological revolution that could jumpstart productivity, boost global growth and raise incomes around the world. Yet it could also replace jobs and deepen inequality.”

The WEF expects 170 million new jobs by 2030, 92 million lost — a net gain of 78 million. So not only loss. The fastest growth is in big-data specialists (+110%), fintech engineers (+95%), AI/ML specialists (+85%); the sharpest decline is in classic office jobs like clerks and cashiers. 39% of the core skills demanded today are considered obsolete by 2030.

One detail that makes me optimistic: the Anthropic Economic Index analyzed around two million AI conversations in early 2026. For the first time, augmentation (52%) — AI helping a person with their work — overtook automation (45%), where it takes the task over entirely. More often an assistant than a replacement. At least so far.

Then vs. now — my own math

Now it gets concrete, because this is exactly where I live. An AI store, a blog — the images on them have to come from somewhere. So I looked at what a classic ad photo shoot actually takes. Not the cliché, the math.

Let’s start with the rights, because most people overlook them. To use an image, a “buy-out” is added on top of the fee — a surcharge that scales with reach and how long you use it, “longer period, higher buyout,” as the agencies put it. Licenses typically run one to five years; after that you may no longer use the image. And in the US you need a model release for every recognizable person — without it, commercial use is simply illegal, celebrity or not.

Then the people on set. An agency takes 15–20% commission from the model and adds a markup on top for the client. A fashion shoot for a single day adds up fast: photographer 1,000–3,500 dollars (in New York or L.A. easily 3,000–8,000+), model 500–3,000, studio 300–1,500, hair and makeup 400–1,200, stylist 400–1,800, retouching per image on top. A lean production lands around 2,600 dollars, a mid-range one around 12,700, high-end between 25,000 and 100,000 dollars — per day. And the shoot day is, as one production company dryly notes, “the smallest part”: for a TV spot, preproduction is 40–50% of the time, a spot takes 6–10 weeks in total, a campaign 8–12. Casting, location scouting with permits and insurance, a crew of 2 to 200 people, weather contingency.

One example that sticks: a celebrity hairstylist charged 12,000 dollars for two shoot days — just for the work, no travel, no agency. Just the hair.

And today? That exact result — the person, the light, the place, the mood — is a prompt. Seconds instead of weeks. No buy-out, no expiring license contract, no weather. This page and the shop are the living example: what you see here as images would have meant a mid-five-figure production and a month of lead time three years ago. How these images come to be is what I show on the AI page — the AI tools are the toolbox behind it.

I’m not saying this to dismiss anyone’s craft. A good photographer, a good stylist — that stays skill. I’m saying it because the threshold has vanished. What used to require budget and connections now requires a good idea and a precise sentence.

The catch

So this doesn’t turn into an ad, here’s the honest part. The state barely uses these models, and neither do many companies — for data-protection reasons. And that’s not stupid, it’s understandable. Anyone typing sensitive data into someone else’s cloud model gives up control they may not be allowed to give up. An agency, a clinic, a law firm can’t just hand personal data to a US provider.

But the answer to that isn’t “don’t,” it’s “here’s how.” Remember Stable Diffusion from August 2022 — that ran locally from the start, on ordinary hardware, without a single byte leaving your machine. This exact direction — models that run on your own computer or in your own building — is the way for public agencies and data-sensitive companies to do this without losing control. The technology for it gets better and smaller month by month. Ignoring it doesn’t solve the problem, it just kicks it down the road.

In closing

The pattern is always the same. First “this will never happen,” then “fine, but it doesn’t count,” and then it’s everyday life and nobody remembers ever doubting it. Go. Translation. Protein folding. Images, voices, video with sound. And yes, back then the internet too, and that “digital money” that was supposedly just rat poison.

I don’t claim to know where this ends. Nobody knows. But there’s one lesson from three and a half years I’ll dare to write down: it comes faster than almost everyone thinks — faster, too, than the smart people think, the ones who say “never.”

That’s why I keep this timeline going. Every time something falls that was called impossible only yesterday, another entry gets added. If you want to grow along with it, drop by now and then — and if you’ve seen something that’s missing here, let me know via the contact page. This list isn’t finished. It’s just getting started.