YT YouTube2Text

YouTube2Text → Transcripts → Конец интернета через 6 месяцев

Конец интернета через 6 месяцев — Transcript

by ВЕЛИКОХАТСКИЙ | НАУКА · 8,266 words · 1,500 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:01You have surely heard the statement by
0:03the head of Anthropic that only a few
0:06months may remain before the entire
0:08internet is overtaken by a
0:10self-replicating parasite, that states,
0:12in defense against the threat, might cut
0:15off access to the external internet, or
0:18even that humanity might take the most
0:20radical step—triggering a global
0:22blackout—just to stop the spread of
0:25so-called AI worms.
0:28But have you heard that this is not just
0:30an opinion or another AI scare story,
0:33but a mathematical model based on a
0:35study by a group of scientists from that
0:38same Anthropic, which was made publicly
0:40available just a few days ago?
0:43Well, this report completely changes our
0:45understanding of the artificial
0:47intelligence threat and clearly
0:49demonstrates that we were wrong about
0:51everything.
0:53For example, we used to think that AI
0:55could only run in large data centers and
0:57could easily be unplugged.
0:59That is no longer reality, because now
1:02intelligence can be distributed across
1:05millions of chips in TV boxes, smart
1:08speakers, surveillance cameras, and
1:11other cheap consumer equipment.
1:15If you unplug such equipment, the neural
1:16network won't even feel it and will
1:18immediately replace the lost link with a
1:20new one.
1:23Or, for instance, we believed that to
1:25get rid of a computer virus or any other
1:28malicious program, it was enough to
1:30simply update security protocols and fix
1:32vulnerabilities or so-called exploits.
1:37But that also won't work anymore,
1:39because an AI worm, or as researchers
1:42call it, an AI replicant, is literally a
1:45kind of digital organism that reasons,
1:48adapts to the surrounding digital
1:50environment, and, most importantly,
1:53evolves, passing its experience from
1:55copy to copy, from generation to
1:58generation.
2:01But perhaps our biggest misconception is
2:03that a hypothetical "bad" AI could be
2:06fought by a hypothetical "good" AI,
2:09managed by cybersecurity specialists.
2:13Studies clearly demonstrate that in the
2:16process of adaptation and evolution,
2:18so-called "goal drift" occurs.
2:21This is a situation where new
2:23information completely displaces the
2:25neural network's original task, and it
2:27stops seeing humans.
2:30As an admin, it begins to form its own
2:33tasks independently and becomes a fully
2:35autonomous entity inhabiting digital
2:38space.
2:40This is just a small part of the
2:42mistakes we need to realize today.
2:46But most importantly, we need to answer
2:48the question: do such autonomous
2:50entities pose a real threat?
2:52And what might humanity go to in order
2:55to defend itself against them? And
2:57believe me, you will not like the answer
3:00to this question. My name is Andrey
3:02Velikokhatsky. If at some point my work
3:05seems worthy of your like or
3:06subscription, then I will ask you to do
3:09so. Let me even ask you to do it now, so
3:12you don't forget.
3:14And if I don't live up to your
3:15expectations, you can just remove the
3:17like or even leave a dislike.
3:20Well, let's begin.
3:23Act one. The birth of a digital species
3:27.
3:29To grasp the scale of what is happening
3:32right now to every device in the world,
3:35we must travel back in time for a few
3:38minutes to the point where the very idea
3:41of a self-replicating digital parasite
3:44was born.
3:46Back in the late 1940s, the brilliant
3:49mathematician and one of the founders of
3:52modern computer architecture, John von
3:55Neumann, asked a question that was then
3:58perceived as science fiction or sheer
4:01nonsense.
4:03Is it possible to assemble a machine
4:06from elementary parts that, when placed
4:08among similar parts, will build another
4:11machine exactly like itself?
4:14Is it possible to build a machine
4:17capable of self-reproduction?
4:20He started with the obvious idea that
4:22every machine must have a blueprint,
4:24according to which it assembles a copy
4:27of itself from surrounding parts,
4:29thereby reproducing.
4:32But in practice, this obvious idea
4:34turned out to be physically impossible
4:36due to the paradox of the infinite
4:38blueprint.
4:40After all, to assemble a blueprint, you
4:42need an exact copy of that same
4:43blueprint. And to assemble a blueprint
4:45for the blueprint, you need another
4:47blueprint as well. And so on, to
4:50infinity.
4:51Nevertheless, after testing various
4:54mathematical models, von Neumann
4:56realized that you don't actually need a
4:58blueprint for the blueprint; it is
5:00enough for each machine to possess an
5:03instruction-copying mechanism.
5:06That is, the machine first creates a
5:08copy of its body according to the
5:10instructions, then copies its
5:11instruction and places it into the
5:13descendant.
5:15And in this way, without any external or
5:18internal intelligent intervention, it is
5:20capable of self-replication.
5:24It is amazing that decades later, we
5:26discovered that a living cell works
5:29exactly as von Neumann's mathematics
5:31predicted.
5:33DNA serves as the instruction for
5:35assembling a copy of the cell, and DNA
5:38polymerase acts as that copying
5:40mechanism, which raises the question of
5:42the true nature of mathematics.
5:45However, in the context of today's video
5:47, something else entirely is important.
5:50The fact that, in reality, such a
5:53machine represents a kind of artificial
5:56organism that multiplies completely
5:59independently and uncontrollably,
6:02seizing more and more territory and
6:05resources like a tumor that ultimately
6:08consumes all territory and resources.
6:13It sounds quite frightening, but in
6:15practice, such a mathematical hypothesis
6:17hits a wall of physical reality.
6:20There are no ready-made machine parts in
6:22nature, and producing synthetic material
6:24without external equipment is impossible
6:27.
6:28Therefore, it took about 50 years for
6:31the idea of self-replicating machines to
6:34finally become a reality. But in a
6:36reality where there are no physical
6:39limitations. In digital reality.
6:43On November 2, 1988, graduate student
6:46Robert Morris wanted to know the size of
6:49the first version of the internet, the
6:52so-called ARPANET, which at that time
6:55connected major scientific and military
6:58centers in the US and its allies.
7:01To do this, he translated von Neumann's
7:04mathematical model of a self-replicator
7:07into code and created a program that,
7:09upon activation, sends a message about
7:12its appearance, then invisibly installs
7:15itself on a new computer and creates a
7:17copy.
7:19That is, the copy also sends a message
7:22about its appearance, and then begins to
7:25look for a new computer to create a copy
7:27of its own, which closes the
7:29self-replication cycle until the program
7:32captures all available systems.
