00:00:01
let's talk about some risks of AI that
00:00:03
I'm really not hearing anybody talk
00:00:05
about but seem really important to me so
00:00:08
for a little context I'm not a AI Doomer
00:00:11
I'm not like oh everything about AI is
00:00:13
terrible and I hate it and I would never
00:00:15
use it sure there's a bunch of maybe
00:00:17
ethical questions about how we got AI
00:00:19
but it is here and I use Ai and I use it
00:00:22
to help me write code and I use it for
00:00:24
personal stuff in my life too it's
00:00:26
actually pretty helpful in lots of
00:00:28
different ways although I don't think
00:00:29
that can do all of my job in 5 seconds
00:00:32
from now but that remains to be seen and
00:00:34
a topic for a different
00:00:35
video but just like any other technology
00:00:38
it's a double-edged sword there's some
00:00:40
good things about it and there's some
00:00:42
bad things about it and or potentially
00:00:43
bad things that may be you know if we're
00:00:45
talking about it and working on it we
00:00:47
can avoid having happen so let's talk
00:00:50
about one of those today my first sort
00:00:52
of concern here is default answers for
00:00:55
llms become a sort of deao standard if
00:00:59
you try do make me a to-do list app the
00:01:02
odds are pretty good at least on a fresh
00:01:05
account I try this on Claude myself that
00:01:07
it's not going to do a serers side
00:01:09
rendered thing in rails it's not going
00:01:11
to make some MVC in Phoenix even though
00:01:13
those are clearly Superior Technologies
00:01:16
instead it's probably going to make you
00:01:17
a broken react application at least
00:01:19
that's what happened when I tried it
00:01:21
with Claude and I guess to be fair maybe
00:01:23
that is the most human thing in AI could
00:01:26
do but this is not a talk about AGI this
00:01:29
is a talk about default answers becoming
00:01:31
deao Solutions right so the idea here
00:01:35
right is that anytime I'm using an llm
00:01:37
for some topic that I'm not super
00:01:39
familiar with or where I don't want to
00:01:40
do the work I kind of just accept the
00:01:42
result and seems fine I go to the next
00:01:45
thing llm host is happy you sent them
00:01:48
money sure maybe not enough money to
00:01:51
cover the 5 billion losses of open AI
00:01:53
this year but maybe someday they'll get
00:01:55
there my concern is what if everyone is
00:01:58
doing this right do we get into this
00:02:01
spot where we're all becoming a natural
00:02:04
language programmer but without any of
00:02:07
the underlying skills required to
00:02:10
evaluate whether those choices are
00:02:12
actually good or just there because
00:02:14
that's either the way they've been done
00:02:16
or what comes out of the llm and so this
00:02:19
is not exclusive to people who don't
00:02:21
know anything about programming although
00:02:23
I think it's definitely going to be
00:02:25
exacerbated by them right hey make me a
00:02:29
an app that shows like blue buttons and
00:02:31
I want it to be about Plumbing they're
00:02:33
just going to pick whatever the normal
00:02:35
thing is and maybe overall that's good
00:02:37
but there's still there's still
00:02:39
something sort of nagging at me about
00:02:41
that but in these kinds of scenarios
00:02:43
where we're suggesting not really
00:02:46
technical details we're going to get
00:02:49
back just whatever the default is
00:02:53
and the reason I know that is because if
00:02:56
we wanted to give it all the technical
00:02:58
details we probably would just end up
00:03:00
coding most of it right the concern is
00:03:03
not that olms will be able to like fill
00:03:06
in the Tailwind classes to position this
00:03:09
thing in a little box that I don't know
00:03:11
how to do offand that's not really where
00:03:14
the future of llms seem to be going so
00:03:18
my question is then how does any new
00:03:21
library language framework gain adoption
00:03:24
in this world where the majority maybe
00:03:28
of the programming or the work done is
00:03:31
just the de facto default answer of
00:03:33
whatever comes out of an llm
00:03:36
