Home
Entrar
Cadastrar
Conteúdo de Estudo
Loading...
The Hidden Reason Anthropic Wants to Slow Down AI - Video học tiếng Anh
Prática de escuta
Prática de escuta
/
Video
/
The Infographics Show
/
The Hidden Reason Anthropic Wants to Slow Down AI
The Hidden Reason Anthropic Wants to Slow Down AI
Selecione o modo de aprendizagem:
Ver legendas
Escolher palavra
Reescrever palavra
Highlight:
3000 Oxford Words
4000 IELTS Words
5000 Oxford Words
3000 Common Words
1000 TOEIC Words
5000 TOEFL Words
Legendas (200)
0:00
People think Anthropic is calling for “responsible scaling” of frontier
0:03
AI to prevent human extinction. At least, that’s what the company says.
0:07
But there’s a more cynical, more strategic viewpoint.
0:10
A global slowdown wouldn’t just slow AI’s march. It would also slow the market,
0:15
locking in today’s leaders. It would make it harder for open-source competitors to
0:19
reach Anthropic’s capabilities. The timing isn’t a coincidence.
0:23
In May 2026, Anthropic just closed a record private fundraising round with a trillion dollar
0:28
valuation then filed for a public listing. Just 3 days later, it released a report warning that AI
0:34
may be out of control and urged governments to consider slowing the entire industry.
0:39
To some, it’s caution. To others, it looks like
0:41
strategy disguised as safety. Chapter 1: The Kill Switch
0:45
All it took was a week for the entire AI industry to be rocked.
0:49
On May 28th 2026, Anthropic closed a new round of investment, its eighth since it emerged
0:54
in 2021. The investors, Altimeter, Dragoneer, Greenoaks, and Sequoia,
1:00
valued the company at roughly $1 trillion. It was a staggering amount. On paper, a company less
1:06
than 5 years old now sits above JPMorgan Chase in valuation, a bank built over 2 centuries.
1:12
Days later, the company told regulators it plans to go public, perhaps before the end of 2026.
1:18
The listing would be one of the biggest in market history. Not a single share
1:22
has even been priced yet. Then came the bombshell.
1:25
On June 4th, a report called "When AI builds itself" was released. It was co-authored by Marina
1:30
Favaro, who leads the Anthropic Institute, and Jack Clark, the company's co-founder.
1:36
Their warning was stark. AI systems are getting close
1:39
to improving themselves with barely any human help. It’s reaching the point of
1:42
no return and the world needs to hit the brakes before it’s too late. Anthropic
1:46
were the guiding light in the industry, willing to sacrifice themselves for the greater good.
1:51
At least that’s what it seemed like. Anthropic would only back an industry-wide
1:55
slowdown under one condition: all major labs across countries agree to pause at the same
2:00
time under verifiable rules. Until then, the race goes on.
2:04
It’s a glaring contradiction. The most valuable AI startup in
2:08
the world is about to go public, and at the same time wants governments to coordinate a pause that
2:13
would freeze every rival in place. Anthropic would become the locked-in leader of the AI industry.
2:18
That isn’t a prediction. The evidence is in its own code.
2:22
Chapter 2: Eighty Percent In May 2026, more than 80%
2:26
of the code going into Anthropic's own systems was written by Claude, the company's own AI.
2:32
Before Claude Code launched in February 2025, that figure sat in low single digits.In just
2:37
under 15 months, Claude went from writing almost none of the company’s code to writing most of it.
2:42
At the same time, Anthropics engineers increased their production. They now ship
2:46
8 times as much code per day as in 2024. A week's worth of work can be done in a day.
2:52
Jack Clark has stated publicly that the company believes it can reach 100% AI written code
2:57
by 2028. AI could soon be responsible for nearly all of its own code, even Anthropic
3:03
executives are issuing warnings about the tech. This changes what it means to be an engineer at
3:08
the company. When the machine writes the code, the human is no longer the author.
3:12
They become the reviewer. It’s not about how fast someone can type but how quickly they can read.
3:17
If a slowdown is implemented, competitors are left permanently trying to catch up.
