Pediatric Surgery  ·  Part one

The Story
of AI

Seventy-six years, four eras, and one word that has never once stood still.
Mohamed Maslooh  ·  2 September 2026
Chapter one  ·  1950 – 1980s

The rule-based era

Forty years of trying to write intelligence down by hand, one rule at a time.
TURING  ·  THE NAME  ·  ELIZA  ·  MYCIN  ·  DEEP BLUE
1950  ·  where the whole thing starts

It started with one man,
sitting and asking himself a question.

Alan Turing, a British mathematician. His 1950 paper opens by putting three words in quotation marks.
Can machines think?
No machine on earth could have answered him. So he swapped the question for a test.
Alan Turing, seated in a garden
Alan Turing 1912 – 1954He wrote the question down seventy-six years ago, when no machine on earth could go near it.
A B ? A B
If he cannot tell which is which, the machine has passed.
Computing Machinery and Intelligence  ·  Mind LIX(236), October 1950, pp 433–460. It opens: “I propose to consider the question, ‘Can machines think?’”↗
1955  ·  where the name comes from

Somebody had to
give it a name.

Dartmouth College  ·  31 August 1955
A Proposal for the Dartmouth Summer
Research Project on Artificial Intelligence
J. McCarthyM. L. MinskyN. RochesterC. E. Shannon
Dartmouth summer research project proposal, 31 August 1955.
1966  ·  the first chatbot

ELIZA had no idea
what you were saying.

All of it, in one rule
if the sentence contains X
reply “tell me more about X”
23%
judged to be the human
five minute conversations, 2025
MIT  ·  PROJECT MAC
Weizenbaum, CACM 9(1), January 1966, 36 to 45↗.  Score: Jones & Bergen, arXiv 2503.23674, 2025↗, a three-party test run sixty years later.
1974  ·  the rules reach medicine

MYCIN chose antibiotics
better than the

RULE 001
IF the stain is gram negative
AND the shape is a rod
AND the site is the blood
THEN suggestive evidence (0.8) that the organism is E. coli
0rules, typed by hand,
one at a time
Therapy judged acceptable by blinded outside experts
Ten infectious-disease cases · higher is better
MYCINStanford faculty
MYCINthe program
65%
Facultyfive specialists
42.5 – 62.5% spread
55.5%
0 25 50 75 100%
Faculty bar is the mean of the five; the band is their full range.
Yu et al., “Antimicrobial selection by a computer: a blinded evaluation by infectious diseases experts”, JAMA 1979;242(12):1279–82↗. MYCIN itself: Shortliffe, Stanford AI Memo AIM-251, 1974.
1974  ·  and yet

It never treated a single
real patient.

Nobody could feed it

Every fact about the patient was typed in by hand, one question at a time.

It did not fit the day

It lived on a shared mainframe at Stanford, and one consultation ran half an hour.

Nobody would sign for it

If the machine is wrong, the responsibility has no owner.

Being better at the task was never the hard part. That is the whole lesson of the rule-based era, and it has not expired.
1997  ·  the rule-based era’s last and loudest win

Deep Blue beat the world champion.

Garry Kasparov at the board
Garry Kasparov, world champion 1985–2000. Six games, New York, the first time a machine took a match from a reigning champion.
Chess positions examined, per second
This is the whole explanation of how it won
Kasparovthe world champion
about 2
Deep BlueIBM, 1997
200,000,000  →
Drawn to scale. Kasparov’s bar is three pixels wide and Deep Blue’s runs off the edge of the slide.
And it learned nothing from a single game it ever played.
Chapter two  ·  1986 – 2015

The machine
learning era

Somebody asked the obvious question: why are we writing the rules at all?
EXAMPLES  ·  2012  ·  A SHORT DETOUR
The single most important change in the story

Stop writing the rules. Show it examples instead.

WHAT A PERSON HAD TO WRITE rule 118 of 450
IF the stain of the organism is gram negative
AND the morphology is rod
AND the aerobicity is aerobic
THEN there is suggestive evidence (0.6) that the class is enterobacteriaceae
Typed by hand, one at a time. A new disease means starting again.
WHAT YOU HAND IT INSTEAD 100,000 labelled slides
Nobody writes a rule. It finds the difference itself.
Nobody tells it what a tumour looks like. It is shown the pictures and the answers, and works out the rule that connects them.
2012  ·  the year it actually turned

A competition to name
what is in a photograph.

Top-5 error · lower is better
Best hand-
written systemrunner-up, 2012
26.2%
Neural networkToronto, 2012
15.3%
a photograph of a cat
cat
a photograph of a cup of coffee
cup
a photograph of a handful of coins
coins
a photograph of a rocket on the pad
rocket
Krizhevsky, Sutskever & Hinton, University of Toronto, 2012↗.
A short detour  ·  2015

AI was trained to read breast pathology slides.

Accuracy on the slides, over fifteen days of training
50% 60% 70% 80% 90% 100% CHANCE 85% day 1 5 10 day 15
85%the learner,
after fifteen days
80%a panel of
radiologists
?
Then what was it
The radiologist figure is from the mammogram set in the same study. No pathologist score is reported for these histology slides.
And now the part I did not tell you

It was a pigeon.

Eight of them, in a box, pecking a touchscreen for a food pellet.
Which tells you something important

Spotting a pattern
is not the same as
understanding it.

