August 21, 2024 • 18 mins
AI is everywhere. And it does everything...or does it?
In this episode we break down the different types of AI, including generative AI like ChatGPT. We'll explain how each AI type operates uniquely, with a focus on their current roles and limitations. And boy, do they have limitations.
Corey provides a clear breakdown of AI, explaining the distinction between weak AI, which performs specific tasks, and the elusive strong AI seen in science fiction. She explores how AI is currently used in healthcare, from interpreting medical imaging to handling large data sets, and how it might influence massage therapy through the development of "rubbing robots."
This episode offers valuable insights into the intersection of technology and massage therapy, balancing excitement with caution. Whether you're a therapist or just curious about the future of AI in health and wellness, this episode will give you a thoughtful perspective on what's ahead.
The definition of massage therapy (Kennedy 2016)
ChatGPT Just Solved Chess (Video)
Google apologises for Photos app's racist blunder (BBC)
Man is to Programmer as Woman is to Homemaker: Bias in Machine Learning
Artificial Intelligence and Its Role in Massage Therapy with Whitney Lowe (ABMP Podcast)
Artificial Intelligence: A Guide for Thinking Humans by Melanie Mitchell
The Alignment Problem: Machine Learning and Human Values by Brian Christian
Send us an email: podcast@healwell.org
Leave us a voice message: 703-468-1799
Check out our interview-style podcast: Interdisciplinary
You can support Healwell and the cool things we make by donating here!
Ways join in:
- Leave us a review on Apple Podcasts
- Check Healwell’s live and online classes
- Continue the conversation with a free 3-day trial of the Healwell Community
- Find a copy of Rebecca Sturgeon's book: "Oncology Massage: An Integrative Approach to Cancer Care"
Thank you to ABMP and AMTA for sponsoring us!
Healwell is a 501(c)(3) non-profit based out of the Washington DC area. Check us out at www.healwell.org
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:07):
Welcome to the Rub a HeelWild podcast about massage
therapy.
I'm your host, kori Rivera,licensed massage therapist and
information magpie, and intoday's episode we're going to
talk about artificialintelligence and rubbing robots.
As a quick reminder, thispodcast uses the word rubbing to
(00:34):
talk about massage activitiesthat do not qualify as massage
therapy.
The academic definition ofmassage therapy is massage
therapy is for both self-careand health maintenance.
Massage therapy includesrubbing, but it also involves
health promotion and educationmessages.
The.
The setting of the treatment,such as a hospital or a spa, may
influence the results, but theresults absolutely depend on the
(00:54):
skill, training and experienceof the therapist and on the
therapeutic relationships andcommunication between therapist
and client.
Okay, let's talk robots.
Recently I read a current modelof how we think dopamine works.
Didn't think I was going toopen an episode about robots
talking about a hormone, did you?
(01:15):
Dopamine has been called afeel-good hormone.
We thought for a long time thatdopamine was a reward.
Do something that gives you anevolutionary advantage, like
eating a Snickers bar.
It's nutrient-dense?
Okay, eat the Snickers, get ahit of dopamine.
But in 1993, a study came outof Switzerland authored by Dr
Wolfram Schultz.
The study found that whenmonkeys were trained to open a
(01:38):
box with food in it, theygenerated dopamine.
However, the rush of dopaminefrom opening the box diminished
over time.
Eventually, there was nodopamine generated when the
monkey found food.
The study also discovered thatif the monkey found food
randomly, instead of every timeit opened the box, the dopamine
came back, which led to thequestion if it's not the food,
(02:01):
then what caused the reaction?
Currently, we think thatdopamine is what happens when
you anticipate that the futurewill be better than the present.
I would call it hope.
Dopamine is what happens whenyou have hope.
Dopamine is the pull of theslot machine, but not the coins
that come out.
Or hitting the button to lookat social media, but not what's
(02:21):
actually in your feed.
Or waiting for that adorablekitten on your screen to do
something adorable, which ittotally will, because why else
would its person have posted thevideo?
There is a 100% chance ofincoming adorableness and for
some of us, dopamine is whatmakes us excited about
artificial intelligence andrubbing robots.
Because I am excited.
(02:43):
Innovation makes me excited.
Novel uses of technology makeme excited.
Accessibility makes me excited.
