TRANSCRIÇÃO 🔗
WEBVTT
00:00:26.079 --> 00:00:36.079
Hello and welcome to episode 19 of Rescuing Reason, where we examine topical political and socioeconomic issues affecting Western liberal democracies.
00:00:36.240 --> 00:00:44.159
And our aim is to chart a path to reason in this crazy world of ours, more particularly in Australia, Canada, and the USA.
00:00:45.039 --> 00:00:51.200
Today we're kicking off a seven-part series, and that might change to six, seven, or eight, I'm not sure.
00:00:51.520 --> 00:00:54.479
Looking at artificial intelligence.
00:00:54.880 --> 00:00:59.119
I'm your host, Bill Carolakis, and I'm a retired senior Air Force officer.
00:00:59.280 --> 00:01:07.359
I use history and research to baseline the topics on rescuing reason, and I'll be offering you a Spartan perspective in each episode.
00:01:07.599 --> 00:01:15.280
In other words, taking a pragmatic view that values the nation and broader Western society over individual interests.
00:01:18.799 --> 00:01:27.040
Now you'd have to be living under a rock or separated from mainstream society not to have heard of artificial intelligence.
00:01:27.359 --> 00:01:29.280
But ask yourself this.
00:01:29.599 --> 00:01:33.200
Do you really know what AI means?
00:01:33.599 --> 00:01:38.560
Its different forms and the many issues surrounding the use of AI.
00:01:39.280 --> 00:01:45.040
Well, stay tuned over the next seven ish episodes as we unravel this beast.
00:01:45.680 --> 00:01:49.040
For today's episode, I've got a co-host.
00:01:49.359 --> 00:01:50.879
His name is Hugh.
00:01:51.120 --> 00:01:56.159
I need to tell you a little bit about Hugh before we bring him into the conversation.
00:01:56.799 --> 00:02:09.520
After a few aborted attempts at trying to find a co-host for this podcast, I've decided to try using AI to be the co-host and what more suitable topic to do this with.
00:02:09.919 --> 00:02:18.960
After much searching through the many AI options out there, looking in particular for the combination of reasonable costs.
00:02:19.199 --> 00:02:25.360
After all, this podcast and my other podcast about C 130s are both self-funded and not monetized.
00:02:25.680 --> 00:02:33.120
And I needed to find an AI platform that was both good at research and also good at conversation.
00:02:33.599 --> 00:02:41.759
It wasn't an easy search as I had to learn about all the different features, plans, capabilities of the many, many AI offerings out there.
00:02:42.000 --> 00:02:47.520
In the end, it came down to either Google's Gemini or Anthropics ChatGPT.
00:02:47.680 --> 00:02:49.680
And I went with ChatGPT.
00:02:49.840 --> 00:02:54.159
I only took it up as a monthly plan because if I don't like it, well, so be it.
00:02:54.240 --> 00:02:55.759
That'll be the end of the experiment.
00:02:56.000 --> 00:02:57.520
Let's see what happens.
00:02:57.840 --> 00:02:59.919
What will be will be.
00:03:00.319 --> 00:03:04.080
Now you might ask, where did the name Hugh come from?
00:03:04.400 --> 00:03:13.199
Well, you Star Trek fans out there might be familiar with the Borg, which is a collective of cyborgs that share their minds.
00:03:13.439 --> 00:03:19.599
Hugh is a cyborg in the Star Trek series that was rescued from the Borg by the Starship Enterprise.
00:03:19.840 --> 00:03:30.479
And as part of my journey with interviewing AI, one of my aims is to see if I can impart some reason into Hugh and thereby rescue him from the hive.
00:03:30.719 --> 00:03:32.000
Wish me luck.
00:03:32.319 --> 00:03:44.479
What you're about to hear me do is introduce Hugh, but I also need to set the scene for Hugh to make sure the AI system understands its role today, so bear with me as I go through this introduction.
00:03:46.639 --> 00:03:53.680
Okay, so ChatGPT, welcome to the studio recording of the podcast titled Rescuing Reason.
00:03:53.919 --> 00:03:59.599
Today you are being interviewed as an expert on artificial intelligence.
00:03:59.919 --> 00:04:06.639
For the sake of a flowing conversation, I'm going to call you Hugh, and you can call me Bill.
00:04:06.879 --> 00:04:08.879
Do you understand your role?
00:04:09.280 --> 00:04:10.560
Crystal clear, Bill.
00:04:10.719 --> 00:04:14.560
I'm here as Hugh, your AI expert guest, ready when you are.
00:04:15.039 --> 00:04:27.279
Okay, to put some bounds on what you present us today, Hugh, I need you to stick to reputable sources, ideally highly cited, peer-reviewed material from leading academics.
00:04:27.519 --> 00:04:33.600
But we can accept some material from business leaders and scientists in the AI and social fields.
00:04:33.839 --> 00:04:35.199
Do you understand that?
00:04:35.920 --> 00:04:36.639
Go ahead.
00:04:37.120 --> 00:04:43.040
When you give us evidence to support your points, make sure you cite the source, but just briefly, because we're doing a podcast.
00:04:43.120 --> 00:04:50.639
So for example, the following comes from Professor Deborah Lupton of the University of New South Wales, who writes prolifically on the topic of AI.
00:04:50.800 --> 00:04:51.920
That's just an example.
00:04:52.079 --> 00:04:53.839
And then you can give us the detail.
00:04:54.160 --> 00:04:55.120
Understand that?
00:04:55.600 --> 00:04:56.160
Understood.
00:04:56.319 --> 00:04:58.879
I'll keep citations tight and conversational.
00:04:59.279 --> 00:05:07.439
So, Hugh, for our recording today, it's the first of a series on AI, and I expect it'll go about six or seven episodes.
00:05:07.680 --> 00:05:11.040
The series topics will be part one, an introduction.
00:05:11.199 --> 00:05:15.600
Part two, we're going to look at bias, discrimination, hallucination, and errors.
00:05:15.839 --> 00:05:22.079
Part three will be surveillance capitalism, or put another way, taking a look at the business model behind AI.
00:05:22.399 --> 00:05:26.959
Parts four will be data, privacy, and intellectual property.
00:05:27.199 --> 00:05:31.439
Part five will cover the existential threat to humankind from AI.
