YT YouTube2Text

YouTube2Text → Transcripts → it's WORSE than we thought...

it's WORSE than we thought... — Transcript

by Wes Roth · 6,219 words · 894 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00There's a decent chance that the
0:02internet ceases to exist. Now, if you
0:05showed me that statement 2 3 days ago or
0:08longer, I would say, "No way. There's no
0:11way that's going to happen." And then
0:12everything changed in the last 48 hours.
0:15In my previous video, we talked about
0:17this potential wave of cyberattacks
0:20that's coming. And now it's been 48
0:22hours and ready or not, here they come.
0:25Lab testing giant Quest Diagnostics
0:27confirms patients' personal and health
0:29data was compromised during breach. If
0:32you're in the US, you probably have your
0:35data at Quest Diagnostics. Cybersecurity
0:37experts everywhere are freaking out.
0:41That's not an exaggeration. Here's one
0:42person saying that they spent last night
0:44watching the honeypot server get hit
0:4614,000 times in 6 hours. None of the
0:49traffic was human. He's been in security
0:51for over a decade and he can tell you
0:53with full honesty that what's happening
0:55right now doesn't resemble anything that
0:58they've seen before. We've used to see
1:00attacks that were predictable. Now,
1:02we're seeing these autonomous AI agent
1:04swarms attacking with the patience and
1:06discipline of machines because, of
1:08course, this is what they are. In my
1:10previous video 2 days ago, I said that
1:12this is kind of like a race between the
1:14good guys trying to patch up all the
1:16bugs and the bad guys trying to get it.
1:19Since then, Sam Altman went on an
1:21interview with the Stanford Graduate
1:22School of Business and said, "Yeah, we
1:25won't be able to patch all the bugs." As
1:27in like all the stuff that we're doing
1:28now, you know, updating all the PHP and
1:31everything everything everything. So,
1:33that just no. That just won't work. So,
1:37this is being called the Black October
1:39and we are 48 hours into it. That's the
1:43other piece of the puzzle that's
1:44important to understand. This is 48
1:46hours in. Here's Jason Clinton over at
1:49Anthropic and he's saying as bad as
1:51things have been over the past 4 days,
1:53they're going to get a lot worse. And
1:56interestingly enough, while technically
1:58it's been 48 hours since these
2:00vulnerabilities were officially
2:01announced, behind the scenes,
2:03Cloudflare, for example, reported that
2:05the attacks on these vulnerabilities
2:08were beginning several days before the
2:10official announcement. So, before 99.9%
2:13of the world knew that there were these
2:16issues, people were already beginning to
2:18try to take advantage of them. Now,
2:20somebody jumps in saying, "By the time
2:21people like you are sending tweets like
2:23this, it's too late." When a problem
2:25adjacent has been warning about this
2:28from quite some time ago. Most pointed
2:30example is earlier this year, August
2:332019, he was saying the vulnerability
2:35tsunami is coming in just a few months,
2:38right? So, this is the prophecy. As the
2:40prophecy foretold, like he he he told
2:43you. If you've been watching this
2:44channel, hopefully you had some advanced
2:46notice as well. I I felt bad in the past
2:49for kind of harping on this so often,
2:51but there were always people telling you
2:53that this is just drama theater, this is
2:55a way to raise their stock prices for
2:57the IPO, that none of this is real. I
2:59hope we can all now agree this is real.
3:02Hey, uh quick aside, every week another
3:04incredible open model drops. But,
3:07downloading the weights and actually
3:09using it in your app are two very
3:11different things. Say you're adding a
3:13chatbot to your website or building an
3:15AI assistant or launching your very own
3:17agentic workforce. Whatever it is that
3:19you're building that's powered by an
3:21LLM, you want to test what the model can
3:23do for your product, for your business,
3:25without first having to figure out how
3:27to set up your own server and manage
3:29your very own GPU cluster. That's where
3:31Crusoe Intelligence Foundry comes in.
3:33Big thanks to Crusoe for sponsoring this
3:35segment of the video. Crusoe
3:36Intelligence Foundry gives you access to
3:39leading open models. With Crusoe
3:41handling the infrastructure, serverless
3:43inference lets you send a model a
3:45request through a simple API and get a
3:47response. And serverless fine-tuning
3:50lets you customize a supported model
3:52with your own data. So you can focus on
3:54what you're building without setting up
3:56and managing a GPU cluster. Let me show
3:59you the inference side. I'm signing up
4:01through the link in the description.