3:21
That’s what Anthropic wants. But first we want to thank the
3:24
sponsor of today’s video — Incogni. Have you ever searched your own name
3:27
online and found way more than you expected? Your home address, phone number, email,
3:32
even information about your relatives can end up on people-search sites and data
3:36
broker databases. And once that information is out there, it can be used for phishing scams,
3:40
identity theft, harassment, or worse. That’s why I decided to try Incogni.
3:45
Getting started was simple. I created my account, selected whose personal data I wanted removed,
3:49
and gave Incogni permission to work on my behalf. From there, Incogni contacts data brokers for you,
3:54
sends removal requests, deals with objections, and keeps you updated along the way.
3:59
What I really like is the dashboard, because it makes the whole process feel real. You can
4:04
actually see the requests going out and track the progress over time. And after more than a week,
4:08
it’s much easier to see just how much work Incogni is doing for you behind the scenes.
4:12
This feels especially important right now because major data breaches keep exposing
4:17
personal information, and that data can be used to make scam emails, calls, and texts seem much
4:22
more convincing. I recently got one of those text messages saying a package couldn’t be delivered
4:26
and that I needed to click a link to fix it. And because I was actually expecting a package,
4:31
it looked way more legitimate than it should have. Even one profile left up can create real risks.
4:36
Incogni also has a new Custom Removals feature on the Unlimited and Family Unlimited plans. So if
4:42
you find your information on a specific people-search site or another website,
4:46
you can flag it, and one of Incogni’s privacy agents will handle the rest for you.
4:50
Go to incogni.com/tis and use code TIS for an exclusive 60% off. Again, that’s incogni.com/tis,
4:59
code TIS, for 60% off. Click the link below and get your personal data off the
5:04
market with Incogni. Chapter 3: The Loop
5:07
There have been a number of concerns surrounding the implementation of AI since it was first
5:11
introduced. And one of the major worries has been the speed in which it’s developing.
5:16
The independent research group METR measures if and when AI systems will become a threat
5:21
to society. In one study, they asked a simple question: How long can an AI finish
5:26
a task on its own before it slips up? They tested a number of frontier models
5:30
going back to 2019, and discovered a pattern. Models could handle increasingly longer tasks,
5:35
with capacity doubling roughly every 7 months. The trend held for 6 years straight.
5:41
Then, it accelerated. METR's newer estimate puts
5:44
the doubling time, since 2023, at roughly 4 months. AI Digest also looked at the
5:50
data and suggested that if the pace holds, AI agents could handle month long tasks by 2027.
5:56
And there’s no ceiling. A couple of years ago, these systems
5:59
could barely manage tasks of a few minutes. Now they can hold a thread of work that takes human
6:04
engineers a day. If this rate holds up, and the timescale shifts from day to weeks and months.
6:09
In April 2026, Anthropic said Claude made more than 800 fixes, reducing one class of
6:15
system errors by a factor of 1,000. A human team might take years to clear something like that.
6:20
The model did it in a single month. Give the system enough computing power,
6:24
and machines could design and build their own replacements. Humans wouldn’t be needed. While
6:29
Anthropic is sounding the alarm, it’s also careful to say that this hasn’t
6:32
happened. It’s "not inevitable." But it "could come sooner than
6:37
most institutions are prepared for." If the model is building its successor,
6:41
the industry needs a rulebook, something that decides when to keep going and when to stop.
6:46
It just so happens Anthropic already has one. Chapter 4: The Trigger
6:51
The Responsible Scaling Policy reads like a warning. Something that all of
6:54
humanity should be aware of. The core idea is simple: an AI model can become powerful
6:59
enough to be genuinely dangerous, and at that point, extra protections have to kick in.
7:04
Anthropic calls those trigger points Capability Thresholds.
7:07
The October 2024 version of the policy begins with the obvious risks: chemical, biological,
7:13
and nuclear weapons. Then it adds one more, that an AI that can conduct its own AI research,
7:18
with little human help. The policy groups this under Autonomous AI Research and Development.
