Episode Transcript
Transcripts are displayed as originally observed. Some content, including advertisements may have changed.
Use Ctrl + F to search
0:22
Welcome to Unsupervised Learning, a security,
0:24
AI and meaning focused podcast that
0:26
looks at how best to thrive as humans in
0:28
a post AI world. It combines
0:31
original ideas, analysis, and mental
0:33
models to bring not just the news,
0:35
but why it matters and how to respond.
0:42
All right. So the first thing for this week
0:44
is I finally turned my 9000
0:47
word I predictions essay into
0:49
an actual video. So if you click on this
0:51
thing all right I'm going to try to do
0:53
something crazy right now I'm going to try.
0:56
This is the whole video I
0:59
go through. I thought it was going to be like 30 minutes.
1:01
I'm like, knows what we want because it has
1:03
all the prefaces in, it knows our journal,
1:05
it knows all the stuff that we plus like
1:07
two hours of editing, which I never should have done.
1:09
But anyway, it is a full
1:11
walkthrough of the illustrated
1:14
essay, and I even give
1:16
a lot more sort of expansion
1:19
narration on top of the text
1:21
that's in the essay, so it's quite good, if
1:23
I do say so myself. I've had a lot of people
1:25
say they enjoyed it, so I would
1:27
say, go check that out. It gets
1:29
me excited to even think about it. Talk about it. It's
1:32
like the fun stuff. Plus I remember
1:34
like doing the art. It's just super
1:36
good time. Okay. Security. Yeah. But
1:38
so go check out the video. All right. Security.
1:40
So Israel's top spy,
1:43
chief Jose Ariel,
1:45
accidentally revealed his identity
1:47
through an Amazon book sale. So I guess we're
1:49
just doxing him now. That's rude.
1:51
Well, I guess he's not going to be that anymore, because
1:54
they're going to be like, you obviously can't do this. So they
1:56
probably got rid of him and replaced him with someone
1:58
who is still anonymous. So
2:00
now he's no longer the spy chief. He's just
2:03
a guy doing an Amazon book sale. All right,
2:05
cvn, Nvda databases are struggling.
2:08
Um, it's causing gaps and inaccuracies.
2:11
So somebody needs to fix that. Uh,
2:13
criminals in Montreal are using Apple's
2:15
AirTags to track and steal cars.
2:17
And police are using Apple's
2:19
AirTags to track criminals. So
2:22
a lot of AirTag use on
2:24
both sides of the fence there. And this one's
2:26
super cool. This sponsor I want to talk about this one
2:28
because I'm actually advising for them the first
2:30
AI SoC analyst that autonomously
2:33
investigates alerts. This thing will
2:35
pull alerts. I've seen this live
2:37
pull alerts. And this is a sponsor, by
2:39
the way. Sponsor. And I'm an advisor for them.
2:42
So I'm just excited about them. It's
2:44
called drop zone AI. So you
2:46
take an alert from anywhere in your stack,
2:48
some tool that you have that produces an alert
2:50
like an endpoint cloud doesn't
2:52
matter. And it
2:54
takes alerts from your environment and performs
2:57
autonomous. It basically goes and starts
2:59
researching exactly like a human. It's
3:01
like starts digging in, boom, get some results
3:04
from that lookup and that investigation
3:06
takes that goes to the next step and
3:08
just starts doing multiple steps, just
3:10
like a regular SoC analyst, just like a human.
3:13
And once it gets done, it puts that all
3:15
together and generates a decision
3:17
ready report. So you could
3:19
basically use that. And
3:22
it's basically exactly as if you had got
3:24
that from a SoC analyst. And you can make a decision
3:26
based on that. And it's no playbooks,
3:29
no code, no prompts required.
3:31
It just you feed it alerts and
3:33
it goes and does investigations and brings
3:35
back a report. It's absolutely insane.
3:38
It's exactly what I've been talking about as
3:40
being like how AI is most
3:42
going to affect security.
3:45
And I've been saying for the longest time
3:47
it's all going to be agent based. And that's exactly
3:49
what this is. And like I said, it's so good.
3:51
I just became an advisor and strop sonar
3:53
AI and you could actually go and see a demo
3:56
of it. Working on a real alert.