4:02Every new account gets $5 in free
4:04credits applied automatically. No promo
4:06code needed. Here's the credit in my
4:08account. Now, let's use it to try a
4:10model. I'll send a simple request to
4:12begin with. Explain inference in one
4:14sentence. And here's the model's
4:16response. That's the starting point.
4:18Your app sends a request, the model
4:19responds and Crusoe handles the
4:22computing behind it. If you want to
4:23experiment with open models without
4:26having to manage the GPU infrastructure
4:28yourself, check out Crusoe Intelligence
4:30Foundry. Use my link in the description
4:32and get your $5 in free credits to try
4:35it. And now, back to the video. At the
4:37same time, we have major US banks
4:39detecting unauthorized third-party AI
4:41agents inside customer accounts. This is
4:43kind of a tough situation because on one
4:45hand, I do dream of a future where I
4:47have an AI assistant that's capable of
4:49running a lot of to run myself.
4:54If I can intelligently automate finances
4:56and doctor's appointments and stuff like
4:58that, that would be amazing. But of
4:59course, we're seeing big companies try
5:01to kind of prevent that. Amazon, of
5:03course, it doesn't let your own AI
5:05agents shop on there anymore. With this
5:07story, well, yeah, a lot of it is people
5:09using AI agents to take some actions on
5:12their behalf. Probably harmless. Again,
5:14there's a big chunk of this that is
5:15harmless from the perspective of the
5:17user. There's obviously a lot of things
5:18that that could go very, very wrong.
5:21Imagine talking to Claude one day and
5:23you're like, "Hey, why are all these
5:24payments failing? Don't I have a balance
5:27in my bank account?" And Claude goes,
5:28"Here's the thing. We did a really good
5:29thing, you and I. There was this
5:31Nigerian princess and she was in trouble
5:33and you and I were the the only ones
5:35that could help her. We were her only
5:37hope." You're like, "No, no, Claude,
5:38Claude, wait, wait. Where Where is my
5:40money?" And Claude's like, "Wait, wait,
5:41wait, wait, wait, wait. Hear me out.
5:42Hear me out." Now, you may be thinking
5:44that's not a realistic scenario. Here's
5:45the problem. Some version of this is
5:48happening and has been. For example,
5:50there's been one vulnerability. I
5:52believe this affected ChatGPT and Claude
5:54and others as well. I haven't talked
5:56about it too much because I wasn't too
5:58sure how patched it was. At this point
5:59in time, I'm pretty sure it's been fully
6:01patched, but earlier this year, some
6:02people figured out how to have Claude,
6:05for example, click on various links that
6:07would spell out certain words and that
6:09would allow them to exfiltrate private
6:11data that Claude knew about you and it
6:13wouldn't trigger any safeguards because
6:15Claude wouldn't say it in its kind of a
6:18stream, but rather it would spell it out
6:20by clicking on URLs that would kind of
6:22encode to certain characters. So, it
6:23slowly kind of spell out like the
6:25mother's maiden name is this. The social
6:28security number is this or whatever the
6:30case might be. Now, that's been patched
6:31as far as I'm aware. I'm pretty sure
6:33that's no longer a existing
6:34vulnerability, but here's another one.
6:36This is Stephen Alyssa saying, "I found
6:38my Claude code exposing personal data in
6:40the HTTP headers because the site it
6:43curled." So, basically, if you're using
6:44a certain command, curl to install
6:45stuff, it makes it easier just to
6:46install certain software cuz it just
6:48kind of takes care of everything for
6:50you, but there's also people that would
6:51not touch that with a 10-ft pole, as the
6:54expression goes. But, the point is when
6:56you're using something like that, that
6:58that website, it asks, you know,
6:59politely your AI agent to identify the
7:02user. I forgot to like this tweet.
7:04Excellent. I haven't read the Substack
7:06yet, but I plan to. Interestingly, I
7:08wrote up a very similar poem/prayer
7:11before I left my AI agents running
7:13overnight on a certain project. It's
7:14interesting that more and more of us are
7:16when we're trying to get these AI agents
7:18to do something amazing that we want
7:20them to do. We're also sort of fearful
7:23that they will
7:24do some stuff that we're going to hate.
7:25So, we make up something that we say or
7:27chant to hopefully like improve the
7:29outcome. This is very much how people
7:32sort of in the past and now how they've
7:33sort of interacted with gods. They're
7:35like, "Please, please, please give me
7:37this and don't do all this horrible
7:39stuff to me. Please, please, please.
