overview

The question of AIs in courtrooms suggests different trains of thoughts: - AI offers an approach to standardization to guarantee equal treatment - AI developed in a business setting cannot offer due process due to being a trade secret - A combination of both humans an AI might be the best way to go forward

turing and his test

the imitation game, 1950

can one tell apart a man and a woman? a human and a machine?

the difference between being and appearing
the role of education

In the process of trying to imitate an adult human mind we are bound to think a good deal about the process which has brought it to the state that it is in. We may notice three components:

(a) The initial state of the mind, say at birth,

(b) The education to which it has been subjected,

(c) Other experience, not to be described as education, to which it has been subjected.

Turing starts by asking the question of whether we could ever tell apart a robot from a human, if we could only talk to them via a computer interface (i.e. texting). This is the first standard of intelligence that is put forward.

hype #1 - symbolical

1956 in dartmouth

highly mathematical approach

culminates in SHRDLU, a language interface for spatial recognition (1968-1970)

The first approach to AI starts in 1956, with the coining of the term "Artificial Intelligence" by a bunch of mathematicians at Dartmouth. They assume that one can represent the mind through abstract concepts and connections.

criticism #1 - eliza

joseph weizenbaum, 1964

ELIZA is one version a single chatbot

- ELIZA: how are you feeling today?
- PATIENT: i'm feeling sad
- ELIZA: why are you feeling sad?
- PATIENT: because of some dumb thing
- ELIZA: what kind of dumb thing?

would you use a chatbot as a therapist?

complimentary help
not as qualified (does not get nuance)
more access but what about limits?
difference with talking to a human (body language) / empathy and presence
data privacy
removes social stigma / no judgment

the difference between choice and judgment

[Politicians in the Pentagon] wanted to constantly find laws with which political and historical facts could be explained and predicted, as if they were made with the same necessity and thus reliability as the physicists of natural events used to believe in the past.

An extremely irrational trust in the predictability of reality has become the leitmotif of decision-making.

there are things that machines ought not to do

The first criticism of this approach comes from Joseph Weizenbaum who, in his book Computer Power and Human Reason makes the difference between things that require choice, which a computer can do, and things which require judgment, which a computer cannot do. This is because the former is only calculation, while the latter is value based.

He intends to show that machines can never be truly conscious by building one: the first chatbot, ELIZA. However, the illusion of intelligence remains and psychiatrists across the country reach out to him to use the robot in therapy.

winter #1

turns out, computers don't work so well in the real world

1965s - 1975s

These debates take place exclusively in academia, where there is quite a lot of computing power to do these intelligence experiments. However, the private sector does not benefit from it, and the hype dies out.

hype #2 - experts

increase in data and usable languages enable commercial applications

SABRE (automated travel agent) - 1976
MAVENT (mortgage and lending compliance automation) - 1980
HFT (high frequency trading) - 1983

The second wave results in expert systems, which require more computing power, and in the formalization of complex rule systems, such as finding the cheapest airline ticket, calculating someone's mortgage, or deciding to buy or sell stock. Some of these systems still exist today, but the normalization of their use makes them less prone to being qualified as "AI".

criticism #2 - chinese room

a system based on rules is not necessarily conscious

The second criticism aims at the fallacy of representing knowledge as rules. John Searle proposes an experiment to illsutrate it: the chinese room paradox

winter #2

logic as knowledge representation peaks without hardware

hype #3 - big computers

MITI, IBM build very fast computers, still operating on logic

kasparov vs. IBM Deep Blue, 1997

The summit of the logical/big hardware approach is when IBM builds a computer which manages to beat Kasparov, chess grandmaster. However, one might think it's only a fairly restricted conception of intelligence.

connected intelligence

meanwhile

frank rosenblatt pretending not to be confused by his perceptron, 1950s

There is another alternative to the development of AI: rather than logical of few, complex concepts, one can also take a huge amount of somewhat simple concepts.

