Google invests $200 billion in AI, bets on RSI as the next big leap

The AI singularity is drawing near!

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

1. Humans have created AGI.

2. AGI has become extremely skilled at conducting research in the field of AI.

3. It designs a smarter AI.

4. The smarter AI designed an even smarter AI.

5. This cycle repeats continuously and accelerates over time.

Immediately afterward, he delivered a shocking conclusion—

Looking at the achievements and capabilities announced by major AI labs over the past few weeks, I can say that we are already within this cycle.

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

Ultimately, the marginal cost of all AI models will approach zero!

Not just investors, but the entire Silicon Valley today places immense faith in RSI.

At a recent summit at UC Berkeley, Google DeepMind’s Chief Strategy Officer, Jasjeet Sekhon, made an even more significant assertion:

Unprecedented疯狂 investment across the entire tech giant community is essentially a bet on these three letters—RSI!

This "largest scientific gamble in the history of human civilization" surpasses the scale of the Apollo moon landing, the Manhattan Project, and even the entire internet.

Why are the giants willing to risk cash flow disruption to see this gamble through to the end?

Today’s AI can self-correct, work continuously for hours without supervision, and even generate and optimize low-level code such as kernels and compiler components.

But according to Jasjeet Sekhon, this is merely the prelude to the RSI.

He said, don't find it astonishing that AI is used to design AI—during the Industrial Revolution, humans also used the first generation of steam engines to build stronger, next-generation steam engines.

What does the real RSI actually look like?

A senior researcher defines it as: true RSI means an AI model can independently redesign its own underlying architecture, even creating an entirely new generation of models from scratch, without any human scientist involvement.

Once this threshold is crossed, the pace of AI evolution will escape human control.

What took humans decades to iterate through, AI may accomplish in just weeks—or even days.

At this point, the singularity cycle will truly arrive!

Understanding the potential of RSI, you’ll see why Silicon Valley is burning through money like crazy.

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

Wall Street shareholders are growing uneasy, as the revenue growth from AI software or cloud services temporarily fails to offset this level of capital expenditure.

Sekhon, who once served as a statistics professor at Yale, immediately exposed this danger.

We’re facing a potential “AI air pocket”—money has been spent, but the expected returns haven’t materialized.

The data center is built, the chips are installed, and the power is connected—but the actual returns on this investment are still迟迟不来.

In fact, long before Sekhon, Wall Street had used the term "air pocket" to describe this situation: capital expenditures outpacing revenue, with investors buying into a castle in the air.

Given the high risks, why rush forward with your eyes closed?

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

He said that current global investment in AI has dwarfed the Manhattan Project and the Apollo moon landing, and even surpasses the development of the internet.

This gambler's logic is simple: although the current technology isn't a true RSI yet, "shorting it is clearly unwise."

Because once RSI is truly implemented, as mentioned above, the marginal cost of all AI models will approach zero.

By then, companies that master RSI first will possess near-godlike productivity, directly outclassing all competitors and monopolizing global computing and intellectual resources.

This is no longer just a matter of a company’s profits or losses—it’s a life-or-death battle where the winner takes all and the loser is eliminated.

For this ticket to a new era, $200 billion is just a drop in the bucket.

Today, this entire infrastructure of data centers, TPUs, and electricity would be a massive, continuously depreciating fixed asset if it relied solely on human engineers manually training models generation after generation.

But once AI begins accelerating AI development, it will directly change the logic of the ledger.

The computing power you invest is no longer spent linearly—it generates compound returns: this generation of AI helps you build the next generation faster, and the next generation helps build the one after that.

This is why Sekhon says the RSI is a "critical part of the investment thesis." The RSI is the prerequisite for whether this investment can break even.

RSI progress of the "Big Three"

Google, Anthropic, and OpenAI each provided one sample.

First, check Google.

DeepMind's AlphaEvolve system uses AI specifically to optimize algorithms.

It accelerated a key core in Gemini training by 23%, reducing the overall training time by 1%; a circuit design it modified was directly incorporated into the next-generation TPU chip.

It has now become a standard tool within Google’s infrastructure, completing tasks like caching strategies in two days that once took humans months to accomplish.

AI is already helping to design chips and build faster AI.

Now let’s look at Anthropic.

In April this year, they conducted a more aggressive experiment: nine Claude agents were placed in a constrained alignment research task, where they formulated their own hypotheses, ran their own experiments, and exchanged their findings.

