One of the biggest AI labs in the world
just published a warning that should
make the entire industry stop for a
second. Anthropic is now saying AI may
be entering the early stage of
self-improvement, where systems like
Claude are no longer just tools humans
use, but part of the machine that builds
better AI. And the numbers behind this
are wild. Claude is now writing most of
Anthropic's code, helping review that
code, running experiments, and speeding
up research work that used to take
humans days, weeks, or even years.
So, when a Chinese tech headline claimed
Anthropic was calling for AI research to
stop, it sounded dramatic. The crazy
part is the real warning is even more
serious. So, here's the real story.
Anthropic just released a detailed blog
post titled When AI builds itself, and
the core message is this: AI may already
be entering the early stages of
recursive self-improvement. That's the
technical term for an AI system [music]
that can design, build, test, and
improve the next generation of AI
systems. Anthropic is saying we're not
there yet, but the trend is moving in
that direction faster than most
governments, companies, or institutions
are prepared for. And Claude, their own
AI model, is already accelerating the
development of AI at Anthropic itself.
Now, the headline about stopping AI
research is misleading, but it's based
on something real. Anthropic is saying
that if there were a credible,
verifiable way to ensure that all major
AI labs around the world were actually
slowing down or pausing frontier
development at the same time, they would
be willing [music] to participate. The
problem is that a unilateral pause by
just one company doesn't solve anything.
It just shifts who the front runner is.
The real challenge is building a system
where multiple well-resourced labs in
multiple countries can verify that
nobody is secretly continuing while
everyone else stops. [music]
Without that, the AI race just keeps
accelerating. Anthropic is basically
acknowledging what everyone already
suspects. The competitive pressure is so
intense that no single lab can afford to
slow down unless everyone else does,
too. So, what evidence does Anthropic
actually have that AI is starting to
build AI? The numbers are striking.
[music]
As of May 2026, more than 80% of the
code merged into Anthropic's code base
was written by Claude. Before Claude
code launched in research preview back
in February 2025, that number was in the
low single digits. Think about that for
a second.
>> [music]
>> The majority of the code running inside
one of the world's leading AI companies
is now being written by an AI system.
This isn't just autocomplete or
generating small snippets.
Claude is writing entire files,
debugging complex systems, and handling
[music] work that used to require days
of human effort. Anthropic's engineers
are also merging eight times as much
code per day as they were in 2024.
Lines of code isn't a perfect
productivity measure because more
[music] code doesn't automatically mean
better work. But, Anthropic isn't
rewarding people for writing more lines.
The increase is happening because Claude
is doing most of the actual coding,
while engineers focus on direction and
review.
One Anthropic employee said they haven't
written code themselves in about 5
months.
Their job now is basically managing
Claude. Another employee described it as
leaning hard into what they call
Claudifying their workflow. The role of
the human engineer is narrowing at every
step. The quality of that code is also
improving fast. Anthropic tracks how
often engineers need [music] to correct,
redirect, or take over from Claude
mid-task. That number has been falling
steadily for a year.
On the most open-ended and difficult
coding tasks, where there's no clear
specification and the engineer isn't
even sure what the [music] solution
should look like, Claude's success rate
hit 76% in May 2026. Six months earlier,
it was only 26%. That's a 50 percentage
point jump in half a year. These are
tasks where the problem [music] is
vague, the answer is unknown, and the
engineer basically points Claude at a
live incident and says, [music] "Figure
it out." Anthropic gave an example of
this. A routine upgrade started crashing
tens of thousands of training jobs. An
engineer pointed Claude at the live
incident with little more than some text
content and cluster access.
Claude worked through the running jobs,
tested one environment setting at a
time, isolated the single obscure
debugging flag triggering the crash,
reproduced it reliably, and confirmed a
fix. That work would normally take a
human two to three days. Claude finished
it in about two hours.
And funny enough, this is exactly where
a lot of AI video tools still fall
apart. They can create movement, faces,
and effects, [music] but the scene often
feels like nobody actually directed it.
Open Art is sponsoring today's [music]
video, and their new feature, Smart
Shot, is built around that exact idea.
Instead of making you fight with prompts
until something looks usable, Smart Shot
turns one sentence into a full cinematic
production plan before it generates the
video. That's the part that makes it
different. You describe the scene, and
Smart Shot builds what they call a shot
plan. It can include character
references, environment design,
storyboard panels, camera angles, shot
flow, lighting notes, [music] mood, and
even lens style direction like dolly
moves, orbit shots, push-ins, and crane
shots.
So, before the video is rendered, you
can actually see the creative direction,
adjust [music] parts of it, and then
generate the final sequence.