I'm thrilled to be able to tellyou about both AI and the
rubbing robots that are beingdeveloped.
However, all that dopamine alsosparks a big old warning sign
in my brain, because myexcitement about the future also
makes me vulnerable tounexpected consequences and hype
(03:06):
cycles, and this is why theseepisodes are taking so long.
My thoughts about artificialintelligence and rubbing robots
swing like a pendulum with everynew piece of information, and
there is a lot of informationcoming through right now.
In this particular episode, I'mgoing to give you a rundown of
AI in general and rubbing robotAI in particular.
(03:26):
We're going to talk about whatAI really is and what it really
does.
If you haven't already listenedto the previous episode,
massage Tools, I recommend it.
Unless you live in completeisolation in the mountains or
the desert or something you'vedefinitely noticed the explosion
of artificial intelligenceeverything desert.
Or something you've definitelynoticed the explosion of
(03:47):
artificial intelligenceeverything.
Ai will write your paper, drawyou a picture, compose music and
soften your irritated tone inthat email conversation you've
been having for three daysstraight with no conclusion in
sight.
That last one might be a mething, but in the endless
headlines about AI is it useful,is it dangerous?
Is it cheating?
Most people haven't seemedconcerned about what.
First, let's talk about whatit's not.
(04:12):
The name artificial intelligencewas coined at the first ever
conference about it in 1956.
The people who work on itsdevelopment admit that it is a
terrible name.
There are a bunch of highlydebated ideas about what
intelligence in human is, letalone a machine.
Is intelligence knowledge,abstract thinking, problem
solving?
What about emotionalintelligence?
(04:33):
Does the situation matter?
How do you measure intelligence?
What makes a human smart?
And we all know smart peoplewho are really dumb.
How can someone be smart anddumb at the same time?
So calling a computerintelligent is confusing off the
bat.
Additionally, there are twoavenues of research in AI.
One group wants to createsomething that can think in big
(04:54):
bunny quotes, and one wants tomake programs that do things
better or faster than humans.
These tasks are related but arevery different if you're the
person trying to accomplish themand spectacularly different if
you're the person trying towrite a grant that funds them.
So what we do not have iscalled general artificial
intelligence or strong AI.
(05:15):
This is the kind of AI thatmovies are interested in.
This is Hal from 2001, a SpaceOdyssey, sunny from iRobot or
Samantha from Her.
It's a computer thatunderstands, can think about
ethics and morality andpotentially decide to destroy
humanity because we make somepretty terrible decisions.
We have not made that.
Depending on who you ask, weare close decades away or it is
(05:36):
completely impossible.
What we do have is weak AI.
This is AI that can do specifictasks and, lately, can do them
really well.
You've been hearing aboutalgorithms for a while.
Algorithms are like toddler AI.
You've probably heard thingslike.
The YouTube algorithm is ahorrible thing that fans the
flames of the fires of bias anddivision in order to keep you
(05:59):
watching longer and make moneyfrom ads.
And if you hadn't heard that,maybe you've heard.
The Netflix algorithm cansuggest movies you might like.
Both are true, by the way.
An algorithm is a strict set ofinstructions.
It's programmed veryspecifically and spits out an
answer that it was designed tofigure out.
The YouTube algorithm isprogrammed to figure out what
keeps you watching and then feedthat to you and your
(06:21):
suggestions.
Artificial intelligence is madeup of a group of algorithms and
it can reprogram thosealgorithms to meet a goal.
Instead of being fed strictinstructions, ai is fed data.
The part humans programmed ishow AI deals with the data.
This means that AI alters itsown algorithms using a simple
technique like trial and errorto complete the task it was
(06:42):
given.
Deep learning is what happenswhen AI is given data and a goal
but no instructions.
Deep learning is kind of likeLittle Red Riding Hood here's a
basket, get to grandma's house,godspeed.
So and this is important all ofthe weak AI we have are
learning to complete their tasksin a completely different way.
(07:02):
There are a lot of methods, butwe're going to talk about three
.
We're going to talk about AIthat play games, ai that
recognize images and generativeAI like ChatGPT.
If someone asked you who comesto mind when you think about
chess, garry Kasparov might popinto your head.