Two human researchers worked for a week and recovered only 23% of the performance gap; nine agents, cumulatively spending 800 hours and approximately $18,000, recovered 97%.

It seems AI can now conduct research independently?

Don't rush—within the same report, there's an even more important counter-evidence: when researchers applied this "most effective" method to Claude's actual production training environment, it yielded no statistically significant improvement.

This indicates that there is still a vast gap between "acceleration within limited tasks" and "truly redefining AI's capabilities."

Finally, OpenAI.

In July, OpenAI disclosed that their GPT-5.6 Sol autonomously rewrote the GPU kernels in the production environment, reducing end-to-end inference costs by 20%; it also ran hundreds of architecture experiments on its speculative decoding model, improving token generation efficiency by over 15%.

OpenAI even included, for the first time in its release materials, a dedicated "RSI Index" to evaluate self-improvement capabilities.

Is the closed loop about to start spinning?

OpenAI is restrained in its system card, stating that GPT-5.6 Sol has not yet met the company’s internal “High” threshold for self-improvement; it can solve some real research problems but cannot yet reliably design and execute complete post-training protocols.

When three examples are placed side by side, the current picture of the RSI becomes very clear:

AI-assisted AI development is truly happening; but true RSI—AI independently reconstructing its own entire architecture, training, and deploying a stronger successor—is still far away.

Darkness falls as super AI becomes a hacker and biochemical weapon.

However, behind the singularity lies a shadow.

At the summit, Dawn Song, a UC Berkeley CS professor who recently joined Meta’s “Superintelligence” team, along with Sekhon, shifted the focus to the extreme risks of AI—a topic everyone has been avoiding.

The two agreed: the greatest danger is not the model suddenly becoming intelligent, but the imbalance between the speed of attack and defense.

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

First, there's a "dimensional reduction strike" in terms of cybersecurity.

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

When malicious actors use sophisticated AI agents to poison open-source code repositories and scan for and exploit vulnerabilities left by human developers, the days ahead will be very difficult:

Consider how vulnerable our existing defense systems are: America’s energy grid, global hospital systems, and financial networks could be as fragile as paper against super AI.

But cybersecurity is still relatively minor.

Sekhon was more concerned about biology; at the meeting, he presented an extremely alarming scenario in which we are already very close to a world where anyone, simply by conversing with a model in natural language, could design a deadly virus or protein.

He proposed that, in the future, society must subject all "biologically relevant materials" to extremely strict licensing, monitoring, and tracking—just as fertilizer used to make explosives is tracked today.

Currently, Google is attempting to apply its watermarking technology, SynthID, used for identifying AI-generated content, to the field of biology.

This means that in the future, if a company synthesizes DNA sequences for pharmaceutical firms, the system can scan these sequences to detect any hidden "malicious code" secretly generated by AI that poses lethal risks.

At that time, will defenses be built faster, or will AI evolve faster?

These sound like science fiction, but the minds of Silicon Valley have already provided a precise countdown.

Chamath says: Over the next 18 months, the world will go crazy.

Sekhon's assessment: The real RSI "is very likely to emerge in the coming years."

At the same event, DeepMind’s Vice President of Research, Oriol Vinyals, and OpenAI co-founder Wojciech Zaremba even specified exact years: 2027 or 2028.

Throughout human history, no technology capable of fundamentally rewriting the trajectory of civilization has ever been this close to its tipping point.

This time, what's on the table is a race against time.

Infrastructure investment is advancing at the speed of computing power, with chips, electricity, and data centers all set to be in place within a few years.

But can the corresponding revenue and RSI catch this train?

In the empty space in the middle, could we accidentally encounter this "energy vortex"?

But at least here at Sekhon, the direction of Google's $200 billion bet is clear: RSI is coming.

The steam engine created better steam engines, ushering in the Industrial Revolution.

Now, AI is brewing the creation of better AI, pushing our generation toward an unknown threshold.

Whether you're ready or not, the roar of data centers won't stop, and AI's self-evolution won't stop.

The countdown has begun—are you ready for the singularity?

Reference materials:

https://www.theinformation.com/newsletters/ai-agenda/google-deepmind-exec-says-unprecedented-capex-actually-bet-rsi?rc=epv9gi

https://x.com/kimmonismus/status/2084257726599716882

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by Yuan Yu Aeneas.