STANDARD EDITION: 28 Open Cyber Jobs, Real-world AI Propaganda Poisoning, MCP Explained, Cline vs. Windsurf, and more...
You are currently listening to the Standard version of the podcast, consider upgrading and becoming a member for the full version and many other exclusive benefits here: https://newsletter.danielmiessler.com/upgrade
Subscribe to the newsletter at:
https://danielmiessler.com/subscribe
Join the UL community at:
https://danielmiessler.com/upgrade
Follow on X:
https://x.com/danielmiessler
Follow on LinkedIn:
https://www.linkedin.com/in/danielmiessle
Become a Member: https://danielmiessler.com/upgrade
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
S1 (00:00):
Unsupervised Learning is a podcast about trends and ideas in cybersecurity,
national security, AI, technology and society, and how best to
upgrade ourselves to be ready for what's coming. All right,
welcome to episode 472. Hope your week is going well.
(00:22):
Updates on this side I've got an RSS feed which
I encourage you to go sign up for. RSS is
a real thing and it's awesome. And I wish it
didn't get knocked off the radar by Google Reader going under,
but I use Feedly. A lot of people use Netnewswire.
Whatever your RSS reader is, go check it out. It's
(00:45):
just slash feed RSS, I think at the end of
the URL. Thanks for all the cooking advice. Around 500
people responded with experience, encouragement, recommendations. It was just insane.
So really appreciate all that. I tried windsurf, a lot
of people are talking about it. I basically try all
these tools that people talk about. I've been using Klein
(01:07):
for a while. I was using cursor before that and
I keep trying cursor every time they have a big
release or when windsurf has a big release. I'll try
it again. The stuff is changing so fast that I
just constantly try different stuff. But Klein has been my
go to. I feel like it manages the functionality of
(01:29):
Lmms better and the UI. I just like better, so
that's what I've been using. It is just an extension
into VS code. I wish it was actually a full editor,
just like cursor and windsurf are. So Klein people. If
you're listening, please make an editor. And my personal AI
infrastructure is getting absolutely insane right now. I'm going to
(01:52):
have something to show on this soon. I'm going to
do a demo of like how this all works together,
but basically it's one main agent and then all my
personal services are sitting behind it that do different things,
and they're all able to access different services that do
output as well. So there's like you send the request
(02:13):
into the agent, the agent, if I tell it, send me,
you know, test a website and then send me an
email for it, it'll do the two different services. One
is to test the website and the other one is
to send an email. And again it just naturally parsed that.
And I just sent that from my phone. Right. So
(02:33):
the whole idea is I'm just using my regular tech infrastructure,
which is my iPhone, and I send a request into
my agent, and this is what it's going to be
in the future too. I'm just sort of hacking it
together with my own services. Pretty soon it'll just be
built into Google and built into the iPhone and stuff
like that. But right now I'm just making it happen
(02:55):
through like shortcuts and a bunch of AI services, but
the overall effect is quite awesome and you have to
see this thing actually working. It's very much in line
with what I've been talking about. Uh, in the predictable
AI post, which is this link here where it's all going,
(03:16):
put out a concise explanation of model context protocol, MCP
servers and why everyone's excited about them. So I got
a full thread on that and you should go check out. And, uh,
I'm enjoying the fact that this is starting to happen,
specifically the MCP server, because in that predictable AI post,
(03:37):
which actually comes from the book in 2016, I talked
about the API ification of everything and specifically how businesses
would all become APIs. So this is text. From that,
most software businesses will become algorithms presented to others through
their business daemons. Many traditional businesses will continue to become
(03:58):
software businesses. So basically, Andreessen talked about software as eating
the world. I think AI is going to eat the world.
Like that's pretty. Everyone's saying that it doesn't really mean anything.
What I think is actually eating the world is APIs.
And the fact that AI is what's going to interact
with all these APIs. The primary interface for doing anything
(04:21):
is going to be APIs, because it's the AI that's
parsing them, right? What humans need to look at a
website or a list of services, or a menu of
food or whatever. What humans need for that is way
different than what a computer needs. A computer can parse
an API or an MCP service directory, you know, basically
(04:46):
instantly and know exactly how to use it. So if
you if you think about SEO, if you think about UI, UX,
if you think about advertisements, if you think about websites,
if you think about News media, all of that. It's
all right now focused on presenting that to humans. The
fundamental change that's happening right now is that's all switching
(05:06):
away from being primarily designed for humans to being primarily
designed for AI. Primarily, these interfaces will be for AI
to find a business, AI to find a service AI
to rank those services. Right. So examples will include and
this is keep in mind this text right here. This
(05:26):
is from 2016 when I wrote that stupid book which
you probably don't need to read anymore because it's all
in that, that post and it's got pictures which took
me like hours to make. So it's a better version
of the book that blog post is. But the key
point here is I was talking about this in 2016,
so I'm looking for a little like, I don't know,
(05:48):
somebody to send me a pat on the back. That
was kind of the whole purpose of writing that stupid
book was to get this stuff locked in, because I
saw this happening so clearly, even though that was way
before this AI stuff popped. Right. So what is 22
the end of 22 versus the end of 2016? That's
six years. Six years. I basically said, we're going to
(06:11):
have digital assistants, which are AI powered. I talked about
the AI component right there in there. It's going to
know everything about us. We're going to talk to our Da.
Our Da is going to read all the services. Every
human is going to have a demon. Not every human,
but most humans. A lot of a lot of humans
are going to have demons. But most importantly, all the
(06:31):
businesses have demons. We talk to our Da. The Da
infers what we probably want. We don't even need to
talk to it. It's the one interacting with the world, right?
And there's a millions upon millions of other DA's out
there working for their principles. Some are working for businesses,
some are working for individuals, some are working for organizations, entities, Whatever.
(06:56):
But the point is, they're the ones doing the hard
work of parsing all these millions of demons. Right. So examples,
finding gifts customized to the exact individual organizing person. Perfect
vacation based on the four people going, navigating all the
logistics and the best way possible of that four person vacation.
Determining the current mood of a person or location based
(07:16):
on what is known about them. Finding the optimal route
from one place to another. Determining the best way to
charge a customer based on evolving competition, logistics, and conditions
on the ground. All these tasks will be services available online.
There will be several or thousands of competitors in any
particular space. There will even be services that consume and
(07:38):
rate those services and present their output through their own demons. Right.
So that's another demon. So my digital assistant, my personal
I will be using those services to find the best service,
to organize a trip or to play the perfect song
or to do basically anything. All these SaaS services, all
(07:58):
these B2C services companies. Forget a human going to the
website and clicking around. That's all done. That is all done.
Our Das will be doing that for us through their demons,
through their business demons, which maybe MCP is that, and
that's where we are. This is essentially what MSPs are
(08:19):
finally starting to happen. MSPs are a way to take
any application and turn it into globally available services. That's
why this MCP thing is so exciting. It's taking any
application and presenting it to I. So I like what
we're doing now is we're taking IDs and we're loading
(08:41):
in these marketplaces. So we have the ability to manually
choose these things. But pretty soon our IDs and more importantly,
our digital assistants are just going to there's going to
be natural, like Vetted, high quality lists of these MSPs
of these demons out there, and they're going to be
super high quality. They're going to be rated for quality.
(09:02):
They're going to be rated for uptime. They're going to
be rated for, you know, how good the results are
and everything. So my Da is just going to be
constantly using all of those, right. So when I'm like
hey I've got sniffles. It's going to be find the best, uh, pharmacy,
find the best medication, do all the research, um, find
(09:23):
the best delivery thing and within, you know, like six minutes,
somebody drops it off on my porch, right? It's it's
going to be that simple because it made 400 different
requests to multiple different demons to be able to do
the research and figure that out and find the best one.