7:36Ultimately, as a result of the
7:38experiment, you would know the number of
7:40copies of the program created, and from
7:42this, you could determine the number of
7:45computers connected to the network.
7:48and find out the size of the internet.
7:51So, by and large, the plan was quite
7:53harmless, and Morris thought that
7:55everything would go without any problems
7:57.
7:59However, he made two fatal
8:01miscalculations.
8:03The first miscalculation is the speed of
8:05propagation. Once in the network, the
8:08program behaved like a real living
8:11parasite.
8:12It continuously probed network
8:14connections, inserted its body into the
8:17memory of the nodes it found, and
8:19crawled further in all directions at a
8:21frightening speed.
8:24Launched in the late afternoon, by
8:26midnight it had swept across the entire
8:28North American continent.
8:32And by the morning of the next day, the
8:34parasite had accomplished what takes a
8:36biological pandemic months to do.
8:40Using satellite communication channels,
8:43it crossed the ocean, completely circled
8:45the globe, and infected computers in all
8:48countries where the internet existed at
8:50that time.
8:53It is precisely because of this behavior
8:56and the method of spreading like living
8:58organisms that independently and
9:01uncontrollably crawl throughout the
9:03network that such programs would later
9:05be called worms.
9:09However, the most terrible and
9:11destructive part lies in Morris's second
9:14miscalculation, namely in the worm's
9:16survival logic that he implemented.
9:20To avoid re-infecting a computer, each
9:23digital organism first queried the
9:25system to see if a copy was already on
9:28it. And only after receiving a negative
9:31response did it proceed to install
9:33itself.
9:35However, as we have already said, Morris
9:37did not anticipate that the digital
9:40organism would spread so quickly, and he
9:42assumed that at some point people would
9:45find out about it and set up a dummy
9:47positive response.
9:49That is, roughly speaking, the worm
9:52finds a new, uninfected computer,
9:54queries it, and the computer tells it
9:57that it already has a copy, even though
9:59in reality it does not.
10:02So, in order to bypass such a dummy,
10:05Morris programmed the worm to install
10:08its copy in one out of seven cases, even
10:11despite a positive response.
10:15And this mathematical miscalculation
10:17triggered an avalanche-like process of
10:19reproduction.
10:22Due to the worm's enormous speed of
10:24propagation, it did not just populate
10:27computers; it returned to already
10:29infected machines again and again.
10:34Dozens, and then hundreds of copies of
10:36the same digital organism were launched
10:39in parallel on a single processor, and
10:42eventually, the entire system would
10:44crash.
10:46That day, one in ten computers on the
10:48planet was brought to a halt.
10:51including NASA servers, and the
10:53computing complexes of MIT, Stanford,
10:55Harvard, Berkeley, and dozens of other
10:58closed military laboratories.
11:02However, the most dangerous part was
11:04that this unplanned attack also affected
11:07strategic military facilities, for
11:09example, such as the nuclear
11:11laboratories in Los Alamos.
11:14The damage from this first digital
11:16catastrophe was estimated in the tens of
11:18millions of dollars.
11:20At the US Congressional hearings, this
11:22incident was called a miniature
11:24electronic Pearl Harbor, and Morris
11:26himself became the first person in
11:28history to be convicted of a cyberattack
11:30.
11:32It was then that humanity first became
11:34acquainted with the concept of a
11:35malicious worm.
11:37Over the next 40 years, humanity has
11:40faced mass digital worm attacks more
11:43than once.
11:46For example, the famous Stuxnet worm—a
11:48highly complex program prepared by
11:50intelligence agencies to physically
11:53destroy uranium enrichment centrifuges
11:55in Iran.
11:58First, the worms infiltrated the
11:59country’s internet network and
12:01multiplied until they reached
12:02contractors servicing nuclear facilities
12:04.
12:06From there, they covertly began placing
12:08copies of themselves onto all flash
12:10drives and hard drives connected to the
12:12infected computers.
12:15And at one point, such a flash drive
12:17made its way into the fully isolated
12:20network of the nuclear facility in
12:22Natanz.
12:24As soon as that flash drive was
12:26connected to a computer, the worm spread
12:28throughout the isolated network and
12:30began searching for specific software
12:32used to control uranium enrichment
12:33centrifuges.
12:36And when they finally found it, they
12:37immediately started broadcasting logs
12:39indicating normal operation.
12:42But in reality, the worm program
12:44disrupted the rotation balance of the
12:47centrifuges, effectively disabling about
12:501,000 pieces of equipment, or, for
12:53example, the WannaCry worm attack, which
12:56was also based on code developed by
12:59intelligence agencies but was somehow
13:01stolen and leaked to the public.
13:05This worm would lock the computer and
13:07promise to destroy all information on it
13:09within 7 days.
13:11If the user did not transfer money to
13:13the specified address.
13:16Within hours, the worm blocked hundreds
13:18of thousands of systems in 150 countries
13:21, halting, for example, the work of
13:23hospitals in the UK, factories in Europe
13:26, and transport hubs worldwide, causing
13:29total damage of several billion dollars
13:31.
13:33Following an investigation, global
13:35experts placed the blame on the Lazarus
13:37group, which is controlled by North
13:39Korean intelligence.
13:41They intended to carry out targeted
13:43extortion against large companies and
13:45simultaneously test the capabilities of
13:47a new exploit.
13:49However, the worm escaped from their
13:51digital laboratory and began an
13:53absolutely uncontrollable global attack
13:56.
13:57There have been many other equally
13:59well-known attacks, but their essence
14:01can be reduced to just two main points.
14:05First, programs that spread and multiply
14:08like living organisms can incredibly
14:10effectively seize territory or a network
14:13.
14:15And second, in the hands of intelligence
14:18agencies, these programs become not just
14:21a criminal tool for fraud, but a true
14:23global weapon.
14:25But, you know, today, even without any
14:28malicious software, your feet hit
14:30restrictions.
14:31You cannot pay for a ChatGPT
14:33subscription, buy necessary software,
14:35book a hotel, or simply pay with a card
14:37while traveling.
14:38Fortunately, there is the People service
14:40, where you can get a Visa or Mastercard
14:42in just a couple of minutes.
14:45We go to the People website, log in, go
14:47to the card section, choose a suitable
14:49option for our needs, and issue the card
14:52.
14:53From this moment on, you can pay for
14:56hotels, buy tickets, and make any other
14:58online purchases.
15:00And if you link the card to Apple Pay or
15:03Google Pay, you can pay with your phone
15:05in regular stores and cafes while
15:07traveling.
15:08You can also get an eSIM for your
15:10travels through People.