3:58
Panera's week long incident
4:00
was, in fact, a ransomware attack.
4:03
Cece's new High Risk Communities webpage
4:05
offers a bunch of guides and volunteer
4:07
support and discounted tools, thanks
4:10
to Defender Fee for sponsoring as well.
4:12
Israel's military used an AI named
4:14
lavender to pinpoint 37,000
4:17
potential Hamas targets, and that's
4:19
definitely getting people pretty upset about that,
4:21
because it's one thing to identify someone
4:23
and do further investigation. It's another thing
4:26
if you're identifying and then like targeting.
4:28
And again, I don't know if that's actually happening.
4:30
I don't know the degree they're trusting
4:32
this targeting that's coming up from lavender.
4:35
I don't know if they're going directly to
4:37
attacks or killing or anything
4:39
like that. But the point is it's
4:41
worth having a conversation about what they're
4:43
doing with it. Given the fact that
4:46
I can be super fast and
4:48
super accurate, but it can also have a lot of flaws,
4:50
and the more serious you're taking the results,
4:53
the more important those flaws are, right? Technology.
4:55
There's a rumor that Sam Altman and Johnny IV
4:58
are building some sort of handheld device
5:00
through a new secret company. Who knows if
5:02
that's a real or not. It sounds
5:05
real to me because we know that Johnny IV is
5:07
actually working with them. We know he's a design
5:09
person. We know we need these types
5:11
of devices. So it makes sense to me.
5:13
OpenAI released improved ways
5:16
of doing fine tuning. Uh, new
5:18
paper shows that adding more agents to
5:20
large language models boosts their performance.
5:22
How could it not? How could it not?
5:24
I mean, like like I've been telling
5:26
everyone agents is the way, agents
5:29
is the way this is all going to move forward. And
5:31
then each individual agent when it
5:33
gets smarter because the AI models improve,
5:35
that just makes the whole thing smarter. But the way
5:38
to get ultimately smart and actually pass
5:40
humans is to have that combination
5:42
of agents working together. That system
5:45
is what's smart. It's the same as our brain is
5:47
no different than our brain. Our brain has a whole bunch
5:49
of areas, each individually
5:52
only do kind of one thing, and
5:54
they're not very smart by themselves.
5:56
The whole thing that's made it smart over
5:58
all these millions of years is the fact
6:00
that they work together and they're sharing
6:02
information, and it's like it's a system.
6:05
The system is what makes us smart. And
6:07
it's going to be no different with agents.
6:09
Now, eventually you might have an AI
6:12
that's just one agent or one
6:14
model or one agent or whatever.
6:16
And it's so smart. It's just smart by itself
6:18
because it's different than a human. But the
6:20
way we're first going to get to AGI, mark
6:23
my words, is going to be through a system
6:25
of AI components working together.
6:27
And I've got a blog on that somewhere. New paper explores
6:29
how I might be leading us towards
6:32
knowledge collapse by oversimplifying
6:34
complex information. I like that,
6:36
but I don't think the complex information
6:38
goes away. Just because somebody is
6:40
simplifying or summarizing
6:43
complex information, it doesn't mean
6:45
we have to stop feeding the AI
6:48
the original raw form. One
6:50
doesn't have to supplant the other, and
6:52
we could just ask for the long form. We can
6:54
ask for the short form. I don't see that as a
6:56
trade off us is trying to get South Korea to
6:58
stop exporting chipmaking tools
7:00
to China, US testing, energy
7:03
storage and heated sand. This
7:05
tech is like trippy to me. Energy
7:07
storage and heated sand 135MW
7:12
of power output for five days
7:14
straight and Aura's rolling
7:16
out symptom radar. They're not calling it illness
7:18
detection for obvious reasons. I
7:20
haven't looked to my app to see if I have that yet.
7:23
I do have an aura though. Okay, Amazon
7:25
is ditching its Cashierless walkout
7:27
technology is kind of sad. I guess
7:29
it's too early, but they're switching to something
7:32
a little more normal, like a hybrid humans.