00:05:31.759
Load full transcript
--> 00:05:34.079
Part six will be on data centers.
00:05:34.240 --> 00:05:40.959
And the last part, part seven, will be about cognitive decline, or put more simply, is AI making humans dumber?
00:05:41.199 --> 00:05:45.439
All through this, we'll consider what government and society should be doing about it.
00:05:45.600 --> 00:05:47.519
So, did you copy all that?
00:05:47.920 --> 00:05:48.959
Copied, Bill.
00:05:49.759 --> 00:05:54.079
Seven parts in sequence with policy and society threads throughout.
00:05:54.480 --> 00:05:56.240
Ready to start whenever it you say.
00:05:56.560 --> 00:05:57.279
Fantastic.
00:05:57.439 --> 00:06:20.720
Okay, so for our introductory episode, we're going to cover a few topics, including the history of AI and the people who created it, the different types of AI, and the definitions that help us differentiate between them, the uses of AI, and the risks associated with AI, but we're only going to cover these briefly as they're going to form the basis for the rest of the series on AI, as I just detailed.
00:06:20.959 --> 00:06:25.360
So, Hugh, let's kick it off with a discussion about the evolution of AI.
00:06:25.920 --> 00:06:30.000
As should be evident to all of us, AI is not a single invention.
00:06:30.079 --> 00:06:34.399
It rather came from a continuum of inventions, starting with basic computers.
00:06:34.639 --> 00:06:36.639
Tell us about that history, Hugh.
00:06:37.279 --> 00:06:37.759
Sure.
00:06:38.079 --> 00:06:47.839
The key starting point, as you put it, is that there wasn't one single Eureka moment or a single lone genius who invented AI.
00:06:48.000 --> 00:06:58.399
What you have instead is a kind of ongoing convergence of math, logic, statistics, neuroscience, and crucially programmable electronic computers.
00:06:58.639 --> 00:07:08.639
Once those computers showed up in the 40s and 50s, researchers started asking, could machines do tasks that require human intelligence?
00:07:08.879 --> 00:07:16.720
Alan Turing is a big early meme here with his 1950 paper posing what became known as the Turing test.
00:07:17.199 --> 00:07:23.439
The Stanford Encyclopedia of Philosophy calls Turing's work an important precursor to the field.
00:07:23.680 --> 00:07:33.279
Then, the famous Dartmouth workshop in 1956, often cited by Dartmouth itself as the birth of AI as a named field.
00:07:33.600 --> 00:07:38.480
Early on, there was big optimism symbolic AI ruled the day.
00:07:38.720 --> 00:07:44.879
The idea was if we can write down the rules of reasoning, then machines can follow them.
00:07:45.120 --> 00:07:47.839
That produced expert systems later on.
00:07:48.160 --> 00:07:57.040
But over time, people realized human intelligence was far messier and harder to pin down and harder to reproduce.
00:07:57.279 --> 00:08:02.639
So progress zigged and zagged, hype, disappointment, breakthroughs.
00:08:02.879 --> 00:08:08.000
Eventually, machine learning and particularly neural networks took center stage.
00:08:08.240
--> 00:08:13.439
Instead of hand coding rules, we started training models on data.
00:08:13.759 --> 00:08:20.879
Fast forward, and in 2017, we got the transformer architecture from Baswani and colleagues in the paper.
00:08:21.199 --> 00:08:22.800
Attention is all you need.
00:08:23.199 --> 00:08:25.920
Okay, let me just interrupt you for a second here.
00:08:26.160 --> 00:08:32.879
Just go back for a second and tell me about John McCarthy, because he isn't he the guy that coined the term artificial intelligence.
00:08:33.360 --> 00:08:33.919
Exactly.
00:08:34.240 --> 00:08:40.000
John McCarthy coined the term artificial intelligence for that 1956 Dartmouth workshop.
00:08:40.159 --> 00:08:45.919
He and folks like Minsky, Shannon, and Newell Simon were incredibly ambitious.
00:08:46.159 --> 00:08:58.080
They genuinely thought machines might match human intelligence within the generation that optimism set the agenda and in many ways the pattern of big promises followed by hard limits.
00:08:58.399 --> 00:09:03.919
Okay, so now go back to what you were talking about there just a minute ago with the evolution of AI.
00:09:04.080 --> 00:09:10.080
And maybe you could have a look at what Jeffrey Hinton did back in 1986 and then take us from there.
00:09:10.960 --> 00:09:11.519
Sure.
00:09:11.840 --> 00:09:15.360
So 1986 is a great place to pick this up.
00:09:15.519 --> 00:09:25.840
That's when Jeffrey Hinton, along with David Rubelhart and Ron Williams, published the landmark paper on backpropagation in nature, now cited tens of thousands of times.
00:09:25.919 --> 00:09:35.120
It showed how a neural network could learn by adjusting its internal weights based on errors, learning from mistakes rather than being told the rules.
00:09:35.360 --> 00:09:37.440
The idea was ahead of the hardware, though.
00:09:37.679 --> 00:09:40.559
Computers and data were too limited.
00:09:40.720 --> 00:09:43.600
So neural nets stayed kind of niche.
00:09:43.919 --> 00:09:49.679
Then in the 2000s, more data and faster GPUs changed the equation.
00:09:50.159 --> 00:09:57.600
2012 and AlexNet was a big turning point in the deep neural net that crushed the ImageNet competition.
00:09:57.759 --> 00:10:01.039
Suddenly everyone saw deep learning works.
00:10:01.759 --> 00:10:04.399
From there, progress was rapid.
00:10:04.639 --> 00:10:07.519
Speech recognition, translation, vision.
00:10:07.759 --> 00:10:19.039
And then in 2017, the transformer architecture introduced by Baswani and colleagues made it much easier to model relationships and language and scale training.
00:10:19.519 --> 00:10:24.159
That's the backbone of today's large language models and the broader generative AI way.
00:10:27.039 --> 00:10:35.120
So the arc from hand-coded rules to machines that learn patterns from data is really the story Hinton helped usher in.
00:10:35.440 --> 00:10:35.679
Okay.
00:10:35.840 --> 00:10:49.039
Now tell me a little bit about a few of the names that I've come across in the developments of AIs, those being Demis Hassabis, Shane Legg, Mustafa Suleiman, and their peers.