7:23
When a model nears that line, the guidelines don’t suggest pulling the plug. Instead it calls
7:28
for more intensive testing and speeds up work on stronger safeguards. Put simply, the company has
7:33
built a structured way to handle the possibility of an AI that starts improving AI itself.
7:39
This is more than an internal memo. A similar idea almost made it into law. California tried to do at
7:45
a state level what this policy does inside one company. It focused only on the biggest models,
7:50
the ones costing over $100 million to train, and asked them to meet a higher bar. That meant
7:56
more safety checks and external testing. Breaking the rules would mean heavy fines,
8:00
starting at 10% of a training run, rising to 30% for repeat violations.
8:05
But there was pushback. And critics saw an opening.
8:08
The venture firm Andreessen Horowitz warned the rules wouldn’t stop at the industry heavyweights.
8:13
Over time, it said, they would hurt “startups and burden small developers." The reason is cost. The
8:18
paperwork, the audits, the legal exposure all land hardest on the companies with the least
8:23
money. A fine worth 10 to 30% of a training run is pocket change for a trillion dollar company. For a
8:29
startup, that equates to the entire company. The rules would mean a dedicated team of specialists
8:34
to oversee the models. Only the firms with the deepest pockets can carry that cost.
8:39
California’s proposal never became law. Governor Gavin Newsom vetoed it in September 2024. But it
8:45
never really went away. It keeps coming back in various guises, in other states and countries.
8:51
Meanwhile, Anthropic’s Responsible Scaling Policy remains in place.
8:54
When one company starts defining the rules, those rules tend to become the baseline for
8:59
everyone else to clear. Chapter 5: The Ladder
9:02
In May 2026, Anthropic’s revenue hit $47 billion. At the end of 2025, it was around
9:09
$9 billion. The Wall Street Journal expects another 130% increase, which would push the
9:14
company into profitability for the first time. Those numbers are reflected in the valuation.
9:19
In September 2025, the company was worth $183 billion. By May 2026, it was almost $965 billion.
9:27
It more than doubled in a single quarter. That puts the company in the same range as Tesla, which
9:33
took 16 years to reach a $1.5 trillion valuation. Anthropic earned its first dollar less than three
9:39
years ago. Since then, its revenue has been growing more than tenfold each year.
9:43
That’s why the global AI slow down would only benefit one company.
9:47
The timing of the public listing and the public warning isn’t a coincidence. Anthropic is pulling
9:52
up the ladder at the right time. Chapter 6: The Moat
9:55
Anthropic’s report describes how fast things are moving and also suggests that leading labs
10:00
should coordinate how development proceeds. In practice, that means the companies at the front
10:05
end up helping define the constraints for everyone behind them. The most advanced systems become the
10:10
new standard everyone else has to meet. Except that standard isn’t written from
10:15
a neutral position. It’s where the leading
10:17
AI companies are at that point. This isn’t a fringe argument. It’s
10:21
coming from senior figures in the industry. Yann LeCun, Meta’s chief AI scientist and a
10:26
Turing Award winner, has been one of the most vocal critics. He is worried that companies are
10:30
overplaying AI risks in ways that push regulators toward restricting certain models. After Anthropic
10:36
published a study claiming its system had been used in a cyberattack with minimal human
10:40
involvement, a US senator cited it as evidence for tighter AI laws. LeCun hit back hard,
10:46
saying critics were being “played” and warned that fear-based claims could end up shaping
10:50
regulation in ways that lock out smaller players. David Sacks, an AI advisor in Washington, has made
10:56
a similar point, accusing Anthropic of using fear of risk as a way to shape regulation in its favor.
11:01
If regulations do align with the bigger companies' outlook,
11:04
everyone else is trying to catch up to a standard they didn’t help set. A frontier lab can support
11:09
restrictions that it is already able to pass. Smaller labs and independent researchers can’t.
11:14
But it's not that simple. Anthropic itself acknowledges
11:17
that a real global pause would require the US and China - and other major AI powers - to
11:23
enforce it together. It’s an ask that might never be met. But in the meantime, nothing has to stop.