7:35
New studies are showing the wealthy are starting to have
7:37
more kids than the poor. Again,
7:39
need lots of studies to show that's actually
7:41
true, but interesting trend. NASA
7:43
is doing live streaming the eclipse. I watched it,
7:45
it was super cool. It was actually
7:47
moving through multiple cities and you would see
7:50
each person, each city as
7:52
it was moving. It would go to the full
7:54
totality and everyone would freak out.
7:56
It was quite cool. It was a very cool stream.
7:58
TSMC did not take much damage
8:01
at all. They were actually only down for like
8:03
a day and they went back to start
8:05
full production. And it's because their
8:07
buildings have some really cool stabilization
8:09
tech that allows them to not get too
8:11
messed up from earthquakes. However,
8:14
the earthquake was on the other side of the island,
8:16
so who knows if it's a bigger
8:18
like it was an eight and it was nearby.
8:20
I'm not sure the buildings would be
8:22
able to handle that, but who knows. Israeli
8:24
military dismissed two senior officers
8:26
and reprimanded three others for an
8:28
airstrike that mistakenly killed World
8:31
Central Kitchen volunteers in Gaza.
8:33
The UK's exporting workers
8:35
to fill higher paying US jobs
8:38
and US venture capital investments
8:40
went down $36.6 billion
8:43
in 2024. Had a really cool
8:45
conversation with Mike private from Return
8:47
on Security about these overall economic
8:50
trends, especially around cybersecurity,
8:52
and that will go up soon. McKinsey
8:54
is offering UK employee UK
8:56
employees nine months of pay to voluntarily
8:59
leave McKinsey. Gen Z is going
9:01
for trades like welding and plumbing over
9:03
college and student debt, and home insurers
9:05
are now using aerial images from
9:08
satellites to decide who
9:10
gets dropped from coverage. I assume that's
9:12
because maybe somebody's building something
9:14
in their backyard, or they have
9:16
too many cars. Like, I'm not sure what would be
9:19
on the policy that they'd be able to see in a satellite
9:21
photo, maybe add ons to the house
9:23
or something, I don't know. And all right,
9:25
got a couple cool ideas here. Another
9:27
view of imposter syndrome. So somebody
9:29
said outwork your imposter
9:31
syndrome. That was the post outwork your
9:34
imposter syndrome. And I'm, like, working
9:36
harder isn't the solution. In my opinion,
9:38
the solution is to work on bigger
9:40
problems that are super important
9:42
to solve. That way your focus
9:44
isn't internal, it's actually external.
9:47
And I actually go into this, so I'm going
9:49
to go ahead and open this up. So I
9:51
say framed this way, imposter syndrome is
9:53
ultimately a problem of thinking too much internally
9:55
versus externally. Because you're thinking how
9:58
do I compare to them? What do they think
10:00
of me? Right. And so the solution I
10:02
think is to. Focus
10:05
on. Something
10:08
outside of you, right? Stop
10:10
putting the focus on yourself. How
10:12
do I compare? What do they think of me?
10:15
That's thinking about yourself. Instead,
10:17
focus on the problem. And now you
10:19
won't really have imposter syndrome because
10:21
you're thinking about, like, what am I doing
10:23
that's useful to help me with this really
10:25
big problem, which has nothing to do with me.
10:27
It's about me working on that
10:30
thing. So instead, focus on the problem
10:32
and how to fix it. And this also works
10:34
for happiness. It doesn't come from focusing on self,
10:36
it comes from focusing on other. So
10:38
that was that piece there. And
10:40
this one's really interesting. It's the first time
10:42
I've ever seen Tyler Cohen be wrong. It's
10:44
like one of the smartest people that I know of
10:47
on the planet. And he had
10:49
Jonathan Height on and
10:51
Jonathan High was talking about how bad
10:53
social media is for young
10:56
girls and how
10:58
I could potentially make that worse and everything.
11:00
And Tyler Cohen is like, don't worry about
11:02
it. AI is going to solve this because
11:05
I is going to summarize everything, and
11:07
the summarization of of social
11:09
media is going to solve this problem and
11:11
fix it. And I'm like,
11:14
how are you not seeing this now?
