00:00:00
hi friends my name is Tris and this is
00:00:01
no Bower plate where I make fast
00:00:03
technical videos today I'm going to
00:00:05
explain my framework for thinking about
00:00:06
AI tools what they're great at what
00:00:08
they're not so good at why they don't
00:00:09
live up to their claims and what to do
00:00:11
about that so Apple intelligence has
00:00:13
been out for a couple of months now but
00:00:15
like a lot of AI promises it's Fallen a
00:00:17
little short right the AI hype train is
00:00:20
driven by the tantilizing promise of AGI
00:00:22
general intelligence like we see in the
00:00:24
movies Maria Hal Marvin Johnny 5 seeo
00:00:27
Rachel and deard Holly Jarvis and Wally
00:00:29
but desite four years of promises Apple
00:00:32
intelligence is the latest example of
00:00:34
these products missing the mark the best
00:00:36
two things that Marquez here has to say
00:00:38
about Apple intelligence are one the
00:00:40
background Aras a tool is pretty good
00:00:42
and two it has bumped up the base Ram
00:00:43
across all of Apple's Hardware lineup
00:00:45
this is well overdue as I mentioned in
00:00:47
my video on the unreasonable
00:00:48
effectiveness of Linux workstations cets
00:00:50
did not hold back on their criticism
00:00:52
either Apple intelligence was announced
00:00:53
at WWDC in June 2024 but didn't ship
00:00:57
with the brand new iPhones and the other
00:00:59
Hardware that was announced then and
00:01:00
only after months with these strangely
00:01:02
mediocre features released to us the
00:01:04
really good stuff is coming we are
00:01:06
promised and I believe we have heard
00:01:08
that before my video scripts are
00:01:10
dedicated to the public domain
00:01:12
everything you see here script links and
00:01:13
images are part of a markdown document
00:01:14
available freely on GitHub at the above
00:01:16
address part one language is important
00:01:20
to understand what is happening with AI
00:01:22
let us tighten up our definitions and
00:01:24
give credit to what does work well
00:01:26
artificial intelligence is a large
00:01:27
discipline containing many fields with
00:01:29
applic that are already so well
00:01:31
integrated with our tools that we forget
00:01:33
about them searching our photos by the
00:01:35
contents of the photo instead of far
00:01:36
name or date is AI near perfect at least
00:01:39
in English voice recognition is AI
00:01:41
generative fill for editing out unwanted
00:01:43
parts of images is also AI these
00:01:46
features are all AI tools but we don't
00:01:49
typically call them that like when
00:01:51
alternative medicine is proved to work
00:01:53
we call it medicine when AI Tech works
00:01:56
we stop calling it AI it fades into the
00:01:58
background of our normal Computing this
00:02:01
video is about generative AI large
00:02:03
language models and GPT the technologies
00:02:05
that the companies promise much with but
00:02:08
deliver surprisingly little large
00:02:10
language models like chat GPT are great
00:02:12
at comprehending language for instance
00:02:15
I've never seen such a great thesaurus
00:02:17
you can just describe the feeling you
00:02:18
want to convey and get 10 reasonable
00:02:21
words or phrases back but start to use
00:02:24
it for knowledge not language and you
00:02:26
get into trouble chat GPT 4 got this
00:02:29
question partially wrong and so did
00:02:31
Claude at the time of writing Gemini got
00:02:33
the right answer a harpsicord but did
00:02:35
not also identify the second instrument
00:02:37
a melron which chat GPT did the more
00:02:40
specific the answers you want the less
00:02:43
reliable large language models are it
00:02:46
reminds me of the demon cat from
00:02:47
Adventure Time which has approximate
00:02:50
knowledge of many things it's very
00:02:52
confident but often inaccurate this
00:02:54
trend exists across all the GPT tools I
00:02:56
have tested from cloud providers such as
00:02:58
open AI to running and tweaking my own
00:03:00
local models with olama but that's fine
00:03:03
there's so much value on the left side
00:03:05
of this graph for initial research and
00:03:07
shallow exploration you can absolutely
00:03:09
use a GPT tool to quickly find areas you
00:03:11
want to look into deeper for yourself
00:03:14
however there are real limits in these
00:03:16
generative techniques that you come up
00:03:17
against very soon after you start using
00:03:19
them for complex work let's talk about
00:03:21
where these limits come from and how to
00:03:23
avoid them it's just me running this
00:03:25
Channel and I'm so grateful for everyone
00:03:26
for supporting me on this wild adventure
00:03:28
if you'd like to see and give feedback
00:03:30
on my videos up to a week early as well
00:03:32
as get private Discord access and even
00:03:33
your name in the credits it would be
00:03:35
very kind of you to check my patreon I'm
00:03:37
also offering a limited number of
00:03:39
mentoring slots if you'd like one toone
00:03:41
tuition on personal organization rust
00:03:43
creative production web Tech or anything
00:03:45
that I talk about my videos do sign up
00:03:47
and let's chat part two the magic beans
00:03:50
don't work because they don't have to
00:03:53
GPT is a Marvel of natural language
00:03:55
processing autocorrect that is trained
00:03:58
on the whole internet can almost always
00:03:59
offer sensible suggestions about what
00:04:01
should come next in a sentence but
00:04:03
language ability as we learned in Star
00:04:05
