The AI bubble dwarfs the scale of the Apollo moon landing program, Google has burned through 200 billion US dollars to bet on the biggest gamble of the 21st century.

The AI Singularity is approaching!

Recently, Chamath Palihapitiya, the renowned investor known as the "Warren Buffett of Silicon Valley", shared on X his ultimate roadmap for the "AI Singularity".

1. Humanity builds AGI.

2. AGI becomes extremely adept at conducting research in the AI field.

3. It designs an even smarter AI.

4. That smarter AI then creates an AI that is far more intelligent than itself.

5. This cycle repeats continuously, with its speed getting faster and faster.

Right after that, he put forward a staggering conclusion —

Looking at the achievements and capabilities released by major AI labs in the past few weeks, I dare to say that we are already in this cycle right now.

In the next 18 months, RSI will rapidly enhance AI capabilities.

Eventually, the marginal cost of all AI models will approach 0!

It is not only investors, but the entire Silicon Valley today that holds unparalleled faith in RSI.

Just recently, at a summit held by UC Berkeley, Jasjeet Sekhon, Chief Strategy Officer of Google DeepMind, raised an even more groundbreaking judgment:

The unprecedented massive investment across the entire tech giant circle is essentially betting on these three letters — RSI!

This "largest scientific bet in the history of human civilization" has a scale that exceeds the Apollo moon landing program, the Manhattan Project, and even the development process of the entire internet.

Why are giants willing to risk a breakdown of their capital chains to push this gamble all the way to the end?

Forget AGI, RSI is the ultimate totem of Silicon Valley

Current AI can self-correct, work continuously for several hours without human supervision, and even generate and optimize low-level code such as kernel and compiler components

But in Jasjeet Sekhon's view, this is only the prelude to RSI.

He said, don't think it's incredible to use AI to design AI. During the Industrial Revolution, humans also used the first generation of steam engines to build the next generation of more powerful steam engines.

So what does real RSI actually look like?

A senior researcher defines it this way: Real RSI means an AI model can completely independently redesign its own underlying architecture, even build a brand-new generation of model from scratch, without any intervention from human scientists throughout the whole process.

Once this critical threshold is crossed, the speed of AI evolution will get out of human control.

Iterations that take humans decades to complete may only take AI a few weeks, or even a few days.

At that point, the Singularity cycle will truly arrive!

The $2000 billion "AI air pocket", the largest scientific bet in the history of human civilization

Once you understand the potential of RSI, you will see why Silicon Valley is burning money like crazy.

This year, Google's total capital expenditure on AI data centers, chips and infrastructure alone has reached an astonishing $195 to $205 billion, and the number will rise significantly next year!

Shareholders on Wall Street are starting to get uneasy, because according to financial reports, the revenue growth of AI software or cloud services cannot cover capital expenditures of this magnitude for the time being.

Sekhon, who used to be a professor of statistics at Yale, pointed out this risk directly on the spot.

We are facing a potential "AI air pocket" (the term refers to a sharp drop in air pressure that causes an aircraft to plummet mid-flight, which he uses to describe failed investments): money has been spent, but the expected revenue has not appeared as scheduled.

Data centers are built, chips are installed, and power supply is connected, but the thing that can deliver returns on these investments is nowhere to be seen.

In fact, long before Sekhon, Wall Street had already used the term "air pocket" to describe this situation: capital expenditure runs ahead of revenue, and investors are buying a castle in the air.

Since the risk is so high, why are people still rushing forward without looking back?

Sekhon's answer is: This is the largest scientific bet in the entire history of human civilization.

He said, the current global investment in AI has dwarfed the Manhattan Project and the Apollo moon landing, and even the development of the internet in the early years cannot compare to it.

This gambler's logic is very simple: although current technology is not real RSI, "it is obviously unwise to short it".

Because once RSI is truly realized, as mentioned above, the marginal cost of all AI models will infinitely approach 0.

At that time, the company that takes the lead in mastering RSI will have nearly god-like productivity, directly defeating all competitors with dimensionality reduction strikes, and monopolizing global computing power and intellectual resources.

This is no longer a matter of profit and loss for a single company, but a life-or-death game of "winner takes all, loser exits".