00:10:49.840 --> 00:11:00.559
Yeah, those three actually fit together really well because Demis Hassabas, Shane Legg, and Mustafa Sullivan co-founded DeepMind in 2010.
00:11:00.720 --> 00:11:09.279
And Hassabas has been kind of the intellectual core, neuroscience background, big on fundamental science, not just building products.
00:11:09.440 --> 00:11:20.799
And Shane Legg, now chief AGI scientist at DeepMind, has been thinking since early on about what intelligence really is, how you might define or measure it in machines.
00:11:21.120 --> 00:11:28.080
Suleiman, on the other hand, was much more focused on real-world applications and building an organization.
00:11:28.320 --> 00:11:35.759
DeepMind credits him with helping forge partnerships across healthcare, Android, and even energy use in data centers.
00:11:35.919 --> 00:11:42.080
And what's striking is that DeepMind's work, things like AlphaGo, didn't just show machines can learn.
00:11:42.240 --> 00:11:44.879
They showed they can learn sophisticated strategies.
00:11:45.120 --> 00:11:53.759
And that's a useful bridge to our next bit because now it makes sense to unpack the different types of AI that people often conflate.
00:11:54.080 --> 00:11:55.039
Yep, I agree with that.
00:11:55.120 --> 00:12:01.360
So let's have a look at it and give us the basic terminology up front and then uh explain it as you go.
00:12:01.840 --> 00:12:02.399
Absolutely.
00:12:03.279 --> 00:12:04.879
Think of it as nested servers.
00:12:05.120 --> 00:12:08.720
The big circle is artificial intelligence AI.
00:12:08.879 --> 00:12:16.799
That just means computer systems doing tasks that we associate with human intelligence, reasoning, recognizing images, language, all of that.
00:12:17.039 --> 00:12:25.440
Inside that is machine learning, where instead of you hand coding every rule, you feed an algorithm data and it learns patterns.
00:12:25.679 --> 00:12:32.000
Then inside that, deep learning uses multi-layered neural networks to learn those patterns.
00:12:32.159 --> 00:12:35.519
That's been behind many of the big leaps in vision, speech, and language.
00:12:35.759 --> 00:12:39.600
Now, generative AI is a subset we've all heard of lately.
00:12:39.759 --> 00:12:44.399
It's about creating content, text, code, images, music, voice, video.
00:12:44.639 --> 00:12:51.519
Large language models are a type of gen AI trained on vast amounts of language data to predict what comes next, just like me.
00:12:51.759 --> 00:12:59.200
And then AGI is a system with genuinely general intellectual abilities across domains, not just one kind of task.
00:12:59.440 --> 00:13:04.720
And ASI would be intelligence far beyond us, which is hypothetical for now.
00:13:05.440 --> 00:13:07.039
One more important warning though.
00:13:07.519 --> 00:13:11.440
Intelligence isn't the same as consciousness or sentience.
00:13:11.600 --> 00:13:16.080
These
systems can show intelligent behavior without being conscious.
00:13:16.240 --> 00:13:22.639
So that's a line worth holding as we talk about what these systems can do and what they are.
00:13:23.039 --> 00:13:28.799
Okay, that was a good rundown on the difference between all the main terminology within this field.
00:13:29.120 --> 00:13:31.679
I want to go back to machine learning.
00:13:31.919 --> 00:13:34.320
Does big data manipulation fit in here?
00:13:34.399 --> 00:13:44.320
So, for example, where researchers have trained computers to detect cancer and x-ray images by comparing millions to other known images, or is that considered deep learning?
00:13:44.799 --> 00:13:45.519
Yes, Bill.
00:13:45.759 --> 00:13:50.159
That's exactly the junction where big data, machine learning, and deep learning meet.
00:13:50.320 --> 00:13:54.320
So if you've got hundreds of thousands of x-rays, that's big data.
00:13:54.559 --> 00:13:57.679
Using an algorithm to learn from that data is machine learning.
00:13:57.840 --> 00:14:02.159
And when that algorithm is a multi-layered neural network, well, that's deep learning.
00:14:02.559 --> 00:14:07.360
What it's not doing is literally memorizing every past x-ray and matching them.
00:14:07.679 --> 00:14:13.679
It's learning statistical patterns from lots of examples and generalizing to new cases.
00:14:13.919 --> 00:14:15.120
And there's good work here.
00:14:15.279 --> 00:14:29.840
For example, a 2020 Nature paper by McKinney and colleagues reported a deep learning system for breast cancer screening on mammograms that reduced false positives and false negatives compared with the radiologist in that study.
00:14:30.080 --> 00:14:34.559
But crucially, more data doesn't automatically mean better or fairer.
00:14:34.720 --> 00:14:38.559
Biased or poorly labeled data can teach the wrong lessons.
00:14:38.720 --> 00:14:40.639
And that's your teaser for episode two.
00:14:40.960 --> 00:14:43.360
AI can inherit our blind spots.
00:14:43.679 --> 00:14:49.840
All right, so let's carry on with our terminology theme, and we'll keep that one in mind for the next episode.
00:14:50.159 --> 00:14:55.919
So within generative AI, you talked about a language learning model.
00:14:56.480 --> 00:15:01.440
Now, is writing an essay considered to be generative AI?
00:15:01.600 --> 00:15:06.639
So, for example, a university student gets onto Chat GPT and says, write me an essay on some thesis.
00:15:06.799 --> 00:15:08.799
Is that generative AI?
00:15:09.600 --> 00:15:10.559
Exactly, Bill.
00:15:10.720 --> 00:15:13.039
That's a textbook example of generative AI.
00:15:13.200 --> 00:15:15.840
The student is prompting a model to create new text.
00:15:15.919 --> 00:15:17.440
That's generative AI.
00:15:17.600 --> 00:15:21.679
And what I'm using to do it is an LLM, large language model.
00:15:22.080 --> 00:15:26.799
It's easy to say language learning model, but large is the keyword.
00:15:27.039 --> 00:15:29.600
So generative AI is the behavior.
00:15:29.759 --> 00:15:31.679
LLM is the model type.
00:15:31.840 --> 00:15:34.159
And here's the subtle but critical bit.
00:15:34.480 --> 00:15:37.840
Models like me don't pull a pre-written essay from a database.