Turns out I paid like a tiny amount of money
(09:43):
for that, or whatever the the amount was. Probably won't
even be a human delivery. It'll be like a Waymo delivery. Anyway,
bottom bottom line is we all get digital assistance. Everything
gets an API, including people, objects and businesses. Our Das
are the ones interacting with the world on our behalf,
because there will be billions of APIs or demons out
(10:04):
there to interact with, and our Das will. Ultimately, this
is the AR piece, which is not quite happening yet,
but everyone's building it. It will display in our are
the best version of of the world to look at, right?
If we're in a security situation, it will turn to
like a security filter for in like a dating situation,
it will turn into like a dating filter. If we're
(10:25):
in like a library, it'll like. Overlay interesting topics based
on the books that we're currently looking at. It'll be like, oh,
you know, Jason really liked this book and Chris really
liked that book or whatever. Or, hey, you were reading
something about this. This book will go into it in
more detail, right? So contextual display of information within your AR.
(10:49):
Everyone's currently building Das MCH and lots of other projects
are APIs for everything, and lots of people are working
on the AR stuff as well, like meta and I'm
sure Apple eventually. All right. Security, cyber security critical PHP, CGI,
remote code execution vulnerability is being mass exploited against windows systems.
(11:09):
It was originally a bunch of Japanese targets. Now it's gone.
Global help Net Security put together a list of 28
cyber security positions across the US. President is nominated Sean Plinky,
Plonky veteran with experience at US Cyber Command in the
previous Trump administration to head Cisa. So Cisa has a
(11:31):
new leader. Malicious poisoning of AI narratives. So TechCrunch put
out a report that Russian propaganda was there was a
basically a a network called Pravda, which is a Russian
propaganda network, is flooding the web with fake news about,
I assume it's like, uh, Russian narratives around Ukraine and
(11:53):
a whole bunch of stuff that just Russia wants everyone
to believe. This is crazy. This is an actual AI
poisoning attack. Real world because everyone was worried about you.
Poison the training data. Oh, but it's like, well, how
are you going to break into the company and poison
the training data? Turns out these bots, like OpenAI, like anthropic,
(12:14):
they're all constantly parsing the internet. You don't have to
poison the original pre-training stuff. Well, actually, that's being poisoned
as well, because every time you do a pre-training, you
have to re harvest from the internet. If you change
the internet, then you have poisoned it, right? So, for example,
if this this group, Pravda is trying to change the
(12:37):
narrative around something. So let's say this company, Pravda successfully
saturates the entire world with the idea that Ukraine attacked first. Okay.
Ukraine came over the border and killed a bunch of
innocent people. And what Russia did was they did a counterattack.
And this is purely defensive. I'm not saying that's that's
(12:58):
what the propaganda said. Let's just take it as an example. Well,
if they put out millions of articles like that across
the entire internet, and then you parse the internet and
you put it into your training data for the next
version of sonnet or the next version of, uh, you know,
O4 or GPT five or whatever it is. Well, how
(13:22):
is the thing? How was I going to know that's
not real? It's it's parsing the internet to figure out
what's real. So this is an actual AI data poisoning attack.
Ten leading chat bots were tested. Chatbots were tested for
susceptibility to Russian disinformation. Pravda published 3.6 million misleading articles
(13:44):
just in 2024, 3.6 million misleading articles and news Knausgaard's
analysis found that the chatbots collectively repeated false information a
third of the time. That's insane. Real world AI data poisoning.
So it used to be about SEO poisoning. Now it's
about flooding the internet with marketing or propaganda that will
(14:06):
get you picked up by up by bots. Now I
say marketing and propaganda. It's the same thing, right? Let's
say you have something. You have a an endpoint product
that's nowhere near as good as CrowdStrike, but you hire
Pravda and you blast out to the internet that, you know,
secure point is way, way better than CrowdStrike. And all
(14:30):
the bots parse this thing. And a third of the time,
Secure Point comes back as better than CrowdStrike. That would
be massive. I mean, that that would get you promoted
on a marketing team, because guess what? The new search
engine is actually AI. A lot, a lot of people
are actually asking AI to compare products and to come
(14:53):
back because. Because it knows that the AI is read
everything and is is summarizing the results of all of that. Well,
that changes that changes significantly when now everyone's going to
switch to this. So, so all these all these LLM
providers are going to have to figure out a way
(15:13):
to filter out poisoning. I don't see an easy way
to do it. You're going to need a lot of
really smart I just to do the data cleansing for
this type of poisoning. All right. Poland's space agency Polska
has taken its systems offline after detecting unauthorized network access. And, uh.
And they're blaming Russian cyber campaigns against the country. Japanese
(15:37):
telecom giant NTT got breached and exposed corporate data for
around 18,000 companies. AI is targeting of enemies. US is
using AI to scan social media and revoke visas of
foreign students it believes support Hamas or other terrorist groups.
The reason I think this is worrying because this this
might be a good use of it, right? But the
(15:58):
problem is, once you use it once and everyone's like, oh,
that's okay. I mean, there's a saying that says first
they come for so-and-so, then they come for so-and-so, and
you say it's okay, and then they come for so-and-so,
and you say, oh, that's still okay because it's not me.
But eventually they come for you. And I'm not saying
that's going to happen. I'm just saying whenever you have
(16:19):
a thing that's like, hey, we're looking for people like
this because they're bad and the people are like, well,
that's okay. And everyone like allows that type of Gestapo behavior.
It's slippery, very slippery, and it's way worse when it
would have taken a lot of research in the past
to actually go and find these people. Right. You got
(16:41):
you got to parse all these videos. You got to
you got to, you know, go through all the blog posts,
you got to find people You basically have to hire
a bunch of researchers, researchers to do this. Well, what
happens if you just spin up a really powerful AI
with a bunch of tools and say, hey, go parse
the internet for the last two years or during this
six month period and find everybody who in the crowd
(17:05):
was waving a Hamas flag, or in the crowd who
was waving whatever flag for the thing that I currently
don't like. And this could be a left leaning thing.
This could be a right leaning thing. Doesn't matter. You
could target an enemy and basically go after them. That
is that is very serious stuff. And the better the
AI gets, the better it's the faster it's going to
(17:26):
be able to do that. I mean, you could be like, look, hey,
we got a Swat team that's ready to go find
me the best possible enemy of me and my friends,
and find me a reason in all of their in
the last six months to go after them. And boom,
you send the Swat team like that. That's a very
(17:46):
powerful weapon, and you have to be very careful who
is actually wielding that thing. DOJ charged 12 Chinese nationals
for state backed hacking operations, including two government officers, eight
employees of a hacking company called I-soon and a few
members of APT 27. Microsoft Threat Intelligence says China linked
(18:06):
Silk Typhoon hacking group is shifting from compromise to actually
going after it. Supply chains as an entry point to
corporate networks, national security. Russia and China and probably others
are targeting Doge fired US federal workers with special focus
(18:27):
on those with security clearances. So basically, a bunch of
people with clearances who know things get laid off and
they're going after them, especially if they're angry, right? That's
a lot of vulnerability. China announced another 7.2% increase to
its military budget, taking it to $245 $45 billion a year.
(18:47):
The Polish president, Andrzej Duda, is pushing NATO, NATO members
to increase defense spending to 3% of GDP. Yeah. China
is trying to bump it by another 7.2%, and we're
trying to get, uh, NATO members to 3%. Oliver Carroll
reports that the US has cut off critical Himars, intelligence, Himars,
(19:11):
Himars intelligence to Ukraine, reducing their ability to defend against
Russian forces. I so manas broke the internet last week.
Big hype is, uh, it's kind of like AGI similar. Um,
they didn't say it was AGI. The video that they
put out was like, hey, it's leaning towards AGI. It's
(19:34):
a little bit like AGI. We're heading in that direction.