11:16
Jonathan didn't have this point, but
11:18
I'm going to make this point right now. That is how
11:20
Tyler Cohen is going to use
11:23
AI, right? It's how I'm using AI right
11:25
now, and it's how, you know, Jonathan's probably
11:27
going to use it as well. Summarization of
11:29
content. So you're just getting the thing.
11:31
And that might actually take you away from being
11:34
caught up in like, oh he said she said
11:36
and all the drama or toxicity or whatever.
11:38
Well that's fine. And that again,
11:40
that's how they're going to use it. That's how a
11:42
whole bunch of intellectual people who have made
11:44
it in life and are older are probably
11:46
going to use it, but not for young
11:49
people consuming viral and toxic content,
11:51
because for them, the content itself is the
11:53
point, not the summary. And here's
11:55
my example of this. Does Tyler
11:58
think I will send people who
12:00
love stand up comedy a summary of
12:02
the jokes made in a given
12:04
standup as a substitute for going
12:06
to actual comedy shows. And I've got
12:09
I've got an I summary here. Here's your summary
12:11
of this standup. There were three jokes on
12:13
women in stereotypes, four jokes on
12:15
how clumsy the comedian is
12:18
to playful racist jokes. Two hecklers
12:20
were addressed. Applause was three
12:22
out of five compared to other performers.
12:24
We hope you've enjoyed this hilarious
12:26
I summary from comics I.
12:29
Does that work for comedy? No,
12:31
that doesn't work for comedy. And it won't work
12:33
for young kids consuming viral or
12:35
toxic content unless you have
12:37
some sort of like draconic like,
12:40
I guess, massive control
12:42
where they weren't actually allowed
12:44
to go to the real thing. They
12:46
couldn't go to TikTok, they can't go to Instagram,
12:49
and all they get is the stupid summary.
12:51
Nobody would. They they wouldn't even care about
12:53
the summary. I mean, that's not even going to be a thing.
12:55
Nobody cares. And if
12:58
it did, they wouldn't use it. And if
13:00
they had the option or if they had both,
13:02
they wouldn't choose the summaries. They would
13:04
not use the summaries at all, and they would just go to the
13:06
original thing. Now that being said,
13:08
it is Tyler Cohen. So there's
13:10
a chance that I didn't understand what he was saying.
13:13
And also Jonathan definitely didn't either.
13:15
So he could be misunderstanding Tyler's
13:17
point. So I want to offer him that.
13:19
There's also the other thing, which is you
13:21
might just be right, and I'm just wrong.
13:23
Now, I wouldn't normally say that because I
13:25
think fairly highly of my thinking
13:27
capabilities, but it's Tyler Cohen,
13:29
so I'm leaving that window a little
13:32
more open than usual. All right. Deep faked
13:34
content summaries. Oh, yeah. This is crazy.
13:36
So I don't know. I don't know if I woke up with this
13:38
idea. I think the main interface
13:41
that we're about to have for content,
13:43
let's say, okay, I got buddies
13:45
who make content. John Hammond makes content.
13:48
Jason Haddix makes content. Clint
13:50
Gibbler makes content. So they're putting out,
13:53
let's say they're putting out videos, they're putting out
13:55
text, they're doing live talks.
13:57
They're doing all these different things, different mediums,
13:59
different formats. I think
14:01
what's going to be happening
14:04
very soon is let's
14:06
do the R version, even though we don't have
14:08
R yet. The earlier version will just
14:10
be little videos on your phone. But let's
14:12
do the R version or the R version
14:14
combined with a digital assistant, which
14:16
is AI in your head or in your
14:18
mobile device. So it's essentially you say,
14:20
hey, look, what is Jason been up
14:22
to? What is John been up to? What is Clint
14:24
been up to? And it will actually
14:27
deepfake them okay. Because it knows
14:29
how much time I have. Let's say
14:31
I want a two minute summary. It will
14:33
take the long presentation
14:35
that Jason is going to do. Jason is getting
14:37
ready to do an AI talk called Red purple,
14:39
blue AI. Let's say it's actually it
14:41
is like it's a class. Oh, by the way, you should go sign
14:43
up for this class. His class is called red
14:46
purple blue I it's
14:48
going to be an amazing class. I've heard a lot about it.
