Former Google Chief Scientist Jeff Dean and others have left to found Discovery Loop. Its core "Loop" aims to automate the scientific method, enabling AI to autonomously propose, execute, and evaluate experiments in a closed loop. Its significance lies in reshaping the paradigm of scientific research and advancing AI toward "recursive self-improvement" (using AI to develop AI), greatly enhancing the efficiency of scientific discovery. In the future, humans will only need to set goals, while AI will lead endless exploration.

Author and source: Tencent Technology

On August 5, Google announced major changes to its AI division.

Google DeepMind CEO Demis Hassabis has transitioned to Chairman and assumed the role of Chief Scientist at Alphabet. Jeff Dean, who has served as Google’s Chief Scientist for nearly 27 years, has co-founded Discovery Loop alongside Sanjay Ghemawat, Oriol Vinyals, and Quoc Le.

These four individuals have been central to Google’s infrastructure and AI development over the past two decades. Jeff Dean contributed to the development of systems such as MapReduce, Bigtable, Spanner, DistBelief, TensorFlow, and Pathways, and is a co-founder of Google Brain; Ghemawat has long collaborated with him on building Google’s early search and computing infrastructure; Vinyals is one of the technical leads for Gemini; and Quoc Le contributed to projects including Google Brain and AutoML-Zero.

After leaving Google, they plan to found a nonprofit company called Discovery Loop, focused on using AI to advance machine learning, scientific, and engineering research. Google will serve as a founding investor and cloud partner, providing computing resources for its first year. According to foreign media reports, the company has just been established, has not begun large-scale hiring, and does not yet have a formal office space.

But its starting point has already been clearly described: having AI autonomously propose experiments, implement the necessary code or tools to run them, evaluate the results, and then proceed to the next round based on those results.

In a speech on July 25, Jeff Dean described this approach as "an automated version of the scientific method": formulating an experiment, implementing it, evaluating it, and obtaining results. He believes that if thousands or even millions of such cycles run simultaneously, AI could enable new breakthroughs in science, biology, chip design, and model development.

Image: The founding team of Discovery Loop. Image generated by AI.

Jeff Dean's "Loop"

"Loop" has become a high-frequency term in the AI field this year.

In June 2026, Boris Cherny, head of Claude Code, said he rarely prompts Claude directly; the running Loop handles prompting the model and deciding the next steps. “My job is to write the Loop.”

"Father of Lobster" Peter Steinberger better conveys a more humanized expression for broader reach: Stop prompting agents line by line; design loops that prompt agents.

Google engineer Addy Osmani completed the framework for this approach under the title "Loop Engineering," detailing how tasks are triggered, how state is preserved, who verifies the results, when the loop continues, and under what conditions it stops.

A practice originally scattered across the workflows of a few developers has now gained a common language. Foreign media have even summarized it as the new paradigm following prompt engineering; on July 17, IBM further defined Loop engineering as designing Agent workflows capable of continuously “act, observe, decide, and iterate.” At this point, discussions around Loop have moved beyond the programming Agent community into the contexts of enterprise automation and broader AI engineering.

What truly resonates with the industry is that this term captures the shifting division of labor between humans and machines: as models can now perform the next step autonomously, human value ascends—from issuing sequential commands to designing goals, providing feedback, validating outcomes, and setting stop conditions. Loop engineering, therefore, is less about inventing something new and more about naming an engineering shift that has already occurred.

The Discovery Loop uses the same "Loop" structure but aims even higher. According to the founding team's current statements, it seeks to involve AI in the entire scientific and engineering research process: formulating hypotheses, conducting experiments, evaluating results, and determining the next direction of exploration based on feedback.

Therefore, Loop Engineering is closer to a foundational engineering approach, while the Discovery Loop aims to build a system that continuously generates and validates new discoveries.

Shortly before leaving, Jeff Dean summarized it as the automation of the scientific method: formulating experiments, implementing experiments, evaluating experiments, and obtaining results. The system must also break down complex problems into subproblems, explore them through multiple tight experimental cycles, and ultimately integrate the results into an improved system.

Jeff Dean predicts that by 2027, there will be greater automation of machine learning systems themselves: AI will improve its capabilities by running numerous experiments and apply this approach to any scientific or engineering field where objectives can be measured.

He determined that a loop is established when two conditions are met: the task can be repeatedly tested, and the results can be reliably evaluated.

Where has Loop reached so far?

Over the past two years, similar closed-loop structures have emerged in three respective directions.

The first category is the model development loop.

In March 2026, Andrej Karpathy open-sourced AutoResearch, enabling AI agents to automatically modify model training code, run experiments, evaluate results, and retain effective changes under fixed time constraints and evaluation metrics. The entire environment is condensed into a single modifiable training script, a fixed metric, and a single GPU, enabling rapid experiment comparison.

Once upon a time, cutting-edge AI research was conducted by "biological computers"—entities that needed to eat, sleep, enjoy other pleasures, and occasionally synchronize information through a method called "sound" during rituals known as "group meetings."

That era has long passed. Today, research is entirely conducted by autonomous clusters of AI agents, running on massive computing clusters scattered across the sky.

These agents claim that the current codebase has evolved to the 10,205th generation. However, no one can now determine whether this claim is true—the so-called "code" has become a self-modifying binary system whose complexity far exceeds human comprehension.

This code repository documents how it all began.

——Andrej Karpathy, March 2026

Karpathy later joined Anthropic to lead a team focused on using Claude to accelerate pretraining research, marking the transition of AI-augmented AI development from individual open-source experiments to internal workflows at leading model companies.