But it's basically like Claude or OpenAI, whatever they have
behind the thing. But it's just way smarter. It's doing
a lot more manual work that that would have needed
to be done by a human in the past. And
it's very much in line with what I was talking
about the other day where, um, I don't think that
(19:55):
we need smarter models, actually, to get to AGI. I
think we just need better orchestration of tooling and more
working memory for the models we already have. And it
seems to me like madness is just is just that. Right?
It's probably Claude underneath. And a lot of people have been, uh,
you know, basically hypothesizing that that's the case. We don't
(20:19):
know for sure, but it could just be Claude with
better orchestration. Really want people to think deeply about this
general intelligence competence might be already possible, given how smart
something like O1 Pro or Sonnet 37 already are. It
already has the intelligence to understand stuff like millions or
(20:40):
billions of questions that you could ask it. It's smart
enough to generalize allies answers, or you can give it
situations like brand new situation and based on everything it
knows about these other things, it can give you the
direction or a solution, you know, for that new problem
that you have. That's generalization, right? There might be smart enough.
(21:01):
The problem is its eyes and its hands and it
getting confused after doing too many things because of context windows. Right.
And just overall coordination and orchestration of the thing that
it's doing to try to accomplish a task. Seems like
Manas did that better, but I think that's all that's
needed is that orchestration. So the point is people are like, well,
(21:24):
AI is not getting smart enough fast enough to actually
move us towards AGI. We don't need that. We don't
need smarter models. We need larger context windows, more tools
for the agents to have and better orchestration. That by itself,
I think gets us there to AGI. And I've got
(21:48):
another post, which yeah, I'm not going to pull it up,
but I've got another post that basically talks about exactly
what AGI is going to look like under my definition,
which is, um, my, my definition is basically can it
do the work of a knowledge worker making $80,000 in
the year 2022? And I say 22 because that was
(22:08):
like pre AI or whatever. So the marketing for the
first AGI is going to be something like welcome to
new hire. New hire will join your team on Monday morning.
The day that they start they will do the onboarding.
They will read all the slack messages. They will read
(22:29):
all the documentation. They will meet with their manager. They
will take direction. They will reach out to the team
members and introduce themselves, and they will start to do
the work that they are assigned. And most importantly, if
that work changes because of a reorg, or because they
bought some new company or whatever. They get new tasks,
new goals, new work they have to do. They will
(22:51):
discard their previous stuff. They will learn the new stuff,
they will read the new docs, they will take new direction.
They will learn what they have to, and they'll start
doing that work. And if their manager is like, hey,
you're doing pretty good here, but you need to improve
on this just like a human. They'll be like, okay, thanks, boss.
I'll make that adjustment. I'll make sure you get a
status report every day and whatever it is they were
(23:13):
asked for this company, new hire, which is not a
real company, by the way, but it's going to be
something exactly like this. That's how they're going to market it.
Treat it exactly like a regular employee. That is going
to be the moment, in my opinion, that we've reached AGI,
because if that tool, by the way, there's going to
(23:33):
be a bunch of tools that say they can do
this and then people are going to implement them and
they're going to be super stupid. They're going to be
more waste than it's worth. It's just not going to work.
But when this happens and let's say it's this company
called New Hire. Um, when that happens, let's say new
hire takes off and it just it goes slow for
a second and all of a sudden everyone starts saying,
(23:56):
oh my God, I employed ten new hires, and they're
doing the work of like 50 regular employees. And I
just increased my subscription because I'm growing my business faster
than ever. I employed my new hires in marketing. I
employed them in sales, I employed them in administration. I
employed them in security administration in my SOC. Whatever the
(24:20):
moment that happens, I'm calling that the AGI moment. I
believe that is the actual AGI moment, because there is
no better definition of general intelligence than humans. That's what
we're basing this whole entire thing on, is humans ability
to generalize, generalize when it's working. Right. The fact that
(24:40):
you could change up context. Say, hey, that's not your
job anymore. Here's your new job. That's the best definition
we have is a knowledge worker constantly dealing with change,
constantly dealing with one offs and asks and adjustments. So
that's the perfect definition also for actual AGI artificial general intelligence.
(25:01):
And if it could do the job of an average
knowledge worker I would say that hits the bar. So
that's why I'm using it for a definition. And I'm
saying that when a product comes out that can do
this and it's actually doing it, that will be the moment.
And my prediction is late 2025 at the earliest. This
(25:22):
is my modified prediction. My overall prediction, which I said
in 23 was 25 to 28. So my modified prediction
is late 25 to early 27. So it's still right
there inside that window. But I think it's getting pretty
close and things like manis are starting to show signs
(25:44):
of it. All right. New model q w q 32
billion parameters. This thing is outperforming or performing just as
good as 671 billion parameter deep seek 32 billion versus
671 billion. And they are right on par. This is
the absolute most impressive local model I've ever used. I
(26:09):
test out all these big ones when they come out.
I've got right in my other room. I've got two
49 seconds over there and I fire them up and
I test this thing. This I mean, it performs honestly
very similar to like a cloud or an OpenAI, um,
(26:29):
like an O one or an O one mini. It's really,
really good. Um, I used it with some, uh, story
writing stuff. I'm trying to write a fiction story about some,
like an eye feature, getting pretty excited about that, but
I was using it to help me write, and it's
doing things that no previous model could possibly do. Um,
(26:50):
just vast jumps in quality in the output that I'm getting.
So really excited about that. You should go check it out.
You could run it with a llama llama run q
w q. That's all you have to do. It'll download it,
it'll run it, and you could test it out. And
it's a thinking model built with reinforcement learning, which means
you're going to see the thinking instructions as part of
(27:12):
the output before it starts outputting the real thing. Myo
addresses eye hallucinations, so doctors are very upset and worried
about hallucinations for good reason because it's life and death
for them. Right? So reverse Rag is a technique that
they're using to basically figure out the sources for the
data that they return. So they don't just take hallucinations
(27:36):
or they don't just take outputs and assume it's good.
What they do is they take each particular one and
they go and research it and find sources to it.
And then they say, okay, this is now trustworthy. And
I've been talking about this for, I don't know, six
months or now or however long. I did a talk
about this at some conference. And basically what we're talking
(27:56):
about here is a pipeline of trust that has to
happen for high trust output. So if you're in medical
or you're in military or you're in, uh, law or something,
if you're in something where results actually matter, you have
to have a system that comes after the AI output,
(28:18):
which is like a series of steps for validation, and
only after it goes through those steps and gets a
green check mark. Then you can trust the thing, right?
I mean, humans need that, by the way. I mean,
humans don't put out stuff that's trustworthy, right? So I
think AI is going to be actually much better at
doing this than, uh, most human doctors. You still don't
(28:40):
get that. Like the doctor touch and the doctor intuition
and the the human aspect. So it's it's not better
in all ways, but in terms of, like, putting out
the best possible diagnosis and stuff like that, I think
it's going to be fantastic. So kudos to Mayo for
doing that. Tyler Cowen told Dwarkesh Patel that he predicts
AI will only boost economic growth by 0.5% per year,
(29:04):
half a percent. I massively disagree with this and so
does Dwarkesh. Black Forest Labs, also known as replicate, is
dominating image generation. I'm using this myself. Um, I've been
using it in conjunction more with Midjourney. Midjourney is like
my go to, but I've been using a lot more
(29:24):
replicate and it is absolutely fantastic. Larry Page is working
on a new AI company that uses artificial intelligence to
design and manufacture optimized products. Ben Buchanan, Biden's AI advisor,
says Washington now believes AGI is possible and realizes they're
competing with tech companies for talent to regulate it. Nirvana
raised 80 million for the AI based trucking insurance platform
(29:48):
that uses telematics and 20 billion miles of driving data
to create better policies for drivers. And I've got a
post here on real time insurance premiums from 2016. Yeah,
I've been talking about this for a while. 2016 essentially, again,
everything with AI comes down to context. So if I'm
wearing a microphone and a camera and I give access
(30:12):
to that, there's like that progressive thing. Remember that progressive
box you put in your car and it can see
how you're driving? Well, imagine that it could see that
you're cussing or yelling at your spouse or kicking garbage
cans or, um, you just came from the bar. And
it knows that because it has the cameras on and monitoring. Well,
(30:33):
guess what? Your insurance premiums might be getting modified in
real time or close to real time. Or you might
be like, hey, if you get in this car right now,
having come from, you know, oceanians, then, um, you your
policy will be instantly cancelled, right? It's going to be
things like that. Um, all possible because of AI and context.