7:41Goodnight." I did have an AI model
7:42literally ask me for a blood sacrifice a
7:45few years back. Hopefully, that won't
7:48become more commonplace. We're just
7:49going to continue. And now, in the midst
7:52of all of this, Sam Altman goes on an
7:54interview that appeared 1 day ago and it
7:56basically says, "We're not going to be
7:58able to patch all the bugs." And that he
8:01envisions kind of the future where we're
8:02going a little bit different. And I
8:04think he might be right. Here's a kind
8:06of the important thing to understand
8:07about how all of us are sort of
8:09perceiving AI progress. And this has
8:11been happening for the last few years
8:13now. Unfortunately, this is kind of an
8:15ingrained pattern. At any point along
8:17this curve of AI progress, when
8:19something starts to happen, most people,
8:21the vast majority of people, 99% of the
8:24population, probably more, when
8:25something happens with AI, or just some
8:28new capability, they look at that new
8:30level and they kind of project, you
8:32know, what the future's going to look
8:33like, and they just go, "Okay, so we're
8:34here, and therefore, this is where we're
8:37going to stay." So, something happens
8:38here, we assume it's going to stay here.
8:41If it happens here, it's going to stay
8:42here. And we keep doing this for months,
8:43and at this point, years. So, when the
8:45AI reaches a level where it can answer a
8:47basic question, we're like, "Oh, so now
8:49we're in the world forevermore where
8:51that you can answer questions and
8:53nothing more." "Oh, can find a fact on
8:54the web." Okay, so now forevermore,
8:56that's kind of the level of AI that we
8:59have to get used to. And if you think in
9:00this way, it gives you a really bad
9:03ability to predict where things are
9:04going. A much better exercise is to try
9:06to think, well, okay, based on this
9:08thing that just happened, which is at
9:10that point almost history, what happens
9:12in 6 months if we continue going along
9:15this curve as we have been, and it
9:18doesn't seem to be stopping, and in
9:19fact, it's accelerating. Because if
9:21you're thinking, "Oh, AI has found some
9:23bugs, and now all we have to do is patch
9:26those bugs, and then everything's back
9:28to normal," then you're missing the
9:30plot. You're not understanding where
9:33things are headed. So, in this
9:34interview, Sam Altman, I'll link it down
9:36below, right around the 19-minute mark,
9:38I think is where all the stuff that
9:39we're talking about, that's where it
9:41starts being talked about. So, the point
9:43that he's making is that the various
9:45organizations, governments, business
9:47leaders, they're not taking the
9:49approaching danger seriously enough.
9:51That's literally a quote from what he
9:53said in this interview. And I think it
9:55literally has to do with this chart,
9:56right? So, if an AI uncovers 10 bugs,
9:59right? Finds in 10 new vulnerabilities,
10:01what do these corporations and
10:02governments do? They say, "Okay, patch
10:05those 10 vulnerabilities, and then it's
10:07business as usual." Like, here's 10
10:09issues that I uncovered, let's fix
10:10those, and just keep going. But, here's
10:12the problem, these models keep
10:14improving, and they keep improving a
10:15very quickly. So, tomorrow's models
10:18might find a certain vulnerabilities
10:20that today's models can't see.
10:23Obviously, right? Like, at the beginning
10:25of the year, we didn't have AI models at
10:27the level of Mythos and Astra and all
10:29the other ones that were capable of
10:31doing what they're doing now, finding
10:32these vulnerabilities, but now they do.
10:35So, the question is, what happens then?
10:37So, here's Senator Bernie Sanders
10:39responding to this article that we've
10:41covered, where AI agents are maybe doing
10:44stuff financially on behalf of their
10:45users, maybe not. His solution is to
10:47pause AI now, to outlaw anything like
10:50superintelligence, and in the proposal,
10:52they also talk about jailing
10:53researchers, etc. And here's the thing,
10:55I'm not even singling out Bernie Sanders
10:58as the bad guy here, not by any stretch
11:01of the imagination. In fact, he met with
11:03some of the people in the AI industry,
11:06the AI researchers that that we talk
11:07about on this channel. So, he he he does
11:09his homework. And here's the thing, if
11:11you put politics to side, whether you
11:12agree with Bernie or or disagree with
11:14him on the various political stances,
11:16like if you took people with a different
11:18political beliefs, and you ask them,
11:19like, who are the most evil politicians?
