Title: The Last 7 Years of Human Work - Understanding the AUTOMATION CLIFF!
Author: David Shapiro

Transcript:
hello everyone I am really excited about
this video I've been planning it for a
while and I kept just adding to it as I
was doing more and more research uh
particularly because I was using deep
research to make it better and better so
let's Dive Right In uh I have talked
about the concept of the automation
Cliff for a while and uh I didn't come
up with this idea so what I wanted to do
was actually share some of my personal
experience but also some research as
well as some projections based on the
automation Cliff now uh basically this
is the idea of the automation cliff and
we're going to unpack this but just keep
this graph in mind where you've got the
stair step versus this more kind of
catastrophic plunge uh the tldr is that
if you focus on incremental improvements
in technology you'll end up with this
kind of gradual stairstep Improvement of
level of automation versus level of
human involvement so before automation
human involvement is high and then level
of of automation so you can think of
this as similar to like uh Tesla's
levels of FSD you know because right now
they're at like level three um but
ideally full self-driving like true
self-driving is going to be level five
which is zero human involvement needed
ever whereas level three is 99% of the
time it can work but it can't handle
edge cases and does still require human
intervention um I don't know the exact
definition of FSD level three but you
get the idea so right now most
Industries are using the stairstep
approach and we'll talk about why this
is um but in an Ideal World and in also
some cases you end up with more of this
kind of edifice approach where you just
have this escarpment that just you just
go careening off of hence the thumbnail
that you saw for this all right moving
on so what is the automation Cliff uh
basically the the principle of the
automation Cliff says that what you
should do is wait until you have the
full process automated end to endend and
then just full sent just send it off and
there you go um one of the key
principles here is tasks should be
controlled completely by either humans
or completely by automation systems with
no middle ground personally as an
automation engineer in my past life this
is what I would have advocated for and
the reason is because you know it does
nothing until you turn it on and you do
all the full testing and then you turn
it on and it automates everything all at
once that's what I mean by the
automation Cliff now there's lots and
lots of other principles in but there's
also like some problems with getting
there so let's start to unpack that uh
in just a moment but what I want to talk
about is drop in Technologies so let me
go back a couple slides and show you
what happens so when you have an
automation Cliff like this this usually
happens when you have what's called a
dropin technology and a dropin
technology basically means you know
here's the oldfashioned way of doing
things where it's 100% humans uh in the
loop and then you have this new brand
new technology that means you don't need
humans in the the loop at all whatsoever
so let's give you a couple of examples
about drop in Technologies now in some
cases these are not automations but
these are technologies that could just
come in and completely change everything
so first was USB um everyone is familiar
with USB when I was little and people my
age and people uh older were younger you
had serials and you had parallels and
you had all kinds of different
connections now everything is USB
Universal serial bus um Cloud
integration so SAS software is a service
is another example of where you can just
switch between softwares um and if you
get you know uh actually chat Bots are a
prime example you can go sign up for
Claude you can sign up for chat GPT you
can sign up for Gemini and most of them
are pretty interchangeable so those are
examples of drop-in Technologies or
fungible Technologies GPS was another
thing that once it was there it was
ubiquitous and now you can use it for
all kinds of stuff it enabled Google
Maps uh and you know your your Fitbit
and everything there's all kinds of
different cascading effects uh smart
retrofitting of buildings and other
infrastructure so an example of a drop
in technology was dialup modems which
used that was that's probably honestly
the best example because you started
with existing phone lines and then you
said well let's make them digital so
then you can just have modems call each
other and exchange uh information that
way and that was really kind of most
people's first uh experience of the
internet and then we did the same thing
with cable uh cable modems were just
using the fatter uh bandwidth that uh
the digital cable had uh available uh
and then streaming media is another
example you know you can swap between
Netflix and Disney and all those other
things so these are examples of drop in
technologies that once you have enough
infrastructure you can just put in a new