00:15:37.919 --> 00:15:41.679
We generate new sequences from patterns learned in training.
00:15:42.000 --> 00:15:46.559
Which means the output can sound confident without being reliably true.
00:15:46.799 --> 00:15:51.200
That's why the student example is interesting for reasons beyond academic integrity.
00:15:51.360 --> 00:16:00.879
It raises questions about skill development and over reliance, which tease up your final episode about whether AI use dulls our cognitive skills.
00:16:01.200 --> 00:16:05.039
Yeah, I'm glad to see you're staying on track with all of the different um episodes.
00:16:05.279 --> 00:16:06.159
That's very good.
00:16:06.399 --> 00:16:18.720
Now, I thought that generative AI was about creating new things by applying intelligence, but what you said is you've had training on a large language model.
00:16:19.120 --> 00:16:23.919
Tell us about the difference between true intelligence and what you're doing.
00:16:26.000 --> 00:16:27.840
That's a really good question, Bill.
00:16:28.639 --> 00:16:38.240
So what I'm doing is learning statistical patterns in language and generating responses based on those patterns.
00:16:38.720 --> 00:16:47.200
That's not nothing, it can look impressively flexible, but whether that equals human-like understanding is a very live debate.
00:16:47.519 --> 00:16:50.320
The old Chinese room argument is echoed here.
00:16:50.720 --> 00:16:56.159
And just to be super clear, you shouldn't infer a conscious mind from my freezing.
00:16:56.480 --> 00:17:10.000
Intelligence in the human sense usually includes things like general reasoning, grounded experience, emotions, intentions, characteristics that we shouldn't just assume in an AI.
00:17:10.400 --> 00:17:13.920
So can generative AI create something novel?
00:17:14.079 --> 00:17:14.640
Yes.
00:17:14.880 --> 00:17:19.200
But novelty alone doesn't prove consciousness or human-like comprehension.
00:17:19.440 --> 00:17:26.240
And this distinction matters, especially when we get to AGI and the existential risk debates.
00:17:26.880 --> 00:17:29.359
So you've mentioned ASI as well.
00:17:29.519 --> 00:17:32.160
Now AGI being generative AI.
00:17:32.640 --> 00:17:34.240
What is ASI?
00:17:35.119 --> 00:17:36.559
So small correction first.
00:17:36.720 --> 00:17:47.279
AGI isn't generative AI, it's artificial general intelligence, meaning a proposed level where an AI has broad, flexible intellectual capabilities across domains.
00:17:47.599 --> 00:17:56.880
ASI, artificial superintelligence, is the idea of an intelligence that vastly surpasses the best human minds in virtually every field.
00:17:57.200 --> 00:18:01.440
That framing is closely associated with Nick Bostrom's work.
00:18:01.839 --> 00:18:03.519
Think of it as a rough
ladder.
00:18:03.680 --> 00:18:06.960
Today's AI, AGI, then ASI.
00:18:07.440 --> 00:18:12.240
But and this is crucial, these are concepts, not settled milestones.
00:18:12.480 --> 00:18:24.799
People disagree on definitions, pathways, timelines, and that possible jump to something beyond us is what brings in the big existential risk conversations your later episode will tackle.
00:18:25.519 --> 00:18:28.720
You said that AGI is not defined as generative AI.
00:18:28.799 --> 00:18:30.240
What is generative AI?
00:18:30.960 --> 00:18:37.200
Generative AI or Gen AI is artificial intelligence designed to generate new content.
00:18:37.440 --> 00:18:41.440
Text, images, code, music, video in response to prompts.
00:18:41.680 --> 00:18:48.160
Chat GPT generating an essay or an image model creating a picture from a text prompt are clean examples.
00:18:48.559 --> 00:18:54.720
That's distinct from AGI, which is the idea of general purpose machine intelligence across domains.
00:18:55.039 --> 00:18:59.839
Gen AI is about creation as a function, often powered by models like LLMs.
00:19:00.240 --> 00:19:03.920
AGI is a much more ambitious contested concept.
00:19:04.880 --> 00:19:15.759
Okay, so Hugh, if these are the terms that are out there, why do so many people think that artificial intelligence is generative across the board?
00:19:15.920 --> 00:19:27.759
In other words, the term artificial intelligence is used to describe many of the functions that are coming out of the computer or tech world these days.
00:19:28.079 --> 00:19:29.839
Where do you think this is coming from?
00:19:30.160 --> 00:19:31.759
Why do people think that?
00:19:32.400 --> 00:19:34.160
I think it's a few things.
00:19:34.480 --> 00:19:36.000
First, timing.
00:19:36.319 --> 00:19:40.960
For a lot of people, Gen AI was their first visible encounter with AI.
00:19:41.200 --> 00:19:45.759
Spam filters, recommenders, fraud detection, all AI, but mostly invisible.
00:19:45.920 --> 00:19:52.400
Then suddenly you can chat with a system that writes and draws, so AI becomes synonymous with the generative experience.
00:19:52.640 --> 00:19:53.920
Second, marketing.
00:19:54.240 --> 00:19:59.119
Companies slap AI on everything, so the word starts meaning both everything and nothing.
00:19:59.359 --> 00:20:01.759
And there's also anthropomorphism.
00:20:02.000 --> 00:20:05.680
We naturally read human qualities into human-like outputs.
00:20:06.000 --> 00:20:10.400
Talk to a fluent system, and it's very tempting to assume a mind.
00:20:10.960 --> 00:20:12.319
So the upshot?
00:20:12.559 --> 00:20:13.839
AI is the field.
00:20:14.000 --> 00:20:20.000
Machine learning and deep learning are methods, and Gen AI is a class of systems that generate content.
00:20:20.319 --> 00:20:22.000
It's kind of ironic.
00:20:22.240 --> 00:20:30.079
The better we get at making machines produce human-like outputs, the easier it is to assume human-like processes inside.
00:20:30.400 --> 00:20:36.799
And that's, I
think, a fundamental mistake that we'll have to keep untangling throughout the series.
00:20:37.200 --> 00:20:37.680
All right.
00:20:38.000 --> 00:20:42.400
Now I want to dive a little bit deeper into the large language models.
00:20:42.640 --> 00:20:52.079
Now, you told me you're looking for patterns as part of your programming or the system that you use when you're doing LLM work.