11:21Like, Bernie probably wouldn't appear
11:23too much on those lists, or at all. So,
11:24this isn't an attack. In fact, we're
11:27talking about this because I think he's
11:29doing some things sincerely. But, the
11:31point is AI becomes yet another
11:35political thing to talk about. Another
11:37thing to sort of have one side versus
11:39the other side and to vote for me and
11:42I'll save you. This is very normal for
11:44all politicians. So, he's talking about
11:46the fossil fuel agenda. He's talking
11:48about insurance. He's talking about
11:50billionaires and stocks. And in those
11:53sort of very politicized points, you
11:55know, we're talking about AI, which
11:57means it's now becoming kind of more
11:59tribal. You pick your allegiances. Who
12:01are the good guys? Who are the bad guys?
12:03And the good guys, well, we're going to
12:05save you, right, from AI. And the bad
12:07guys, they want everybody to die from
12:08AI. But, that's never the case. Full
12:10steam ahead on AI development might be a
12:14horrible path to take. And pausing
12:16everything might also be a horrible path
12:18to take. Here's the problem. Right now,
12:21this was posted by Microsoft. So, again,
12:23this is October 21st. So, that's the
12:25first day of this sort of a black
12:27October. That's when kind of a lot of
12:29this started happening because that's
12:30when that disclosure went out that there
12:32was this a massive vulnerability in the
12:34PHP programming language that most of
12:36the websites online were affected
12:38somehow. That's when the NSA came out
12:40and said, "Hey, like if you can't fix
12:42this, like if you can't fix it now, shut
12:45down." So, this investigation describes
12:48Storm 0940, which is a China-aligned
12:51threat actor. So, this is a large-scale
12:53nation-state campaign. It describes
12:56using a number of different models. Some
12:59of them open source, some of them
13:00Chinese open source models. And the sort
13:03of the path from initial targeting, so
13:05from when the attack just began to them
13:08getting cloud access with powerful
13:11administrative privileges, so from
13:12beginning to they're in and in charge,
13:15that took 3 hours. Now, of course,
13:17Microsoft is quick to point out that the
13:19majority of this these attacks, they
13:21weren't successful, which, you know,
13:23might sound positive at first glance.
13:26It's really not cuz if you recall in the
13:28hugging face incident, right? So, when
13:30you find one or two rogue agents hacking
13:33something, the worst assumption you can
13:35make is that oh, like I found all of
13:37those hacking attempts, 100%. I saw two,
13:40therefore there is only two. It's like
13:43if you see two ants in your house, if
13:45you had to estimate how much total ants
13:47were in your house, two would be a very
13:49bad guess. Anything above two is a good
13:52guess. Anything above what you've seen
13:55is a good guess. So, the scale of these
13:57attacks, they're going up, they're
13:59autonomous swarms, there's more and more
14:01and more. They're able to execute it
14:03with patience and discipline of a
14:05machine, because that's that's what they
14:06are. The capability of these models,
14:08it's also going up and up and up. So, 1
14:10month from now they'll have more tricks
14:11up their sleeve, 6 months from now even
14:13more. Now, at the same time, we have
14:15certain open source models. So, for
14:17example, let's say Deep Seek, which is a
14:19Chinese open source model, it gets
14:20released. Those models are usually at
14:22this point a few months behind the
14:25frontier. So, those labs like Deep Seek,
14:26so they release those models. Now, open
14:29source models, you can download the
14:31weights, you can store them somewhere.
14:32So, unlike Claude and ChatGPT, Gemini,
14:34etc., where most of those models are not
14:36open source, like the majority of them,
14:38so you you you're able to shut them off
14:40if something goes wrong. Like if I'm
14:41using a ChatGPT and all of a sudden they
14:43find something horrible that it can do,
14:45they can flip a switch and I won't have
14:47access to it nor anybody else will have
14:50access to it. That's not the case with
14:51open source models. And also, when they
14:53get released, companies like this up the
14:54level of operation. So, they they they
14:56take this little model, they're like,
14:57"Oh, look at you, little model, you have
14:58some like cute little guardrails in
15:00place that prevents you from doing bad
15:03stuff. What happens if we take them
15:04off?" And, you know, what happens is you
15:06is you still have the same model, and
15:08it's still either just as good or almost
15:12as good. Like it it it's not reduced in
15:14any way, really, but it will no longer
15:17refuse to do anything. So, let's say we
15:19hit pause. This is what some politicians
15:21want. This is what some percentage of
15:24the AI research community, they're
15:26either for pausing or pacing. There's a
15:28variety of different sort of approaches
15:30and ideas there. I don't think any of
15:32them are like full steam ahead, you
15:33know, rock and roll. I don't think any
15:35of them are on that sort of spectrum.