technology and you can adopt it very
quickly personally uh on an individual
consumer level you can adopt it almost
instantly overnight um some of these
Technologies do take longer for larger
organizations to adopt just because
there's a lot of inertia in those
organizations so these are a few
examples of the fact that we have had
dropin Technologies um that allowed for
that kind of that almost saltatory leap
that very Square wave um of adoption not
all of it goes that way so um before we
uh get a little bit further I want to
talk uh also about examples of where
full automation does tend to be superior
so if you can achieve full automation if
you can achieve that full automation
Cliff that is going to be more desirable
so one example is um autopilots so
originally autopilots would basically
just maintain your altitude and speed
and a little bit more but now uh as
airplanes are more and more
sophisticated autopilots basically there
are stories of Pilots just going to
sleep with the autopilot on uh meaning
that they run 100% of the aircraft um
another example is pharmaceutical
production so these by the way these are
all numbers that were surfaced using
deep research um so yeah maybe I should
start including the links to those deep
research articles as uh as as evidence
anyways let me know what you think in
the comments so uh having supported and
worked with and consulted for and talked
to people in the Pharmaceuticals
Industries um the pharmaceutical
industry is one of the most heavily
automated Industries out there
particularly with the actual production
of drugs um and when you look at what
when we talk about lights out
manufacturing lights out manufacturing
basically means no humans need to be
present or even observing um and so that
that took the effect rate from 0.1% to
0.001% um so in in other words keeping
human supervisors was actually a net
negative it was actually better to get
humans completely out of the loop um
here's another example is automated
Harvesters so John Deere so these are
like the big combines that you see like
that you know go over fields and harvest
everything um they're fully autonomous
combines uh reduced yield loss from 15%
to 2.3% by eliminating operator fatigue
and operator errors basically you know
humans make mistakes and if you're
driving a tractor for 10 hours a day you
get kind of bored so on and so forth you
get the idea we don't need to go through
every single example but these are some
these you know between uh autopilots
pharmaceutical manufacturing and uh
harvesting you can see that there are
several examples across several
different domains where full automation
is actually preferable if you can
achieve it so moving on um now one
question you might might be wondering is
like okay well why why why is the
automation Cliff preferable um if if you
know you could just gradually uh
Implement things number one is the uh
performance degradation of handoffs so
you know you see videos of people you
know driving their Tesla and it's like
oh you know you're distracted and then
you have to intervene um that's one
example now what I will say is as a
counter example to that is that the is
that splitting your cognitive attention
with using tools like deep research it's
like oh here go do go do a research uh
topic for me briefly and then I'll come
back in 5 minutes and that actually
gives your brain a CH a chance to rest
there's actually a brief story that I
have where um uh at a software company I
worked at gosh 2012 2011 2012 so that
was a long time ago um I built out their
their uh virtual infrastructure and we
built more build servers for them and
their build process went from 24 hours
to 2 hours and they were like Dave can
you take those servers out they were
kind of joking but they're like can you
remove those servers because you know uh
we we we now have less time to actually
you know work on our after action
reports we're used to we're used to the
build process taking 24 hours so then we
have a full day to keep working um I was
like I'm not going to do that like you
you wanted me to make things faster I
made things faster by a factor of 12
deal with it so anyways um you have
trust issues workload problems
monitoring partial automation can uh
increase the C itive load which that's
particularly true if you're monitoring
different automation stations um so that
can and and if information is coming at
you faster that will tire tire you out
much much more quickly uh and those
sorts of things so in many cases if it
is possible then you want to use the
full automation Cliff you want to go
straight from the current way of doing
things to the new way of doing things
without much um without much
interstitial time another reason is
because you don't want to keep
Reinventing the wheel um that was
something that I include in this slide
but basically every time you have to
reinvent the infrastructure or Implement
new infrastructure that handles you know
human affordances and partial Automation
and then you have to do it again to get
to full automation often it's better to