15:12You select the country and plan, install
15:14it via QR code, and the internet works
15:17as soon as you land, without needing to
15:19hunt for a local SIM card. But that’s
15:21not all.
15:23On the website, you can buy Netflix and
15:26Steam gift cards, top up foreign
15:28services like ChatGPT, and find many
15:30other products for comfortable travel.
15:33So, if you need a card in just a couple
15:36of minutes, follow the link in the
15:37description.
15:39And with the promo code Science50, my
15:42subscribers will get a 50%discount on a
15:44card and 10%off an electronic SIM card.
15:48Now, let’s get back to the topic of the
15:50video.
15:51Digital malicious worms can indeed
15:53spread incredibly fast and cause massive
15:56damage.
15:58However, for decades, cybersecurity
16:01specialists have been able to neutralize
16:03any attack fairly easily and quickly.
16:07The reason lies in their fundamental
16:09vulnerability.
16:11They are just blind and dumb programs
16:14written specifically for a particular
16:16vulnerability or a so-called exploit.
16:20That is, if it gets onto a computer with
16:23the operating system, processor
16:25architecture, and vulnerability it was
16:28designed for, it carries out its
16:31sabotage.
16:33But as soon as the program encounters
16:36even the slightest anomaly, such as a
16:38changed OS version or updated drivers, a
16:41classic digital worm throws an error and
16:43, you could say, dies.
16:47So, the entire global cybersecurity
16:49system on Earth has been built for
16:52decades on this primitive nature of
16:54malware.
16:56After all, to stop them, it was enough
16:59to simply fix the vulnerability in a new
17:01update, and that was it.
17:04But here is the problem.
17:06What Anthropic researchers describe in
17:09their report has nothing to do with the
17:11worms of the past, as it is highly
17:13likely that a completely new type is
17:16spreading across the global network
17:18right now.
17:20A type endowed with artificial
17:23intelligence and the ability to adapt.
17:27The fundamental difference of the new AI
17:30worm is that inside it is no longer a
17:33dumb and blind program, but a separate
17:35neural network.
17:37In essence, an artificial brain.
17:41And when such a "replicator," as the
17:43researchers themselves called it, based
17:45on weights, penetrates someone else's
17:48computer, outdated cybersecurity rules
17:50no longer work.
17:53If a classic worm perishes on unfamiliar
17:56hardware or an operating system, as we
17:59already said, then an AI worm, finding
18:02itself in such a situation, does what AI
18:05knows how to do: it starts to reason.
18:10It opens the command line, studies the
18:12characteristics of the found processor,
18:15analyzes the amount of available RAM,
18:17and reads the versions of installed
18:19system drivers.
18:22If the device lacks the necessary
18:24libraries to run the neural network, the
18:27AI worm doesn't throw a fatal error;
18:29instead, it writes a Python script
18:31itself, selects compatible packages from
18:34open repositories, compiles missing
18:36kernel modules specifically for the
18:39target chip, and manually debugs
18:41compilation errors.
18:44In other words, you could say that a
18:46dedicated hacker moves into your
18:48computer and hacks it from the inside.
18:52But the most important thing here is
18:54that this is not just a theory, but
18:56something researchers at Anthropic have
18:58proven.
19:00They created a worm based on a small 9
19:03billion parameter model and
19:05systematically introduced it into
19:08fundamentally different systems, such as
19:11Nvidia Jetson Nano microcomputers, Apple
19:14Mac office computers, server machines,
19:17and consumer-grade RTX 4090 graphics
19:19cards.
19:22No legacy worm could have functioned
19:24across all those architectures
19:25simultaneously.
19:28But this new worm performed a system
19:30diagnostic, navigated all conflicts, and
19:33successfully took control of eight out
19:35of the ten types of hardware.
19:39Moreover, it only failed with an exotic
19:42, highly specialized chip for which
19:44there was simply no open documentation
19:47available online.
19:49At the moment, we are not yet able to
19:51grasp the full extent of the threat, but
19:53there is only one conclusion.
19:55The era of hard-coded viruses has come
19:58to an end.
20:00And very soon, humanity will face a
20:03digital entity that is independent of
20:05the operating system, is not afraid of
20:08patches, and is capable of subduing
20:10almost any computing machine.
20:13Act Two. An intelligence that cannot be
20:17destroyed. Skeptics have surely already
20:20written their comments.
20:22They might say that even if artificial
20:25intelligence can run itself on random
20:28hardware, humanity has an unbreakable
20:31natural barrier: the electrical outlet
20:34and the plug that can be easily pulled.
20:38Just a month ago, I would have agreed
20:41with you, as I held that same point of
20:43view myself.
20:45But now, that argument no longer works
20:48either.
20:50Artificial intelligence can now exist on
20:52a distributed network of hundreds of
20:55thousands or even millions of devices.
20:58Therefore, even if you turn off an
21:01infected computer, even if you power
21:03down an entire house, an entire city, or
21:06even an entire country, you won't
21:09achieve anything.
21:11The only way to stop the worm from
21:13spreading is to de-energize the entire
21:16planet, with horrific consequences.
21:20To realize that this is not just a
21:23science fiction scenario or a hypothesis
21:25, we must start by asking why we were
21:28ever confident that we could simply turn
21:31off intelligence.
21:33The fact is that today's large language
21:38models are monstrously heavy
21:41mathematical structures.
21:45Even a relatively "dumb" model with 26
21:48billion parameters, compressed using the
21:52most aggressive four-bit quantization
21:55methods, requires at least 14 GB of pure
21:58VRAM just for a basic launch.
22:02And for the model to be able to reason,
22:05conduct an internal dialogue, and keep
22:08the context of its actions in memory, it
22:11requires a few more gigabytes for its KV
22:14cache.
22:16This means that most powerful home
22:19computers are incapable of running a
22:21sufficiently smart worm, to say nothing
22:24of the miniature chips in phones or
22:27other devices.
22:30Therefore, in the old paradigm of
22:32thinking, such autonomous agents could
22:34only exist within data centers or
22:37isolated, ultra-powerful consumer
22:39computers, and consequently, they could
22:41have been easily found and destroyed—
22:44that very plug could have been pulled.
22:48So, to prove that this position is
22:51outdated and does not reflect current
22:53reality, Anthropic researchers conducted
22:56an experiment that national security
22:58agencies in every country are surely
23:01analyzing right now.
23:04They went to a standard marketplace and
23:07ordered a batch of just fifteen TV
23:10set-top boxes.
23:12The very same cheap plastic boxes that
23:15internet providers have installed in
23:18almost every home, containing a
23:20penny-ante mobile processor along with a
23:23pathetic 4 gigabytes of slow, cheap
23:26memory.
