Artificial stupidity


Codemotion Madrid 2019-09-24

Your Host Today

@pinchito from Influencity

What We Will Cover


Early AI


Coming of Age


Case Studies


Analysis


Conclusions

Why Me?


20+ years in the industry


Interested in the area since the 90s


Read a bunch of blog posts


Extremely muscular body

In Memoriam Sergio Leone

Roma 1929 - Roma 1989

Sad HIll Unearthed

Now on Netflix

Early AI

Pioneers



Quite Optimistic

In the literature (US)


 

In the Literature (EU)

Coming of Age

Deep blue

1997-05-11: Kasparov defeated

Extra, Extra! Humans Defeated!


But were we?

Definitely human!

So rather: algorithm beats person

Automated Theorem Proving

Neural Networks

In the Literature (us)

In the Literature (Japan)


Case Studies


1: Robots

Industrial Robots

space Robots


BattleBots

Don't give them the kill switch!

You know software has bugs, don't ya?

Actual robots

Not so Intelligent


Not real bots

Not so versatile after all


It is probe or move apparently

Exo drives

Humans are great value after all

Why the emphasis on robots?


The robot myth has been used for centuries to discipline the workforce.

Workers might be replaced first by a steam machine, next by an industrial machine and now by an intelligent machine.

2: Deep learning



How Deep learning works

Many hidden layers

Train on large dataset


Verify on even larger dataset

Two Neurons (2004)

Alphazero


Beat Stockfish after 9 hours of training


The game of chess represented the pinnacle of AI research over several decades
-- DeepMind


Alphazero on your laptop?



Running on 44 cores + 4 TPUs


Stockfish had quite weird settings

The End of the Line?


I think we need to consider the hard challenges of AI and not be satisfied with short-term, incremental advances.


Realistically, deep learning is only part of the larger challenge of building intelligent machines.

Maximization

The power of a neuron

Roughly equivalent to a complete neural network

We have ~100 billion neurons

3: Face Recognition


Pattern Recognition too!

Scrambled Images



Hard for a human, easy for an AI

"Bag of features" model

"Live" Demo


How to fool a neural network

One pixel attack for fooling deep neural networks

A tragedy in four acts

 

Easy to Misuse

Amazon recommends 99% accuracy at least

Too Many Giraffes?


Not easy to escape?

4: self-Driving car

Improvements are awesome

state of the art sensors

vs primate in metal box

Ape wins

Visionaires?


The automobile of the day-after-tomorrow will not be driven by the owner, but by itself.

It may be one day a serious offence to drive an automobile on a public highway

-- Arthur C Clarke, 1962

Try 1830


Self-Driving Car Manufacturers

Waymo
GM
Nissan
BMW
Tesla
...
Seat?
Dacia?
Perodua??
Saipa???

Tesla?


That Tesla?


Analysis

Optimization Problems


AIs are great at optimizing multiple parameters in many-d

Everything becomes an optimization problem

Politics as an Optimization problem


Maximize votes


Minimize inequality?


Maximize power?


Minimize dissent?

Ethics as an optimization problem


Minimize suffering?


Maximize happiness?


Minimize inequality?


Maximize ...?

Philosophy as an optimization problem


Maximize coherence?


Minimize first principles?


What should we ask?

Religion as an optimization problem?


...


Don't even know where to start

Maximization of unfairness


Algorithms perpetuate systemic bias
Perpetuate unfairness

AI as tools


Employment

Algorithms are the definitive bogeyman

AI uprising

Conclusions

Conclusions


AI is awesome, but not that awesome


We are many years away from science fiction AIs


Using AIs is not a blank check


Think about what you are doing


You can Never be sure



Dear robot overlords:
I was just trying to maximize happiness!

Thanks!

Artificial Stupidity - Codemotion Madrid 2019

By Alex Fernández

Artificial Stupidity - Codemotion Madrid 2019

Presentation for Codemotion Madrid 2019-09-24

  • 3,266