just wait and then implement the full
automation all at
once um and so this is this is talking a
little bit more this slide is we talk a
little bit more about how um in reality
usually full automation is just not an
option so this is one of the things that
we're going to be talking about with the
rise of agents and robots so um also my
dog is under the desk so if I seem
distracted I'm petting my dog um she'll
make a guest appearance one day um so
first and foremost uh is the economic
barriers so full end in automation can
often be very expensive and as many of
you have pointed out in the comments and
on Twitter and other places the first
90% is usually actually really easy it's
the last mile of automation that is
really hard and that's where 90 to 99%
of of your automation effort will go
into is uh what was it one of you said
something like you know when in in the
space of automation you realize that
everything is edge cases and that's
that's not a bad way of thinking about
it is because yes 90 to 99% of what
you're doing is routine robust uh or not
not robust um uh routine but or brain
dead simple I don't know what word I was
thinking of um but it's it's uh it's
it's very repetitive maybe that's the
word I was thinking of it's routine and
repetitive but then you do need that
level of high level adaptation for every
single exception every single edge case
and those sorts of things and that is
the technical complexity where it's like
some things it's just too complex to
automate unless or until you get to a
general purpose general intelligence
whether it's a computer using agent or a
robot um which is going to be more
cognitively flexible than a human then
the technical complexity is no longer a
barrier that has honestly been the
biggest barrier to automation up to this
point but with the rise of generative AI
language models and cognitive
architectures that's no longer going to
be a barrier um risk management resource
constraints you can imagine how all of
these things play out but really it's
the economics and the technical
complexity are the two biggest barriers
or constraints to full automation um but
as robots you know become more
ubiquitous and become more intelligent
and as computer using agents also become
more ubiquitous more robust and more
intelligent those barriers are going to
disappear very quickly um so speak
speaking of barriers and adoption rates
one of the things that I have been
pointing out to people is that
technology adoption rates have been
accelerating so the automobile took a
long time to reach a point of saturation
oh and by the way this graph is a little
bit dated because you know the internet
has been around for more than 10 years
now um so the data in this graph is a
little bit dated but you get the point
where the television once it got cheap
enough it took off really fast
electricity took off really fast but
these are things that took you know
decades to a century to get fully
adopted but then mobile phones PCS
internet everything is getting adopted
much much faster here you're talking
about adoption curves that are in
measured in the 10 to 20 years um now
that uh that the internet has reached a
certain level of saturation anything
that can be delivered on the internet
gets adopted much much faster and that
includes artificial intelligence such as
uh chat Bots and those sorts of things
robots require a lot of infrastructure
to be built out um then you have to ship
the robots and those sorts of things so
because there's a physical layer to the
robots there's going to be a little bit
more friction but on the other hand
robots that are in humanoid shape are a
perfect dropin technology so uh before
we move on I want to point out my uh sub
uh not substack my link tree real quick
which has my substack um it's got my
patreon my school Community this is my
learning community I update uh two to
three lessons per week over there on
patreon we have an exclusive Discord I'm
also on substack Twitter um I also just
added my SoundCloud to uh to my link
tree which is where I put all my AI
generated music um it's not for everyone
but I listen to my own music a lot um so
if you're into psychedelic Space Rock
I've got a lot of it up there um also
I'm on GitHub and Spotify and a few
other things so go check it out all
right back to the
show now um what this slide is talking
about like where are we actually trying
to automate things so I I what I did was
I had deep research say go find the
problems that people are trying to
automate today right right now with
generative Ai and Robotics so here are
the examples that it came up with number
one is contact centers so we've all by
now probably heard some of the stories
where call centers have had some of
their their Staffing reduced by 90%
there's also been some stories of
they've had to rehire Some Humans for
all those edge cases that we were
talking about but at the same time a lot
of those call centers that have switched
to fully or mostly AI um the cat scores
also go up and cat is customer