Garry Kasparov, aside frombeing the world chess champion
(07:23):
for 15 years, was the firstgrand chess master to be beaten
by an IBM computer called DeepBlue in 1997.
You might be less likely toremember Fan Hui, a world
champion of the game Go, who wasbeaten by the Google computer
AlphaGo in 2015.
If you want to come to theHeelwell community, I will
happily explain how thecomputers won, but for this
(07:43):
episode, I just want you to knowthat they didn't win.
Using the same method, go hasexponentially more possible
moves than chess, but the botthat won at Go wasn't a more
advanced version of the chessbot.
It was a completely differentlearning system.
The other thing you should knowis that both Go and chess are
games of perfect information.
That means that both playersknow what is happening all of
(08:04):
the time.
In games like hide and seek,you don't know where people are.
In chess and Go, you knowexactly where your opponent is
and how many pieces they haveand which kinds.
Chess and go, you know exactlywhere your opponent is and how
many pieces they have and whichkinds.
So Deep Blue and AlphaGo aretotally different programs doing
what looks like the same taskbeating a human at a game of
perfect information but one ofthem got there using a submarine
(08:26):
and one of them got there usinga hot air balloon.
Now let's talk about imagerecognition.
Identifying images is way morecomplicated than playing a game.
Interestingly, these programswork a lot like human eyes.
They were trained to recognizeedges of things and then simple
shapes, and then complex shapesand then whole objects and faces
.
This is the same order oureyeballs use to understand
(08:48):
visual information.
The training used to recognizevisual images is a lot like
playing Marco Polo.
The training used to recognizevisual images is a lot like
playing Marco Polo.
Marco Polo.
The system makes a guess, getsfeedback, changes the system a
little bit, makes another guess.
This loop of guessing andcorrecting eventually allows the
computer to understand theimage.
That is the super simplifiedversion and we are going to stop
(09:08):
there.
But the GameBots and theImageBots Not the same.
Okay, chatgpt ChatGPT was aninterface released by the
company OpenAI in November of2023.
What made ChatGPT unique was itsability to reply to a human and
sound like a human when it did.
You might remember Watson, theIBM computer that won at
(09:29):
Jeopardy in 2011.
Now we come to Watson, who isBrad Stoker.
While it was very impressive,recalling facts and using a
specific way to answer is apretty narrow task compared to a
completely open response.
Chatgpt is what is known as alarge language model, or LLM.
It also fits into the categoryof generative AI.
This means that it isn't simplyreporting information or
(09:50):
meeting a goal.
It is producing a unique answerevery time you ask it something
.
Large language models work byassociation.
It learns what words arerelated to other words, which is
truly impressive when you thinkabout it.
The word run has over 600definitions, including idioms.
You can run cross-country andalso have a run in your socks.
You can run a company or run areport, but the thing is ChatGPT
(10:15):
doesn't understand anything.
It associates, and it does itincredibly well.
And while you could argue thathumans are also association
machines, humans have a greatmany more things involved,
including, but not limited to,life experience, emotion, a body
and social pressure.
Writing is more than wordsstrung together, and words are
(10:35):
more than sounds strung together.
However, for a machine thatstrings things together without
understanding what it's saying,it does a great job making me
sound less irritated.
It saves me from having to combmy writing for SEO keywords and
it can write humorous things ifyou ask it, but it can't play
chess.
In 2023, a user in the AnarchyChess subreddit pitted ChatGPT
(10:55):
against the current best chessbot, which is called Stockfish,
in a hilarious game.
Chatgpt proceeded to change thecolor of the pieces, move rooks
diagonally and ultimately losebecause it put itself into check
.
This is because ChatGPT is aword association AI.
It found patterns in writtenchess games and then imitated
that.
It never knew the rules ofchess and if you tried to
(11:18):
explain them, it wouldn'tunderstand.
Chatgpt cannot play chess andStockfish cannot write a
sentence.
Generative AI like ChatGPT takesdata and a model and generates
a solution.
Did I say generative AI takesdata?
It takes data, a lot of data.
So it takes data, a lot of data.
(11:42):
So so much data An estimated10,000 data points on the low
end for each piece of a puzzleyou're trying to analyze.
So even if there were agenerative AI model that could
create massage therapy protocolfrom an intake form, there
wouldn't be enough data for itto work.
