--- title: "Google DeepMind executive: Unprecedented capital expenditure, essentially a bet on recursive self-improvement (RSI)" type: "News" locale: "en" url: "https://longbridge.com/en/news/294806070.md" description: "Google DeepMind Chief Strategy Officer Jagjit Sethi pointed out that the unprecedented capital expenditure in the industry is essentially a bet on Recursive Self-Improvement (RSI). RSI is seen as the core of investment logic and the next milestone after AGI. Although AI revenue has yet to cover the current massive investments, tech giants continue to ramp up data center construction, raising shareholder concerns about cash consumption" datetime: "2026-08-04T10:22:43.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/294806070.md) - [en](https://longbridge.com/en/news/294806070.md) - [zh-HK](https://longbridge.com/zh-HK/news/294806070.md) --- # Google DeepMind executive: Unprecedented capital expenditure, essentially a bet on recursive self-improvement (RSI) Google DeepMind Chief Strategy Officer Jasjeet Sekhon Last weekend, Jasjeet Sekhon, Chief Strategy Officer of Google’s artificial intelligence division DeepMind, explained with a concept: what kind of results can support the unprecedented capital investment in the industry — **Recursive Self-Improvement (RSI)**. This currently hot term refers to artificial intelligence that can automatically iterate and generate stronger versions of itself. At the AI summit held at the University of California, Berkeley, Sekhon stated that Recursive Self-Improvement (RSI) is “a core component of the entire investment logic.” Now in the industry, RSI has taken over the heat from Artificial General Intelligence (AGI) and has become the next milestone that everyone is discussing. Whether RSI will arrive before AGI or be achieved simultaneously with AGI — that is, creating AI that surpasses humans in most economically valuable tasks — is a completely independent topic. Objectively speaking, RSI has long been a goal pursued by the industry. I first heard this term from researchers in early 2023, but it wasn’t until the past few weeks that major tech companies began discussing this concept in public. **The AI field may face “stalling in mid-air”** Tech giants, including Google, which have heavily invested in AI, are seeing their software and cloud service-related sales revenue continue to accelerate, but Google shareholders are increasingly anxious: the company plans to invest about $200 billion this year in building AI data centers and related hardware, with cash consumption continuously expanding, and capital expenditures will further increase next year. Sekhon joined Google from Bridgewater Associates in April this year, having previously served as a statistics professor at Yale University. He candidly stated: “Currently, the revenue generated by AI is not enough to cover our current capital investment.” This means there is a huge risk: **we may face an “in-air stall” in the AI industry — massive investments continue to occur, but expected revenues are delayed in materializing.** Sekhon bluntly pointed out that the current capital investment across the industry is “the largest scientific gamble in human civilization history,” far exceeding the U.S. Apollo moon landing program, the Manhattan Project, and early internet research investments. He mentioned that the total investment in railway construction might be higher, but “humanity had already mastered the mature technology of building railways at that time,” so it cannot be considered a scientific gamble aimed at the unknown. He stated that the current technology in the industry has not yet reached true RSI, but “betting that RSI cannot be achieved does not seem like a wise choice.” The industry can now see the embryonic form of RSI: major AI companies are using models to assist in designing the components of other models "In fact, this is not new; humanity relied on steam engines to create a new generation of steam engines." **True Recursive Self-Improvement (RSI)** Many readers may wonder whether RSI will end up being as hollow as the concept of general artificial intelligence in the past. However, it is undeniable that AI has made tremendous leaps in just a few years: From static models that could only converse, could not acquire new information, and could not take proactive actions; to evolving into systems capable of autonomously searching the internet for information, leveraging massive computational power to deduce complex problems and output answers; capable of reading and writing multilingual code; able to manipulate browsers, computers, and various software to execute tasks. Modern AI agents also possess these capabilities: Integrating multiple data sources to complete tasks; recording execution processes for developers to troubleshoot errors; learning human