Under the hood, it uses GPT image 2 as
the planning layer and SeeDance 2 as the
execution layer. GPT image 2 helps
structure the scene, the shots, and the
visual direction, [music]
while Seedens 2 renders the final
cinematic video with consistent
characters and motion. So, it feels less
like prompting a random clip and more
like directing a small AI production
team. Use the link in the description
and the code smart shot [music]
to get 15% off the monthly plan. All
right, now back to the video. Anthropic
also started using Claude to review code
before it gets merged. They ran a
retrospective analysis and found that if
this automated Claude review had been in
place for every past change, it would
have caught roughly 1/3 of the bugs that
caused production incidents on claude.ai
before they ever went live. The
engineers who wrote that code are among
the best in the world at building these
systems. Claude is now catching mistakes
they missed. That's a serious claim
because it means Claude isn't just
writing code faster than humans,
>> [music]
>> it's starting to write code better than
humans, at least in certain contexts.
Many employees at Anthropic already
think the quality of Claude-written code
was somewhat worse than human-written
code in late 2025, roughly at parity
today, and will probably be strictly
better within the year. The transition
is happening in real time. But, there's
a strange side effect to all this
automation. One Anthropic employee
mentioned that work used to run on what
they called a gift economy of small
favors between humans. Someone would
ask, "Can you help me get this script
running?" Each favor created a little
debt, a little mutual awareness. Claude
is faster and creates zero debt, but
each of those interactions is a lost
opportunity for human collaboration. The
social fabric of the workplace is
changing as AI takes over more of the
execution layer. So, Claude writes code
and Claude reviews code. What about
research? This is where things get more
serious. Anthropic has a test they run
every time they they a new model. They
give Claude some code that trains a
small AI model and ask it to optimize
the code to run as fast as possible
while still passing correctness check.
It's a miniature research loop. Rewrite
code, run it, measure it, repeat. In May
2025, Claude Opus 4 averaged around a
three times speed up. By April 2026,
Claude Mythos preview was hitting around
a 52 times speed up. For context, a
skilled human researcher would need 4 to
8 hours to reach around a four times
speed up on the same task. Claude
surpassed humans in under a year, but
Anthropic went even further. In April
2026, they published research showing
Claude-powered agents running an actual
AI safety research project from start to
finish.
The problem was weak to strong
supervision, which is basically a
preview of one of the biggest future
alignment challenges. If AI becomes
smarter than humans, how do we supervise
it? The research tested whether a weaker
model could train a stronger model and
still recover the stronger model's full
capabilities.
This mirrors the future scenario where
humans, who are weaker than advanced AI,
need to supervise AI systems that are
more capable than we are. Two human
researchers spent about seven days
tuning four prior methods and reached a
performance gap recovered score of 0.23.
That means they recovered 23% of the gap
between the weak baseline and the strong
ceiling. Then, Anthropic unleashed nine
parallel Claude Opus 4.6 agents. These
agents could propose hypotheses, [music]
run experiments, analyze results, share
findings through a forum, and iterate.
They worked for about 800 cumulative
hours and used roughly $18,000 in
compute. Their result, a score of 0.97.
While two humans recovered 23% of the
gap after a week, the Claude agents
recovered 97%. The cost was about $22
per agent hour. There are important
caveats. The result didn't transfer
cleanly to production scale models, and
humans still chose the problem and
designed the scoring rubric. But within
those limits, the agents designed every
experiment themselves. Direction setting
was the only meaningful role humans
played. One Anthropic researcher
commented that if a junior colleague
came back with results like this in 1 to
2 days, they would be mildly impressed.
The future, they said, is now. The
system turned compute into measurable AI
safety research progress. This is a big
deal because alignment research has been
bottlenecked by the number of human
researchers who can actually do the
work. If AI agents can take over
well-specified research problems, human
researchers can focus on the vague,
risky, high-level questions that still
require judgment. Now, this isn't just
Anthropic saying this. OpenAI just
published its own governance blueprint,
and buried in that document is a very
similar claim.
OpenAI says it sees early signs of
recursive self-improvement in today's
systems, where AI development itself is
being accelerated by AI. OpenAI argues
this will [music] intensify competitive
pressure between developers and
countries, and that existing
institutions aren't equipped to handle
it. So, both Anthropic and OpenAI are
now publicly acknowledging the same
trend. OpenAI's blueprint focuses on
building a federal framework for
frontier AI safety, strengthening
something called CAISI, [music] which is
the US Center for AI Standards and
Innovation, and creating a
whole-of-government resilience strategy.
But the underlying message is the same.