As in all things data-related,more data means greater accuracy
, unless your data is inaccurate.
You've probably heard some ofthe stories about bad datasets.
(12:04):
In 2015, google's fancy newauto-caption software tagged
people with dark skin asgorillas.
An earlier word associationtool called Word2Vec would allow
you to do word math.
You could input king minus manand it would return queen, but
it also gave answers like man isto computer programmer, as
woman is to homemaker.
Ouch Massage therapy researcherDr Anne Blair Kennedy asked
(12:28):
generative AI to create imagesabout massage therapy, and first
the generator told Dr Kennedyit couldn't do that because the
images might be sexual in nature.
So there's that, after somemore specific prompting, it
created a person with their facesmooshed into a table while two
different people touched theirback, a therapist with three
arms, and it couldn't figure outhow to put a client face up.
(12:49):
Each of the images also hadsome less than subtle Asian
cultural overtones.
In the gorilla example, thedataset used to train the auto
caption had a lot of problems.
Like it had more images ofGeorge W Bush than Black women.
That means it had more imagesof one man than 14% of the
population of America.
In the programmer versushomemaker example well, how do
(13:14):
you remove bias from a language?
And in the third man, if youwant to know what the general
population thinks aboutsomething, ask a generative AI
that's been trained on theentire internet.
I'm going to finish up thisepisode talking about how AI is
(13:35):
being used in healthcare and howit is being used in the rubbing
robots that are being developed.
In healthcare.
The most common use of AI is toread imaging and interpret test
results.
The second most common use isto deal with large amounts of
data and third, in a limitedfashion, to help design
treatment plans.
First, the imaging AI is ableto spot pathological issues in
(13:56):
MRIs and other scans faster thana human.
It is particularly good inoncology and dermatology, both
specialties that rely on imaging.
One study found that the mostaccurate results came from a
human and a program workingtogether rather than one working
independently.
The program could identifypotential problems, but it had a
hard time incorporating otherinformation from the patient's
chart to make a decision.
(14:16):
Ai can also help interpret testresults to detect events like
heart attacks.
In oncology, ai can be used todecide what course of treatment
is most optimal.
The more we learn about cancerand disciplines like genetics,
the more we are able topersonalize treatment to
individuals.
But it takes a lot of work andtime to understand all the
information necessary.
Potentially, ai programs willhelp clinicians navigate the
(14:38):
large amount of data in patientcharts.
The advent of electronic healthrecords has allowed us to keep
better track of patients'medical histories.
But keeping track isn't thesame as organizing and it
definitely isn't the same aseffective utilization.
So that's healthcare.
What about rubbing?
A lot of the new robots thatare being advertised toss in the
fact that they use AI, but Ithink that's misleading.
(14:59):
The availability and wide useof ChatGPT might lead you to
believe that the robot cangenerate massages the same way
ChatGPT generates sentences.
It can't.
The artificial intelligence inthe rubbing robots is used to
aim the robot arms.
Aim the robot arms.
You know how when you sit in arubbing chair at the mall and
(15:20):
the tracks that are supposed torub next to your spine rub on
your spine because you're 5'2and not 5'11 and your rib cage
sits almost directly on yourpelvis, so your waist is like an
inch tall Just me, okay.
The AI in the new rubbingrobots solves those problems.
Navigating in three dimensionsis challenging and all bodies
are different and, even better,are different on any given day.
I have ribs that shift out ofplace, which results in my
(15:43):
shoulder blades wandering aroundon my back.
These robots can scan your bodyand adjust its pre-programmed
routines to fit your special,unique shape.
It uses artificial intelligenceto do it, but the AI is less
chat, gpt and more chess.
But the robot doesn't read aclient's intake form and ask
them about their decision toclean out their flower beds for
(16:04):
nine hours the day before,notice how far they can rotate
their trunk without wincing,make a treatment plan and then
appreciate the pictures of theready-to-be-planted flower beds,
because that would be massagetherapy, and the robots only do
rubbing.
So why should you, a massagetherapist who works with humans,
care about how AI works?
(16:27):
Mostly because, although some ofthe claims might be more hype
and less truth, ai isn't goingaway.
I wanted to give you a peekbehind the chrome curtain so
that statements made by techcompanies would be less
mysterious and, hopefully, lessthreatening in that unknown,
foggy future kind of way.