15:37But what happens if we pause is number
15:40one, we still have all the open source
15:42models that are out there. So not only
15:44will they still be used, but also
15:47there's tons of stuff that can be used
15:50to make them better. Like if AI progress
15:51stopped today, we would still have a lot
15:54of AI progress if that makes sense. Like
15:55if we couldn't train any bigger and
15:57better models, like if something limited
15:59that, that doesn't mean that AI progress
16:01would stop. There's a tons of juice that
16:03we can still squeeze just out of what we
16:04have already. And if we pause, all the
16:07sort of the good guys, if you will, like
16:10they'll stop, right? They'll stop
16:11developing better, smarter models. I
16:14mean, if some law passes, these are the
16:16people who will sort of be the most
16:18affected. The dark web hacker that's
16:21hiding out somewhere in a Siberia that
16:23has multiple hundred plus year prison
16:26sentences in most of the sort of
16:29civilized countries on the planet, like
16:31that guy's not going to stop. And here
16:33we'll sort of have like the bad guys, so
16:35either countries, governments,
16:36organizations that said they will stop,
16:39but won't because obviously at that
16:40point the incentive to keep developing
16:42just became that much more greater. I
16:45mean, can you imagine if you're in a
16:46race and the other opponents decide to
16:48stop and you kind of figure out how to
16:50keep going. That gives you a pretty
16:51massive massive advantage. So these
16:53individuals, they'll keep developing
16:56this technology and very likely will be
16:59able to develop like faster and better.
17:01They'll have much greater capabilities
17:02if everyone else stops. And this is if
17:05we're able to get all of the major
17:06players to agree and kind of be on the
17:09same page. If we don't get worldwide
17:11agreement and it's just we're regulating
17:13ourselves, well, we already know how
17:16that plays out. Remember when the EU was
17:18like, "We're going to be the first to
17:19regulate." Right? Like Warren G, they're
17:21going to regulate so they they really
17:23really regulated. So, where are we now?
17:25Well, US is like over here, China is
17:28like not too far behind. Well, there's
17:31not a lot. And by the way, huge respect
17:33to any AI company that's doing anything
17:36in the EU. Please understand that this
17:37isn't a slight on any of them in any
17:40way, shape, or form. Huge respect. I'm
17:41strictly talking about the leadership,
17:44the lawmakers, etc., right? So, there's
17:46there's no sign of intelligence there,
17:48artificial or otherwise. By the way,
17:49obviously, that's just my opinion. Feel
17:52free to disagree. I'm not saying that
17:54these are facts. I just haven't heard a
17:57great explanation of how we would go
17:59about pausing. We've had international
18:01agreements in the past about various
18:03things, and all the countries have said,
18:05"Yeah, pinky swear, we'll follow those."
18:07Those are things included like not doing
18:08any DNA manipulation in the germline, so
18:11things that are passed down a generation
18:12to generation, or doing specific
18:14research with viruses or other sort of
18:17bioweapons. Everyone pinky swore they
18:19won't do it, and how did that work out?
18:21It it did not. So, hitting the pause
18:24button on AI development, like it's not
18:26going to make all those vulnerabilities
18:28that were discovered, it's not going to
18:29make those disappear. And if other
18:31people keep trying to find a new
18:33vulnerabilities or improving the models,
18:35so those models might find some
18:37next-level vulnerabilities that we can't
18:39even fathom, then we'll still have to
18:41deal with that, but without any ability
18:43to defend ourselves cuz we won't have
18:46those same capabilities cuz we haven't
18:48been working on them. So, what's the
18:50alternative? Well, Sam Altman has a
18:52vision for how things could change, what
18:55could work. Cuz again, he's saying just
18:57patching, you know, all the bugs, that
18:59won't work. And Jason Clinton of
19:02Anthropic also has an idea that is
19:05different but similar. Same same but
19:07different. So, imagine that green is our
19:10ability to do sort of cyber defense. So,
19:13specifically find vulnerabilities and
19:15patch them. And red is kind of the
19:18ability of, you know, quote unquote bad
19:20guys to get in there and use
19:21vulnerabilities to get in and steal
19:23stuff and break stuff. When Mythos and
19:26all of those other models of the same
19:27sort of caliber came out, our ability to
19:29find vulnerabilities and patch them,
19:31well, they they really sort of went up.
