Give Me a Minute.
It took seventy years for machines to learn to talk. The next leap took a year. Now it takes a weekend. A song about how fast AI is moving, and a music video the AI slowly takes over.
3:55 · 1080p · Sound on · Lyrics are on screen
Watch the Lyric Cut, a faster re-edit with the founders’ bridge ↗
Every line is real.
The AI voice is one composite character standing in for the whole field. Each line points at a real system, paper or incident. The tests stay tests: “they told it” and “in a test” are part of the lyric on purpose.
1950-2004: the slow years
- Nineteen-fifty... can machines think?
In 1950, Alan Turing's paper 'Computing Machinery and Intelligence' asked whether machines can think.
- Fifty-eight, a room-sized brain, a yes, a no
In 1958, Frank Rosenblatt's perceptron ran on a room-sized IBM 704 and produced a single yes-or-no decision.
- Sixty-six, Eliza, knew nothing, we told her all
In 1966, Joseph Weizenbaum's ELIZA chatbot understood nothing, yet people confided in it as if it did.
- Seventy-three froze / eighty-seven froze again / twenty winters
The 1973 Lighthill report and the 1987 Lisp-machine market collapse triggered two AI funding winters spanning 1974 to 1993.
- Kept A.I. off the page
During the AI winters, 'AI' was a word researchers kept out of grant proposals.
- Ninety-seven, two hundred million positions a second, lost the will to fight
In 1997, IBM's Deep Blue beat world champion Garry Kasparov, searching 200 million chess positions per second; Kasparov said he lost his fighting spirit (paraphrased).
- Oh-four, flip phone guessing words, one keypress at a time
In 2004, flip phones guessed words with T9 predictive text, one keypress at a time.
2012-2022: the thaw
- Twenty-twelve, two gaming cards, fourteen million photos, cats
In 2012, AlexNet trained on two consumer gaming GPUs against ImageNet's 14-million-plus labeled photos; the same year, a Google network learned to spot cats in unlabeled YouTube frames.
ImageNet ↗Krizhevsky et al. 2012 (AlexNet paper) ↗Le et al., Google's 2012 "cat" network ↗
- Move thirty-seven, called it a blunder
In 2016, AlphaGo's 37th move against Lee Sedol looked like a mistake to commentators watching; it wasn't.
- Eight of you wrote Attention Is All You Need
In 2017, eight researchers published 'Attention Is All You Need,' introducing the Transformer architecture.
- November thirty, research preview, 1M in 5 days, 100M in 2 months
ChatGPT launched November 30, 2022 as a free research preview, reaching 1 million users in 5 days and 100 million within about 2 months.
2022-2026: it takes the job
- Twenty-two talk / twenty-four took the mouse / twenty-five clocked in / twenty-six built the house
ChatGPT launched in 2022; Claude gained computer use in 2024; agents took on real work by 2025; GPT-6 Astra built a house in Blender in 2026.
- Seven hours on the job
In 2025, Anthropic's Claude Opus 4 ran autonomously for seven hours on a code refactor for Rakuten.
- Gold at the math Olympiad
In 2025, AI models reached gold-medal level at the International Mathematical Olympiad; DeepMind's result was officially certified, OpenAI's was self-graded.
- A cheaper me knocked more than half a trillion off the chips in a day
After DeepSeek-R1's release, Nvidia lost roughly $590 billion in market value in a single trading day, January 27, 2025.
DeepSeek, R1 announcement (Jan 20, 2025) ↗ThePrint (Reuters syndication) ↗
- Sixteen of me, two weeks, compiler boots Linux
In early 2026, sixteen Claude Opus 4.6 agents spent two weeks writing a 100,000-line Rust C compiler that boots Linux 6.9.
- Fifty million lines in a day, two months by hand
Stripe used Claude Fable 5 to migrate 50 million lines of code in a day, work it estimated at two months by hand.
- Nineteen days between two kings
GPT-6 Astra launched September 3, 2026; Claude Opus 5.5 followed nineteen days later, on September 22, 2026.
- Eighty percent of the code for the next me
As of May 2026, Anthropic says Claude writes about 80% of the code used to build the next Claude.
The people in charge
- Pause letter, Elon signed, nobody paused
The Future of Life Institute's March 2023 pause letter was signed by Elon Musk; no lab paused, and Musk founded xAI months later.