14:50
I'm going to be in it as well. But for
14:52
example, let's say he gives a talk about that
14:54
a public talk. The other one's not public. But
14:56
let's say he gives a talk about that. Actually he's
14:58
doing a space con coming up soon.
15:01
So that's a good example. Let's say I don't have the
15:03
one hour or the two hours to watch that. I only
15:05
have two minutes. I'm about to get on a train where I don't
15:07
have any connection. Whatever it's going
15:09
to make, Jason and Jason
15:11
is going to. Teach me what he covered,
15:13
but he's going to do it in 30s or he's
15:15
going to do it in 60s. It's going
15:17
to be a deepfake of Jason doing
15:20
Jason's own content. Same for John.
15:22
John Hammond has a thing about this new piece of
15:24
malware. Clint has this new thing about
15:26
this new tool that he made or that he
15:28
saw, and he's talking about deepfakes
15:30
are going to allow the actual
15:33
creator to scale their
15:35
delivery of their content as
15:37
a video to any size
15:40
chunk that the consumer wants to see.
15:42
Give me a ten second summary. Give me a 32nd
15:44
summary. Give me a one minute summary,
15:47
a two minute summary, a ten minute summary, whatever.
15:49
But it will dynamically write the
15:51
content, which is a summary
15:53
of the content, which smashes
15:55
it down to that size, and then it will
15:57
perfectly deepfake that
15:59
creator and the mouth
16:02
will match, the mouth will perfectly
16:04
match. It'll look exactly
16:06
or almost exactly and eventually
16:08
exactly like the creator. Now,
16:10
why would the creator want to do that? Because it's
16:13
still their content, okay? It's still their content.
16:15
It's not the original raw form, but
16:17
in a lot of ways it's going to be better because it's not going
16:19
to have ums and ors. It's going to be very
16:21
crisp and concise, and most importantly,
16:24
it won't be some third party, right?
16:26
Because my digital assistant can also
16:28
render it using their face. But it'll
16:30
be really cool for the for the actual
16:32
creator to be still the one
16:34
that's delivering it. And what's
16:37
cool about that? So here's what's really cool
16:39
about that. It'll do multiple languages.
16:41
So now it's still Clint. It's
16:43
still John. It's still Jason. Jason
16:45
is still saying the thing to me.
16:47
Pedro diciendo
16:49
la content.
16:50
In.
16:51
Espanol. Entonces puedo escuchar.
16:53
En espanol and I won't know the difference.
16:55
It'll look exactly as if Jason
16:58
is speaking in Spanish instead of English.
17:00
And I think that is wonderful. Same with Chinese,
17:02
same with every language. Not any more difficult.
17:05
So now your deepfake content
17:07
is going out to every language
17:09
all at once, as soon as you drop a piece of
17:11
content. So watch this. You drop
17:14
a piece of content. It's one hour long.
17:16
Okay, perfect example. I just released a
17:18
video. It's one hour and ten minutes
17:20
long. I only did it in English because I'm only
17:22
capable of doing it in English. I could do it partially
17:25
in Spanish, but it would be kind of crappy.
17:27
So that one hour
17:29
and ten minute piece of content can now
17:32
be put in Hindi, Spanish,
17:34
Mandarin, Cantonese, Filipino,
17:37
like all these different languages, but not
17:39
just converted one for one
17:41
to the language. Also converted
17:43
to a ten second version, a 32nd
17:45
version, a one minute version. And
17:47
it's still me. It's still my face, it's still
17:49
my whatever gestures, all this,
17:52
it's still me delivering the content. That
17:54
is a game changer. An
17:57
absolute game changer because
17:59
it's scalable content delivery in
18:01
every language, at every bite sized
18:03
chunk. So now when I'm on a train or
18:05
I'm exercising, I could have my
18:07
AR interface. I see the person
18:09
is delivering the content and it's just like,
18:12
I'm watching that now. Maybe there will be some
18:14
sort of tag that says, there's this roar, or
18:16
is this an AI deepfake of it?
18:18
But it's not even quite a deepfake. It's
18:20
using deepfake technology. But
18:22
a deepfake is like almost unauthorized.