instructions to master task methods; autonomously correcting errors after failures, continuously solving problems over long periods with minimal human intervention; deriving "sub-agents" to break down complex tasks. It is noteworthy that AI is already capable of generating and optimizing kernels, compiler components, and other low-level software, enhancing the operational efficiency of models on dedicated servers. Researchers can also use the model's own output results as training data to update weight parameters and continuously optimize subsequent outputs. But **the true standard for RSI is much stricter**: the system must be able to iteratively optimize the software architecture and training processes that enable its upgrades, forming a positive reinforcement feedback loop. Some researchers define RSI as the model's ability to independently reconstruct the entire architecture and evolve into a completely new model. Humanity is still far from this step, let alone relying on AI to independently make new significant scientific discoveries. However, based on recent technological advancements, Sai Hong predicts that RSI "is likely to emerge in the next few years." During the same roundtable discussion at the summit, his DeepMind colleague Oriol Vinyals and OpenAI co-founder Wojciech Zaremba both believe that RSI is expected to be realized between 2027 and 2028. **Biological Attack Security Risks** Sai Hong stated that while the realization of RSI can bring enormous benefits—AI is expected to tackle diseases and assist humanity in exploring the universe—he and Berkeley computer professor Song Xiaodong, who recently joined Meta's "Superintelligence" research department, focused on discussing various risks derived from AI, especially its use in cyberattacks and biological attacks. Song Xiaodong pointed out: **In the short term, AI is more advantageous to attackers**. There is an inherent asymmetry in offense and defense: attackers only need to find one vulnerability to break through, while defenders must secure all potential attack entry points. Sai Hong agreed with this. Attackers continuously pollute open-source code repositories and exploit vulnerabilities in human-written code to launch attacks, stating, "The situation is not optimistic in the short term. Many systems have weak links, such as the U.S. power grid, medical institutions, and so on." In the long run, defenders will also be able to rely on AI to counter AI-driven attacks, and more enterprises and institutions are expected to build verified, vulnerability-free system codes However, Sai Hong believes that compared to the biological attack risks spawned by AI, cybersecurity issues are almost a minor challenge. "Humanity is currently very vulnerable in the face of biological risks. We are very close to a dangerous scenario: anyone can design viruses and modify proteins just by conversing with AI in natural language. In this scenario, the long-term advantage still lies with the attackers." To defend against such threats, a comprehensive access, monitoring, and tracking system must be established for "materials with biological application value," similar to the current societal controls on fertilizer materials that can be used to make explosives. He also revealed that Google is adapting its AI content provenance watermarking technology, SynthID, for the biological field. For example, it can help biopharmaceutical companies that synthesize DNA screen for potentially dangerous biological sequences generated by AI ### Related Stocks - [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md) - [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md) - [GGLL.US](https://longbridge.com/en/quote/GGLL.US.md) - [GGLS.US](https://longbridge.com/en/quote/GGLS.US.md) - [GOOW.US](https://longbridge.com/en/quote/GOOW.US.md) - [GOOY.US](https://longbridge.com/en/quote/GOOY.US.md) - [GOU.US](https://longbridge.com/en/quote/GOU.US.md) - [GOOX.US](https://longbridge.com/en/quote/GOOX.US.md) - [GOOP.US](https://longbridge.com/en/quote/GOOP.US.md) - [MAGX.US](https://longbridge.com/en/quote/MAGX.US.md) - [META.US](https://longbridge.com/en/quote/META.US.md) ## Related News & Research - [Google nixes its Earth AI feature one day after launch, amid criticism it would spread misinformation](https://longbridge.com/en/news/294549600.md) - [Gemini agent digs a 13-year-old sandbox escape out of Chrome's code](https://longbridge.com/en/news/294501470.md) - [Brazil's Mombak delivers early to Google, other buyers in boost for carbon removal credits](https://longbridge.com/en/news/294810673.md) - [Alphabet Inc. $GOOGL is Saturna Capital Corp's 6th Largest Position](https://longbridge.com/en/news/294578714.md) - [Banks in talks to lend $15 billion for Anthropic data center backed by Google, WSJ reports](https://longbridge.com/en/news/294401941.md)