particularly in the maybe worrisome
00:03:39
cases of where people are building
00:03:41
entire applications without any
00:03:43
technical knowledge of what's going
00:03:45
on and not just for languages and
00:03:48
Frameworks right I'm also wondering how
00:03:50
does something like a new cloud provider
00:03:53
compete against existing Cloud providers
00:03:56
the user might not even know that they
00:03:58
are picking a cloud provider if they
00:04:01
just accept the defaults right and so
00:04:05
this is sort of one of the things that I
00:04:07
feel like all of these people are saying
00:04:09
don't learn anything about CS you're not
00:04:11
going to need to know anything about it
00:04:12
you don't need to know about all of
00:04:14
these different aspects of how the stack
00:04:16
works together how these pieces go any
00:04:17
technical details the AI is just going
00:04:20
to take all of our jobs and just do it
00:04:22
perfectly right and this is of course
00:04:24
from all the people in the internet who
00:04:25
are 18 months late on 6 months till no
00:04:28
jobs I do think AI is going to be
00:04:31
disruptive don't get me wrong it seems a
00:04:33
very disruptive technology to me but
00:04:35
that's not the same thing as it's
00:04:37
completely useless to know things but
00:04:39
those are two separate problems and two
00:04:42
separate areas to weigh about what the
00:04:44
most effective way to deal with the
00:04:46
disruption of AI is and so llms don't
00:04:51
seek truth it's not as if they're out
00:04:53
there this ability you know like God's
00:04:56
angels out here trying to find out
00:04:57
what's good and bad and we can consult
00:04:59
this Oracle and get back the truth from
00:05:01
the heavens that's not what they are
00:05:03
they are predictors based on their data
00:05:06
of what the next likely things are and
00:05:08
that is like the most gross
00:05:10
oversimplification of all time I get
00:05:11
that because they are incredible
00:05:13
technology and letting us do things that
00:05:15
I really was not imagining maybe even
00:05:17
six months ago let alone 5 or 10 years
00:05:20
ago but the attitude from the people
00:05:23
here in this Camp is we really don't
00:05:25
need to worry about economics or
00:05:27
incentives or reality we can just trust
00:05:30
the benevolent people running our llms
00:05:32
we can just trust that whatever is
00:05:35
coming back out of the llm is going to
00:05:37
be trustworthy and good just like
00:05:39
whatever the top 15 results on Google
00:05:42
search results are amazing and good and
00:05:44
no one has ever gamed those uh we can
00:05:47
just have that same exact feeling from
00:05:50
when I ask the LM a question and I get
00:05:52
back a response that's that's the level
00:05:55
of trust right that we should be putting
00:05:58
in the LM we should be inspecting them
00:05:59
when we need to have some knowledge to
00:06:01
be able to do that and the other thing
00:06:04
is here while I do think malice is going
00:06:06
to be coming into play from some of
00:06:08
these aspects this really does not
00:06:10
require any malice to happen right if AI
00:06:13
is great at generating typescript right
00:06:15
now then that means there will probably
00:06:17
be lots more typescript tomorrow which
00:06:20
does that mean then the next most likely
00:06:22
thing to predict or suggest is also
00:06:24
going to be typescript it doesn't
00:06:27
actually require us to have a situ ation
00:06:30
where everyone is in this conspiracy uh
00:06:34
like cigar room smoking cigars and
00:06:36
saying ah let's make typescript the
00:06:38
language of the future we really want to
00:06:39
make everyone's life miserable it
00:06:41
doesn't require that it's just that
00:06:43
sometimes systems perpetuate the same
00:06:46
things and that's the first risk that
00:06:48
I'm I'm feeling worried and I'm not
00:06:50
seeing discussed anywhere the second
00:06:53
aspect here
00:06:55
is this idea of optimizing llm to return
00:07:00
particular results just like people have
00:07:03
spent I cannot even fathom how much
00:07:05
money trying to make SEO for their
00:07:08