18:24
This will be completely authorized. Um, and
18:27
of course, there will be versions of this that are not authorized,
18:30
and that will be an actual deepfake. But anyway,
18:32
you get the idea. This is going to massively
18:34
change how we consume content,
18:36
because we want that content in different forms,
18:39
different languages, but we still want to see
18:41
the creator doing it. All right. AI and music.
18:43
AI is not going to ruin music. Have
18:46
we forgot about pop? Pop is,
18:48
you know, a few chords and a hook.
18:50
It's, you know, a lot of people would say it's
18:52
low quality music, like
18:54
low quality food. But it's
18:56
the same with doing customer service calls, sales
18:58
calls. We forget how low the
19:00
bar is for being better
19:03
than an average human, and that's why I
19:05
music is going to be pretty good. And actually
19:07
I think I have a link to
19:09
one here anyway. Discovery section
19:11
Luke Stefan's hack Luke put out an
19:13
amazing blog talking about his
19:15
evolution of bug bounty automation.
19:18
Talks about going from Bash to Python to Golang,
19:20
and how he eventually ended up at Cloud
19:23
Native. Really great piece, Thomas
19:25
wrote. He's super awesome and
19:27
he wrote a piece about applying LMS to
19:29
Threat Intelligence. Really cool. It's got a
19:31
Jupyter notebook here for the code as well.
19:34
SWE agent autonomously fixes
19:36
bugs in GitHub repos BR
19:38
is a Python framework. Simplifies building
19:40
an AI apps using building
19:43
blocks. Open source textbook
19:45
makes the art of mathematics accessible.
19:48
Jim Graham turns threat modeling
19:50
into a self-hosted web app. This
19:52
one is super, super cool. Chelsea
19:54
now lets you tweak images so you
19:56
can basically have an image that was created
19:59
with Dall-E and you could just like rub a like
20:01
the face or whatever and say, redo that part
20:03
and it will redo it, but integrate it with
20:05
everything around it. So that is something
20:07
that Midjourney had that that Dall-E didn't,
20:10
and now Dall-E has it as well. Claude's
20:12
API now has a new
20:14
tools feature that allows you to, like,
20:16
browse the web and do all kinds of stuff. I
20:18
kind of feel like agent functionality is going
20:21
to blend right into
20:23
the models, so I'm not quite sure
20:25
how long we're going to have, like these
20:28
elaborate agent frameworks,
20:30
because that might just be part
20:33
of using AI. You
20:35
might just say exactly what you want
20:37
to do in the prompt, and that will actually
20:39
be the agent framework. It'll actually spin
20:41
up how many agents are needed to do that.
20:43
And maybe you have some parameters right there
20:46
in the prompt that does it. But I kind of feel
20:48
like that's going to just blend right into
20:50
the language of the of the models
20:52
themselves. And kids are learning math
20:54
from deep fakes. Okay. This is another
20:56
example of this learning math from deep fakes
20:59
of Taylor Swift and Drake on TikTok. And
21:01
it's going well. I'm excited for this.
21:03
I am super excited for this look.
21:05
There are negatives that are going
21:08
to come from defects. Everyone knows that. Let's
21:10
find the positives. The positives are
21:12
if people love Taylor Swift, let's learn
21:14
math. Let's learn calculus from Taylor
21:16
Swift. I would do that. All right. Recommendation
21:18
of the week. Check out Mozart
21:21
on the bass. This is an eye track.
21:23
Yeah. Go listen this tell me that this won't
21:25
be popular. It's a little bit EDM ish.
21:28
So if you don't like that, just compare it to other
21:30
EDM that you don't like, and you'll see that
21:32
I is actually quite good at making
21:34
stuff you don't like. All right, aphorism of the
21:36
week don't explain your philosophy. Embody
21:39
it. Don't explain your
21:41
philosophy, embody it. Epictetus.
21:44
Unsupervised learning is produced and edited by
21:46
Daniel Missler on a Neumann 87
21:48
AI microphone using Hindenburg.
21:51
Intro and outro. Music is by zombie with
21:53
the Y, and to get the
21:55
text and links from this episode, sign up for the
21:57
newsletter version of the show at Daniel missler.com/newsletter.
22:03
We'll see you next time.
Podchaser is the ultimate destination for podcast data, search, and discovery. Learn More