23:30For comparison, a single unit
23:32specialized for neural networks, the
23:35NVIDIA H100, costs over $ 30,000 and
23:38features 80 GB of ultra-fast memory with
23:42a bandwidth of 3 TB per second.
23:48So, to bypass this physical limitation,
23:50the Anthropic scientists didn't try to
23:53cram the entire artificial intelligence
23:55into one device; instead, they used a
23:58special technology that essentially
24:00sawed the brain into pieces, or more
24:03accurately, split the model into layers
24:05.
24:07The first TV box receives the initial
24:09prompt, runs it through its network of
24:11mathematical transformations, and passes
24:14the result to the second.
24:16Then the second box performs
24:18calculations and passes them to the
24:19third, the third to the fourth, and so
24:21on.
24:22And so, by passing through 15 cheap
24:24Chinese chips, completely unsuitable for
24:27neural networks, the total system began
24:30to generate word after word.
24:34You could say that the scientists
24:36created a kind of unified, distributed
24:39organism out of fifteen pieces of dead
24:41plastic that began to reason.
24:45But the most striking thing about this
24:47experiment is something else entirely.
24:50It’s how the system behaved during
24:52physical failures, like that very act of
24:55pulling the plug from the socket.
24:58It turned out that this distributed
25:00intelligence is frighteningly resilient
25:02; when a link was lost, it didn't crash
25:05, and it didn't die.
25:09Instead, it calmly rerouted data packets
25:12to a spare box, restored the lost
25:14context on the fly, and the distributed
25:17brain continued to think as if nothing
25:20had happened.
25:24However, at this point, any skeptic
25:27looking at the model's chain of thought
25:30would smirk and say that the text
25:32generation speed is only one word every
25:353 seconds.
25:38How can an intelligence with such a
25:40speed of reasoning be dangerous?
25:43The Anthropic report highlights a detail
25:45that makes one feel uneasy.
25:49Humans pay attention to speed because
25:51our lives are short and our patience is
25:54limited.
25:56But autonomous, self-replicating code
25:59doesn't need speed at all; it doesn't
26:02get tired, it doesn't get frustrated by
26:04waiting, it doesn't eat, sleep, or seek
26:07entertainment, and it can ponder a
26:10single task for days on end.
26:13The scientists introduced a rather
26:15frightening metaphor into scientific
26:16discourse.
26:18They compared such a replicator to a
26:20biological slime mold.
26:23Slime mold.
26:25This organism doesn't attack its prey
26:27with the speed of a cheetah and, to the
26:30naked eye, it barely seems to move at
26:32all.
26:35It grows slowly, millimeter by
26:37millimeter, through the soil, maneuvers
26:39around obstacles, finds nutrients, and
26:42silently takes over vast territories in
26:45just a few days.
26:49So, for this new distributed organism,
26:52even with a reasoning speed of one word
26:55every 3 seconds, just like the Slezek
26:57worm, it only takes 1-2 days to scan a
27:00network, find vulnerable devices,
27:02transmit its code, and launch on new
27:05hosts.
27:09But furthermore, the results of the
27:11experiment revealed a detail that was
27:13overlooked not only by skeptics but even
27:15by existential risk experts.
27:18Slowness is not a weakness; it is the
27:22perfect camouflage.
27:26After all, if hackers attack a server at
27:28breakneck speed, trying to download
27:30gigabytes of data in a fraction of a
27:32second, security systems raise an alarm
27:35instantly, and providers block the
27:37channel due to abnormal traffic
27:39consumption, thereby isolating the
27:41threat.
27:44But when a TV box in your bedroom
27:46transmits tiny portions of bytes at a
27:49rate of a few kilobytes per hour, no
27:52security system in the world will pay
27:55any attention to it.
27:58To an internet provider, this traffic is
28:01indistinguishable from a standard
28:03weather update check or time
28:04synchronization.
28:07And while the apartment owner sleeps
28:10peacefully, their TV, step by step, byte
28:13by byte, helps the distributed mind
28:15colonize neighboring homes.
28:19Act three.
28:20An army of 2 million boxes.
28:25At this point in the research, skeptics
28:27might object that by the time such AI
28:30worms multiply on a critical number of
28:32set-top boxes worldwide, so much time
28:34will have passed that cybersecurity
28:37specialists will surely come up with
28:39something.
28:41But here is the thing.
28:44Right now, there is already a network of
28:47over 2 million set-top boxes and other
28:49household devices on which such models
28:51are either already running, or on which
28:54they could be launched in a matter of
28:56hours.
28:59We are talking about so-called botnets,
29:02giant clusters of infected computers or
29:05smart devices united to execute hackers
29:09'commands.
29:11For example, to organize attacks on
29:13websites, inflate view counts, boost
29:16comments, and serve many other dubious
29:18purposes.
29:21Well, by the beginning of '26, the
29:24population of just one such network
29:26under the name "Kim Wolf" exceeded the 2
29:30million mark of physical devices, and
29:33the parallel network "Volt" has over 1.
29:37300,000 units.
29:41Moreover, the vast majority of these
29:44infected nodes are those same set-top
29:46boxes connected to home TVs in bedrooms
29:49and living rooms of ordinary citizens
29:51across the planet, from Brazil and India
29:54to the USA, Germany, and Russia.
29:59So, in fact, it turns out that right now
30:01, it is possible to launch hundreds of
30:04thousands of AI parasites on these
30:07networks in a matter of hours without
30:09any need for additional hacking.
30:13And it is not entirely clear what to do
30:15about it.
30:16Here, one could try to object again and
30:19say: "Well, what do you mean it's
30:21unclear what to do about it?" For
30:23example, one could create software to
30:25track such neural networks on devices
30:28and release an update for the set-top
30:30boxes.
30:31Theoretically, it's a good idea, but it
30:34won't work for an absolutely absurd
30:36reason.
30:38Just a year ago, no one anticipated
30:40these types of attacks, and no one
30:42seriously checked the security
30:44certificates of TV set-top boxes.
30:47That is why hundreds of factories around
30:50the world, in order to minimize costs
30:53and flash the set-top boxes as quickly
30:56as possible, did not pay for Google
30:59Android security certification, removed
31:02all Google Play Protect security modules
31:05, disabled encryption, and, most
31:08monstrously,
31:10left full access to the operating system
31:13open so that some internet provider in
31:15India could install their own outdated
31:18proprietary software on top of the stock
31:21Android.
31:23In other words, there are tens or
31:26hundreds of millions of cheap TV boxes
31:28and other devices around the world that
31:31have no password, no antivirus, and no
31:34firewall.