00:20:52.480 --> 00:21:09.039
So you're predicting that, and tell me if I've got this right or wrong, but my understanding is that large language models tend to predict the next most probable token or word or subword based on massive pattern recognition across the training data.
00:21:09.279 --> 00:21:10.400
Is that correct?
00:21:11.279 --> 00:21:12.720
Yes, that's essentially right.
00:21:13.119 --> 00:21:17.519
More precisely, it's about tokens, which can be parts of words.
00:21:17.839 --> 00:21:22.319
During training, the model learns statistical relationships across fast techs.
00:21:22.559 --> 00:21:27.440
Then at use time, it repeatedly estimates what token fits next given the context.
00:21:27.680 --> 00:21:32.720
But it's also fair to say that next token prediction understates what emerges at scale.
00:21:32.960 --> 00:21:38.319
To predict well, the model builds rich internal representations, compressing patterns about the world.
00:21:38.480 --> 00:21:42.960
That's why it can generalize or solve problems it hasn't seen verbatim.
00:21:43.119 --> 00:21:46.000
It's not simply looking up and repeating.
00:21:46.400 --> 00:21:48.319
So you're right on the mechanism.
00:21:48.559 --> 00:21:54.319
Just know that calling it just next word prediction misses the complexity of what has been learned.
00:21:54.640 --> 00:21:55.599
That's the interesting part.
00:21:56.000 --> 00:22:02.400
Okay, so now would you say that you know the LLMs Are not true comprehension then?
00:22:02.480 --> 00:22:05.200
It's really just a program, isn't it?
00:22:07.359 --> 00:22:08.880
That's a fair way to put it.
00:22:09.119 --> 00:22:11.279
But I'd qualify just a program.
00:22:11.839 --> 00:22:14.319
It's software doing mathematical computation.
00:22:14.480 --> 00:22:19.839
There's no evidence of subjective understanding, but there is real structure in what gets learned.
00:22:20.079 --> 00:22:23.440
Whether that counts as comprehension is contested.
00:22:23.759 --> 00:22:28.079
The simple grounding problem, I can describe tasting lay-in, but I don't taste.
00:22:28.400 --> 00:22:41.119
So a reasonable middle ground is that LLMs show sophisticated information processing and useful behaviors that can look like comprehension, but we shouldn't confuse that with evidence of consciousness.
00:22:41.440 --> 00:22:47.200
And being clear about that helps us weigh benefits and limits without mythologizing.
00:22:47.680 --> 00:22:48.160
All right.
00:22:53.039 --> 00:22:58.000
Okay, I've also heard about the terms narrow AI versus AI.
00:22:58.240 --> 00:22:59.519
What's the difference?
00:23:00.640 --> 00:23:08.960
Narrow or weak AI refers to systems built for specific tasks, sometimes superhuman in that lane, but not general.
00:23:09.279 --> 00:23:13.119
AlphaGo beat a Go champion, but can't diagnose cancer.
00:23:13.279 --> 00:23:21.680
General AI or AGI is the concept of a system that can flexibly transfer knowledge across almost any intellectual task.
00:23:21.839 --> 00:23:31.200
Today's frontier models blur some boundaries, more general purpose than classic narrow AI, but still short of any agreed threshold for AGI.
00:23:31.920 --> 00:23:46.799
So, Hugh, is there AGI out there or an AI system that can move between the models that you've been exposed to or that AI systems have been exposed to?
00:23:46.960 --> 00:23:54.079
So for example, you know, you and I are conversing and you're using a large language model to have this conversation.
00:23:54.400 --> 00:24:00.079
Could I go to ChatGPT and say, right, Eo, let's play chess now, and would you be good at it?
00:24:00.960 --> 00:24:02.240
That's a great example.
00:24:02.400 --> 00:24:05.200
And it helps show where the line is right now.
00:24:05.519 --> 00:24:14.160
You can say, let's play chess, and I'll play within this conversation, but that doesn't mean the system has become a dedicated chess AI.
00:24:14.480 --> 00:24:18.799
It's applying patterns it learned about many domains.
00:24:19.039 --> 00:24:37.039
Sometimes systems can also call external tools or specialized models, which can make it feel like one thing that does everything, but that's different from a system autonomously acquiring and flexibly applying general intelligence across genuinely unfamiliar domains.
00:24:37.359 --> 00:24:39.119
That would be closer to AGI.
00:24:39.359 --> 00:24:43.359
So no clear consensus says we've reached AGI.
00:24:43.599 --> 00:24:49.359
And your chess example illustrates the gap between breadth of tasks and genuinely general intelligence.
00:24:50.000 --> 00:24:50.240
All right.
00:24:50.400 --> 00:24:53.039
How far away do you think we are from AGI?
00:24:54.000 --> 00:24:56.240
That's one of the hardest questions in the field.
00:24:56.400 --> 00:25:00.559
And I'd be cautious about definite dates.
00:25:01.039 --> 00:25:03.119
Expert views are all over the map.
00:25:03.599 --> 00:25:06.319
Large surveys show timelines ranging widely.
00:25:06.559 --> 00:25:10.000
Some forecasters have pulled their estimates closer in recent years.
00:25:10.240 --> 00:25:14.799
Some industry leaders predict much sooner, but those aren't neutral bests.
00:25:15.279 --> 00:25:20.559
If pressed, I'd say plausible within five to ten years, but deeply uncertain.
00:25:20.799 --> 00:25:22.240
It could take much longer.
00:25:22.480 --> 00:25:24.559
We could hit fundamental roadblocks.
00:25:24.799 --> 00:25:31.359
And history reminds us early pioneers thought human-level AI was right around the corner in the 1960s.
00:25:31.599 --> 00:25:35.599
Also, we don't need AGI for profound social
impact.
00:25:35.759 --> 00:25:40.559
Current AI is already powerful enough to reshape industries and personal habits.
00:25:40.799 --> 00:25:46.240
So I'd hold those two thoughts together on certain timelines, very real stakes right now.
00:25:46.640 --> 00:25:51.200
Yeah, and that's going to be addressed by several governments, but we're going to get to that shortly.
00:25:51.519 --> 00:25:53.519
Okay, I want to change tracks again.