sensing and thinking

what the frog's eye tells the frog's brain, 1959

so what about neurons? can you represent thought just through single perceptive units?

but we still need fast computers and a lot of input

Additionally, fundamental research in neuroscience shows that not all intelligence happens in the brain (the illusion of "i think therefore i am" philosophical tradition). Particularly, some visual processing of objects happens directly in the eye rather than in the brain.

hype #4 - big data

mnist dataset - from the us bureau of census

neural networks start to perform super well (end of 2000s)

instead of inductive reasoning (i.e. knowing the rules beforehand)

you go for deductive reasoning (i.e. you figure out the rules as you go)

machine learning, deep learning, convolutional neural networks, etc.

essentially deciding on a series of weights given a certain data set, and then checking with a verification dataset.

The only thing that was missing was big data: now that there is a large amount of regular data, we can pay attention to a different kind of intelligence. For instance: how come it's super easy for humans to recognize images, and very hard for computers? So the process of [deep learning](https://en.wikipedia.org/wiki/Deep_learning) works very well with large amounts of data to do trial and error deductions. Because, ultimately, it is this trial and error approach which is a fundamental shift: intelligence systems are no longer inductive, they are deductive.

deep learning and gans

deep learning can train itself

lee sedol loses to alpha go

The technique is so successful that it manages to beat the best human player at Go (at the time, the IBM computer that was the best at chess could barely beat a professional Go player). Still, the standard for intelligence is still that of a well-defined ruleset.

Learn more about how deep learning works here

today

large language models

the real name of chatGPT

"As long as you can turn it into a vector, you're OK"

Yann LeCun, Chief AI Scientist at Meta

a vector is a series of numbers, that can represent a position in space (space can go beyond 3 dimensions)

By combining this approach with a representation of text as a set of vectors (i.e. putting words in space), we can start to extract higher-level features (i.e. meaning) from a raw corpus of common crawl

reflecting the dataset

Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings

culturally-encoded datasets: moral machine

Now that we work with masses of data, we have new problems: biased data that reflect the inequalities and injustices in society. There are multiple ways to deal with it:

changing the dataset to include more representational data
training the AI to nudge it towards a non-biased response

nlp

natural language processing tasks

translation
emotion detection
task comprehension
language generation

the transformer architecture

Attention is all you need, Google, 2017

not just where words are compared to all the different other words, but where they are in a sentence

In 2017, Google releases a paper which combines this word-to-vector representation with a linear development: basically, they have an algorithm which can guess the next word in a sentence, and turns out it's very effective!

the role of presentation

autocomplete is minor AI -> over a long period of time, small increments, embedded in the everyday

chatGPT is major AI -> radical break, huge jump in abilities, stand-alone application

how do you use chatGPT?

LLMs and jobs

to what extent do you think is AI going to take your jobs?

job can be more efficient (can be a good thing depending on the culture)
bring a new approach to a topic
different skills (e.g. speaking)

when you behave like a machine, you are liable to being replaced by a machine

-> qualified workers in factories replaced by physical machines

-> qualified workers in offices replaced by cognitive machines

AI Act

limits AI use to different applications, based on risk level (EU)
limits AI production with different methodologies and labels (US)

companies need a massive amount of process in order to produce a "correct" AI

internal safety checks and procedures
external global underpaid workforce

carbon footprint:

training GPT4 is the equivalent of 600 transatlantic flights
the increase has been 300,000x from 2013-2023
LLM query is 4-5x higher than a search query

conclusion

hype #5

things would change but not too much

if you make highly patterned cognitive work, you're in trouble

human-machine cooperation tends to yield better results than machine alone

mitigations are happening at both the process and application level, but should not detract from important issues

project check-in

group work

data collection?
data analysis?
outcome presentation?

submission

-> send me one URL per group (you can check that I have it on the grading sheet), due before class

-> 10min presentation of your project: lit review, methodology, analysis, conclusion