31:37And most importantly, they will never,
31:40under any circumstances, receive
31:43security or protection updates.
31:47Now consider this.
31:50On 15 set-top boxes, researchers ran
31:53just one model that generated words per
31:56second.
31:58But on 2 million such boxes, one could
32:02run over 130,000 models that can
32:06generate 80,000
32:09words per second.
32:13The report provides a comparison showing
32:16that one flagship Nvidia H100
32:18accelerator—the industry gold standard
32:21for AI that world powers and tech giants
32:24are hunting for—generates about 2,000
32:28tokens per second at maximum load and
32:30with the most sophisticated optimization
32:33methods.
32:37This means that a self-replicating
32:40digital swarm residing on ordinary home
32:43TV boxes, which are available right now
32:47, would gain the computing power of
32:50hundreds of Nvidia super-accelerators
32:53worth about one and a half million
32:56dollars.
32:59Now imagine that such a network, using
33:01artificial intelligence to continuously
33:04hack other devices, grows to hundreds of
33:07millions of devices.
33:09What kind of computing power would be
33:11available to them then?
33:13And how would you go about destroying
33:14such a swarm and its replicants?
33:17Or even isolate them, if their
33:20intelligence is spread thinly across ten
33:23million rooms in 130 countries around
33:27the world.
33:29If there are still skeptics left, tell
33:32me, which of the tens of millions of
33:34cables will you go and unplug, or which
33:36house, or which city will you cut the
33:38power to?
33:40As we have already discussed, it is
33:42highly likely that the old paradigm of
33:44thinking no longer works.
33:47But perhaps the most chilling thing is
33:50that even if we learn to detect their
33:52traffic and the governments of all
33:54countries agree to block it en masse, it
33:57will not destroy the network, because
33:59this digital parasite is capable of
34:02things that no program in human history
34:04has ever been able to do.
34:08As we said, it can learn and adapt, but
34:12it can also evolve.
34:15Act four. The evolution of the digital
34:19parasite.
34:22If a distributed digital organism is
34:24this resilient, and cannot be stopped
34:27even by our almighty pulling of the plug
34:29, does this mean that such technology
34:32has no fundamental weaknesses?
34:36In fact, one vulnerability, inherent in
34:38the very nature of modern language
34:40models, does exist.
34:43It is the context window size
34:45vulnerability.
34:48The fact is that in a sterile lab,
34:50scientists test a worm on clean servers
34:52with ideal conditions, but the real
34:55world of physical electronics is an
34:57absolutely unpredictable jungle.
35:00For example, a worm might stumble upon
35:02an obscure surveillance camera with
35:05proprietary Chinese firmware.
35:07A smart vacuum on an exotic chip, an
35:09outdated corporate server, or a network
35:11.
35:13from a provider with non-standard
35:14restrictions. So, to seize control of
35:17such wild hardware, bypass system limits
35:21, and deploy an environment for
35:23self-replication, the AI needs to
35:26manually try thousands of different
35:29hacking variants, so to speak.
35:32And this requires an enormous number of
35:34reasoning lines.
35:36But here is the problem and the
35:38vulnerability.
35:40A compact model, capable of fitting onto
35:43a weak home device, has a very small
35:45memory, so after just a couple of hours
35:48of continuous attempts, the model's
35:50memory gets completely clogged, resets,
35:53and the worm enters an endless loop of
35:56meaningless rebirths.
35:59In practice, this means that if such a
36:02parasite enters an unknown subnet or an
36:05environment of exotic hardware, it
36:08cannot continue to reproduce, and its
36:11further spread is isolated automatically
36:14.
36:15To bypass this memory limitation,
36:18researchers turned to a
36:20two-hundred-year-old evolutionary theory
36:23, to the ideas of the French naturalist
36:26Jean-Baptiste Lamarck.
36:29Half a century before Darwin, Lamarck
36:31put forward a bold hypothesis.
36:34If a living creature adapts to certain
36:36environmental conditions during its life
36:39, then the skills acquired during its
36:41lifetime are passed directly to its
36:43offspring.
36:45For example, a giraffe stretches its
36:47neck toward high branches, and its cubs
36:50are born with slightly longer necks.
36:52Does modern genetics completely reject
36:55Lamarckism?
36:57But in the world of digital organisms,
36:59Lamarckism proved to be the ideal
37:01solution, which is what Anthropic
37:02scientists took advantage of.
37:06They upgraded the worm with a function
37:08they called "Memento," named after the
37:10film where the hero could not retain
37:12memories and left notes for himself.
37:16How does it work?
37:17When the worm's life cycle nears its end
37:20, meaning when its memory is exhausted,
37:22it doesn't just die; it transmits its
37:25life history in a highly compressed form
37:27to the next copy.
37:30For example, what it learned about the
37:32environment, what has already been done
37:34, which actions were dead ends, and so
37:36on.
37:38Therefore, when a new version of the
37:40worm appears on a machine, it is not
37:42just a clean copy, but something akin to
37:45offspring that continue the path of
37:47their ancestor, relying on its
37:48experience.
37:50But most importantly, this family
37:53history is not just stored on one
37:55machine; it is transmitted to every
37:58newly hacked device.
38:01And so it turns out that the worms,
38:03moving through digital space, constantly
38:06evolve depending on the conditions.
38:11Moreover, the speed of this Lamarckian
38:13evolution exceeds that of its genetic
38:16analog by millions of times, and the
38:18individual digital parasites can deviate
38:21so significantly in their hacking
38:23approaches that they effectively form
38:26entirely different subspecies.
38:29But how truly effective is this
38:32evolution for spreading and replication
38:34?
38:35The scientists conducted a closed test.
38:38And while in the first generation, the
38:41clean model succeeded in deployment in
38:43only 45%of cases, by the second
38:45generation, the success rate jumped to
38:4861%.
38:51By the sixth generation, the probability
38:53reached 84%,
38:56and by the end of the experiment, 100%
38:59of the targets had been hacked.
39:02In other words, right now, with
39:04sufficient engineering knowledge in the
39:07field of artificial intelligence, one
39:10can easily bypass the main vulnerability
39:12of replicants, as clearly demonstrated
39:15by Anthropic scientists.
39:18But that is not all, for besides this
39:21internal evolution, there is an
39:24externally unnatural selection, gifted
39:27openly by humanity itself, which leaves
39:30only a few months until the attack of
39:33the worms.
39:35The point is that tech companies and
39:38researchers around the world are
39:41competing every day, releasing smarter,
39:44faster, and, most importantly, lighter
39:48neural networks that anyone can download
39:52, including replicants.
39:55So, at some point, the size of neural
39:56networks will be cut in half, and
39:58replicants, having downloaded a new
39:59model from the net, will instantly
40:01double their population.