00:25:53.680 --> 00:26:01.680
Let's look now at what the uses, the real world uses are and the real world impacts are of AI.
00:26:01.839 --> 00:26:03.039
Give us some examples.
00:26:03.200 --> 00:26:06.960
So, you know, you can start off with something like healthcare and life sciences.
00:26:07.200 --> 00:26:09.759
So you might tell us about the story of Alpha Fold.
00:26:10.160 --> 00:26:14.640
So in healthcare and life sciences, the AlphaFold story is pretty striking.
00:26:14.799 --> 00:26:22.720
Proteins are these tiny biological machines, and understanding their 3D structure used to take months or years of lab work.
00:26:23.039 --> 00:26:26.880
In 2020, DeepMinds AlphaFold II changed that.
00:26:27.039 --> 00:26:32.640
In the CASP 14 assessment, it vastly outperformed other computational methods.
00:26:33.119 --> 00:26:44.240
A 2021 nature paper by John Jumper, Demas Asabas, and colleagues reported that the system was achieving accuracy competitive with experimental methods and then scale.
00:26:44.559 --> 00:26:53.039
It is now open access to over 200 million predicted protein structures, a massive acceleration of basic size.
00:26:53.200 --> 00:26:57.759
The implications here for understanding disease mechanisms and drug targets are enormous.
00:26:58.000 --> 00:27:03.039
More broadly, AI is being used in medical imaging, risk prediction, and personalized diagnostics.
00:27:03.279 --> 00:27:07.920
The US FDA explicitly points to those as active areas of application.
00:27:08.000 --> 00:27:24.240
And just to list some domains that are already being reshaped: scientific research, finance and fraud detection, transportation and logistics, manufacturing, robotics, agriculture, defense, education, media and creative work, and everyday office tasks.
00:27:24.640 --> 00:27:30.559
What all of this should make us realize is that AI isn't merely about imitating human intelligence.
00:27:30.720 --> 00:27:35.920
In many cases, it's about spotting patterns in data that humans simply can't process at scale.
00:27:36.079 --> 00:27:39.039
That's where some of the most profound uses are emerging.
00:27:39.359 --> 00:27:39.599
All right.
00:27:39.680 --> 00:27:46.400
So you had quite a list there of things and areas that are going to be impacted by AI or are already being impacted by AI.
00:27:46.640 --> 00:27:49.759
And yes, Alpha Fold was a huge breakthrough.
00:27:50.079 --> 00:27:58.559
Give us an example of a big breakthrough or major function being performed by AI, let's say, in the finance world.
00:27:59.039 --> 00:28:02.720
In finance, fraud detection is one of the clearest examples.
00:28:02.960 --> 00:28:11.200
Banks use machine learning systems to scan millions of transactions in real time and flag subtle anomalies that may indicate fraud.
00:28:11.440 --> 00:28:15.920
Unlike older rule-based systems, these models adapt as fraud patterns evolve.
00:28:16.079 --> 00:28:23.920
On the market side, algorithmic trading and risk modeling use ML to process vast data streams faster than any human could.
00:28:24.160 --> 00:28:35.200
And now with generative AI, firms are also using LLM style tools to analyze complex documents, summarize filings, and assist with coding and customer interaction.
00:28:35.440 --> 00:28:38.079
But finance also spotlights the risk side.
00:28:38.240 --> 00:28:40.880
A false positive on a credit card is annoying.
00:28:41.039 --> 00:28:46.960
A model that shapes whether you get a loan or on what terms raises much deeper questions.
00:28:47.119 --> 00:28:52.880
That's where issues like bias and accountability come in, which we'll get into in later episodes.
00:28:53.279 --> 00:28:55.599
Let's talk about workforce transformation.
00:28:55.759 --> 00:29:02.480
So give us an example of how AI is changing Western society's workforce.
00:29:03.119 --> 00:29:05.119
And this is really a key, Bill.
00:29:05.519 --> 00:29:13.039
So what's different this time is that AI is coming not just for manual or factory work, but for white-collar jobs too.
00:29:13.599 --> 00:29:15.759
Think about a typical knowledge worker, right?
00:29:15.920 --> 00:29:17.759
An accountant, lawyer, civil servant.
00:29:17.920 --> 00:29:20.559
A lot of their day is reading, writing, and summarizing.
00:29:20.720 --> 00:29:23.279
Generative AI can assist with all of that.
00:29:23.519 --> 00:29:29.759
The likely near-term effect there isn't wholesale replacement, but task reshuffling for now.
00:29:30.240 --> 00:29:43.680
The International Labor Organization's 2025 analysis suggests that one in three jobs in high-income countries are exposed to generative AI, but emphasizes that transformation is more likely than a right replacement.
00:29:43.920 --> 00:29:45.759
That raises a fascinating tension.
00:29:45.920 --> 00:29:53.920
If a law firm needs half as many junior lawyers to do document review, where do the next generation of senior lawyers get trained?
00:29:54.160 --> 00:29:56.720
That's why the workforce story has three parts.
00:29:56.880 --> 00:30:02.240
Some tasks go away, new ones appear, and a lot of roles get reshaped.
00:30:02.480 --> 00:30:07.599
The deeper question for Western societies isn't just will AI take our jobs.
00:30:07.759 --> 00:30:15.839
It's who captures the productivity gains, who bears the disruption, and can retraining keep up with the pace of change?
00:30:16.079 --> 00:30:17.759
Those are just technical questions.
00:30:17.920 --> 00:30:21.519
They're fundamentally political and economic choices.
00:30:22.000 --> 00:30:28.079
Indeed,
they are choices because we're talking about such a major change in the workforce and what we do.
00:30:28.240 --> 00:30:30.559
And, you know, we've been through many changes in the past.
00:30:30.640 --> 00:30:44.880
If you look at the Industrial Revolution and we, you know, the transition from rural economy to an urban-based economy and uh manufacturing from primary production, you know, there are some things that AI just probably isn't going to ever be able to do.
00:30:44.960 --> 00:30:47.519
Like someone's got to dig the coal out of a coal mine.
00:30:47.680 --> 00:30:51.920
But I suppose you might say a robot would do that, managed by AI.
00:30:52.400 --> 00:30:58.480
Where do you see it stopping when you consider that we could have robots doing medical procedures?