The pigeon has no idea what cancer is. It got very good at telling one kind of picture from another. That is exactly what the machines in this chapter do.
A pigeon in a test chamber facing a touchscreen
The actual apparatus. The bird pecks blue or yellow to call the slide benign or malignant, and a correct answer drops a food pellet.
Levenson, Krupinski, Navarro & Wasserman, “Pigeons (Columba livia) as trainable observers of pathology and radiology breast cancer images”, PLOS ONE 10(11): e0141357, 18 November 2015↗.
Chapter three  ·  2017 – 2022

Generative AI

One paper out of Google, and almost everything you have heard of since.
THE TRANSFORMER  ·  GPT  ·  THE GUESS  ·  NOVEMBER 2022
June 2017  ·  the most important paper in this story

Almost every model you have heard of
is built on the same design.

Google  ·  12 June 2017  ·  eight authors↗

Attention Is All You Need

The design in this paper is called the transformer. Google built it, then published it for anyone to use.

ChatGPTOpenAITRANSFORMER
DeepSeekChina
GeminiGoogle
KimiMoonshot
GrokxAI
QwenAlibaba
ClaudeAnthropic
MistralFrance
GLMZ.ai
CommandCohere
LlamaMeta
MiniMaxChina
FalconTII
HunyuanTencent
every one of them  ·  a transformer
2017  ·  what the paper actually added

Every word gets to look at everything before it.

BEFORE · ONE WORD AT A TIME, LEFT TO RIGHTthecatdidnotcrossthestreetbecauseitwastired2017 · IT WEIGHS EVERY WORD BEFORE IT, ALL AT ONCEthecatdidnotcrossthestreetbecauseitwastiredwide
Vaswani et al., “Attention Is All You Need”, arXiv:1706.03762, 12 June 2017↗.
What it is actually doing

It is not looking the answer up.
It is guessing, one word at a time.

Why is the sky blue?
THE WORDS SO FARFIXED WEIGHTS THE NEXT WORD
If you photograph one slide, make it this one

Not four competing things. Circles inside circles.

ARTIFICIAL INTELLIGENCE Anything a machine does that would need intelligence from a person. MACHINE LEARNING It learns from examples instead of being handed the rules. DEEP LEARNING It learns in layers This is what reads a scan. GENERATIVE It produces, rather than labels The newest and smallest circle of the four.
So when somebody says “AI did this”, there is one question worth asking: which circle do you mean?
30 November 2022

Then somebody put
a text box on it.

chat.openai.com
ChatGPT
0
people signed up
5days
Days taken to reach one million users
Shorter is faster
Netflix1999
1,278 days
Facebook2004
304
Instagram2010
91
ChatGPT2022
5 days
All four bars are on the same scale.
Chapter four  ·  2023 – now

Agentic AI

The newest of the four, and the least settled. It stops answering you and starts doing things.
THE ONE YOU WILL BE SOLD NEXT
What actually changed

From answering to acting.

A CHAT ASSISTANT you ask, it answers, it stops
YOU Arrange dinner for the department on Friday evening.
IT Of course. Here are a few restaurants you could try, and a suggested message you might send to everyone…
Then it stops. You still have to do all of it yourself.
AN AGENT same request, four different tools
YOUR CALENDAR Opens itand finds that nine people are free after six
A WEB BROWSER Searchesfor a table for nine near the hospital
THE BOOKING FORM Fills it inand confirms the table for eight o’clock
YOUR EMAIL Tells everyoneand puts it back in the calendar
if the table is gone, it goes back and finds another one
It is no longer telling you what to do. It is reaching into your things and doing it.
As of 2026 this is the least mature of the four. Treat any number you are quoted as provisional.
Today  ·  so where has it actually got to

In 1997 it beat one man at one game. This is where it is now.

TURING’S OWN TEST2025
Judged to be the human, five minute conversation
ELIZA1966
23%
GPT-4.52025
73%
Jones & Bergen, arXiv 2503.23674, 2025↗
BREAST SCREENING2026
Cancers found, 105,934 women, randomised
Two radiologistsdouble reading
73.8%
The same twowith AI
80.5%
MASAI, The Lancet, 2 Feb 2026↗
DIAGNOSTIC REASONING2024
Median score on real cases, 50 physicians
Doctorsas they work now
73.7%
AI aloneno doctor
92%
Doctors with AIgiven the tool
76.3%
Goh et al., JAMA Netw Open 2024;7(10):e2440969↗
The one thing to take out of the room

Three different technologies,
wearing one name.

IF feverAND culture positive THEN treat
Rules we wrote
Deep Blue · MYCIN

Does exactly what it was told. Predictable, and it cannot surprise you.

Trust it inside its rules. Nowhere else.
Patterns it learned
Bone age from a hand film

One narrow task, learned from labelled films and measured against radiologists before anyone used it.

Trust it where it has been measured.
the next ?
A machine that guesses
Claude · ChatGPT

Produces the next likely word. Fluent everywhere, verified nowhere.

Never trust it unchecked on a fact.
And one last thing about the word itself

Every one of these was called artificial intelligence the year it arrived.

Chess1997
Spam filterslate 1990s
Finding a route2000s
Autocorrect2000s
Face unlock2017
Voice assistants2011
“As soon as it works, no one calls it AI any more.”
Often attributed to John McCarthy, though only after his death; the earliest documented form is Larry Tesler’s “AI is whatever hasn’t been done yet” (1979). Quote Investigator↗
Back to where we started

1974.
Remember these?

01
Nobody could feed it Today it reads your notes in a second.
02
It did not fit the day Today it sits in your pocket.
03
Nobody would sign for it Fifty two years later, still nobody has an answer.
Pediatric Surgery  ·  The story of AI

Thank you.

Mohamed Maslooh

Speaker notes

→ NEXT  ·  ← BACK  ·  N NOTES  ·  F FULL SCREEN
This is an interactive presentation. You are looking at a preview that cannot run it. Tap the share button and choose Open in Safari or Open in Chrome, and it will work.
Turn your phone sideways
← Lecture page