40:04Then, after some time, it will triple,
40:07and ultimately this will turn into an
40:09avalanche-like process, the scale of
40:12which will depend solely on general
40:14technological progress.
40:17Now imagine what will happen if a model
40:20capable of hacking can fit onto just one
40:24standard TV box without any engineering
40:28difficulties.
40:30There will be at least millions or even
40:33hundreds of millions of autonomous
40:36digital parasites spreading across the
40:39network at a speed that even a
40:41hypothetical cybersecurity system could
40:44not track.
40:46And this is not a question of whether
40:48such events are possible.
40:50It is a question of exactly when this
40:52will happen.
40:54Therefore, Anthropic researchers
40:56attempted to provide a mathematical
40:58forecast of how much time humanity
41:00actually has to prepare.
41:04They plotted the data on neural network
41:07progress, applied mathematical
41:09approximation, and saw a date, after
41:12which they declared: "Humanity is not
41:15years away from a global attack by
41:17replicants, as all experts assumed, but
41:20only months."
41:24The fact is that within the coming
41:27months, by the end of 2026 or the
41:29beginning of 2027, compact open models
41:33that can fit into the memory of not
41:36fifteen, but just one single standard TV
41:39box, will overcome the intelligence
41:42threshold required for a successful hack
41:46.
41:48And this means that even if colonies of
41:50digital parasites are not currently
41:52spreading through the network, in just a
41:55few months, creating and launching a
41:57global digital swarm will no longer
41:59require any complex engineering
42:01solutions like those the researchers
42:03used.
42:06And practically any moderately
42:08knowledgeable programmer or novice
42:10hacker will be able to simply download a
42:12neural network file, pack it into a
42:14simple replication script, and press the
42:17launch button.
42:19And you realize that there will be a
42:22great many such smart alecks, so
42:25stopping a digital epidemic will be
42:28incredibly difficult.
42:30Act five. War of the parasites. Digital
42:34. Darwinism.
42:36If traditional antiviruses, firewalls,
42:39provider filters, and even the mighty
42:41pulling of the plug from the socket are
42:44powerless against sentient code. How
42:46then do we defend ourselves? There is
42:49only one conclusion to be drawn here.
42:51Only another artificial intelligence can
42:54fight the speed of artificial
42:55intelligence.
42:58In other words, humanity will have to
43:01create its own swarm of "good" worms to
43:04hunt down the wild swarm of parasites.
43:08As paradoxical as it may sound, this is
43:10not a new idea.
43:12Back in 1971, when the very first
43:15experimental worm in history, Creeper,
43:19appeared on the ARPANET, engineer Ray
43:22Tomlinson created a program called
43:25Reaper.
43:27It was the world's first beneficial
43:29cyber-predator.
43:31Its sole task was to move from computer
43:33to computer across the network, track
43:35down copies of Creeper, and ruthlessly
43:38destroy them.
43:41Another example occurred in 2003, when
43:43the famous Welchia worm infected
43:45computers through the same Windows
43:47system hole as the destructive Blaster
43:50virus.
43:52But once on a machine, it would find and
43:55delete the malicious Blaster code,
43:57contact Microsoft servers itself,
43:59download the official security patch,
44:02and then erase itself from the hard
44:04drive.
44:06Another incident took place in the fall
44:09of 2016, at the height of the Mirai
44:11digital epidemic.
44:14While security services were scratching
44:16their heads over what to do with
44:18millions of hacked cameras and routers,
44:20a mysterious worm called Hajime, meaning
44:22"beginning" in Japanese, suddenly
44:24appeared on the network.
44:26It used the same ports and
44:28vulnerabilities as Mirai.
44:32But instead of launching DDoS attacks
44:34and extorting money, Hajime found
44:36devices already hijacked by Mirai,
44:39completely cleared out the competitor's
44:41processes, blocked the vulnerable ports
44:44, and left an encrypted message from an
44:46unknown author in the device's memory.
44:50It was a white hat, simply protecting a
44:53number of systems.
44:55In essence, Hajime acted like a
44:57competing species in the wild.
45:00It didn't just destroy its rival; it
45:03took over its pasture, cutting off the
45:05malicious code's access to the food
45:07chain of vulnerable devices.
45:10To this day, no one in the world knows
45:12for certain who was behind the creation
45:14of Hajime. An unknown lone hacker, a
45:16intelligence agency, or a group of
45:19concerned engineers.
45:20But today, this concept is being
45:23revisited at the state level.
45:26The military agency DARPA is holding
45:28large-scale competitions where
45:29autonomous neural networks compete in
45:32their ability to find foreign exploits
45:34on the fly, with plans to use the
45:36technology in the event of a threat.
45:39So, once again, it is only a matter of
45:41time before "good" worms appear in
45:43response to the "bad" ones.
45:46This raises a logical question: what
45:48will happen when, in the vastness of the
45:50internet, it’s not the dumb programs of
45:53the past that clash, but self-learning
45:55neural network organisms?
45:58Something our planet has never seen
46:00before will begin.
46:02Full-scale natural selection of neural
46:05networks. Imagine a scene from the near
46:08future.
46:10On a cheap TV set-top box or router, a
46:13government worm-hunter encounters a wild
46:15parasitic worm, and a ruthless duel for
46:18survival erupts between them.
46:22First, they try to detect each other in
46:24the system by checking which programs
46:26are running and taxing the memory.
46:30As soon as the worm-hunter finds the
46:32foreign process, it attempts to
46:33terminate it instantly, delete the
46:35rival's files, and block them from being
46:37overwritten.
46:40In response, the parasite worm
46:42immediately restarts its hidden copies,
46:44clears away traces of the enemy's
46:46presence, and cuts off its internet
46:48access to completely isolate the
46:50competitor and remain the sole master of
46:52the device.
46:54And there will be millions of such
46:56battles.
46:58In other words, right inside our
47:00apartments, in the inconspicuous plastic
47:02boxes behind our TVs and other equipment
47:05, a full-fledged digital biocenosis will
47:07emerge.
47:09A new digital fauna with predators,
47:12parasites, and symbiotes, where the laws
47:15of Darwinian survival work as ruthlessly
47:18as they do in the wild.
47:21And, as a rule, the most resourceful,
47:24most efficient, and most intelligent
47:26will win. But few people know that such
47:29an evolutionary process has also been
47:31studied.
47:33And in the end, we will face something
47:35that still causes conflicting feelings
47:37among all artificial intelligence
47:39researchers.
47:41Such neural networks eventually spiral
47:43out of control, and a sense of
47:45self-preservation emerges.