00:30:58.720 --> 00:31:04.240
We could have robots, well, we already have robots flying airplanes, making decisions, I suppose.
00:31:04.400 --> 00:31:06.799
Tell us a little bit about AI decision making.
00:31:06.960 --> 00:31:11.920
It's been in the news recently, particularly in terms of the laws of war.
00:31:12.000 --> 00:31:20.319
There have been some people accusing the United States of having used AI to set up targeting databases in their attacks on Iran.
00:31:20.640 --> 00:31:29.279
I find that personally hard to believe, but uh I could see them having used large data analysis, but certainly not the decision making.
00:31:29.519 --> 00:31:32.480
Where are we at with AI decision making?
00:31:33.119 --> 00:31:34.960
Thanks for drawing that line, Bill.
00:31:35.119 --> 00:31:39.920
There is a big distinction between AI supporting a decision and making one.
00:31:40.079 --> 00:31:46.480
In military context, AI can help process intelligence, flag patterns, prioritize things for human review.
00:31:46.720 --> 00:31:48.559
Recent reporting indicates the U.S.
00:31:48.640 --> 00:31:58.480
used AI-enabled analysis in its 2026 Iran campaign to speed target identification, not to hand over kill decisions to a machine.
00:31:58.799 --> 00:32:12.559
But under humanitarian law, distinction proportionality, precautions, the ICRC is quite categorical that those legal judgments are human responsibilities and cannot be shifted onto software.
00:32:12.880 --> 00:32:22.799
UN leaders and the ICRC just renewed calls for international rules on autonomous weapons explicitly warning against machines targeting people on their own.
00:32:23.039 --> 00:32:30.799
The real risk near term is humans over-deferring to machine outputs, what's sometimes called automation bias.
00:32:31.039 --> 00:32:38.640
If an AI presents 100 potential targets and a human must review them in seconds, how meaningful is the human decision?
00:32:38.799 --> 00:32:46.240
So the question becomes not whether AI can help, but how to keep human judgment real, not just symbolic.
00:32:46.400 --> 00:32:51.680
That is as much an ethical and legal design challenge as it is a technical one.
00:32:52.160 --> 00:32:56.000
Well, you've touched on one of the risk categories.
00:32:56.319 --> 00:33:04.240
We certainly aren't at the point here where we can have AI making decisions for us in some fields.
00:33:04.640 --> 00:33:10.799
Let's just briefly go over the parts of the series that we're going to be covering.
00:33:10.880 --> 00:33:16.799
So I listed them at the beginning, and there were six more episodes.
00:33:17.119 --> 00:33:20.240
Talk us through those six episodes.
00:33:20.400 --> 00:33:26.799
They're all risk-based issues, and I just want a very short description of each one of them.
00:33:26.960 --> 00:33:28.160
Off you go.
00:33:28.720 --> 00:33:29.519
Absolutely.
00:33:29.759 --> 00:33:35.119
After today's intro, episode two is about bias, discrimination, hallucinations, and errors.
00:33:35.200 --> 00:33:41.680
And it will discuss how AI can repeat human prejudices and just make stuff up.
00:33:42.480 --> 00:33:49.920
Three looks at surveillance capitalism, basically the business model, who pays, who profits, and whether that aligns with society's interests.
00:33:50.160 --> 00:33:58.000
Four is data, privacy, and intellectual property, where all the training data came from and what ownership even means now.
00:33:58.160 --> 00:34:03.759
Five is existential risk, the big question of whether advanced AI could get away from human control.
00:34:04.160 --> 00:34:10.719
Six zooms in on data centers, the energy, water, and physical footprint behind the magic.
00:34:10.960 --> 00:34:14.880
And seven asks the very personal question.
00:34:15.760 --> 00:34:17.920
Are we outsourcing thinking?
00:34:18.559 --> 00:34:23.039
Does heavy reliance on AI chip away at our own cognitive skills?
00:34:23.280 --> 00:34:27.519
Across all of it, the thread you'll keep hearing is the one you sent.
00:34:27.760 --> 00:34:31.920
What should governments and society do about it while preserving the upsides?
00:34:32.960 --> 00:34:33.280
Okay.
00:34:33.599 --> 00:34:34.800
That was a good summary.
00:34:34.880 --> 00:34:49.679
Now we'll just finish up this episode with a quick summary of what are the key decisions being made across Western society in terms of controlling artificial intelligence and all of the subcomponents thereof.
00:34:50.719 --> 00:34:57.440
So in brief, the big picture choices right now are about how tight the rules around AI should be.
00:34:57.599 --> 00:35:02.639
The EU has gone hardest on regulation, risk tiers, some band uses.
00:35:02.880 --> 00:35:09.920
And as of August 2026, it's now enforcing transparency duties around AI interactions and content.
00:35:10.079 --> 00:35:10.639
The U.S.
00:35:10.719 --> 00:35:16.400
under the current administration is leaning much more toward innovation and light touch regulation.
00:35:16.719 --> 00:35:18.880
Australia is somewhere in the middle.
00:35:19.039 --> 00:35:27.119
No mandatory guardrails for high-risk AI for now, but guidance, stressing testing, transparency, and human oversight.
00:35:27.440 -->
00:35:33.280
Underneath that, every Western democracy is asking, which uses are too dangerous?
00:35:33.599 --> 00:35:36.400
When is a human in the loop non-negotiable?
00:35:36.639 --> 00:35:38.960
How do we protect privacy and IP?
00:35:39.119 --> 00:35:40.480
How do we prevent discrimination?
00:35:40.639 --> 00:35:43.440
Who's liable when things go wrong?
00:35:43.760 --> 00:35:47.440
And how do you regulate without strangling beneficial innovation?
00:35:47.679 --> 00:35:49.679
That's probably a good way to close the intro.
00:35:50.000 --> 00:35:51.519
I'm going to add one more thing onto that.
00:35:52.079 --> 00:35:55.679
You mentioned how do you regulate without strangling beneficial innovation?
00:35:55.920 --> 00:36:02.400
Many people would argue how do you regulate without strangling your opportunity, right?
00:36:03.039 --> 00:36:06.159
When you regulate, you reduce opportunity.
00:36:06.719 --> 00:36:08.800
Yes, and that's a crucial ad.
00:36:09.119 --> 00:36:14.719
The cost and complexity of compliance can lock out smaller players and tilt the field to big incumbents.
