To provide you with ideas and strategies to make your own choices.
You know what works for your life/circumstances better than me.
4 of 52
Agenda
What is AI?
What is ChatGPT?
What is the good, the bad, and the ugly of all this?
What does history teach us?
How should we think about this?
What does the future hold?
5 of 52
Search as we know it…
If a query can be answered in one interaction, Google can probably do it.
An “AI” is like a librarian who’s read all the books and just tells you the answer via a conversation.
It’s a big time-saver.
6 of 52
What is “artificial intelligence” (AI)?
Creating and training machines to perform intelligent tasks.
Something that, if a person was doing it, we would call it intelligence.
Theoretical work dates from Alan Turing’s 1950 paper.
The term was coined in 1956 by Marvin Minsky and John McCarthy.
Improvements proceeded incrementally for decades.
7 of 52
Jargon Alert!
Almost all of the tech jargon in this talk is on the next slide…
8 of 52
AI: ML, LLM, Neural Network, GPT
Machine learning (ML) and neural networks (NN)
“Learn” patterns and predict outcomes by analyzing massive amounts of data.
Supervised learning through large amounts of manually-labelled data.
Large language model (LLM)
Uses ML to learn to write and converse with users; “chatbot” tech.
They find statistical relationships between words.
New unsupervised models can absorb nearly the entire Internet.
Generative pre-trained transformer (GPT)
A type of semi-supervised LLM able to generate novel human-like content.
9 of 52
The new AI “Humanness”
It feels like conversing with another human being.
An LLM has no body, no visceral connection to the real world.
This can lead to some dumb pronouncements.
Unlike a human, it doesn’t go on vacation for a week, wonder where you are, or get irritated.
10 of 52
Key Idea
Extraordinary things happen when you can hold the entire picture in your brain at once while interacting with a human.
These tools can do that.
We can’t.
11 of 52
ChatGPT embodies that key idea
GPT architecture was invented in 2017.
First GPT model was widely deployed in 2020.
GPT-3 was deployed in March, 2022.
ChatGPT was first deployed in November, 2022.
Natural extension of decades of ML and LLM improvements.
Adds a general comprehensive natural human interface to AI.
All hell broke loose in March, 2023.
Adopted by 100M global users in two months, fastest in history.
12 of 52
What is ChatGPT?
It’s a very sophisticated mimic with very human-like interactions.
Based on an utterly inhuman technology.
More model complexity equals more humanness.
It’s imperfect and capable of creating nonsense.
It’s frozen in September, 2021, for hard technical reasons.
Data model coverage limitations.
GPT-3 limited to about 3,000 input words; GPT-4: about 30,000 words.
GPT-4 understands images as well as words.
It’s a thought-partner.
It’s a gift to transparency.
13 of 52
What’s it good for?
It’s a very useful tool, but more of a proof of concept at this point.
We need to work out how to fit it into our lives.
It’s like having infinite helpful interns or graduate students.
It’s a powerful oracle, a muse for ideas and inspiration.
It’s an innovative form of social collaboration moderated by statistics.
You must trust but verify its responses; it can be “usefully wrong.”
It has limited ability to take action, but that will change.
14 of 52
Interacting with AI: Prompts
Treat it as if it can effectively understand your question and its response.
Prompts are all-important. It’s a conversation.
More art than science.
Repeat questions generate different responses.
Chain of Thought (CoT) prompting.
“Let’s work this out in a step by step way to be sure we have the right answer”
The ability to effectively talk to machines might be one of the most vital skills of the 21st century.
15 of 52
Demo Time!
16 of 52
Anatomy of a Megaprompt
17 of 52
Hallucinations
ChatGPT isn’t an end-to-end solution.
It’s a perfectly accurate reflection of its training set.
It’s responses are always “this is the kind of answer that people tend to give to questions like that.”
The more information relevant to the question that existing in the training set, the more accurate the answers.
The problem is that people will believe the inaccuracies.
18 of 52
“Usefully Wrong”
“Hallucinations” can’t be solved.
It’s absolutely fundamental to what it’s doing.
Inaccuracy is not that relevant.
It’s a general purpose reasoning engine.
It doesn’t need to be truthful or even reliable.
It just need to be challenging. It’s suggestions are within the bounds of credibility.
Don’t just copy and paste responses; reflect on what it’s telling you.
Use Google and other tools to verify facts.
It’s usefully wrong.
Like an intern who makes a five-page sales document, but you have to check it.
The back and forth drives value.
19 of 52
What could possibly go wrong?
Just about everything we think is bad about hacking and social media posts will likely get worse.
Copyright issues.
Regulation can’t solve anything.
Existential fears.
“Climate change should hire AI’s PR firm.”
Let’s look at a bit of history…
20 of 52
Electricity Panic
Anti-electricity cartoon from 1889.
The great Thomas Edison was behind this.
This tragic chapter of technology history became famous as "War of Currents" between Edison and Westinghouse.
21 of 52
Horseless Carriages
Here's an old anti-car advertisement by a horse carriage company.
Gave rise to the “horseless carriage” fallacy.
22 of 52
No words necessary…
23 of 52
Some Historical Context
How new technology unfolds always takes everybody by surprise.
Until recently, technological change proceeded slowly.
There have always been trade-offs, mostly unforeseen, with technological change.
Any general-purpose technology brings huge opportunities.
And absolutely no one has any real idea of the changes ahead. It’s pointless to talk ourselves into an illusion of predictability.
24 of 52
Life can only be understood backwards,
but it must be lived forwards.
—Søren Kierkegaard
25 of 52
We’ve been here before…
We need to embrace AI, take the plunge, adapt.
Operate from a place of opportunity and respect for the limitations of the tech.
Find a way to use it to get better at what you do.
We respond to experience, not
something that’s not yet happened.
The process is going to take longer
than you think…
26 of 52
Think EV’s are new?
Electric trucks charging at London St Pancras Station 100 years ago.
27 of 52
AI and Employment
AI won’t replace your job. People who use AI will.
AI has broad application that will impact a wide range of workers.
AI will increase efficiency and accelerate innovation.
AI will mimic top performers and push low performers to act more like them.
AI will teach us new, non-obvious ways of doing things (AlphaGo).
AI will help humans reason from scratch.
28 of 52
Lump of Labor Fallacy�
29 of 52
The Lump of Labor Fallacy
Lump of Labor Fallacy
The mistaken belief that there is a fixed amount of work or jobs in an economy, implying that if one person (or a machine) gains employment, another person must lose it. Zero-sum game thinking.
Because increased productivity leads to…
Higher demand for goods and services, which generates even more employment opportunities, income, and buying power.
Because innovation equals job growth
We’ve never been able to predict the new jobs that will be created.
30 of 52
Jevon’s Paradox
31 of 52
Jevon’s Paradox
Jevon’s Paradox
Increases in the efficiency of resource use can lead to higher overall consumption of those resources rather than expected reductions.