I encourage you to try some AItools, because the ones that
already exist can really saveyou time, particularly at boring
(16:49):
tasks.
I'm going to direct you to theABMP podcast, where Whitney Lowe
has done an episode on AI thattalks more specifically about
using the robots for good.
Where Whitney Lowe has done anepisode on AI that talks more
specifically about using therobots for good.
If you're looking for a placeto start right now, I second
Whitney's suggestion ofperplexityai, a generative AI
that cites and links its sources, along with providing a brief
answer to your questions.
In the next episode of the Rub,we're going to take a closer
(17:12):
look at the robots themselvesand what they might mean for the
profession of massage therapy.
I hope you learned somethingcool today and thank you for
listening.
Welcome to the Rub a HeelWild podcast about massage
therapy.
I'm your host, kori Rivera,licensed massage therapist and
information magpie, and intoday's episode we're going to
talk about artificialintelligence and rubbing robots.
As a quick reminder, thispodcast uses the word rubbing to
(00:34):
talk about massage activitiesthat do not qualify as massage
therapy.
The academic definition ofmassage therapy is massage
therapy is for both self-careand health maintenance.
Massage therapy includesrubbing, but it also involves
health promotion and educationmessages.
The.
The setting of the treatment,such as a hospital or a spa, may
influence the results, but theresults absolutely depend on the
(00:54):
skill, training and experienceof the therapist and on the
therapeutic relationships andcommunication between therapist
and client.
Okay, let's talk robots.
Recently I read a current modelof how we think dopamine works.
Didn't think I was going toopen an episode about robots
talking about a hormone, did you?
(01:15):
Dopamine has been called afeel-good hormone.
We thought for a long time thatdopamine was a reward.
Do something that gives you anevolutionary advantage, like
eating a Snickers bar.
It's nutrient-dense?
Okay, eat the Snickers, get ahit of dopamine.
But in 1993, a study came outof Switzerland authored by Dr
Wolfram Schultz.
The study found that whenmonkeys were trained to open a
(01:38):
box with food in it, theygenerated dopamine.
However, the rush of dopaminefrom opening the box diminished
over time.
Eventually, there was nodopamine generated when the
monkey found food.
The study also discovered thatif the monkey found food
randomly, instead of every timeit opened the box, the dopamine
came back, which led to thequestion if it's not the food,
(02:01):
then what caused the reaction?
Currently, we think thatdopamine is what happens when
you anticipate that the futurewill be better than the present.
I would call it hope.
Dopamine is what happens whenyou have hope.
Dopamine is the pull of theslot machine, but not the coins
that come out.
Or hitting the button to lookat social media, but not what's
(02:21):
actually in your feed.
Or waiting for that adorablekitten on your screen to do
something adorable, which ittotally will, because why else
would its person have posted thevideo?
There is a 100% chance ofincoming adorableness and for
some of us, dopamine is whatmakes us excited about
artificial intelligence andrubbing robots.
Because I am excited.
(02:43):
Innovation makes me excited.
Novel uses of technology makeme excited.
Accessibility makes me excited.
I'm thrilled to be able to tellyou about both AI and the
rubbing robots that are beingdeveloped.
However, all that dopamine alsosparks a big old warning sign
in my brain, because myexcitement about the future also
makes me vulnerable tounexpected consequences and hype
(03:06):
cycles, and this is why theseepisodes are taking so long.
My thoughts about artificialintelligence and rubbing robots
swing like a pendulum with everynew piece of information, and
there is a lot of informationcoming through right now.
In this particular episode, I'mgoing to give you a rundown of
AI in general and rubbing robotAI in particular.
(03:26):
We're going to talk about whatAI really is and what it really
does.
If you haven't already listenedto the previous episode,
massage Tools, I recommend it.
Unless you live in completeisolation in the mountains or
the desert or something you'vedefinitely noticed the explosion
of artificial intelligenceeverything desert.
Or something you've definitelynoticed the explosion of
(03:47):
artificial intelligenceeverything.
Ai will write your paper, drawyou a picture, compose music and
soften your irritated tone inthat email conversation you've
been having for three daysstraight with no conclusion in
sight.