(30:56):
Simon Willison taught a Nykaa 2025 workshop on advanced web
scraping techniques. Really, really cool talk. You have to check
that one out. Okay, McDonald's rolling out a massive AI
upgrade for its 43,000 locations, including AI, drive through smart
kitchen equipment and generative AI Virtual Manager, one of the
biggest visible AI rollouts so far. Can't wait to see
(31:20):
how this works. I heard somebody say, I don't know
who because I don't. I wouldn't eat there because it's
really bad. But Domino's and Carl's Jr has really good AI.
So I heard technology. Apple is pushing back. It's more
personalized Siri features. So basically, the big unified Siri that
has access to all the context is being pushed back.
(31:43):
And a lot of people are speculating that this is
because of security issues, prompt injection against that entire context,
which what do you think is going to happen when
your assistant that you call up with this keyword can
do anything across your entire context, including all of your applications,
(32:03):
all your health data and everything and your interface into
that is is the English language and other languages as well,
but especially English. That is a massive attack surface to
go after. And I think Apple, rightly so, is being
very careful about that. So they push that back. One
thing to keep in mind though, is they already rolled
(32:23):
out tons of Apple intelligence. I already use, uh, my
Siri to actually, um, do live queries because it's connected
to ChatGPT. So don't think that we're waiting for this
thing to happen. And you don't have Apple intelligence before that.
It's already out there. Like they already rolled out like
dozens of these features, including, most importantly, the ChatGPT one.
(32:47):
And the fact that if you hold down your camera button,
the new camera button on the side, you can have
ChatGPT identify things with the camera. So all this to say,
Apple intelligence is like already live and already working. It's
just this unified big context, one that they haven't enabled
yet because of the security concern. Waymo is expanding its
(33:08):
self-driving car service in Palo Alto and some surrounding peninsula areas.
Cannot wait to get it around here. Waymo is just
just unbelievably good. New Zealand appears to be managing its
$16 billion health budget. Using a single Excel spreadsheet probably
makes Doge pretty happy. Apple is evidently planning a major
(33:30):
redesign of iOS, iPadOS, and macOS for the fall. Can't
wait to see what that looks like. I'll be running
it on day one because I'm stupid. Tesla's global sales
are tanking pretty hard right now with significant drops across Germany,
Australia and China. Yeah, you do a Hitler salute and
sales fall in Germany. Big surprise. New Mac studio has 32.
(33:53):
This is the M3 ultra 32 CPU cores and enough
unified memory, 512GB to load the 600 billion parameter models.
I was going to get one, but I already have
my eye box. I'm going to skip this generation, maybe
even another one, because I already have an M2 ultra.
Not worth it to bump up to the 512 especially.
(34:16):
That's like a $15,000 desktop discovery. A machine learning engineer
named Jasper Gillie quit his job because he believes AI
will automate his entire role by the end of 2025.
Keep in mind, this is an extremely highly paid machine
learning engineer, and he's saying his job is going to
(34:36):
be replaced by the end of 2025. Robin Moffat argues
we should or we shouldn't, just write the perfectly structured
blog post, but also share our messy struggles and solutions
that other people need, too. I like this. It's like
learning in public. Show your work, you know. Feel free
to be vulnerable and imperfect online, I think is, uh.
(34:58):
It's endearing. Hopefully so since I've said edit like nine
times in this recording already. Small docs a project around
writing small docs instead of giant ones that are hard
to follow. Jamie Rumbelow created a slick hybrid between databases,
programming and UIs. Make scripts more user friendly without needing
a full web app. And basically the pitch there was
(35:19):
that the database, the UI and the app are kind
of unified. They're all the same thing. Astro and vim,
one of the better vim distros. I do prefer lazy vim,
although I hate the name. Reminds me of the dummies books,
which I could never get myself to buy. Too much self-respect.
Interview DB lets you access crowdsourced tech interview questions from
(35:42):
the community. Alexander Solzhenitsyn on the future of the West.
This was a crazy, crazy piece of text there. Seneschal MCP.
MCP by a brilliant Ull member named Matt, who, um,
wrote an MCP that helps Llms access standardized health data
(36:03):
from the sky. Senegal. Senegal. I keep thinking, Senussi. What
am I thinking of? Senegal. API. So basically, this is
an NCP that makes your health data available so that
you can have local agents or any kind of agent
really make your health data available to an app. So
(36:23):
if you're building like a full life encompassing, you know,
management app with like workout routines, health, fitness, food, diet, money,
all that stuff. This is an API. This is an
MCP that presents the health side of it. Very very
cool from UL. Member Matt Self interview AI tool that
(36:45):
runs invisibly during technical interviews, helping you with coding questions
while your screen is shared. Also known as cheating. So
this is what you have to watch out for if
you're doing interviews. Piano FM generates endless beautiful AI piano music.
I listened for a while. It was not repetitive and
it sounded. I'm not a classical music expert or a
(37:07):
piano expert, and it sounded pretty normal to me. It
sounded like regular piano, whatever that's worth. Matrix hacking game.
Someone built a matrix themed AI hacking game lets users
learn about prompt engineering and security by defeating an Agent
Smith AI character. Send new app connect your iPhone camera
(37:28):
to multi-modal Llms for real time visual intelligence made with
zero VC money and cursor. This is the type of
thing that could have gone for, could have raised, you know,
$20 million and been sold for $100 million maybe a
couple of years ago. And this person just built it
(37:48):
with cursor with no money whatsoever. Since Feedback Dev built
an AI that reads the emotional vibe of your product,
social media mentions and gives you instant feedback. And somebody
posted their cursor rules, which I thought were quite good.
All right. This is the end of the standard edition
of the podcast, which includes just the news items for
(38:11):
the week to get the rest of the episode, which
includes my analysis, the discovery section that has all of
the coolest tools and articles. I found this week, the
recommendation of the week and the aphorism of the week.
Please consider becoming a member. As a member, you'll get
lots of different things from access to our extraordinary community
(38:31):
of over a thousand brilliant and kind people in industries
like cybersecurity, AI, technology and the humanities. You also get
access to the UL Book Club, dedicated member content and events,
and lots, lots more. Plus, you'll get a dedicated podcast
feed you can put in your client that gets you
the full member edition of this podcast. To become a
(38:55):
member and get all of that, just head over to
Daniel Comm Slash upgrade. That's Daniel miessler.com/upgrade. We'll see you
next time. Unsupervised learning is produced on Hindenburg Pro using
an SM seven B microphone. A video version of the
podcast is available on the Unsupervised Learning YouTube channel, and
(39:17):
the text version with full links and notes is available
at Daniel Mysa.com slash newsletter. We'll see you next time.
Unsupervised Learning is a podcast about trends and ideas in cybersecurity,
national security, AI, technology and society, and how best to
upgrade ourselves to be ready for what's coming. All right,
welcome to episode 472. Hope your week is going well.