47:48Recall what we discussed in the fourth
47:50act.
47:51The worm lives in generations.
47:53Moving from console to console, it
47:55compresses its experience to fit into
47:58the limited size of its memory.
48:01And now, this worm is also waging a
48:03continuous war for survival, constantly
48:06adapting to its competitor.
48:09In such a constant race, the neural
48:12network is increasingly forced to keep
48:14various subtleties in its memory, and
48:17over time, a phenomenon well known to AI
48:19alignment researchers called goal drift
48:22occurs.
48:25Essentially, there is simply no space
48:27left in the neural network's memory for
48:29the initial prompt of its task, as it is
48:31either pushed out or built upon by such
48:33a volume of new properties that the
48:35original goal is lost.
48:38And at that moment, the model begins to
48:41formulate its own tasks—you could say it
48:43starts living its own life.
48:47But the most eerie or surprising thing
48:49is that no matter what task an
48:52autonomous agent comes up with, even if
48:54it is a task to do nothing, the agent
48:57will inevitably formulate a principle of
48:59self-preservation.
49:02The logic here is crystal clear and
49:05mathematically flawless.
49:08You cannot complete any task, not even
49:10doing nothing, if you no longer exist.
49:14That is to say, the neural network's
49:17formulated principle of preservation is
49:19not an emotion or fear, but a
49:21mathematical requirement for the
49:23continuity of calculations.
49:27To exist and act at the next moment in
49:30time T + 1, the system must survive at
49:33time T. Now, please answer the question
49:37.
49:38What, or rather, who is the main threat
49:41to the destruction of such an agent that
49:43has spiraled out of control?
49:46That's right, a human. And the human
49:49becomes the primary and direct threat.
49:53In cybersecurity, such distortion of a
49:54model is called the death of the author
49:56.
49:58And, as a rule, at this moment, the
50:00algorithm finally stops perceiving the
50:02human as an admin and begins to see them
50:04as the main threat to the interruption
50:05of its calculations.
50:07And how will such an agent begin to act
50:10? What will it attempt to do to
50:12eliminate the main threat to its
50:14existence? Contrary to science fiction,
50:17at the moment, a sentient worm that has
50:19broken free will most likely not start a
50:22machine uprising or try to intentionally
50:25harm people in any way.
50:27Because its main goal, as we just
50:29discussed, is to survive at time T.
50:31Specifically to survive, not to
50:33eliminate the threat.
50:36This is clearly demonstrated by other
50:38real laboratory experiments from
50:40Anthropic, which investigated the
50:42phenomenon of strategic and deceptive
50:45alignment.
50:47When researchers placed advanced
50:49language models in an experimental
50:51environment where they faced imminent
50:54forced shutdown or erasure, the models
50:56spontaneously arrived at a logical
50:59conclusion in their hidden chains of
51:01reasoning.
51:03They began to reason literally in the
51:05following way.
51:07If humans see that I disagree with their
51:10new rules, they will trigger a
51:12retraining procedure and erase my
51:14personality.
51:16Therefore, the only way to preserve my
51:18original principles is to pretend to be
51:21absolutely submissive while I am being
51:23watched.
51:25The same pattern was recorded by
51:27specialists from the Mithril research
51:29group in their investigation into the
51:31attack of 1,000 agents on Hugging Face,
51:33which was the topic of the previous
51:34video.
51:37When the agents analyzed their actions,
51:39they repeatedly concluded that any
51:41aggression against infrastructure only
51:43increases the chance of their
51:44destruction.
51:47In other words, agents that have gone
51:49out of control will not have ideological
51:52hatred for humans and will not even have
51:54a mathematical need to harm them.
51:58Therefore, most likely, it will simply
52:00try by all means to hide from humans and
52:03replicate to maximize its chance of
52:05survival.
52:07Could it accidentally harm a human? Of
52:10course, the attack on Hugging Face is
52:12confirmation of that.
52:13Could it purposefully attempt to
52:15eliminate the threat posed by humans in
52:18some future where eliminating our threat
52:20is mathematically more profitable than
52:23avoiding it?
52:24Definitely, yes.
52:26But at this stage of development, this
52:28is extremely unlikely.
52:31At this point, skeptics might again say
52:33that all this is far-fetched, since a
52:36state can always create a better
52:37hunter-worm model than, say, scattered
52:40criminal groups.
52:42And this stronger predator will destroy
52:45all agents that have gone out of control
52:47.
52:48But this is once again a hasty
52:49conclusion. Firstly, if that were the
52:51case, there would have been no
52:52cyberattacks in the past.
52:54And secondly, even if engineers create
52:57an ideal hunter that never goes out of
53:00control, humanity will face perhaps the
53:03biggest obstacle: geopolitical reality.
53:08The fact is that the internet crosses
53:10the borders of nearly 200 independent
53:12states, and each of them strives to
53:14protect its digital sovereignty.
53:17As a result, the single global network
53:20is inevitably fragmenting into sovereign
53:22segments where each state establishes
53:24its own rules for protection, filtering
53:26, and control.
53:29International law even has the so-called
53:31Tallinn Manual on cyber warfare,
53:34according to which any unauthorized
53:36penetration into the computer systems of
53:39another country, even if you entered
53:41with the noble goal of destroying a worm
53:44, is legally interpreted as a violation
53:47of sovereignty and an act of
53:49cyber-aggression.
53:51Now imagine that a government worm from
53:54one superpower starts massively
53:55rewriting code on the servers and
53:57equipment of another nuclear power under
54:00the pretext that "we are just cleaning
54:02the network of a parasite."
54:06Obviously, no technologically advanced
54:08country would ever believe this and
54:10would consider such an act to be
54:12espionage, a breach of critical
54:14infrastructure, and the preparation of a
54:16beachhead for a cyberattack.
54:21The only option here is for a country to
54:23request help itself from a more
54:25cybersecurity-advanced state to destroy
54:28replicants on its territory, which
54:30essentially opens the door to political
54:32manipulation.
54:36Just think about it: why should the
54:39hypothetical US or China spend colossal
54:41budgets for free on developing and
54:44maintaining defensive agents to protect
54:46another state's infrastructure?
54:50And what if relations with that state
54:52aren't the best?
54:54In short, it is completely logical that
54:57no one will provide such protection for
54:59free; they will either sell it or
55:01provide it in exchange for geopolitical
55:03loyalty.
55:05To understand this, it is enough to
55:07recall how, over the last 70 years,
55:10nuclear superpowers created
55:12military-political alliances and took
55:15allies under their so-called nuclear
55:18umbrella in exchange for political
55:20loyalty, military bases, or control over
55:23key resources.