That last one might be a mething, but in the endless
headlines about AI is it useful,is it dangerous?
Is it cheating?
Most people haven't seemedconcerned about what.
First, let's talk about whatit's not.
(04:12):
The name artificial intelligencewas coined at the first ever
conference about it in 1956.
The people who work on itsdevelopment admit that it is a
terrible name.
There are a bunch of highlydebated ideas about what
intelligence in human is, letalone a machine.
Is intelligence knowledge,abstract thinking, problem
solving?
What about emotionalintelligence?
(04:33):
Does the situation matter?
How do you measure intelligence?
What makes a human smart?
And we all know smart peoplewho are really dumb.
How can someone be smart anddumb at the same time?
So calling a computerintelligent is confusing off the
bat.
Additionally, there are twoavenues of research in AI.
One group wants to createsomething that can think in big
(04:54):
bunny quotes, and one wants tomake programs that do things
better or faster than humans.
These tasks are related but arevery different if you're the
person trying to accomplish themand spectacularly different if
you're the person trying towrite a grant that funds them.
So what we do not have iscalled general artificial
intelligence or strong AI.
(05:15):
This is the kind of AI thatmovies are interested in.
This is Hal from 2001, a SpaceOdyssey, sunny from iRobot or
Samantha from Her.
It's a computer thatunderstands, can think about
ethics and morality andpotentially decide to destroy
humanity because we make somepretty terrible decisions.
We have not made that.
Depending on who you ask, weare close decades away or it is
(05:36):
completely impossible.
What we do have is weak AI.
This is AI that can do specifictasks and, lately, can do them
really well.
You've been hearing aboutalgorithms for a while.
Algorithms are like toddler AI.
You've probably heard thingslike.
The YouTube algorithm is ahorrible thing that fans the
flames of the fires of bias anddivision in order to keep you
(05:59):
watching longer and make moneyfrom ads.
And if you hadn't heard that,maybe you've heard.
The Netflix algorithm cansuggest movies you might like.
Both are true, by the way.
An algorithm is a strict set ofinstructions.
It's programmed veryspecifically and spits out an
answer that it was designed tofigure out.
The YouTube algorithm isprogrammed to figure out what
keeps you watching and then feedthat to you and your
(06:21):
suggestions.
Artificial intelligence is madeup of a group of algorithms and
it can reprogram thosealgorithms to meet a goal.
Instead of being fed strictinstructions, ai is fed data.
The part humans programmed ishow AI deals with the data.
This means that AI alters itsown algorithms using a simple
technique like trial and errorto complete the task it was
(06:42):
given.
Deep learning is what happenswhen AI is given data and a goal
but no instructions.
Deep learning is kind of likeLittle Red Riding Hood here's a
basket, get to grandma's house,godspeed.
So and this is important all ofthe weak AI we have are
learning to complete their tasksin a completely different way.
(07:02):
There are a lot of methods, butwe're going to talk about three
.
We're going to talk about AIthat play games, ai that
recognize images and generativeAI like ChatGPT.
If someone asked you who comesto mind when you think about
chess, garry Kasparov might popinto your head.
Garry Kasparov, aside frombeing the world chess champion
(07:23):
for 15 years, was the firstgrand chess master to be beaten
by an IBM computer called DeepBlue in 1997.
You might be less likely toremember Fan Hui, a world
champion of the game Go, who wasbeaten by the Google computer
AlphaGo in 2015.
If you want to come to theHeelwell community, I will
happily explain how thecomputers won, but for this
(07:43):
episode, I just want you to knowthat they didn't win.
Using the same method, go hasexponentially more possible
moves than chess, but the botthat won at Go wasn't a more
advanced version of the chessbot.
It was a completely differentlearning system.
The other thing you should knowis that both Go and chess are
games of perfect information.
That means that both playersknow what is happening all of
(08:04):
the time.
In games like hide and seek,you don't know where people are.
In chess and Go, you knowexactly where your opponent is
and how many pieces they haveand which kinds.
Chess and go, you know exactlywhere your opponent is and how
many pieces they have and whichkinds.
So Deep Blue and AlphaGo aretotally different programs doing
what looks like the same taskbeating a human at a game of
perfect information but one ofthem got there using a submarine
(08:26):
and one of them got there usinga hot air balloon.