(00:22):
Updates on this side I've got an RSS feed which
I encourage you to go sign up for. RSS is
a real thing and it's awesome. And I wish it
didn't get knocked off the radar by Google Reader going under,
but I use Feedly. A lot of people use Netnewswire.
Whatever your RSS reader is, go check it out. It's
(00:45):
just slash feed RSS, I think at the end of
the URL. Thanks for all the cooking advice. Around 500
people responded with experience, encouragement, recommendations. It was just insane.
So really appreciate all that. I tried windsurf, a lot
of people are talking about it. I basically try all
these tools that people talk about. I've been using Klein
(01:07):
for a while. I was using cursor before that and
I keep trying cursor every time they have a big
release or when windsurf has a big release. I'll try
it again. The stuff is changing so fast that I
just constantly try different stuff. But Klein has been my
go to. I feel like it manages the functionality of
(01:29):
Lmms better and the UI. I just like better, so
that's what I've been using. It is just an extension
into VS code. I wish it was actually a full editor,
just like cursor and windsurf are. So Klein people. If
you're listening, please make an editor. And my personal AI
infrastructure is getting absolutely insane right now. I'm going to
(01:52):
have something to show on this soon. I'm going to
do a demo of like how this all works together,
but basically it's one main agent and then all my
personal services are sitting behind it that do different things,
and they're all able to access different services that do
output as well. So there's like you send the request
(02:13):
into the agent, the agent, if I tell it, send me,
you know, test a website and then send me an
email for it, it'll do the two different services. One
is to test the website and the other one is
to send an email. And again it just naturally parsed that.
And I just sent that from my phone. Right. So
(02:33):
the whole idea is I'm just using my regular tech infrastructure,
which is my iPhone, and I send a request into
my agent, and this is what it's going to be
in the future too. I'm just sort of hacking it
together with my own services. Pretty soon it'll just be
built into Google and built into the iPhone and stuff
like that. But right now I'm just making it happen
(02:55):
through like shortcuts and a bunch of AI services, but
the overall effect is quite awesome and you have to
see this thing actually working. It's very much in line
with what I've been talking about. Uh, in the predictable
AI post, which is this link here where it's all going,
(03:16):
put out a concise explanation of model context protocol, MCP
servers and why everyone's excited about them. So I got
a full thread on that and you should go check out. And, uh,
I'm enjoying the fact that this is starting to happen,
specifically the MCP server, because in that predictable AI post,
(03:37):
which actually comes from the book in 2016, I talked
about the API ification of everything and specifically how businesses
would all become APIs. So this is text. From that,
most software businesses will become algorithms presented to others through
their business daemons. Many traditional businesses will continue to become
(03:58):
software businesses. So basically, Andreessen talked about software as eating
the world. I think AI is going to eat the world.
Like that's pretty. Everyone's saying that it doesn't really mean anything.
What I think is actually eating the world is APIs.
And the fact that AI is what's going to interact
with all these APIs. The primary interface for doing anything
(04:21):
is going to be APIs, because it's the AI that's
parsing them, right? What humans need to look at a
website or a list of services, or a menu of
food or whatever. What humans need for that is way
different than what a computer needs. A computer can parse
an API or an MCP service directory, you know, basically
(04:46):
instantly and know exactly how to use it. So if
you if you think about SEO, if you think about UI, UX,
if you think about advertisements, if you think about websites,
if you think about News media, all of that. It's
all right now focused on presenting that to humans. The
fundamental change that's happening right now is that's all switching
(05:06):
away from being primarily designed for humans to being primarily
designed for AI. Primarily, these interfaces will be for AI
to find a business, AI to find a service AI
to rank those services. Right. So examples will include and
this is keep in mind this text right here. This
(05:26):
is from 2016 when I wrote that stupid book which
you probably don't need to read anymore because it's all
in that, that post and it's got pictures which took
me like hours to make. So it's a better version
of the book that blog post is. But the key
point here is I was talking about this in 2016,
so I'm looking for a little like, I don't know,
(05:48):
somebody to send me a pat on the back. That
was kind of the whole purpose of writing that stupid
book was to get this stuff locked in, because I
saw this happening so clearly, even though that was way
before this AI stuff popped. Right. So what is 22
the end of 22 versus the end of 2016? That's
six years. Six years. I basically said, we're going to
(06:11):
have digital assistants, which are AI powered. I talked about
the AI component right there in there. It's going to
know everything about us. We're going to talk to our Da.
Our Da is going to read all the services. Every
human is going to have a demon. Not every human,
but most humans. A lot of a lot of humans
are going to have demons. But most importantly, all the
(06:31):
businesses have demons. We talk to our Da. The Da
infers what we probably want. We don't even need to
talk to it. It's the one interacting with the world, right?
And there's a millions upon millions of other DA's out
there working for their principles. Some are working for businesses,
some are working for individuals, some are working for organizations, entities, Whatever.
(06:56):
But the point is, they're the ones doing the hard
work of parsing all these millions of demons. Right. So examples,
finding gifts customized to the exact individual organizing person. Perfect
vacation based on the four people going, navigating all the
logistics and the best way possible of that four person vacation.
Determining the current mood of a person or location based
(07:16):
on what is known about them. Finding the optimal route
from one place to another. Determining the best way to
charge a customer based on evolving competition, logistics, and conditions
on the ground. All these tasks will be services available online.
There will be several or thousands of competitors in any
particular space. There will even be services that consume and
(07:38):
rate those services and present their output through their own demons. Right.
So that's another demon. So my digital assistant, my personal
I will be using those services to find the best service,
to organize a trip or to play the perfect song
or to do basically anything. All these SaaS services, all
(07:58):
these B2C services companies. Forget a human going to the
website and clicking around. That's all done. That is all done.
Our Das will be doing that for us through their demons,
through their business demons, which maybe MCP is that, and
that's where we are. This is essentially what MSPs are
(08:19):
finally starting to happen. MSPs are a way to take
any application and turn it into globally available services. That's
why this MCP thing is so exciting. It's taking any
application and presenting it to I. So I like what
we're doing now is we're taking IDs and we're loading
(08:41):
in these marketplaces. So we have the ability to manually
choose these things. But pretty soon our IDs and more importantly,
our digital assistants are just going to there's going to
be natural, like Vetted, high quality lists of these MSPs
of these demons out there, and they're going to be
super high quality. They're going to be rated for quality.
(09:02):
They're going to be rated for uptime. They're going to
be rated for, you know, how good the results are
and everything. So my Da is just going to be
constantly using all of those, right. So when I'm like
hey I've got sniffles. It's going to be find the best, uh, pharmacy,
find the best medication, do all the research, um, find
(09:23):
the best delivery thing and within, you know, like six minutes,
somebody drops it off on my porch, right? It's it's
going to be that simple because it made 400 different
requests to multiple different demons to be able to do
the research and figure that out and find the best one.
Turns out I paid like a tiny amount of money
(09:43):
for that, or whatever the the amount was. Probably won't
even be a human delivery. It'll be like a Waymo delivery. Anyway,
bottom bottom line is we all get digital assistance. Everything
gets an API, including people, objects and businesses. Our Das
are the ones interacting with the world on our behalf,
because there will be billions of APIs or demons out
(10:04):
there to interact with, and our Das will. Ultimately, this
is the AR piece, which is not quite happening yet,
but everyone's building it. It will display in our are
the best version of of the world to look at, right?
If we're in a security situation, it will turn to
like a security filter for in like a dating situation,
it will turn into like a dating filter. If we're
(10:25):
in like a library, it'll like. Overlay interesting topics based
on the books that we're currently looking at. It'll be like, oh,
you know, Jason really liked this book and Chris really
liked that book or whatever. Or, hey, you were reading
something about this. This book will go into it in
more detail, right? So contextual display of information within your AR.