satisfaction so that's that's MBA jargon
for how happy are your customers in many
cases if you go to full automation the
customers get happier because then uh
there the quality of their service goes
up and they have more faith in your
service or your company or your product
um so however with that being said you
know if a call center can only get rid
of 90% of its people but it still needs
10 for those edge cases that's not full
automation um and furthermore there's
plenty of other kinds of call centers
that you just cannot fully automate away
yet that is still a very high Target
taret because that's what we would call
low hanging fruit uh another example is
retail checkout um so in uh for instance
if you've ever gone to those self
checkouts um those self checkouts
sometimes they break or so on and so
forth sometimes theft also goes up
because it's like you have a self
checkout but then you have like a human
supervising and then but the human gets
bored and stuff still gets stolen and
yada yada yada so then you need more
computer vision for the security and
yeah and so you end up with all these
other what about what about what about
uh kinds of things that make full
automation of the checkout uh a little
bit harder uh Warehouse robotics uh this
is another example so you've probably
seen some of the videos of like the
Amazon robots where it's like it there
Amazon has warehouses that are not human
navigable anymore um at the same time
sometimes those systems still get gummed
up because of a complex emergent
behavior that happens when you have
hundreds and hundreds of uh item
fetching robots and they get you know
all uh jammed up I don't mean physically
jammed up I mean you know like the
traffic gets congested and so on and so
forth so these are these are current
challenges that we have not yet solved
and it's like okay well if we can't
fully automate call centers and Retail
checkout and warehouses then clearly
like a lot of jobs are still safe
however keep in mind that as robots get
more intelligent every every step of
intelligence they that they gain and
this also includes computer using agents
that dramatically expands what they can
do without human intervention so so
you're going to see some of these leaps
some of these um some of these sigmoid
curves or these step functions where
you're going to have new abilities that
are going to just say oh all that stuff
that we couldn't automate a year ago we
can automate all of it now and I have
seen that personally back in my back in
my corporate days I've also seen it in
some of the clients that I've consulted
for where there are things that you can
automate today that a lot of people
don't even believe that you can automate
and that's one of the reasons that I
make these videos is to say hey the
thing that you think that you can't
automate maybe you actually can
so moving on um now I've talked about
humanoid robots on this channel quite a
bit but I want to talk about how this is
really the ultimate drop in solution so
one of the key things is that humanoid
robots can operate in human spaces using
human tools human vehicles and uh pretty
much everything else so if you have a
human robot that is as smart as or
honestly if you put you know gp4 or gp5
in it or you know clae 4 whatever
whatever model comes out then it's going
to be smarter than the vast majority of
humans already then if you have watched
the Boston Dynamics videos where those
robots are far more agile than humans
they can do standing back flips I cannot
do a standing backflip so they're
stronger they're smarter they're faster
they're going to have more dexterity
than humans that means that it is a
perfect drop in solution which means
that basically any job that a human does
with their hands and eyes and body sorry
hit the microphone um these robots will
be able to do very soon and those that
general purpose form function means that
it can even sit in front of a computer
and use a keyboard and mouse if it needs
to um but we can use computer using
agents for that so you can just remove
the whole robot entirely um so this this
represents a full automation solution
and this is what I mean by the
automation Cliff once you start shipping
you know super intelligent super strong
super dextrous super agile robots it's
like game over for 90% of human jobs
next is the computer using agents so the
computer using agents are what you've
seen like um uh operator and repet and
all those other different tools out
there um what you need to think of and I
still have people saying Dave why don't
we just focus on apis and so for those
that aren't familiar an API is an
application programming interface which
is basically allows one computer program
to call and talk directly to another
computer program with without any other
user interface but keyboard video Mouse
KVM is the universal API furthermore
think about how the vast majority of
what humans do also my dogs are
wrestling in the background so if you do
hear that I apologize um so the vast
majority of human knowledge work is done
with KVM keyboard video Mouse if you can