Now let's talk about imagerecognition.
Identifying images is way morecomplicated than playing a game.
Interestingly, these programswork a lot like human eyes.
They were trained to recognizeedges of things and then simple
shapes, and then complex shapesand then whole objects and faces
.
This is the same order oureyeballs use to understand
(08:48):
visual information.
The training used to recognizevisual images is a lot like
playing Marco Polo.
The training used to recognizevisual images is a lot like
playing Marco Polo.
Marco Polo.
The system makes a guess, getsfeedback, changes the system a
little bit, makes another guess.
This loop of guessing andcorrecting eventually allows the
computer to understand theimage.
That is the super simplifiedversion and we are going to stop
(09:08):
there.
But the GameBots and theImageBots Not the same.
Okay, chatgpt ChatGPT was aninterface released by the
company OpenAI in November of2023.
What made ChatGPT unique was itsability to reply to a human and
sound like a human when it did.
You might remember Watson, theIBM computer that won at
(09:29):
Jeopardy in 2011.
Now we come to Watson, who isBrad Stoker.
While it was very impressive,recalling facts and using a
specific way to answer is apretty narrow task compared to a
completely open response.
Chatgpt is what is known as alarge language model, or LLM.
It also fits into the categoryof generative AI.
This means that it isn't simplyreporting information or
(09:50):
meeting a goal.
It is producing a unique answerevery time you ask it something
.
Large language models work byassociation.
It learns what words arerelated to other words, which is
truly impressive when you thinkabout it.
The word run has over 600definitions, including idioms.
You can run cross-country andalso have a run in your socks.
You can run a company or run areport, but the thing is ChatGPT
(10:15):
doesn't understand anything.
It associates, and it does itincredibly well.
And while you could argue thathumans are also association
machines, humans have a greatmany more things involved,
including, but not limited to,life experience, emotion, a body
and social pressure.
Writing is more than wordsstrung together, and words are
(10:35):
more than sounds strung together.
However, for a machine thatstrings things together without
understanding what it's saying,it does a great job making me
sound less irritated.
It saves me from having to combmy writing for SEO keywords and
it can write humorous things ifyou ask it, but it can't play
chess.
In 2023, a user in the AnarchyChess subreddit pitted ChatGPT
(10:55):
against the current best chessbot, which is called Stockfish,
in a hilarious game.
Chatgpt proceeded to change thecolor of the pieces, move rooks
diagonally and ultimately losebecause it put itself into check
.
This is because ChatGPT is aword association AI.
It found patterns in writtenchess games and then imitated
that.
It never knew the rules ofchess and if you tried to
(11:18):
explain them, it wouldn'tunderstand.
Chatgpt cannot play chess andStockfish cannot write a
sentence.
Generative AI like ChatGPT takesdata and a model and generates
a solution.
Did I say generative AI takesdata?
It takes data, a lot of data.
So it takes data, a lot of data.
(11:42):
So so much data An estimated10,000 data points on the low
end for each piece of a puzzleyou're trying to analyze.
So even if there were agenerative AI model that could
create massage therapy protocolfrom an intake form, there
wouldn't be enough data for itto work.
As in all things data-related,more data means greater accuracy
, unless your data is inaccurate.
You've probably heard some ofthe stories about bad datasets.
(12:04):
In 2015, google's fancy newauto-caption software tagged
people with dark skin asgorillas.
An earlier word associationtool called Word2Vec would allow
you to do word math.
You could input king minus manand it would return queen, but
it also gave answers like man isto computer programmer, as
woman is to homemaker.
Ouch Massage therapy researcherDr Anne Blair Kennedy asked
(12:28):
generative AI to create imagesabout massage therapy, and first
the generator told Dr Kennedyit couldn't do that because the
images might be sexual in nature.
So there's that, after somemore specific prompting, it
created a person with their facesmooshed into a table while two
different people touched theirback, a therapist with three
arms, and it couldn't figure outhow to put a client face up.
(12:49):
Each of the images also hadsome less than subtle Asian
cultural overtones.
In the gorilla example, thedataset used to train the auto
caption had a lot of problems.
Like it had more images ofGeorge W Bush than Black women.
That means it had more imagesof one man than 14% of the
population of America.