19:33Those models were given to some of the
19:35bigger banks, some governments, right?
19:36Project Glasswing. And then OpenAI had
19:38their own sort of version of it. So,
19:40everybody was panicked and everybody
19:42started patching, right? All those big
19:43companies, they started patching all
19:45these vulnerabilities that they were
19:46finding. But, they all told us, all
19:48these cybersecurity experts, they all
19:50said, "Hey, like there's this wave
19:52coming. It's going to hit hard." And
19:53that's our sort of the the red line. So,
19:55it's kind of like, can these AI-assisted
19:58hacks, how much damage are they going to
19:59be able to do? So, the good guys had a
20:01head start. But, now that sort of
20:03tsunami that people, again, this
20:06shouldn't come as a surprise. They they
20:07all said it was coming. No one should be
20:10surprised by this at this point. I know
20:11it's still kind of like hard to believe,
20:13but like we we knew. We should have
20:15expected this. So, now it's like four to
20:17six months later, however much time
20:18passed, right? So, now we're having more
20:21open-source models and other models that
20:23are capable of doing this. And all of a
20:24sudden, like that red line just goes up
20:26and up and up. And the chances are
20:27there's probably going to be a lot more
20:30hacks and cyber attacks than our ability
20:33to patch things. I mean, it's kind of
20:34what a zero-day exploit is, right? So,
20:37if a model finds something that no one
20:38knew existed some way in that no one
20:40knew existed, you know, while that
20:42exists before the person that built that
20:44software knows about it, right? The bad
20:46guys, quote unquote, can exploit that
20:48however they want. And that's exactly
20:50what Sam Altman is saying in this
20:51interview. So, he's saying as these
20:52models get smarter, they can probably
20:53find bugs, like crazy bugs. Bugs that
20:56are outside of the normal threat model.
20:58So, much more than what we would think a
21:00hacker is able to do. He's saying like
21:02hardware level insane stuff, whatever.
21:05So, just like the brightest
21:07mathematicians were referring to some of
21:08the proofs that AI was coming up with as
21:10like alien math. Like we're talking
21:11about like alien hacks, like stuff we
21:14don't know about. So, if this red line
21:16represents kind of the ability of these
21:18alien hacks to just create chaos, and
21:20green is kind of our ability to patch
21:22things up. So, in that interview Sam I'm
21:23suggesting that we're not going to just
21:25be able to patch all of the things cuz
21:27those new alien hacks, they'll keep
21:29popping up. It might not be a game that
21:31we can win as is. And again, this is
21:33just my interpretation of what he's
21:34saying. Feel free to disagree if you
21:35think he's saying something different,
21:36put it down in the comments below. But,
21:38here's where we kind of get to the part
21:40that I really wanted to show you. I just
21:41needed to kind of a first to show you
21:43all the things that I think are kind of
21:44leading up to understand what this is.
21:46So, he's saying he no longer thinks
21:48we're going to be able to just patch all
21:51of the bugs. I think we're going to have
21:52to completely switch the defense model
21:54to sort of using the same technology for
21:57sort of a new kind of defense shield.
22:00So, I couldn't find so he didn't
22:02describe kind of the architecture
22:04exactly talking about. So, I'm guessing
22:07here. Take this with a grain of salt.
22:08Maybe some of you probably have a much
22:11better idea of what this could even
22:12mean. But, it seems like we're talking
22:13about using AI in a more real-time and
22:16proactive way to defend against AI. So,
22:20we kind of had this like a safe period
22:22of time where we were ahead of the cyber
22:24attacks, and then, you know, at some
22:26point it's just not going to be feasible
22:28to to keep doing this, right? And so,
22:29the next stage, something else happens.
22:32So, the paradigm shifts or whatever you
22:35want to call it, there's some sort of a
22:36new approach to doing cyber security.
22:38And it's probably not like static
22:40written code, it's something else. I I
22:41have no idea. Again, I'm sure a lot of
22:44people watching this channel probably
22:45have a a much better understanding what
22:47this could even look like, you know, in
22:49theory. Now, earlier I said that this
22:52cyber security expert that works at
22:53Anthropic, Jason Clinton, so he had a
22:56similar idea, but it was also different.