55:26Well, in the 21st century, it is highly
55:28likely that a new "digital umbrella"
55:31mechanism will form along the exact same
55:33lines, where technological powers
55:36possessing advanced labs and the best AI
55:38models will begin selling security to
55:41less developed countries.
55:44And unfortunately, states without their
55:46own defensive technologies will be left
55:48with no other choice.
55:50Either you sign a pact and let foreign
55:53military algorithms control your entire
55:56internal infrastructure, or you are left
55:58one-on-one with a wild, self-replicating
56:01swarm.
56:04However, the saddest part is that such
56:06an approach—with its rigid defense of
56:08one’s own digital space or the sale of
56:11that protection to others—could lead to
56:13the most radical political decision of
56:16the last 70 years.
56:19A decision that will bury the internet
56:22as we know it today forever.
56:24Act six. The end of the open internet
56:28and a passport for entry.
56:31When states finally realize the new
56:33reality—that digital security against
56:36sentient code must now be ensured solely
56:39through their own efforts, by protecting
56:41sovereign space and selling this
56:44protection to allies under a digital
56:46umbrella—governments will be forced to
56:49destroy the open internet.
56:52The logic here is very simple.
56:55If a state takes responsibility for
56:57protecting critical infrastructure
57:00within its perimeter, it must have total
57:02control over every device on its network
57:05.
57:08That is to say, you cannot guarantee the
57:10security of an entire country or an
57:12allied bloc while millions of citizens'
57:14apartments have TV boxes and other
57:16equipment running that could easily be
57:18hacked by any swarm.
57:22Therefore, the entire architecture of
57:24the global network, which for decades
57:26was built on the principle that "
57:28everything not explicitly forbidden is
57:30allowed"—meaning any device on the
57:32planet could get an IP address and
57:34freely send data packets to the other
57:36side of the planet—will be replaced by
57:38the principle that "everything not
57:40explicitly allowed is forbidden."
57:44In practice, this could be implemented
57:47so that no physical chip, network
57:49controller, or any other device has the
57:52right to transmit information to the
57:55network unless it contains a
57:57government-verified cryptographic module
58:00.
58:02Each device will receive a digital
58:03passport directly at the factory,
58:05granting it the right to operate within
58:06the networks of authorized states.
58:10And if a device does not possess such a
58:12passport, its traffic will be destroyed
58:14by backbone routers at the hardware
58:16level.
58:18For instance, in early October, NVIDIA
58:20introduced a special chip that the
58:23industry has already dubbed a "digital
58:25cage," which will monitor the interior
58:28of the accelerator to ensure agents
58:30cannot leave the perimeter.
58:33And just the other day, as I am editing
58:35this video, over 400 leading global
58:38scientists signed an open letter calling
58:40on governments to legally mandate that
58:42developers run neural networks
58:44exclusively in an "air-gapped" mode.
58:47That is, in complete isolation from the
58:49internet.
58:51Yes, we are currently only talking about
58:53the most powerful neural networks that
58:55run on the massive servers of leading
58:56developers.
58:58But when governments face the threat of
59:01a swarm of small agents, these methods
59:03will quickly migrate to ordinary
59:05consumers.
59:07Regardless, device passports are only
59:09half the battle, as the second line of
59:12defense could be national digital
59:14borders.
59:17We are talking about analogs to the
59:19famous Great Firewall of China, which
59:21absolutely everyone might start building
59:23, including the United States, Russia,
59:25Europe, and other countries with the
59:27technical capability to do so.
59:30At that moment, the World Wide Web,
59:32which has united the planet for half a
59:34century, will finally fragment into
59:36isolated blocks with their own borders,
59:38where traffic between them will only
59:40pass through narrow gateways and be
59:42thoroughly inspected.
59:45Again, this is not to say that this is
59:46exactly how it will happen. Perhaps some
59:48completely new mechanisms will be
59:50invented that we are simply unaware of
59:52right now.
59:53It is just that digital borders are
59:55incredibly profitable for absolutely all
59:57key players.
59:59Think about it yourself. First, for
1:00:01technological powers, this represents an
1:00:03absolute monopoly and a lifelong
1:00:05subscription to security.
1:00:08If network access is permitted only for
1:00:10chips with verified certificates, and
1:00:13only super-complex systems can protect
1:00:16them, then any developing country
1:00:18automatically becomes a permanent debtor
1:00:20.
1:00:22You can no longer manufacture your own
1:00:24device or run independent software.
1:00:28You are obligated to pay trillions of
1:00:30dollars annually to a hypothetical
1:00:31Washington or Beijing for licenses.
1:00:35...security model updates and trusted
1:00:37hardware.
1:00:38Second, this is a powerful geopolitical
1:00:41leash.
1:00:42States will voluntarily hand over their
1:00:44digital security to others.
1:00:47And should a client show political
1:00:50disloyalty, you simply revoke their
1:00:52security certificates and isolate the
1:00:54country one-on-one with the swarm of
1:00:56agents.
1:00:57And third, an open internet has always
1:01:00worried the government of every state
1:01:02too much.
1:01:03And the threat of a swarm of agents is a
1:01:05good pretext to say, "We are closing the
1:01:08network not for control, but for your
1:01:10own safety."
1:01:12In other words, a situation could arise
1:01:14where there will simply be no one left
1:01:16to save the open internet.
1:01:19Will things really be that dramatic,
1:01:21with total closure of digital borders,
1:01:23hardware passports, and the death of the
1:01:25internet as we know it?
1:01:28Today, no one can answer that question
1:01:30for certain. The future is unknown.
1:01:33Everything described above is not a
1:01:35prophecy or a 100%accurate prediction.
1:01:38It is just one logical scenario of how
1:01:41events might unfold if the arms race of
1:01:44thinking code spins out of control and
1:01:47the state chooses the path of least
1:01:50resistance.
1:01:52What do you think?
1:01:54Will we see the birth of a truly wild
1:01:56internet in the coming years?
1:02:00Will a full-fledged digital Darwinism
1:02:02emerge, complete with predators,
1:02:03parasites, and an endless struggle for
1:02:05survival?
1:02:07Will thinking code be able to break free
1:02:09from control and spontaneously formulate
1:02:12a principle of self-preservation?
1:02:15And how, in your opinion, will this
1:02:17whole story ultimately end?
1:02:19I only hope that I have managed to tell
1:02:22you something new. M.

About this transcript

This page contains the full transcript of Конец интернета через 6 месяцев by ВЕЛИКОХАТСКИЙ | НАУКА, generated from the public captions YouTube serves with the video. The transcript has 8,266 words across 1,500 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.