In the programmer versushomemaker example well, how do
(13:14):
you remove bias from a language?
And in the third man, if youwant to know what the general
population thinks aboutsomething, ask a generative AI
that's been trained on theentire internet.
I'm going to finish up thisepisode talking about how AI is
(13:35):
being used in healthcare and howit is being used in the rubbing
robots that are being developed.
In healthcare.
The most common use of AI is toread imaging and interpret test
results.
The second most common use isto deal with large amounts of
data and third, in a limitedfashion, to help design
treatment plans.
First, the imaging AI is ableto spot pathological issues in
(13:56):
MRIs and other scans faster thana human.
It is particularly good inoncology and dermatology, both
specialties that rely on imaging.
One study found that the mostaccurate results came from a
human and a program workingtogether rather than one working
independently.
The program could identifypotential problems, but it had a
hard time incorporating otherinformation from the patient's
chart to make a decision.
(14:16):
Ai can also help interpret testresults to detect events like
heart attacks.
In oncology, ai can be used todecide what course of treatment
is most optimal.
The more we learn about cancerand disciplines like genetics,
the more we are able topersonalize treatment to
individuals.
But it takes a lot of work andtime to understand all the
information necessary.
Potentially, ai programs willhelp clinicians navigate the
(14:38):
large amount of data in patientcharts.
The advent of electronic healthrecords has allowed us to keep
better track of patients'medical histories.
But keeping track isn't thesame as organizing and it
definitely isn't the same aseffective utilization.
So that's healthcare.
What about rubbing?
A lot of the new robots thatare being advertised toss in the
fact that they use AI, but Ithink that's misleading.
(14:59):
The availability and wide useof ChatGPT might lead you to
believe that the robot cangenerate massages the same way
ChatGPT generates sentences.
It can't.
The artificial intelligence inthe rubbing robots is used to
aim the robot arms.
Aim the robot arms.
You know how when you sit in arubbing chair at the mall and
(15:20):
the tracks that are supposed torub next to your spine rub on
your spine because you're 5'2and not 5'11 and your rib cage
sits almost directly on yourpelvis, so your waist is like an
inch tall Just me, okay.
The AI in the new rubbingrobots solves those problems.
Navigating in three dimensionsis challenging and all bodies
are different and, even better,are different on any given day.
I have ribs that shift out ofplace, which results in my
(15:43):
shoulder blades wandering aroundon my back.
These robots can scan your bodyand adjust its pre-programmed
routines to fit your special,unique shape.
It uses artificial intelligenceto do it, but the AI is less
chat, gpt and more chess.
But the robot doesn't read aclient's intake form and ask
them about their decision toclean out their flower beds for
(16:04):
nine hours the day before,notice how far they can rotate
their trunk without wincing,make a treatment plan and then
appreciate the pictures of theready-to-be-planted flower beds,
because that would be massagetherapy, and the robots only do
rubbing.
So why should you, a massagetherapist who works with humans,
care about how AI works?
(16:27):
Mostly because, although some ofthe claims might be more hype
and less truth, ai isn't goingaway.
I wanted to give you a peekbehind the chrome curtain so
that statements made by techcompanies would be less
mysterious and, hopefully, lessthreatening in that unknown,
foggy future kind of way.
I encourage you to try some AItools, because the ones that
already exist can really saveyou time, particularly at boring
(16:49):
tasks.
I'm going to direct you to theABMP podcast, where Whitney Lowe
has done an episode on AI thattalks more specifically about
using the robots for good.
Where Whitney Lowe has done anepisode on AI that talks more
specifically about using therobots for good.
If you're looking for a placeto start right now, I second
Whitney's suggestion ofperplexityai, a generative AI
that cites and links its sources, along with providing a brief
answer to your questions.
In the next episode of the Rub,we're going to take a closer
(17:12):
look at the robots themselvesand what they might mean for the
profession of massage therapy.
I hope you learned somethingcool today and thank you for
listening.
Popular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Therapy Gecko
An unlicensed lizard psychologist travels the universe talking to strangers about absolutely nothing. TO CALL THE GECKO: follow me on https://www.twitch.tv/lyleforever to get a notification for when I am taking calls. I am usually live Mondays, Wednesdays, and Fridays but lately a lot of other times too. I am a gecko.