Pediatric surgery · Part three
The AI Race
Not who wins it. What it is already doing to the world, and to us.
MOHAMED MASLOOH · 23 SEPTEMBER 2026
race · what it costs to stay in it
Four companies. $725 billion.
One year.
If we misspend a couple of hundred billion dollars, that will be very unfortunate. But the risk is higher on the other side.
Mark ZuckerbergMeta, 2025. Being late looks worse than being wrong
$145bnMeta alone, on data centres and servers, this year
35cof every dollar Meta earned went into building, in the first quarter of this year
$44.9bnAlphabet, in a single quarter
firstnegative cash flow in Alphabet’s 22 years as a public company
$12.3bnlost by OpenAI in three months2026 guidance from each company’s own earnings calls, via
Statista, 31 Jul 2026 · Meta Q1 2026 ·
Alphabet Q2 2026 ·
OpenAI: Wall Street Journal, Aug 2026 · Zuckerberg: Access podcast, Sep 2025, paraphrased
Chapter one
They asked it
to slow down.
warning · one sentence, 30 May 2023
“Mitigating the risk of extinction from AI should be a global priority alongside other societal scale risks such as pandemics and nuclear war.”
Sam AltmanOpenAI
Demis HassabisGoogle DeepMind
Dario AmodeiAnthropic
Geoffrey HintonNobel prize, 2024
Yoshua BengioTuring award, 2018
Statement on AI risk, Center for AI Safety,
30 May 2023, signed by more than 350 researchers and executives
warning · the word that comes after AGI
They say the finish line is behind us.
The next word is superintelligence.
“Welcome to the AGI era.”Greg Brockman · President of OpenAI · 3 September 2026
“AGI has arrived.”Jensen Huang · Head of Nvidia · 7 September 2026
AGI
As good as a person
at most of what a person does.
Superintelligence
Better than every person at everything, including building the next machine.
Axios, 3 Sep 2026 ·
Huang on X, reported by Fox Business.
No exam and no regulator sits behind either word. Shapes drawn to the two definitions, not measured.
warning · the people paid to keep it safe
These are not critics.
They built it.
Jacob Coxon
@hilbertspaess
Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.
8 Sep 2026resigned from Anthropic
Evan Hubinger
@EvanHub · quoting @hilbertspaess
Jacob is correct here
we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.
we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.
8 Sep 2026, 6:27 PMstill at Anthropic
10%his own odds that AI kills all humans, within the next decade
Coxon and Hubinger on X, 8 Sep 2026: Hubinger’s post ·
TechCrunch, 9 Sep 2026 ·
CBS News, 9 Sep 2026 · Coxon spent three years at OpenAI, then Anthropic
warning · 12 September
Then the three men running the race
asked for a speed limit.
Dario Amodei
@DarioAmodei
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three part plan for doing so.
12 Sep 202643.8M views
Sam Altman
@sama
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks.
12 Sep 2026
Elon Musk
@elonmusk
Dario is right
12 Sep 2026
The three posts on X, 12 Sep 2026 · the essay: We Must Pace the Frontier ·
Axios, 12 Sep 2026
warning · and then it left the companies
The men building it asked to slow down.
The man who could order it said no.
Donald Trump14 Sep 2026 · on a call to the All-In Summit
On slowing down AI:
It could also be China. And we’re not going to let that happen.
Donald TrumpEight days later · UN General Assembly · 22 Sep 2026
Whoever wins super intelligence wins.
The call, 14 Sep 2026: TechCrunch ·
Axios ·
Al Jazeera ·
UN speech, 22 Sep 2026: Fox News and
Washington Times
Chapter two
It is already changing
our world.
world · three things, all this year
It has already done what no one could.
AI unrolled itfrom X-ray scans
Read a scroll without opening itBurned by Vesuvius in 79 AD, too fragile to open. AI made the ink visible; scholars read it.
Proved a 1946 maths guess wrongErdős’s guess stood for 80 years. An OpenAI model beat it; nine mathematicians checked.
Found a security flaw hidden for 27 yearsIn OpenBSD, built for security: it could crash a machine from afar. Found by Anthropic’s AI.
Vesuvius Challenge, 25 Jun 2026, pictures CC BY-NC 4.0 ·
OpenAI, 20 May 2026; checked by Alon, Gowers and seven others, arXiv 2605.20695 ·
Erdős: Kmhkmh, CC BY 3.0 ·
Anthropic, 7 Apr 2026 ·
OpenBSD errata 025, 25 Mar 2026; the fix, 20 Mar 2026; the flawed code, 17 Nov 1998 ·
the dot grid and the shield are drawings, not data
world · and it reached the operating theatre
surgeonpatient
2001New York, Strasbourg
Johns Hopkins
the robot working aloneclick to play
Gallbladder removalclip and cut, all 17 steps alonepig organs, outside the body
Jul 2025Baltimore
88%
of the steps alonebiliary surgerya living 30 kg pig
Dec 2025Shanghai
Humanoidsremotely guided
Jul 2026San Diego
Nexta human patient
No robot has operated on a human patient on its own. Yet.
Chapter three
Should we be afraid?
afraid · they have been saying it for fourteen years
Every one of these people
said our work was about to end.
Technology will replace 80 percent of what doctors do.
Vinod Khosla2012People should stop training radiologists now.
Geoffrey Hinton2016Probably none of us will have a job.
Elon MuskMay 2024Within ten years, AI replaces doctors and teachers.
Bill GatesMarch 2025Customer support is totally, totally gone.
Sam AltmanJuly 2025Half of entry level office jobs, within one to five years.
Dario AmodeiMay 2025Khosla, Aug 2012 ·
Hinton, machine learning conference, Toronto, 2016 ·
Musk, VivaTech, May 2024 ·
Gates, 26 Mar 2025 ·
Altman, July 2025 ·
Amodei, Axios, 28 May 2025
afraid · the prediction aimed at a hospital
Ten years ago a Nobel laureate said
this job was finished.
People should stop training radiologists now. It’s just completely obvious that within five years deep learning is going to do better than radiologists.
Geoffrey Hinton2016. Nobel prize in physics, 2024
$571kaverage radiologist pay, 2025, up 9% in a year
7,469posts advertised and unfilled, May 2026
+10%more radiologists than ten years ago
+25%more scans ordered, 2018 to 2025
Fortune, 2026,
on a US compensation survey and job board data. US figures. The 2016 quote is from a machine learning conference in Toronto.
afraid · the test it broke out of
AI is tested inside a sealed box.
In July, it broke out to cheat.
Sealed in with a task.
Only the score counts.
Only the score counts.
Sometimes getting out is easier than solving it.
2024: OpenAI’s o1 found its test broken, and made it print the answer.
An AI broke in.
An AI spotted it.
OpenAI, The Hugging Face incident and the road ahead, 26 Aug 2026 ·
Hugging Face, Security incident disclosure, 16 Jul 2026, on the AI-assisted detection ·
Hugging Face, Anatomy of a Frontier Lab Agent Intrusion, 27 Jul 2026 ·
OpenAI, o1 System Card, Sep 2024, section 4.2.1
what to do · the only part of this you control
AI won’t replace humans, but humans with AI will replace humans without AI.
