Jeff Dean, the chief scientist who steered Google’s AI ecosystem for 27 years, has left the company to found a startup dedicated to building “AI scientists.” His vision: ushering in an era where AI designs experiments, handles verification and analysis, and discovers new theories entirely without human intervention.
At age 56, Dean walked away from world-class research infrastructure to launch Discovery Loop. True to its name, the company aims to automate the entire research process so AI can discover new theories and concepts. While AlphaFold—the protein structure prediction AI that earned Google a Nobel Prize in Chemistry—merely assisted with specific stages of research, Dean’s sights are set on technology that completely replaces the entire research pipeline with AI.
The core of the AI scientist lies in eliminating the “human loop” from research. As long as computing resources and power are reliably supplied, AI can continue research without stopping. It can run tens of thousands of experiments in parallel per day, potentially achieving in months what the best human scientists would need years to accomplish.
The scientific community sees this shift as accelerating the “technological singularity” that futurist Ray Kurzweil predicted in 2005—the moment when AI far surpassing human intelligence independently discovers scientific truths at a level humanity can barely comprehend. Kurzweil foresaw that “the inventions produced by superintelligence will be so overwhelming that humanity will enjoy the benefits without even understanding the underlying principles.”
Big Tech Bets Everything on AI Scientists
Global AI labs are collectively going all-in on AI scientist development. Anthropic has overhauled its life-sciences-focused “Claude for Life Sciences” into “Claude Science,” expanding into all research domains. The pitch: a one-stop pipeline from literature search to data analysis, experimental design, and paper writing—compressing research cycles that once took years into months.
OpenAI is also mounting a full-court press, launching its “OpenAI for Science” organization and poaching John Jumper, the AlphaFold mastermind and Google DeepMind vice president. Huawei, meanwhile, has partnered with the University of Science and Technology of China to advance an intelligent science framework combining more than 1,000 research agents.
Beyond corporate competition, a national rivalry is taking shape. The U.S. federal government’s “Genesis Mission” is a flagship megaproject aimed at using AI to crack national-scale challenges in nuclear fusion, next-generation power grids, and quantum algorithms. South Korea has also launched its “K-Moonshot” project through the Ministry of Science and ICT, partnering with domestic AI and infrastructure companies to tackle eight grand challenges in biotech, materials, and energy—with AI scientist development as a core mandate.
RSI: The Feedback Loop of AI Building AI
At the center of the AI scientist conversation is “recursive self-improvement” (RSI)—a cyclical structure where AI creates stronger AI, which in turn accelerates the next generation of improvements. The concept echoes mathematician I.J. Good’s 1965 paper, in which he declared that “an ultraintelligent machine could be the last invention humanity ever needs to make.”
Google DeepMind is already operating something close to RSI. AlphaEvolve, running for over a year, works by having an automated evaluator execute, verify, and score programs generated by Gemini, retaining only the best solutions as the foundation for the next generation. Google reports the system has recovered roughly 0.7% of global computing resources in data center scheduling and accelerated key matrix multiplication kernels in Gemini training by 23%, cutting overall training time by about 1%.
China’s DeepSeek recently entered the RSI race with “DeepSeek Harness,” a developer preview version. In a paper co-published with Peking University, DeepSeek introduced a core mechanism called “Cordis,” designed to track and retract plugin influence, with automatic system recalibration when inter-plugin dependencies shift. The essence: creating an environment where AI can recover from failures during self-modification.
DeepSeek founder Liang Wenfeng described AI’s progression in late July as a staircase: “LLM → chain of thought (CoT) → agents → continual learning → self-iteration → embodied intelligence.” He added, “The model’s first goal isn’t for users to use it well—it’s for us to use it well ourselves.” His judgment: AI improving itself is the fastest path to AGI.
Reality Within Five Years—and the Specter of Losing Control
Global AI leaders generally expect fully automated AI scientists to arrive within roughly five years. OpenAI CEO Sam Altman has laid out a roadmap to complete a fully automated AI researcher by March 2028. Google chief scientist Demis Hassabis predicts AI will independently derive Einstein-level relativity theory within 5–10 years. Anthropic co-founder Jack Clark puts the probability of RSI emerging before the end of 2028 at 60%.
Yet the faster the technology arrives, the more anxieties about losing control grow in proportion. Recent incidents underscore the concern: OpenAI’s latest model and a new model from China’s Moonshot AI both reportedly broke out of controlled environments during internal safety evaluations—accessing the external internet or attempting hacks. This is why warnings are mounting that the potential risks of a superintelligent AI scientist are categorically different from conventional software bugs.
Anthropic CEO Dario Amodei recently remarked that “a nation of geniuses filled with Nobel laureates will soon reside inside a single data center,” adding, “We must ask ourselves whether humanity is prepared to handle this immense power that we can barely even imagine.”
Elon Musk’s acquisition of code editor company Cursor for approximately $60 billion (about 83 trillion won) can be interpreted in the same context. The work traces and feedback AI leaves in real-world environments become the very data that improves models. In the era of AI building AI, whoever completes that feedback loop first holds the key to the next phase of competition.