Wars does not equal intelligence the
00:04:09
problem is that we are extremely
00:04:11
language Centric creatures and we
00:04:14
mistake language proficiency for
00:04:15
intelligence which causes us to misuse
00:04:18
this technology or for this technology
00:04:21
to misuse us you're not chatting to an
00:04:24
intelligent agent it's autocom
00:04:26
completing your questions like a
00:04:28
sociopath getting under your skin by
00:04:30
saying what it thinks you want to hear
00:04:32
large language models can only learn
00:04:33
topics where there is a large amount of
00:04:35
language available to train them for
00:04:37
example the reason llms can't
00:04:39
autocomplete maths is because apart from
00:04:41
trivial examples the state space of all
00:04:44
numbers is too great to expect much
00:04:46
existing training data contrast how
00:04:48
often 1+ 1 equals 2 is written in
00:04:50
textbooks easy for chat GPT to complete
00:04:53
with how often say 2 e^2 + 5 j = 0 is
00:04:57
one of these has a large amount of
00:04:59
existing natur language data available
00:05:00
for training one does not this is why
00:05:03
tools like chat GPT seem good at first
00:05:06
when you ask it simple questions but as
00:05:08
you dig deeper they fall apart and get
00:05:11
increasingly inaccurate or hit
00:05:12
artificial guardrails and only provide
00:05:14
surface level responses it's not that
00:05:17
the technology is new and will
00:05:19
eventually get better it's that this
00:05:21
incredible language ability can only
00:05:23
work after being trained on large
00:05:25
amounts of data by definition there
00:05:27
might be only a single PhD written about
00:05:29
a very Niche topic so GPT will never
00:05:32
learn that information because a single
00:05:34
PhD paper is not a large amount of
00:05:36
language and if you don't have a large
00:05:38
amount of language you can't train a
00:05:40
large language model AI companies can't
00:05:43
fulfill their wild promises so why do
00:05:46
they make them part three we are not the
00:05:49
audience I was confused by the distance
00:05:51
between the hype and reality until I
00:05:53
realized we're not the audience for all
00:05:55
this breathless hype as I've shown if
00:05:58
you take the claims at face value these
00:06:00
Technologies simply don't work and
00:06:02
things that don't work can't solve
00:06:04
problems and you can't sell someone
00:06:06
something that doesn't solve their
00:06:07
problem not twice anyway so why do the
00:06:10
companies keep making these promises
00:06:13
well it's not demand from the customers
00:06:16
nor direction from the engineers nor
00:06:17
even really by choices made by their
00:06:19
CEOs it's because the real decision
00:06:21
makers in these companies are their
00:06:23
wealthy investors broken or mediocre AI
00:06:26
tools that we all hate have been crammed
00:06:28
into everything we use now even our
00:06:30
notifications because the companies have
00:06:32
to impress investors with AI features
00:06:35
even if they don't work well when you
00:06:37
work in Tech startups as I have over the
00:06:38
past 15 years you get to know the
00:06:40
startup Runway very well I wasn't always
00:06:43
as a lowly engineer privy to the actual
00:06:44
amount of funding coming in and salaries
00:06:46
going out each month but we would all be
00:06:48
able to feel when the end of the runway
00:06:50
was in sight you can typically extend
00:06:52
your runway in two main ways one selling
00:06:55
products and services to users for money
00:06:57
or two persuading investors to part with
00:07:00
more of their money selling products is
00:07:03
hard they have to work but selling a
00:07:06
promise that's easy plange perhaps I'm
00:07:10
not so sure it feels different this time
00:07:13
with AI gen AI is this perfect tool for
00:07:16
tricking investors out of their money
00:07:17
because often enough the people asking
00:07:19
for the money and their customers think
00:07:21
it works too llms are great at basic
00:07:24
stuff and in the past when a computer
00:07:26
could automate basic tasks to a good
00:07:27
degree it only required time time
00:07:30
improvements and of course money to
00:07:32
perfect as an investor surely you'd
00:07:33
better get in on the ground floor of
00:07:35
this marvelous new technology just as
00:07:37
you have before what can I conclude from
00:07:40
this from colonies on Mars to
00:07:42
democratizing money it's always easier
00:07:44
to promise a bright future than build a
00:07:46
better present what I remind myself to
00:07:48
do whenever I see these bizarre products
00:07:50
that no one needs is to pay less
00:07:52
attention to what these companies say
00:07:54
their Tech will do in the future and far
00:07:57
more to what they actually can do today
00:08:00
thank you if you would like to support
00:08:02
my channel get early adree and tracking
00:08:04
free videos your name in the credits or
00:08:05
onetoone mentoring head to my patreon or
00:08:07
Kofi if you're interested in
00:08:09
transhumanism and hope Punk please check
00:08:11
out my weekly sci-fi audio Fiction
00:08:13
podcast lost terminal season 2 of the
00:08:14
fosen catalog is broadcasting now if you
00:08:16
like Mysteries and art check it out
00:08:18
transcripts and compiled checked
00:08:19
markdown source code are available on
00:08:21
GitHub links in the description and
00:08:22
Corrections are in the pinned irata
00:08:24
comment thank you so much for watching
00:08:26
talk to you on Discord