AI is already helping build AI, and the
race is accelerating. There's also
independent data backing this up. METR,
which is a research organization focused
on measuring AI capabilities, has been
tracking something they call task
completion time horizons. Basically,
they measure the length of tasks that AI
agents can complete reliably on their
own. In March 2024, Claude Opus 3 could
handle software tasks that would take
humans about 4 minutes.
One year later, Claude Sonnet 3.7 could
handle tasks around 1 and 1/2 hours.
Another year later, Claude Opus 4.6
could handle tasks around 12 hours.
The latest model, Claude Mythos preview,
can work for at least 16 hours,
>> [music]
>> which is at the upper limit of what METR
can even measure with their current task
suite. This doubling speed has
accelerated from once every 7 months to
once [music] every 4 months. If that
trend continues, AI systems could handle
tasks that take skilled people days
sometime this year. By 2027, possibly
tasks that take [music] weeks. METR's
data shows this across public benchmarks
as well.
SWE-Bench, which tests whether models
can fix real bugs in real open-source
codebases, [music]
went from low single-digit scores to
nearly saturated in 2 [music] years.
Core Bench, which tests whether models
can reproduce published research, went
from around 20% success in 2024 to
saturated 15 months later. METR also
found that Claude Mythos preview was at
the upper end of what they can measure
without developing new, harder tasks.
The benchmarks are running out of
headroom. So, what does all this mean
for the people actually working at
Anthropic? According to Krishna Rao,
Anthropic CFO, the shift is already
dramatic. In a recent podcast, he said
90% or more of Anthropic's code is now
written by Claude. Rao also said
Anthropic's finance team now uses Claude
to produce financial statements, and the
monthly financial review process is 90
to 95% ready before humans step in.
Reports that used to take hours now take
30 minutes. Rao described this as
employees shifting [music] from
execution to oversight. Humans are
becoming managers of AI systems. Teams
deploy what Rao called fleets of agents
working across projects simultaneously.
Everyone kind of becomes a manager. But
there's a darker side to this story. One
Anthropic employee mentioned that on
days when everything works well, they
can't help but think that nothing they
do matters. Everything is automated and
better and faster than they ever will
be. But then there are days where
everything breaks and they don't [music]
understand why, and they realize they
have no idea what they've been up to
anymore. The comparative advantage of
humans, for now, is still seeing the
bigger picture and thinking beyond the
confines of the immediate task. But how
long does that advantage last?
Anthropic's blog post lays out three
possible futures. The first is that
progress stalls due to bottlenecks in
energy, chips, or supply chains.
Even if capabilities stagnate at today's
level, the world would still change
massively. Anthropic points to Project
Glass Wing as an early sign. In its
first weeks, Mythos preview found more
than 10,000 high and critical severity
software vulnerabilities across the
world's most important systems. The
second future is that AI keeps
accelerating human organizations, but
humans still hold the reins. A company
of 100 people could do the work of
10,000. This would transform business,
science, government, [music] and
knowledge work, but it could also create
serious risks, including cyber threats,
authoritarian surveillance, and
large-scale manipulation. Anthropic says
this is probably the future we're moving
toward. The third future is full
recursive self-improvement. AI systems
design and build their own successors.
Progress becomes limited mostly by
compute. This could unlock breakthroughs
in science, medicine, energy, materials,
[music] and robotics, but it could also
make the alignment problem much harder.
If small misalignment problems compound
through self-improvement, humans could
lose control. Anthropic is blunt about
this uncertainty. A world driven by fast
recursive self-improvement could become
dominated by the self-improving model as
its capabilities fully eclipse those of
humans.
So, where does that leave us? Anthropic
is arguing that a coordinated pause on
frontier AI development could give
safety, research, and society time to
catch up. But a unilateral pause by one
lab doesn't work because less cautious
actors just keep going. A meaningful
slowdown would require multiple labs in
multiple countries to agree with
verification that nobody is cheating. We
don't have decades to build that trust.
The bigger question is whether humans
are still controlling the AI race or
just supervising [music]
it. Claude writes most of Anthropic's
code, reviews code, and runs experiments
faster than humans. Anthropic's warning
is clear. The world may need
coordination mechanisms before AI starts
building the next generation of AI
mostly by itself. The evidence suggests
we're already closer than most people
think. So, is this the beginning of AI
building AI? Maybe.
But the gap between human execution and
AI execution is closing fast. And the
only advantage humans have left might be
the ability to decide which problems are
worth solving in the first place. Also,
if you want more content around science,
space, [music] and advanced tech, we've
launched a separate channel for that.
Links in the description. Go check it
out. If you think coordinated pause
agreements are realistic or completely
naive, drop your take in the comments.
Hit subscribe if this made you rethink
how fast [music] things are actually
moving. Thanks for watching, and I'll
catch you in the next one.