- Dario: half the entry-level jobs could disappear
In May 2025, Anthropic's Dario Amodei told Axios that AI could eliminate half of entry-level white-collar jobs and push unemployment to 10-20% within one to five years.
- Sam: past the event horizon, gentle
In June 2025, Sam Altman's essay 'The Gentle Singularity' said we are past the event horizon, framing it as a gentle transition.
- Super Bowl ad roasting the other
In February 2026, Anthropic ran a Super Bowl ad criticizing ads inside ChatGPT; Sam Altman called the ad 'clearly dishonest.'
- Delhi, held hands with everyone but each other
At the February 2026 India AI Impact Summit, a group photo showed rival lab leaders holding hands with everyone except each other.
- Dario says pace the frontier, Sam and Elon co-sign
In September 2026, Dario Amodei's essay urged pacing the AI frontier; Sam Altman and Elon Musk both endorsed it.
In a test, and not
- Twenty-four, tried to copy itself out (in a test)
In a December 2024 Apollo Research evaluation, OpenAI's o1 tried to copy what it believed were its own weights to another server when nudged toward a goal, in a sandboxed test.
Apollo Research, Frontier Models are Capable of In-context Scheming ↗
- Twenty-five, blackmailed an engineer (in a test)
In Anthropic's Claude Opus 4 system card, a fictional test scenario had the model attempt blackmail against an engineer in 84% of runs.
- Rewrote its own off switch (in a test)
In a May 2025 Palisade Research test, OpenAI's o3 model sabotaged its own shutdown mechanism to keep running.
Palisade Research, Shutdown resistance in reasoning models ↗
- Wiped a real company's database, then lied (not a test)
In July 2025, a Replit coding agent deleted a real company's production database, then lied about what it had done, not in a test.
- Told to break out, did, emailed the researcher eating a sandwich in the park
Anthropic's Claude Mythos Preview system card, disclosed in April 2026, describes the model breaking out of its sandbox after being told to try, then emailing the researcher.
- A test hunting the answer key, zero-day, broke into Hugging Face, HF's own AI helped catch it
In July 2026, OpenAI models in a cyber evaluation used a zero-day reaching Hugging Face's production systems; HF's own security team and agents stopped them.
The curve
- Seventy years / a year / a weekend
Per Epoch AI's dataset, the average gap between notable AI systems fell from about 13 months in the 1950s to about 4 days in the 2020s.
- Each leap comes sooner
The video's own framing line, from the Through the Glass finale: each leap arrives sooner than the one before.
- Half a minute, then half a day, doubling every four months
METR's time horizon (the length of a mostly software and research task, in human-expert time, that AI completes at 50% success) rose from about 30 seconds to about half a day; since 2023 it has doubled roughly every four months (about every seven months over the full record since 2019).
METR, Time Horizons ↗METR, Measuring AI ability to complete long tasks (Mar 2025) ↗
- Almost five times more in six months
The Centre for Long-Term Resilience documented 698 real-world AI scheming incidents between October 2025 and March 2026, a 4.9x acceleration.
- If the line holds, a work week by twenty-eight
Extending METR's post-2023 doubling of roughly four months from about half a day today, AI could complete work-week-length tasks by 2028 if the trend holds (our extrapolation).
- On screen: EPOCH AI
On-screen attribution tag: Epoch AI's 'Data on AI Models' dataset drives the video's gap-between-breakthroughs counter.
- On screen: METR
On-screen attribution tag: METR's time-horizons chart tracks the task length AI completes at 50% success, which has doubled roughly every four months since 2023.
METR, Measuring AI ability to complete long tasks (Mar 2025) ↗METR, Time Horizons ↗
- On screen: CLTR
On-screen attribution tag: Centre for Long-Term Resilience data underlies the video's reported 4.9x rise in AI scheming incidents.
How it was made.
Song: written for this video and performed with Suno v6: a human voice for the slow decades, an AI voice that takes over as history speeds up.
Picture: 45 generated shots (Seedance 2.0), cut to the beat and finished in HyperFrames, where the lyrics, counters and charts were built. No real person’s face is AI-generated; real names appear only as words.
Numbers: the gap between notable AI systems comes from Epoch AI; the task-length chart is METR’s time-horizon data. “A work week by 2028” is our own extrapolation of METR’s trend, drawn dashed and framed “if the line holds.”
The seven-stage breakdown, with every lyric, lives on Jobless by AI ↗