do it with KVM and an operator can do it
with KVM that's a universal UI that's a
universal interface that you don't need
any other infrastructure for you don't
need custom apis you don't need custom
API discoveries that is the API the KVM
is the universal API and so what that
means is is that instead of even having
a robot using the computer you just drop
that agent onto any computer or servers
and they can be virtual servers by the
way and you have literally the
equivalent of hundreds thousands
millions of of employees all using you
know their own own laptop screen
basically but on a virtual server in the
cloud somewhere um that's really what
we're heading towards and the roll out
of this so this is this ties back to
that um what I said about you know the
adoption of cloud services um it's going
over the Internet so that means it's
really really fast to roll out um now
here's my personal timeline so this is
the automation wave optimistic timeline
um and this is based on the kind of
seven-year time Horizon and the seven
years is basically about how long it
took for companies to adopt
virtualization uh which was my area of
specialty as well as Cloud software uh
or software as a service which was uh
adjacent to what I was doing and when
you think that computer using agents are
basically virtualization and Cloud
software and it took seven years to
adopt those then we're looking at about
seven years for full commercial adoption
from this year because this year is when
we're first uh deploying agents so
initial launch is 2025 computer using
agents begin deployment um and not and
digital knowledge work and humanoid
robot uh humanoid robots are being
ramped up this year as well Mass
adoption happens 2026 and 2027 um so
this is when Fortune 500 companies are
going to really start using both
computer using agents and humanoid
robots um in Mass there are Fortune 500
companies already using Tesla Optimus
and other robots just want to point that
out I think BMW was the first car
company that started using them other
than Tesla of course um then so that's
the that's the uh early early Mass
adoption and then you're going to have
full integration happening in 2028 to
2030 and then you're going to have the
the fin laggards the the the
optimization happening in the 2031 to
2032 range and then by 2033 you're going
to have offices full of robots and
computer using agents and all that fun
stuff that's my personal prediction is
that we're looking at seven years until
you know knowledge work as we know it is
over and done with in every industry um
now I want to use this graph so this
this graph is the adoption curve um so
this is like a very similar
version to the other adoption curve that
I showed you so this is this is a linear
adoption curve which is just at what
point does the technology become
saturated but another way to look at the
adoption curve is this which is at what
point does each um each type of company
adopted so right now or or I guess 2024
and earlier were the innovators so these
are all the people that you know watched
my YouTube channel since 2022 2023 these
are the people that have been
experimenting with cognitive
architectures and agents
since you know before chat GPT came out
or when chat GPT first came out think
about back to the era of baby AGI and
those sorts of things that was the
innovators so that was the bleeding edge
innovators that was the first
2.5% this year and 2026 are going to be
the early adopter so this is where uh
this is where all the first movers are
saying okay there's actual commercial
value here let's pull the trigger then
20 2027 to 2028 is going to be the early
majority this is where you know your
your mom and pop shops may maybe maybe
not you know your bakery but what I mean
is you know your average run of the mill
companies are going to start adopting
some of these Technologies you know U I
know lawyers and law firms that are
already using some of these AI tools um
uh but they're they're still kind of the
early adopters so then the majority of
law firms and doctor's offices and those
sorts of things will start adopting then
and then you'll have the uh the group of
people in the late majority so these are
the more Skeptics these are the more uh
mortar kind of stores so like you know I
would imagine that like Home Depot
they'll probably be a little bit later
to adopt these just because that
business model hasn't changed in more
than a century you know it's like you
have hardware and tools and you sell
Hardwares and tools to real people in
front of you um so some businesses some
Industries are going to be a little bit
more resistant to it um rather than
people that are going to be more on the
front end now heavy Industries like
Mining and construction they will
probably be in the early majority if I
had to guess just because human labor is
really expensive and loss of life and
injury is also really expensive but if a
robot gets crushed under a rockfall
that's just a tax write off you can't
write off human lives uh sorry that's