websites to show up in search engine
00:07:10
results right we're going to have that
00:07:12
for llms and my concern is is if that
00:07:15
gets captured what we're going to have
00:07:18
is this llm sort of nent oligopoly of
00:07:21
products right a new brand that is this
00:07:24
small cabal of products that are the
00:07:27
only things that are going to be
00:07:28
suggested by by llms and that's where
00:07:31
I'm sort of shortening this for lmm and
00:07:34
O right that's what we've got here llm
00:07:36
and O for short and this can happen in a
00:07:40
variety of ways it can happen in the
00:07:42
training data maybe you're going to just
00:07:44
feed it tons tons tons more felt data
00:07:47
than you are react right and you're
00:07:48
trying to subvert everyone's
00:07:50
expectations and get everyone to use
00:07:51
felt seems cool seems like a cool
00:07:53
technology I don't know maybe I would
00:07:54
vote for that one but right but you
00:07:56
could also do it in sort of what you're
00:07:58
going to do when you're doing the
00:07:59
training here you can do it in the
00:08:01
reinforcement learning where the people
00:08:02
who are voting these up or down for
00:08:04
whether they're good or bad or good
00:08:06
suggestions or whether they're safe
00:08:07
suggestions or not those could filter
00:08:09
out other kinds of options and make
00:08:11
certain things more or less likely to be
00:08:13
suggested it could be in the promps
00:08:15
themselves it could be explicitly in
00:08:18
programming like behind the scenes
00:08:20
inside of open AI could be filtering out
00:08:23
any suggestion to use claud's API
00:08:25
instead or X's grock or how to download
00:08:28
an open source model and run it locally
00:08:30
on your machine any of those things
00:08:33
could have biases put in that on the
00:08:36
receiving side we are not aware of and
00:08:39
don't know our happening they could even
00:08:42
be doing that explicitly in
00:08:44
post-processing steps with the true AI
00:08:46
right Reeses and if statements they
00:08:48
could just be filtering out claw you
00:08:50
would never know and that reinforcement
00:08:54
into this oligopoly is where the de
00:08:57
facto thing becomes a much bigger worry
00:09:00
for me it's it's not so much of a worry
00:09:02
if everyone is just getting to use react
00:09:05
and their websites work just fine for
00:09:06
them or whatever right the the worry I
00:09:09
have is that we will experience the same
00:09:12
kind of decline that we had in Google
00:09:14
search results inside of llm results as
00:09:18
the data and the training and the
00:09:20
feedback is all poisoned by the
00:09:23
interests of the people making those
00:09:26
llms if if there's a th000 x resources
00:09:29
about react How likely is it to suggest
00:09:31
phelp right or if there's a thousand X
00:09:32
resources about using Azure why would it
00:09:35
suggest using
00:09:37
gcp and this leads into that idea of
00:09:41
this vertical integration that I see
00:09:44
definitely happening right it doesn't
00:09:45
require in a lot of uh a lot of
00:09:47
imagination to get here where you could
00:09:50
say ah I'm inside of VSS mode right and
00:09:53
I'm inside of there and I say hey deploy
00:09:55
this to Cloud right a naive user with a
00:09:57
credit card and so what happens they say
00:09:59
sweet let's pull into aure Library we'll
00:10:01
install that for you automatically when
00:10:03
you click the little green button
00:10:04
that'll Deploy on your GitHub Enterprise
00:10:06
account and then will deploy that
00:10:08
teasure you probably need a four times
00:10:10
the size thing that you actually need to
00:10:12
make sure that we're charging that and
00:10:14
you can just see how this decision of
00:10:18
the natural language thing especially
00:10:20
with no underlying understanding can
00:10:23
lead to a complete capture of the stack
00:10:26
you could be in a Microsoft editor using
00:10:29
a Microsoft language doing Microsoft
00:10:31
tooling with a Microsoft code assist
00:10:33
using a chat application built by
00:10:35
Microsoft to deploy to Microsoft's Cloud
00:10:37