(10:49):
Everyone's currently building Das MCH and lots of other projects
are APIs for everything, and lots of people are working
on the AR stuff as well, like meta and I'm
sure Apple eventually. All right. Security, cyber security critical PHP, CGI,
remote code execution vulnerability is being mass exploited against windows systems.
(11:09):
It was originally a bunch of Japanese targets. Now it's gone.
Global help Net Security put together a list of 28
cyber security positions across the US. President is nominated Sean Plinky,
Plonky veteran with experience at US Cyber Command in the
previous Trump administration to head Cisa. So Cisa has a
(11:31):
new leader. Malicious poisoning of AI narratives. So TechCrunch put
out a report that Russian propaganda was there was a
basically a a network called Pravda, which is a Russian
propaganda network, is flooding the web with fake news about,
I assume it's like, uh, Russian narratives around Ukraine and
(11:53):
a whole bunch of stuff that just Russia wants everyone
to believe. This is crazy. This is an actual AI
poisoning attack. Real world because everyone was worried about you.
Poison the training data. Oh, but it's like, well, how
are you going to break into the company and poison
the training data? Turns out these bots, like OpenAI, like anthropic,
(12:14):
they're all constantly parsing the internet. You don't have to
poison the original pre-training stuff. Well, actually, that's being poisoned
as well, because every time you do a pre-training, you
have to re harvest from the internet. If you change
the internet, then you have poisoned it, right? So, for example,
if this this group, Pravda is trying to change the
(12:37):
narrative around something. So let's say this company, Pravda successfully
saturates the entire world with the idea that Ukraine attacked first. Okay.
Ukraine came over the border and killed a bunch of
innocent people. And what Russia did was they did a counterattack.
And this is purely defensive. I'm not saying that's that's
(12:58):
what the propaganda said. Let's just take it as an example. Well,
if they put out millions of articles like that across
the entire internet, and then you parse the internet and
you put it into your training data for the next
version of sonnet or the next version of, uh, you know,
O4 or GPT five or whatever it is. Well, how
(13:22):
is the thing? How was I going to know that's
not real? It's it's parsing the internet to figure out
what's real. So this is an actual AI data poisoning attack.
Ten leading chat bots were tested. Chatbots were tested for
susceptibility to Russian disinformation. Pravda published 3.6 million misleading articles
(13:44):
just in 2024, 3.6 million misleading articles and news Knausgaard's
analysis found that the chatbots collectively repeated false information a
third of the time. That's insane. Real world AI data poisoning.
So it used to be about SEO poisoning. Now it's
about flooding the internet with marketing or propaganda that will
(14:06):
get you picked up by up by bots. Now I
say marketing and propaganda. It's the same thing, right? Let's
say you have something. You have a an endpoint product
that's nowhere near as good as CrowdStrike, but you hire
Pravda and you blast out to the internet that, you know,
secure point is way, way better than CrowdStrike. And all
(14:30):
the bots parse this thing. And a third of the time,
Secure Point comes back as better than CrowdStrike. That would
be massive. I mean, that that would get you promoted
on a marketing team, because guess what? The new search
engine is actually AI. A lot, a lot of people
are actually asking AI to compare products and to come
(14:53):
back because. Because it knows that the AI is read
everything and is is summarizing the results of all of that. Well,
that changes that changes significantly when now everyone's going to
switch to this. So, so all these all these LLM
providers are going to have to figure out a way
(15:13):
to filter out poisoning. I don't see an easy way
to do it. You're going to need a lot of
really smart I just to do the data cleansing for
this type of poisoning. All right. Poland's space agency Polska
has taken its systems offline after detecting unauthorized network access. And, uh.
And they're blaming Russian cyber campaigns against the country. Japanese
(15:37):
telecom giant NTT got breached and exposed corporate data for
around 18,000 companies. AI is targeting of enemies. US is
using AI to scan social media and revoke visas of
foreign students it believes support Hamas or other terrorist groups.
The reason I think this is worrying because this this
might be a good use of it, right? But the
(15:58):
problem is, once you use it once and everyone's like, oh,
that's okay. I mean, there's a saying that says first
they come for so-and-so, then they come for so-and-so, and
you say it's okay, and then they come for so-and-so,
and you say, oh, that's still okay because it's not me.
But eventually they come for you. And I'm not saying
that's going to happen. I'm just saying whenever you have
(16:19):
a thing that's like, hey, we're looking for people like
this because they're bad and the people are like, well,
that's okay. And everyone like allows that type of Gestapo behavior.
It's slippery, very slippery, and it's way worse when it
would have taken a lot of research in the past
to actually go and find these people. Right. You got
(16:41):
you got to parse all these videos. You got to
you got to, you know, go through all the blog posts,
you got to find people You basically have to hire
a bunch of researchers, researchers to do this. Well, what
happens if you just spin up a really powerful AI
with a bunch of tools and say, hey, go parse
the internet for the last two years or during this
six month period and find everybody who in the crowd
(17:05):
was waving a Hamas flag, or in the crowd who
was waving whatever flag for the thing that I currently
don't like. And this could be a left leaning thing.
This could be a right leaning thing. Doesn't matter. You
could target an enemy and basically go after them. That
is that is very serious stuff. And the better the
AI gets, the better it's the faster it's going to
(17:26):
be able to do that. I mean, you could be like, look, hey,
we got a Swat team that's ready to go find
me the best possible enemy of me and my friends,
and find me a reason in all of their in
the last six months to go after them. And boom,
you send the Swat team like that. That's a very
(17:46):
powerful weapon, and you have to be very careful who
is actually wielding that thing. DOJ charged 12 Chinese nationals
for state backed hacking operations, including two government officers, eight
employees of a hacking company called I-soon and a few
members of APT 27. Microsoft Threat Intelligence says China linked
(18:06):
Silk Typhoon hacking group is shifting from compromise to actually
going after it. Supply chains as an entry point to
corporate networks, national security. Russia and China and probably others
are targeting Doge fired US federal workers with special focus
(18:27):
on those with security clearances. So basically, a bunch of
people with clearances who know things get laid off and
they're going after them, especially if they're angry, right? That's
a lot of vulnerability. China announced another 7.2% increase to
its military budget, taking it to $245 $45 billion a year.
(18:47):
The Polish president, Andrzej Duda, is pushing NATO, NATO members
to increase defense spending to 3% of GDP. Yeah. China
is trying to bump it by another 7.2%, and we're
trying to get, uh, NATO members to 3%. Oliver Carroll
reports that the US has cut off critical Himars, intelligence, Himars,
(19:11):
Himars intelligence to Ukraine, reducing their ability to defend against
Russian forces. I so manas broke the internet last week.
Big hype is, uh, it's kind of like AGI similar. Um,
they didn't say it was AGI. The video that they
put out was like, hey, it's leaning towards AGI. It's
(19:34):
a little bit like AGI. We're heading in that direction.
But it's basically like Claude or OpenAI, whatever they have
behind the thing. But it's just way smarter. It's doing
a lot more manual work that that would have needed
to be done by a human in the past. And
it's very much in line with what I was talking
about the other day where, um, I don't think that
(19:55):
we need smarter models, actually, to get to AGI. I
think we just need better orchestration of tooling and more
working memory for the models we already have. And it
seems to me like madness is just is just that. Right?
It's probably Claude underneath. And a lot of people have been, uh,
you know, basically hypothesizing that that's the case. We don't
(20:19):
know for sure, but it could just be Claude with
better orchestration. Really want people to think deeply about this
general intelligence competence might be already possible, given how smart
something like O1 Pro or Sonnet 37 already are. It
already has the intelligence to understand stuff like millions or
(20:40):
billions of questions that you could ask it. It's smart
enough to generalize allies answers, or you can give it
situations like brand new situation and based on everything it
knows about these other things, it can give you the
direction or a solution, you know, for that new problem
that you have. That's generalization, right? There might be smart enough.