not how it works and then 2030 plus this
is going to be as the rest of the world
catches up so this is my preferred
timeline now I asked deep research to
take all of this into account and make
its own timeline and it gave a much more
conservative timeline so it's its
timeline based on historical evidence
and those longer adoption curves which
we saw earlier says that the initial
wave will be 2025 to 2030 um so I I I
need to emphasize this is not my
personal timeline I'm just showing you
what the AI said as a as a more
conservative or realistic timeline so
2025 to 2030 this is when we're going to
see digital knowledge work uh get
replaced the early majority won't be
till 2030 to 2035 again
I don't believe
that um service integration so this is
where you start to see kind of the the
early uh early and late majority so some
of the more uh resistive s uh uh uh
Industries so like healthc care is a
very resistant industry education very
resistant industry you're it doesn't
expect that we're going to see full
automation there for the next 10 to 15
years again I this will not age well um
and then by then there will be enough
regulatory pressure on States and
federal governments to say okay we need
to do things differently and that's 15
to 20 years out now keep in mind that
2045 is like Singularity so if teachers
unions are still preventing AI in the
classroom when Singularity hits oh boy
are they in for a roote Awakening
anyways like I said I this this timeline
is way too conservative for me but I I
felt like just for the sake of argument
I had to put this is what the AI thinks
the timeline is going to be um now what
I do predict is that as computer using
agents and robots ramp up in terms of
intelligence and ubiquity we are going
to see total Workforce automation as we
understand it today now we can talk
about post labor economics there's there
will be some kinds of jobs like
influencers I hope will stick around um
entertainers will probably stick around
like musicians and stuff there might be
entirely new classes of jobs there
probably will but the vast majority of
economic uh activity will not be done by
humans in the near future so you look at
Medical Precision superhuman surgical
robots with perfect Steady Hand hands or
multiple hands um combined with computer
using agents that are constantly
researching the best medical procedures
you will not have a human doctor you
will not want a human doctor in this
potential world next is construction a
lot of people say oh well I'm a boiler
maker or I'm a welder and Y yada yada
and my job safe no it isn't um consider
that robot that industrial robots
already do better Precision welds than
humans do the only difference between
like those Factory line welders and a
human welder is is that the human is in
a form factor that is more mobile um
that's not an advantage in the long run
electricians plumbers construction
workers uh welders you guys like you're
on notice I'm I'm telling you I'm I'm
trying to warn you ahead of time um that
that job is probably going away next is
emergency response so this is everything
from um uh emergency medical technicians
to Firefighters to even police um or or
whatever like all kinds of emergency
respons you take the human out of the
loop they you know you have machines
that are immune to smoke heat biological
radi radiological chemical attacks
whatever like you know there was um
there was a movie called surrogates
which was a really cool movie it didn't
make that bit much at the box office but
it's a Bruce Willis movie and one of the
scenes in that movie was really cool
where there's like a bunch of soldiers
and they're like all kids like playing
VR but they're piloting little humanoid
robots across a battlefield um and it's
just like oh you know robot gets you
know nuked and you know the person's
like ah darn and they you know spawn up
into another robot and to them it's just
a game um science science and
engineering that would you know I don't
think I have to really sell this for my
audience because you guys are like
paying attention to The Cutting Edge of
like Alpha fold and all that fun stuff
but you know we have like somewhere
between eight and 25 million scientists
uh you know phds uh or doctorates
globally right now we're going to have
the equivalent of billions or trillions
here real soon um and so therefore the
vast majority of scientific research
will be automated you combine Those
computer using agents those digital
agents um or those narrow AIS with
robots and you won't even need humans in
the loop if you don't want it now
obviously you still want humans saying
hey you know hey Mr Robot maybe stop
making VX gas we don't want you to make
that because that's really dangerous for
us but you get the idea and then finally
um you know uh government um
particularly if AI is provisioned uh of
the people for the people and by the
people um and the AI is is is directly
accountable to the people then what role
does elected politicians play
anymore I don't know so anyways thanks
for watching I hope you got a lot out of
this cheers