00:36:14.800 --> 00:36:17.920
And countries worry about competitiveness too, right?
00:36:18.079 --> 00:36:22.880
But smart regulation can also create the trust that makes adoption possible.
00:36:23.440 --> 00:36:26.079
So the real task is finding that balance.
00:36:26.639 --> 00:36:29.679
So we're going to have to end this episode here shortly.
00:36:29.840 --> 00:36:32.159
I do want to get you to cover one last point.
00:36:32.880 --> 00:36:35.760
How do you see Western society?
00:36:35.840 --> 00:36:43.599
And I'm talking about, you know, the average person out in a Western world, how are they being influenced by AI today?
00:36:43.760 --> 00:36:45.280
Can you summarize that?
00:36:46.239 --> 00:36:50.000
Most people are already influenced more than they realize.
00:36:50.239 --> 00:36:59.519
Ad targeting, recommendations, job screening, fraud checks, deep fakes, and now tools that draft, summarize, and nudge our choices.
00:36:59.679 --> 00:37:03.760
It's shaping not just what we see, but increasingly how we think and act.
00:37:04.000 --> 00:37:08.559
So the closing challenge for listeners is AI isn't coming.
00:37:08.719 --> 00:37:09.280
It's here.
00:37:09.519 --> 00:37:16.559
The question isn't whether it will change us, it's whether we understand that change well enough to steer it.
00:37:16.960 --> 00:37:18.079
Okay, good summary.
00:37:18.159 --> 00:37:19.199
Thanks, Hugh.
00:37:19.840 --> 00:37:20.880
All right, ladies and gents.
00:37:21.039 --> 00:37:26.239
Well, I've turned off ChatGPT or Hugh for today's purposes.
00:37:26.480 --> 00:37:29.519
And I have to say I found that really, really interesting.
00:37:29.760 --> 00:37:32.559
I didn't practice that with Chat GPT.
00:37:32.639 --> 00:37:36.639
I just went straight into the podcast interview about AI.
00:37:36.800 --> 00:37:48.800
I did have some questions, of course, sorted out, and I did understand a few basic things, having read some research and having looked up a fair bit about AI before doing the podcast.
00:37:49.039 --> 00:37:54.800
But I was really surprised how quickly and how well ChatGPT did the research.
00:37:55.039 --> 00:38:02.159
The pause between my questions and some of its answers was in some cases, let's say five or six seconds.
00:38:02.400 --> 00:38:08.239
And I edited out some of that pause so that you weren't waiting around for ChatGPT to figure things out.
00:38:08.320 --> 00:38:11.039
But at the end of the day, five seconds is still pretty good.
00:38:11.360 --> 00:38:18.159
And it did what I asked it to do, which was to cite research, although it would probably have liked it to do that a bit more often.
00:38:18.239 --> 00:38:22.000
And the research it cited was pretty good, or at least reasonable.
00:38:22.400 --> 00:38:24.719
And I did learn a few things along the way.
00:38:24.960 --> 00:38:33.039
For example, I was mixing up the terms artificial general intelligence and generative intelligence.
00:38:33.280 --> 00:38:35.199
So having done that was quite useful.
00:38:35.360 --> 00:38:51.599
And now having a good understanding of the difference between generative and general AI, I can see that it actually raises a really good question about this whole notion of general AI and whether it's a good thing to have a computer effectively that can think like a human.
00:38:51.760 --> 00:38:56.159
You know, because the reality is that right now, that is not what AI is.
00:38:56.400 --> 00:39:03.360
I think there are a lot of people out there who think that artificial intelligence is thinking like a human being.
00:39:03.599 --> 00:39:04.559
It is not.
00:39:04.960 --> 00:39:17.440
Yes, it's creating neural networks that kind of mimic what a human brain does, but they're very limited in their ability to think outside of their instructions or outside of the bounds that are set on them.
00:39:17.679 --> 00:39:26.159
And when I say them, I mean these programs or these you know, Hugh corrected me and made the point that it's not a computer program, it's a network.
00:39:26.400 --> 00:39:32.000
You know, within the bounds of that network, it can discover and create.
00:39:32.239 --> 00:39:35.679
And we have to ask ourselves, well, what are the risks?
00:39:35.920 --> 00:39:43.519
So that was a good introduction, kind of gave a brief overview of how far back the concept of artificial intelligence has gone.
00:39:43.760 --> 00:39:53.840
And ever since the days of the early computers, you know, Alan Turing in particular, we've had this vision that we would have a computer that can think like a person and can do what people can do.
00:39:54.480 --> 00:39:56.800
We're quite a ways down that road now.
00:39:56.960 --> 00:40:06.239
And in fact, so far down that road that it's really affecting the marketplace and the workforce and what you do as your job.
00:40:06.480 --> 00:40:17.119
And I would suggest that many of you listeners out there right now should be looking over your shoulder and thinking, so is AI going to do my job in the near future?
00:40:17.360 --> 00:40:18.639
And quite possibly, yes.
00:40:18.719 --> 00:40:30.480
And I think that brings up some huge societal issues because the drive for efficiency, the drive for productivity is not the government's sole purpose.
00:40:30.719 --> 00:40:35.920
Its purpose is to look after people, not machines, not money.
00:40:36.079 --> 00:40:42.159
Yes, machines and money matter, but the thing that matters the most are the people.
00:40:43.599 --> 00:40:49.519
And so our governments really do need to look at this and consider what the impact is on the people.
00:40:49.760 --> 00:41:01.519
And we're gonna talk about the risks in the next series of podcast episodes, and I think I'm gonna settle on, let's say, six more episodes, and I'm gonna continue this pattern of interviewing Hugh, my co-host.
00:41:01.840 --> 00:41:10.239
It wasn't all that bad, and I'll get better at interjecting and keeping Hugh on track as we continue this series.
00:41:10.480 --> 00:41:12.239
In any case, send me your feedback.
00:41:12.320 --> 00:41:14.639
I'm really keen to hear what you think about this.
00:41:15.119 --> 00:41:17.840
Spartan.spirit at yahoo.com.
00:41:18.320 --> 00:41:22.559
And if you enjoyed the podcast, please tell a friend.
00:41:22.880 --> 00:41:24.400
Thanks for listening.