(21:01):
The problem is its eyes and its hands and it
getting confused after doing too many things because of context windows. Right.
And just overall coordination and orchestration of the thing that
it's doing to try to accomplish a task. Seems like
Manas did that better, but I think that's all that's
needed is that orchestration. So the point is people are like, well,
(21:24):
AI is not getting smart enough fast enough to actually
move us towards AGI. We don't need that. We don't
need smarter models. We need larger context windows, more tools
for the agents to have and better orchestration. That by itself,
I think gets us there to AGI. And I've got
(21:48):
another post, which yeah, I'm not going to pull it up,
but I've got another post that basically talks about exactly
what AGI is going to look like under my definition,
which is, um, my, my definition is basically can it
do the work of a knowledge worker making $80,000 in
the year 2022? And I say 22 because that was
(22:08):
like pre AI or whatever. So the marketing for the
first AGI is going to be something like welcome to
new hire. New hire will join your team on Monday morning.
The day that they start they will do the onboarding.
They will read all the slack messages. They will read
(22:29):
all the documentation. They will meet with their manager. They
will take direction. They will reach out to the team
members and introduce themselves, and they will start to do
the work that they are assigned. And most importantly, if
that work changes because of a reorg, or because they
bought some new company or whatever. They get new tasks,
new goals, new work they have to do. They will
(22:51):
discard their previous stuff. They will learn the new stuff,
they will read the new docs, they will take new direction.
They will learn what they have to, and they'll start
doing that work. And if their manager is like, hey,
you're doing pretty good here, but you need to improve
on this just like a human. They'll be like, okay, thanks, boss.
I'll make that adjustment. I'll make sure you get a
status report every day and whatever it is they were
(23:13):
asked for this company, new hire, which is not a
real company, by the way, but it's going to be
something exactly like this. That's how they're going to market it.
Treat it exactly like a regular employee. That is going
to be the moment, in my opinion, that we've reached AGI,
because if that tool, by the way, there's going to
(23:33):
be a bunch of tools that say they can do
this and then people are going to implement them and
they're going to be super stupid. They're going to be
more waste than it's worth. It's just not going to work.
But when this happens and let's say it's this company
called New Hire. Um, when that happens, let's say new
hire takes off and it just it goes slow for
a second and all of a sudden everyone starts saying,
(23:56):
oh my God, I employed ten new hires, and they're
doing the work of like 50 regular employees. And I
just increased my subscription because I'm growing my business faster
than ever. I employed my new hires in marketing. I
employed them in sales, I employed them in administration. I
employed them in security administration in my SOC. Whatever the
(24:20):
moment that happens, I'm calling that the AGI moment. I
believe that is the actual AGI moment, because there is
no better definition of general intelligence than humans. That's what
we're basing this whole entire thing on, is humans ability
to generalize, generalize when it's working. Right. The fact that
(24:40):
you could change up context. Say, hey, that's not your
job anymore. Here's your new job. That's the best definition
we have is a knowledge worker constantly dealing with change,
constantly dealing with one offs and asks and adjustments. So
that's the perfect definition also for actual AGI artificial general intelligence.
(25:01):
And if it could do the job of an average
knowledge worker I would say that hits the bar. So
that's why I'm using it for a definition. And I'm
saying that when a product comes out that can do
this and it's actually doing it, that will be the moment.
And my prediction is late 2025 at the earliest. This
(25:22):
is my modified prediction. My overall prediction, which I said
in 23 was 25 to 28. So my modified prediction
is late 25 to early 27. So it's still right
there inside that window. But I think it's getting pretty
close and things like manis are starting to show signs
(25:44):
of it. All right. New model q w q 32
billion parameters. This thing is outperforming or performing just as
good as 671 billion parameter deep seek 32 billion versus
671 billion. And they are right on par. This is
the absolute most impressive local model I've ever used. I
(26:09):
test out all these big ones when they come out.
I've got right in my other room. I've got two
49 seconds over there and I fire them up and
I test this thing. This I mean, it performs honestly
very similar to like a cloud or an OpenAI, um,
(26:29):
like an O one or an O one mini. It's really,
really good. Um, I used it with some, uh, story
writing stuff. I'm trying to write a fiction story about some,
like an eye feature, getting pretty excited about that, but
I was using it to help me write, and it's
doing things that no previous model could possibly do. Um,
(26:50):
just vast jumps in quality in the output that I'm getting.
So really excited about that. You should go check it out.
You could run it with a llama llama run q
w q. That's all you have to do. It'll download it,
it'll run it, and you could test it out. And
it's a thinking model built with reinforcement learning, which means
you're going to see the thinking instructions as part of
(27:12):
the output before it starts outputting the real thing. Myo
addresses eye hallucinations, so doctors are very upset and worried
about hallucinations for good reason because it's life and death
for them. Right? So reverse Rag is a technique that
they're using to basically figure out the sources for the
data that they return. So they don't just take hallucinations
(27:36):
or they don't just take outputs and assume it's good.
What they do is they take each particular one and
they go and research it and find sources to it.
And then they say, okay, this is now trustworthy. And
I've been talking about this for, I don't know, six
months or now or however long. I did a talk
about this at some conference. And basically what we're talking
(27:56):
about here is a pipeline of trust that has to
happen for high trust output. So if you're in medical
or you're in military or you're in, uh, law or something,
if you're in something where results actually matter, you have
to have a system that comes after the AI output,
(28:18):
which is like a series of steps for validation, and
only after it goes through those steps and gets a
green check mark. Then you can trust the thing, right?
I mean, humans need that, by the way. I mean,
humans don't put out stuff that's trustworthy, right? So I
think AI is going to be actually much better at
doing this than, uh, most human doctors. You still don't
(28:40):
get that. Like the doctor touch and the doctor intuition
and the the human aspect. So it's it's not better
in all ways, but in terms of, like, putting out
the best possible diagnosis and stuff like that, I think
it's going to be fantastic. So kudos to Mayo for
doing that. Tyler Cowen told Dwarkesh Patel that he predicts
AI will only boost economic growth by 0.5% per year,
(29:04):
half a percent. I massively disagree with this and so
does Dwarkesh. Black Forest Labs, also known as replicate, is
dominating image generation. I'm using this myself. Um, I've been
using it in conjunction more with Midjourney. Midjourney is like
my go to, but I've been using a lot more
(29:24):
replicate and it is absolutely fantastic. Larry Page is working
on a new AI company that uses artificial intelligence to
design and manufacture optimized products. Ben Buchanan, Biden's AI advisor,
says Washington now believes AGI is possible and realizes they're
competing with tech companies for talent to regulate it. Nirvana
raised 80 million for the AI based trucking insurance platform
(29:48):
that uses telematics and 20 billion miles of driving data
to create better policies for drivers. And I've got a
post here on real time insurance premiums from 2016. Yeah,
I've been talking about this for a while. 2016 essentially, again,
everything with AI comes down to context. So if I'm
wearing a microphone and a camera and I give access
(30:12):
to that, there's like that progressive thing. Remember that progressive
box you put in your car and it can see
how you're driving? Well, imagine that it could see that
you're cussing or yelling at your spouse or kicking garbage
cans or, um, you just came from the bar. And
it knows that because it has the cameras on and monitoring. Well,
(30:33):
guess what? Your insurance premiums might be getting modified in
real time or close to real time. Or you might
be like, hey, if you get in this car right now,
having come from, you know, oceanians, then, um, you your
policy will be instantly cancelled, right? It's going to be
things like that. Um, all possible because of AI and context.