11:28
So you end up with calls for restraint on one side, and continued acceleration on the other.
11:33
Both exist at the same time, without resolving each other.
11:36
Chapter 7: The Resistance Not everyone is buying the case
11:39
for a pause, and the loudest objections come from one corner of the industry.
11:43
Open Source. For years, Meta has released its most capable
11:46
models for free, so anyone could download and use them. Its Llama models have been downloaded
11:51
hundreds of millions of times, and some companies now build their products directly on top of them.
11:56
That’s the version of AI the open-source camp is pushing for:
11:59
models that aren’t locked behind a paywall, but available for anyone to use and modify.
12:04
A pause looks less like a protective measure and more like economic warfare.
12:08
Yann LeCun wants the opposite. He says regulating AI research itself,
12:12
rather than specific harmful uses, is "extremely counterproductive." He says it’s built on "false
12:17
ideas about the potential dangers of AI." His deeper concern is what happens to open-source
12:22
AI. Set a danger threshold or limit who can access the data makes it harder to release an open model.
12:28
So the pushback is about more than the right to give software away.
12:31
It is about stopping a wall from going up between two groups. The handful of firms that
12:36
can run at the new speed. And everyone who can’t. Once something like that is written into global
12:41
rules, it doesn’t usually come back down. Chapter 8: The Shockwave
12:45
Up to now this has been about Anthropic’s code. But it’s bigger than that.
12:49
It eventually affects everyone… And their paycheck.
12:52
Anthropic has already admitted engineers don’t write most of the code. They review
12:56
it. The moment that becomes the industry standard, the human workforce collapses.
13:00
The first jobs to feel it sit at the bottom of the ladder. Junior coding work has always
13:04
been built on volume and the cost of one more block of code has fallen to almost nothing.
13:09
That closes the door into the industry. An entire generation of coders and engineers are locked out.
13:14
The Stanford Digital Economy Lab tracked real employer data through 2025 and the evidence
13:19
is alarming. For coders aged 22 to 25, jobs fell nearly 20% from the late 2022 peak. Workers over
13:27
30 in the same roles saw their numbers grow. The entry positions are being culled.
13:32
It’s happening across the industry. Salesforce froze junior hiring for
13:36
a year. Google and Meta took on roughly half as many new graduates as in 2021. Total tech
13:42
layoffs soared past 140,000 in the opening months of 2026 alone. That pace runs about 40%
13:48
ahead of the same stretch in 2025. More and more companies are naming AI as the reason for this.
13:55
A pause or slow down can't undo what has already happened. E
13:58
Even if every major AI company stopped training new frontier models tomorrow,
14:02
the tools already exist. Claude Code would still be writing code. AI agents would still
14:07
be running inside thousands of companies. That matters because the economic effects
14:11
are already underway. The question is no longer whether AI will enter the workplace,
14:15
because it’s already here. The question is how far the current generation of
14:19
systems can go before the next one arrives. Today's models can’t go back in the box.
14:24
They’re here to stay. Chapter 9: The New DNA
14:28
There’s no denying that AI performance has exploded.
14:31
AI systems are taking on longer and more complex tasks. At the same time,
14:35
Anthropic is handing over its coding tasks to Claude, reducing the human
14:39
element to a reviewer. Even if the largest labs agreed to stop training bigger models,
14:43
the systems they already have would keep working. They would keep writing code
14:47
and help build the next generation of tools. That's what makes the whole debate so strange.
14:52
The company calling for caution is also showing how fast the technology
14:55
is improving. The same research that sounds the alarm is measuring the acceleration.
15:00
Maybe a global agreement arrives. Maybe it doesn't. Either way, AI is
15:04
already building the next generation of AI. The future people are arguing about is not
15:09
somewhere ahead of us. It's already here.
15:11
The AI problem is getting bigger with each passing day. And it’s not just the public that are getting
15:16
worried. The corridors of Silicon Valley are starting to feel the pressure. Find out what’s
15:20
going on in ‘Why AI Researchers Are Quitting and Panicking on the Way Out’. Or watch this.