For this ticket to the new era, $2000 billion is just a small cost.

If the current pile of data centers, TPUs and power resources can only rely on human engineers to manually train models generation after generation, they will be a huge fixed asset that keeps depreciating.

But once AI starts to accelerate AI R&D, it will directly change the logic of the accounting book.

The invested computing power will not be consumed linearly, but generate compound returns: this generation of AI helps you build the next generation faster, and the next generation then helps you build the generation after that.

This is why Sekhon says RSI is the "key part of the investment logic". RSI is the prerequisite for this sum of money to generate returns.

The RSI progress of the "Big Three"

Google, Anthropic and OpenAI have each provided a sample of their progress.

Let's start with Google.

DeepMind's AlphaEvolve system is dedicated to using AI to optimize algorithms.

It has speeded up a key kernel in Gemini training by 23%, reducing the overall training time by 1%; one of the circuit designs it modified was directly incorporated into the chips of the next generation of TPU.

Now it has become a regular tool in Google's infrastructure. For tasks such as caching strategy, it only takes 2 days to finish work that used to take humans months to complete.

AI helping to build chips and faster AI has already become a reality.

Next is Anthropic.

In April this year, they conducted a more radical experiment: nine Claude agents were assigned to a limited alignment research task, to raise hypotheses, run experiments, and exchange findings on their own.

Two human researchers working for a whole week only closed 23% of the performance gap; the nine agents spent a total of 800 hours and cost about $18,000, closing 97% of the gap.

It seems that AI can already conduct independent research?

Wait, the same report also hides a more important counter-proof: when researchers moved this "most effective" method to Claude's actual production training environment, it did not bring any statistically significant improvement.

This shows that there is still a huge gap between "acceleration in limited tasks" and "the real capability to restructure AI".

Finally, let's look at OpenAI.

In July, OpenAI disclosed: Their GPT-5.6 Sol independently rewrote the GPU kernel in the production environment through Codex, cutting end-to-end inference costs by 20%; it also ran hundreds of architecture experiments for its speculative decoding model, boosting token generation efficiency by more than 15%.

OpenAI even published a dedicated "RSI Index" that evaluates self-improvement capabilities for the first time in its release materials.

Is the closed loop about to start running?

OpenAI stated very conservatively in its system notes that GPT-5.6 Sol has not yet reached the company's custom "High" self-improvement threshold. It can solve some real research problems, but cannot reliably design and execute a complete post-training scheme.

Putting the three cases together, the current picture of RSI is very clear:

AI-assisted AI R&D is indeed happening; but real RSI — where AI independently restructures its full architecture, trains and deploys a stronger successor, is still some distance away from us.

Darkness falls, when super AI becomes a hacking tool and biochemical weapon

However, there are also shadows behind the Singularity.

At this summit, Dawn Song, a CS professor at UC Berkeley who just joined Meta's "Superintelligence" department, together with Sekhon, turned the focus of discussion to the issue that everyone has been avoiding: the extreme risks of AI.

The consensus between the two is: the most dangerous thing is not that the model suddenly becomes smarter, but the imbalance between the speed of attack and defense.

When AI gains RSI capabilities, it can cure cancer and help humans explore the universe, but it can also become the ultimate weapon to destroy humanity.

The first is the "dimensionality reduction strike" in the cybersecurity dimension.

Dawn Song warned that in the short term, the development of AI will "favor attackers more", which creates an extremely asymmetric battlefield: hackers only need to find one vulnerability to succeed, while defenders must defend against all possible attacks.

When malicious attackers use smart AI agents to poison open source code libraries, scan and exploit vulnerabilities left by human programmers, the days ahead will be extremely tough:

Just think about how fragile our existing defense systems are: the US energy power grid, global hospital systems, and financial networks may be as fragile as paper in front of super AI.

But cybersecurity is still "child's play".

Sekhon is more worried about biological risks. He raised an extremely terrifying scenario at the meeting: we are already very close to a world where "any person, by simply talking to the model in natural language, can design a lethal virus or protein."

He proposed that future society must impose extremely strict licensing, monitoring and tracking on