(30:56):
Simon Willison taught a Nykaa 2025 workshop on advanced web
scraping techniques. Really, really cool talk. You have to check
that one out. Okay, McDonald's rolling out a massive AI
upgrade for its 43,000 locations, including AI, drive through smart
kitchen equipment and generative AI Virtual Manager, one of the
biggest visible AI rollouts so far. Can't wait to see
(31:20):
how this works. I heard somebody say, I don't know
who because I don't. I wouldn't eat there because it's
really bad. But Domino's and Carl's Jr has really good AI.
So I heard technology. Apple is pushing back. It's more
personalized Siri features. So basically, the big unified Siri that
has access to all the context is being pushed back.
(31:43):
And a lot of people are speculating that this is
because of security issues, prompt injection against that entire context,
which what do you think is going to happen when
your assistant that you call up with this keyword can
do anything across your entire context, including all of your applications,
(32:03):
all your health data and everything and your interface into
that is is the English language and other languages as well,
but especially English. That is a massive attack surface to
go after. And I think Apple, rightly so, is being
very careful about that. So they push that back. One
thing to keep in mind though, is they already rolled
(32:23):
out tons of Apple intelligence. I already use, uh, my
Siri to actually, um, do live queries because it's connected
to ChatGPT. So don't think that we're waiting for this
thing to happen. And you don't have Apple intelligence before that.
It's already out there. Like they already rolled out like
dozens of these features, including, most importantly, the ChatGPT one.
(32:47):
And the fact that if you hold down your camera button,
the new camera button on the side, you can have
ChatGPT identify things with the camera. So all this to say,
Apple intelligence is like already live and already working. It's
just this unified big context, one that they haven't enabled
yet because of the security concern. Waymo is expanding its
(33:08):
self-driving car service in Palo Alto and some surrounding peninsula areas.
Cannot wait to get it around here. Waymo is just
just unbelievably good. New Zealand appears to be managing its
$16 billion health budget. Using a single Excel spreadsheet probably
makes Doge pretty happy. Apple is evidently planning a major
(33:30):
redesign of iOS, iPadOS, and macOS for the fall. Can't
wait to see what that looks like. I'll be running
it on day one because I'm stupid. Tesla's global sales
are tanking pretty hard right now with significant drops across Germany,
Australia and China. Yeah, you do a Hitler salute and
sales fall in Germany. Big surprise. New Mac studio has 32.
(33:53):
This is the M3 ultra 32 CPU cores and enough
unified memory, 512GB to load the 600 billion parameter models.
I was going to get one, but I already have
my eye box. I'm going to skip this generation, maybe
even another one, because I already have an M2 ultra.
Not worth it to bump up to the 512 especially.
(34:16):
That's like a $15,000 desktop discovery. A machine learning engineer
named Jasper Gillie quit his job because he believes AI
will automate his entire role by the end of 2025.
Keep in mind, this is an extremely highly paid machine
learning engineer, and he's saying his job is going to
(34:36):
be replaced by the end of 2025. Robin Moffat argues
we should or we shouldn't, just write the perfectly structured
blog post, but also share our messy struggles and solutions
that other people need, too. I like this. It's like
learning in public. Show your work, you know. Feel free
to be vulnerable and imperfect online, I think is, uh.
(34:58):
It's endearing. Hopefully so since I've said edit like nine
times in this recording already. Small docs a project around
writing small docs instead of giant ones that are hard
to follow. Jamie Rumbelow created a slick hybrid between databases,
programming and UIs. Make scripts more user friendly without needing
a full web app. And basically the pitch there was
(35:19):
that the database, the UI and the app are kind
of unified. They're all the same thing. Astro and vim,
one of the better vim distros. I do prefer lazy vim,
although I hate the name. Reminds me of the dummies books,
which I could never get myself to buy. Too much self-respect.
Interview DB lets you access crowdsourced tech interview questions from
(35:42):
the community. Alexander Solzhenitsyn on the future of the West.
This was a crazy, crazy piece of text there. Seneschal MCP.
MCP by a brilliant Ull member named Matt, who, um,
wrote an MCP that helps Llms access standardized health data
(36:03):
from the sky. Senegal. Senegal. I keep thinking, Senussi. What
am I thinking of? Senegal. API. So basically, this is
an NCP that makes your health data available so that
you can have local agents or any kind of agent
really make your health data available to an app. So
(36:23):
if you're building like a full life encompassing, you know,
management app with like workout routines, health, fitness, food, diet, money,
all that stuff. This is an API. This is an
MCP that presents the health side of it. Very very
cool from UL. Member Matt Self interview AI tool that
(36:45):
runs invisibly during technical interviews, helping you with coding questions
while your screen is shared. Also known as cheating. So
this is what you have to watch out for if
you're doing interviews. Piano FM generates endless beautiful AI piano music.
I listened for a while. It was not repetitive and
it sounded. I'm not a classical music expert or a
(37:07):
piano expert, and it sounded pretty normal to me. It
sounded like regular piano, whatever that's worth. Matrix hacking game.
Someone built a matrix themed AI hacking game lets users
learn about prompt engineering and security by defeating an Agent
Smith AI character. Send new app connect your iPhone camera
(37:28):
to multi-modal Llms for real time visual intelligence made with
zero VC money and cursor. This is the type of
thing that could have gone for, could have raised, you know,
$20 million and been sold for $100 million maybe a
couple of years ago. And this person just built it
(37:48):
with cursor with no money whatsoever. Since Feedback Dev built
an AI that reads the emotional vibe of your product,
social media mentions and gives you instant feedback. And somebody
posted their cursor rules, which I thought were quite good.
All right. This is the end of the standard edition
of the podcast, which includes just the news items for
(38:11):
the week to get the rest of the episode, which
includes my analysis, the discovery section that has all of
the coolest tools and articles. I found this week, the
recommendation of the week and the aphorism of the week.
Please consider becoming a member. As a member, you'll get
lots of different things from access to our extraordinary community
(38:31):
of over a thousand brilliant and kind people in industries
like cybersecurity, AI, technology and the humanities. You also get
access to the UL Book Club, dedicated member content and events,
and lots, lots more. Plus, you'll get a dedicated podcast
feed you can put in your client that gets you
the full member edition of this podcast. To become a
(38:55):
member and get all of that, just head over to
Daniel Comm Slash upgrade. That's Daniel miessler.com/upgrade. We'll see you
next time. Unsupervised learning is produced on Hindenburg Pro using
an SM seven B microphone. A video version of the
podcast is available on the Unsupervised Learning YouTube channel, and
(39:17):
the text version with full links and notes is available
at Daniel Mysa.com slash newsletter. We'll see you next time.
Popular Podcasts
Amy Robach & T.J. Holmes present: Aubrey O’Day, Covering the Diddy Trial
Introducing… Aubrey O’Day Diddy’s former protege, television personality, platinum selling music artist, Danity Kane alum Aubrey O’Day joins veteran journalists Amy Robach and TJ Holmes to provide a unique perspective on the trial that has captivated the attention of the nation. Join them throughout the trial as they discuss, debate, and dissect every detail, every aspect of the proceedings. Aubrey will offer her opinions and expertise, as only she is qualified to do given her first-hand knowledge. From her days on Making the Band, as she emerged as the breakout star, the truth of the situation would be the opposite of the glitz and glamour. Listen throughout every minute of the trial, for this exclusive coverage. Amy Robach and TJ Holmes present Aubrey O’Day, Covering the Diddy Trial, an iHeartRadio podcast.
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
The Breakfast Club
The World's Most Dangerous Morning Show, The Breakfast Club, With DJ Envy And Charlamagne Tha God!