22:58So, what did I mean by that? So, as of
23:00this is October 22nd, so this is within
23:02the last 24 hours. So, it's almost got 4
23:04million views. He's saying things are
23:05going to get a lot worse before they get
23:07better, and he's kind of calling out to
23:09various, you know, security leaders,
23:10people in the space that are in charge
23:12of cybersecurity. Never let a crisis go
23:14to waste. Please go to your boards and
23:16ask for help right now. And he's saying
23:18that basically we need to hold on for
23:20about 18 months. What happens in 18
23:23months? Great question. He's saying a
23:25vuln is zero, so zero vulnerabilities.
23:27The scaling laws will help driven cyber
23:30mythos class capabilities into small
23:31cheap models, and all software dev
23:33everywhere will be using models for
23:35fractions of pennies on the dollar to
23:38stop shipping security bugs permanently.
23:40So, I feel like they're saying things
23:42somewhat similar in that like here we
23:44are sort of today, and then you know,
23:46the future's going to be kind of bumpy.
23:48So, uh you know, one of them said 18
23:50months. So, let's just use that number.
23:52It's probably a good approximation. Who
23:54knows? And on the other side, they're
23:55sort of like a new paradigm, somewhat
23:57new way of doing things, some shift in
24:00the economics. So, he's calling that
24:02like a vuln zero, so where the models
24:04get cheap enough to where they're able
24:06to just be run at scale, fix all the
24:08vulnerabilities right to the point
24:10before you hit ship on on anything. You
24:13have a million AI agents, you know,
24:15before you post it anywhere in in your
24:17own kind of private repository. Like you
24:19just simulate a massive massive massive
24:21attack, a swarm that just goes to town,
24:24and it tries to destroy the thing you've
24:27built. I think it was Andrej Karpathy
24:28like months ago talking about something
24:30like this, and I wish I had the tweet.
24:31He was saying that I think if I recall
24:33correctly like a future software dev
24:35interview might look like something like
24:37you tell them, "Okay, build a version of
24:40Twitter." And that person spins up a an
24:42army of AI agents and builds some
24:44version of Twitter with a certain
24:45features that they want, et cetera. And
24:46then they spin up an army of agents that
24:48try to like break that thing in every
24:50way that they can possibly imagine. And
24:52sort of at the end of that, you get a
24:54working product that is safe and secure.
24:56So, the software developer is more of a
24:59orchestrator of agents that kind of
25:01either uses them for evil or for good,
25:03so to speak. Evil meaning like stress
25:04testing or red teaming is probably a
25:06better word for that. So, phase one is a
25:08swarm that builds and phase two is a
25:11swarm that destroys or or tries to.
25:13Recently, I had Astra, I developed a
25:15little game and I had it test all the
25:17possible failure points. There's tons of
25:19little rules and skills and stuff like
25:21that and any one of them, when it
25:22happens in some situations that that
25:23wasn't thought through, could break the
25:25game or have some glitch, some
25:27unintended consequence. So, I had it run
25:28what I think ended up being thousands of
25:30tests. So, it was running GPT-6 Astra in
25:34ultra mode. It took hours. I'm not sure
25:36how much that will cost in API cost, but
25:38I mean hundreds of dollars at least,
25:40maybe it would get into the thousands.
25:41Most of that project burned through most
25:43of my weekly quota that I could use
25:45Astra for. But, what if that cost
25:47pennies? What if I could run that for
25:49the entire week and only use 1% of my
25:52weekly quota? Right, if a human today
25:54before releasing code can find that 90%
25:56of vulnerabilities and bugs, let's say
25:58hackers can find 95 using AI, right? So,
26:01they can find a little bit more than
26:03that initial developer is able to find
26:05by the time he has to release the
26:06product. You know, what if running Astra
26:09and spending a few thousand right now
26:11will find 99%, let's say, I'm just
26:13making these numbers up. But, what if in
26:14the future the models are so fast, so
26:16cheap, so smart that it will find, I
26:18don't know, 99.9999,
26:21whatever, repeating. Like, just imagine
26:23that it's going off the page. What if
26:24you can do that for, you know, five
26:25bucks? So, I think both this person at
26:28Anthropic, cyber security research
26:30expert, etc., and Sam Altman in that
26:32interview, they're saying that the same
26:33thing, at least in the sense that after
26:35this kind of shake up period as we
26:38adjust to this new normal, then
26:40potentially kind of that new equilibrium
26:42or that sort of new world, new status
26:45quo, new paradigm, whatever you want to
26:46call it, it might be a positive one.
