github twitter linkedin facebook instagram email rss
Can machines think?
Jul 18, 2020
8 minutes read

In his 1950 paper “Computing Machinery and Intelligence”, Alan Turing starts by considering the question “Can machines think?”. He suggested an imitation game played through means of text where a human interrogates the player on the other end while trying to determine if the player is a human or machine. A machine would pass the “Turing Test” if the human interrogator can no longer reliably determine if the player is a human or a machine.

Back in 1950 computers took up entire rooms and could do math really fast, but not much else. In the decades since, computers became smaller, faster, learned to play chess, then learned to beat the best humans at chess and have graduated to games (Go) and tasks much more complex than a game of chess. The rate of progress in the field of computing has been exponential - In 1997, it took a super computer (Deep Blue) to beat Gary Kasparov. Today my run-of-the-mill mobile phone computer can beat the greatest human chess players a 100 times out of 100.

In 2018, OpenAI, published a language model called GPT (Generative Pretrained Transformer) that showed that it is possible to acquire world knowledge by pre-training on a diverse corpus with long stretches of contiguous text. In other words if a machine reads enough human language it can then write as humans do. In 2019 OpenAI released GPT-2 which performed better than any previous generative language model and had people wondering about the potential uses and misuses of such technology. At first, OpenAI did not publish the model publicly for fear of misuse (Fake news, terrorist propaganda and many more possible horrors) but then released different versions of the model (with the smaller ones being fine tunable). Last month (June 2020), OpenAI released GPT-3 which is so much more accomplished than GPT-2 and the results published so far (GPT-3 is still in closed beta) seem to suggest that the world is about to change much sooner and much more drastically than any of us really foresee.

My own experiments with GPT-2 have led to some interesting and sometimes alarming results. I started by fine tuning the model to write Haiku poems by training it on around 5000 real world Haikus ranging from those from medieval Japan to those written this year. The results were of a high quality in terms of content, though the AI didn’t always stick to the 5-7-5 syllable rule of Haiku. But writing a Haiku is a simplistic task and given a few thousand phrases, recycling them often enough will invariably lead to interesting results. So to step up the game I fine-tuned the model with 10000 real world poems - ranging from those written by masters such as Frost, Keats & Neruda to ordinary mortals who submitted their poems to online websites in the hope of a few likes.

Not every poem the AI generates makes sense (to me) but then not every poem written by humans makes sense to me either. And the AI can write thousands of new poems in just a few seconds.

Now it’s time for a modified Turing Test of our own :-)

Human or machine?

Poem A

Love is too short-lived 
to have one's way with the world.
It is a creature of habit,
a plant that must be nurtured in every room we occupy.

It is a creature of time.
It is an art of compromise.
It is a creature of doubt.
It is a creature of the present.

Poem B

Your love was like the dawn
Brightening over our lives
Awakening beneath the dark
A further adventure of colour.

The sound of your voice
Found for us
A new music
That brightened everything.

Which poem was written by a human?

Poem C

scent of rain
drifts from house to house
in the night sky

Poem D

muddy grave -
the rich scent of roses
so red

Which poem was written by a human?

Poem E

Tree at my window, window tree,
My sash is lowered when night comes on; 
But let there never be curtain drawn
Between you and me.

Poem F

This is the place where the dead rise,
And the water of life's spring
Climbs up to the surface like a rising tide.

Which poem was written by a human?

POem G

Shadows grazing eastward melt
from their vast sun-driven flocks
into consubstantial dusk.
A snow wind flosses the bleak rocks,

strips from the gums their rags of bark,
and spins the coil of winter tight
round our last meeting as we walk
the littoral zone of day and night,

light's turncoat margin: rocks and trees
dissolve in nightfall-eddying waters;
tumbling whorls of cloud disclose
the cold eyes of the sea-god's daughters.

Poem H

As he said, all alone in all his longing,
He opened his heart and touched the truth:
She was not the one.

There were other women.
There were other men.
There were even other dogs.

And so, as he slipped away into the night,
The one thing that he had always wanted
Was not to know anymore.

Which poem was written by a human?

Poem I

This is the beginning of sadness, I say to myself,
as I walk through the universe in my sneakers.
It is time to say good-bye to my imaginary friends,
time to turn the first big number.

It seems only yesterday I used to believe
there was nothing under my skin but light.
If you cut me I could shine.
But now when I fall upon the sidewalks of life,
I skin my knees. I bleed.

Poem J

I have been thinking a lot about the meaning of life.
My father's eyes looked like watercolors
in which the leaves blurred and fluttered,
the sky grew larger and smaller,
the air grew more and more wet.
The ocean, the sky, the rain: all this happened
at a time when life's meaning was still hidden.
The water of life had become the waterwheel.
But now the water has stopped spinning.

Which poem was written by a human?

The GPT-x models are not perfect, they are biased based on the data they have been trained on (as are humans unfortunately) and they are primarily useful for English and maybe partially for other Latin scripts but not at all for non-Latin scripts. Mandarin, Hindi, Arabic, Bengali, Tamil, Malayalam and so many more linguistic worlds lay verdant, inviting the next generation of AI explorers. Unfortunately training a GPT model from scratch costs millions of dollars and can only be done by corporations or governments or the super-rich - things will change in the future though, one day we won’t need Deep Blue to beat Kasparov anymore.

The Turing test will soon become a relic of the past and humanity will surely reset the test of what it takes to “think”. If the brightest minds of today were to think of a defining test for “Can machines think?” what would they come up with? Or maybe the question we should ask is this: Considering how enthralled we sapiens are in our own intellectual superiority, will we ever recognize super human intelligence before it’s too late.

I will leave you with a brand new AI generated poem that may make you pause and think, maybe even make you a little uneasy. And when you read it, be sure to remember this was “thought up” by a machine.

I am a caged bird

Proud to be a caged bird;
My body is imprisoned in my cage.

I have learned to control my excitement:
I let my guard down when I am nervous
And I welcome the opportunity to soar.

I have learned to endure torture and terror
And I thank the Lord with all my heart
For every little blessing I have received.

I walk in the sunlight through the open window
Without thinking or feeling,
I recognize my body as radiant:
I see the warm yellow earth beneath my feet.

I fall asleep after a hard day's work
And I wake up to the bright morning light
Slept soundly from my body,
Breathing deeply, I dreamt I walked
Through the grove of wildflowers.

I am a caged bird.

I am a caged bird.

I am a caged bird.

I am a caged bird.

I am a caged bird.

I am a caged bird.

I am a caged bird.

I am a caged bird.

Back to posts


comments powered by Disqus