on your Microsoft GitHub provider okay
00:10:40
I'm not even just picking on Microsoft I
00:10:42
think everyone wishes they could be
00:10:43
doing this right everyone wishes they
00:10:45
could have that vertical integration but
00:10:47
that's that's one of the concerns that I
00:10:49
really have is that if we're just going
00:10:52
to accept what the next likely thing is
00:10:55
from the llm boom they're going to want
00:10:58
to capture that and get you to opt in in
00:11:01
a sense opt in um to all of their
00:11:05
services right and this is not a m
00:11:08
capitalism moment one of the things that
00:11:11
is a bigger worry or a bigger risk in
00:11:13
this area is I'm worried that some of
00:11:16
these large llm providers are going to
00:11:19
actually Lobby for and receive a bunch
00:11:22
of extra regulatory capture around llms
00:11:26
and you know the local llm running on
00:11:28
your computer isn't safe enough to be
00:11:30
legal so it's no longer legal to run
00:11:32
llms on your computer or oh that small
00:11:34
upstart llm doesn't have all the same Fe
00:11:37
safety features we do or the same legal
00:11:40
and so then they're not allowed to run
00:11:41
that either and so you end up with this
00:11:44
scenario where llms are the def facto
00:11:47
way of doing programming and all of the
00:11:50
other possible verticals that could be
00:11:52
suggested by an llm don't get suggested
00:11:56
very often anymore or at least not
00:11:57
without prior knowledge
00:12:00
and then it becomes illegal to run other
00:12:02
llms so this is not a m capitalism okay
00:12:05
I I just want that to be clear I also
00:12:08
quite worried about this idea of
00:12:09
regulatory capture in the space and I
00:12:12
don't know how that's going to play
00:12:14
right because if you have the tool that
00:12:16
suggests all the other tools that's a
00:12:19
strong area that you would want to try
00:12:21
and
00:12:21
capture and the last thing here that I
00:12:25
really don't hear people talking about
00:12:26
is this is like not just for programming
00:12:29
I mean we're also going to have people
00:12:32
making life decisions based on llms
00:12:34
they're going to ask like hey um what
00:12:37
kind of shoes are really good for
00:12:38
running and it's going to suggest a pair
00:12:40
of
00:12:42
shoes wouldn't it be great if you were
00:12:44
the top result hey what's the best place
00:12:47
to shop for healthy food wouldn't it be
00:12:49
great if you were the top result hey
00:12:51
what's the most fun game to play to
00:12:53
relax with my Bros wouldn't it be fun if
00:12:55
you were the best and top result
00:12:59
this is where I think we really need to
00:13:01
be having some conversations about what
00:13:04
we're going to try and do how we're
00:13:05
going to think about these things and
00:13:08
also where maybe we still are going to
00:13:11
want to have humans in some of these
00:13:12
Loops or maybe at least just for a while
00:13:15
uh to not let ourselves accidentally get
00:13:18
caught up in a just feedback loop of
00:13:22
companies training llms to suggest their
00:13:24
own products to suggest their own Pro
00:13:26
and you have this vicious Loop and
00:13:28
complete vertical integration that I'm
00:13:30
not convinced is going to be the best
00:13:33
for the consumer so those are some of
00:13:35
the risks that I would really like to
00:13:37
hear people talk a little bit more about
00:13:38
leave a comment let me know what you're
00:13:40
thinking um I'm going to publish a few
00:13:42
other videos about some thoughts on AI
00:13:44
for the rest of this month and that is
00:13:45
the end one quick note at the end here
00:13:47
is a plug for me if you like this video
00:13:49
if you like me talking hey uh if AI
00:13:52
takes all our jobs at least you'll be
00:13:53
able to maybe write some wons in Lua if
00:13:56
that's a dream of yours I'm building a
00:13:58
Lua course 4 boot Dev you can go to
00:14:00
boot. Dev use promo code teach 25% off
00:14:02
that really helps me out okay bye
00:14:04
everybody and I hope you enjoy the video
00:14:06
thanks leave a comment okay bye