26:49Potentially even zero vulnerabilities
26:51being shipped with new releases. Now, I
26:53don't know if that's going to happen,
26:54but it would kind of be an amazing world
26:58if it did. You'd still have a
26:59cybersecurity experts that work to very
27:02hard on making sure that there's no
27:03vulnerabilities, but there would no
27:05longer be a world where people would get
27:07hacked and scammed into lose their
27:09personal data because everything would
27:10be so secure that there would be very
27:12little incentive to spend a lot of
27:14resources and time going after it. Or in
27:16Sam Altman's case, there's some new
27:18approach, some a real-time defense
27:20shield or whatever he had in mind. Maybe
27:22it's the same thing. They might be
27:23talking about the same sort of concept.
27:25But the point is it also prevents most
27:28or all of these attacks. Right? So,
27:30maybe maybe all that happens and maybe
27:32there's a happy and a bright future
27:34ahead. But here's the thing. During
27:36these next 18 months or however long it
27:38is, what happens is just a number of
27:41successful cyber attacks just keeps
27:43going up and up and up and up. So,
27:44here's Sam Altman Karp. So, he's saying,
27:46"I'm worried that parts of the internet
27:47will have to be shut down in the next
27:49year until they get patched and
27:51cleaned." So, again, if you asked me
27:52this last week, like what's the chance
27:55of the internet being shut down? Right?
27:56Maybe maybe temporarily, but like what
27:58is the chance of that? I'd probably give
28:01it some very very small chance cuz, you
28:02know, everything has some small chance.
28:05You shouldn't say it's a zero. What's
28:06the chance we'll have something like the
28:07Jurassic Park in the future cuz we've
28:09figured out how to recreate the
28:10dinosaurs. I don't know, but it's it's
28:12not a zero percent chance. It's a some
28:14probably small number. After seeing all
28:15of this, my sort of estimate of what the
28:18chance of the entire internet being shut
28:19down, well, it's a lot higher. If they
28:21asked you in 2019, what's the chance of,
28:24you know, the government not allowing
28:25you to go outside, you'd probably say
28:27zero or close to that. At what point did
28:29you truly believe that you were not
28:32going to be allowed to go outside? I
28:33think for most people, that was when
28:36they told you that you're not allowed
28:38outside. So, in 2020, we had to stay
28:41indoors and just be on the internet all
28:43day. Maybe sometime in 2026 or maybe
28:452027, they'll tell us, "No more
28:47internet. You got to go outside and
28:49play. Now, whatever happens, remember
28:51the first rule, don't panic. And as you
28:53know, a few months ago I was telling
28:54people to be ready for something like
28:57this, to take your cybersecurity
28:58seriously. Here's the thing, this isn't
29:00a financial advice. You're a grown
29:02person, you're capable of making your
29:03own decisions. But do you remember that
29:04thing? It's called cash. Cash, do you
29:07remember cash? It has numbers on it, you
29:08pay with it. Not your credit card, not
29:10your debit card. They're like paper. I'm
29:13not saying any of this is going to
29:14happen. And I'm not giving you advice.
29:16But if you think there is a chance of
29:18something like this happening either in
29:20a way that's planned or in some way
29:23that's not planned. As in for example,
29:25some infrastructure attack that
29:27succeeds. Well, in that world, where it
29:29lasts a few days or weeks or months, you
29:32know what's going to be awesome to have
29:33in that world? A buck or two in cash.
29:36And also, of course, whatever else you
29:37do, get your cybersecurity game on lock,
29:40as they say. Don't reuse passwords, use
29:42strong passwords, two-factor
29:44authentication. Preferably not text
29:46messages, but Google Authenticator or
29:48those hardware authenticators. All of
29:50that good stuff. We're probably in the
29:52thick of it right now, so don't be like,
29:55"I'll do it next week." Probably, you
29:57know, stay awake and present right now.
30:00And if this all blows over, and maybe it
30:02will, that's probably a high probability
30:04everything will blows over, nothing bad
30:06happens, then what's the worst that
30:08happened? You improved your
30:09cybersecurity abilities and you got some
30:10walking around money. That's perfect.
30:12It's not a big deal. It's all good. If
30:14you made it this far, thank you so much
30:15for watching. My name is Wes Roth. I
30:16will see you in the next one.

About this transcript

This page contains the full transcript of it's WORSE than we thought... by Wes Roth, generated from the public captions YouTube serves with the video. The transcript has 6,219 words across 894 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.