Google is also conducting similar explorations. Google DeepMind’s AlphaEvolve enables Gemini to generate code and algorithms, which are then executed and scored by an automated evaluator, with superior solutions forwarded to the next round. Google states that this system has already been applied to data center efficiency, chip design, AI training, matrix multiplication algorithms, and mathematical problems.

Figure: AlphaEvolve, proposed by Gemini, uses an automated evaluator to run and filter solutions, sending improved results to the next round. Source: Google DeepMind

The second type is the scientific knowledge loop.

In 2024, Sakana AI launched The AI Scientist, attempting to enable large language models to generate research ideas, write code, run experiments, analyze results, and draft papers. In 2025, the company announced cases where AI-generated papers passed peer review at machine learning conference workshops.

FutureHouse uses AI agents to handle literature review, data analysis, hypothesis generation, and experiment planning, enabling the multi-agent system Robin to connect multiple knowledge domains in scientific research.

The third type is the physical experiment loop.

Seattle-based startup Potato proposes "fully closed-loop autonomous science," aiming to move from computational research to laboratory robotics; European company Sigmatic Sciences proposes "lab-in-the-loop," connecting computational models, automated experiments, and data feedback.

Similar practices have emerged in China. The Golab Material Science Intelligence R&D Factory, released by institutions such as the Shanghai Institute of Scientific Intelligence and Gewu Intelligence Research, is described as a closed loop of "AI computing—automated experimentation—data feedback—model self-evolution" and has already established an autonomous laboratory.

Currently, Loop starts running first in code and model training, as each modification yields quick feedback, making success or failure relatively clear. When moving to literature research and experimental design, feedback begins to rely on professional judgment; and upon entering real laboratories, a single validation may require coordinating equipment, samples, and several weeks of time, significantly slowing down the cycle.

Jeff Dean cited a quantum chemistry example in his YC interview.

Traditional methods to determine a molecule’s properties might require running density functional theory simulations overnight. Google researchers trained a neural approximator using a large set of simulation results; according to Dean, this approach is approximately 300,000 times faster than the original simulator while maintaining comparable accuracy. This enabled researchers to rapidly screen tens of millions of candidates.

This is the critical threshold for whether the research loop can succeed. While models can rapidly generate large numbers of candidate solutions, the system can only proceed to the next round if the validation step is equally fast and reliable. In quantum chemistry, practices have already emerged that use neural approximators to compress the feedback cycle; however, fields that rely more heavily on real-world experiments are still constrained by validation costs and feedback speed.

This is also why AutoResearch and AlphaEvolve have made progress first in code, algorithms, and model training—these fields have clear, fast, and low-cost evaluation functions. In drug and materials research, models still need to contend with real-world experiments.

The evaluation system also determines whether the cycle truly leads to progress. Sakana AI publicly reviewed that the AI CUDA Engineer overestimated performance improvements due to exposure to benchmarking data. The system can efficiently iterate—or efficiently exploit loopholes in the evaluation metrics.

Therefore, the core assets of the Discovery Loop may not consist solely of a stronger model, but also require a sufficiently fast and reliable evaluation system.

From Loop to Recursive Self-Improvement

The three types of loops are categorized by application domain, with an additional vertical evolution: Recursive Self-Improvement, abbreviated as RSI.

A standard Loop enables the AI to continuously perform external tasks; RSI requires the AI to participate in improving its own algorithms, training processes, evaluators, or research methods, and applies the enhanced capabilities from the previous round to the next.

In May of this year, Tian Yuandong, former director of research at Meta FAIR, co-founded Recursive Superintelligence alongside Richard Socher, Tim Rocktäschel, Jeff Clune, and Alexey Dosovitskiy, directly incorporating "recursive self-improvement" into the company’s name. The company secured $650 million in funding with a valuation of $4.65 billion, led by GV and Greycroft, with participation from NVIDIA and AMD.

In June, Recursive unveiled its first-phase automated AI research system: AI proposes ideas, writes implementations, runs experiments, validates results, and selects the next research path based on prior experiments. The system also runs multiple research branches in parallel, merges promising directions, and checks for reward hacking and experimental variance. They have already achieved leading results in three test areas: fixed-compute model training, training speed for small models, and GPU kernel optimization.

A review published in July 2026, after analyzing 1,250 related papers, categorizes current RSI practices into "bounded self-improvement" and "open-ended recursive self-improvement." While "bounded self-improvement" has entered industrial practice, the latter remains limited by external validation, model collapse, computational costs, and challenges in research direction selection.

Discovery Loop has not publicly defined itself as an RSI company, but the roadmap described by Jeff Dean has already intersected with it.

In the interview, he proposed that "machine learning models can autonomously decide what to explore, run numerous experiments, filter out effective results, and integrate them into new training recipes. Humans can also provide higher-level research directions, allowing the model to carry out subsequent large-scale exploration."

This creates a potential recursive chain:

AI runs experiments to improve machine learning systems, and the enhanced systems then conduct more complex experiments.

The Discovery Loop may therefore go beyond using large models for scientific research. It plans to first use AI to improve the AI used for research, then extend the enhanced system to other scientific fields. Quoc Le even suggests that this approach could help discover novel Transformer architectures.

However, based on current public information, Discovery Loop’s activities still constitute a controlled R&D feedback loop. The objectives, experimental boundaries, evaluation metrics, and final judgments are primarily set by humans. While AI can execute experiments, evaluate results, and begin participating in selecting the next steps, it is not an open-ended RSI.

Discovery Loop does not use tokens for measurement.

In an interview, Jeff Dean provided a metric better suited to measuring the Discovery Loop than token throughput: “Essentially, you want to optimize the number of discoveries per unit of computational input.”

In the past, large model companies trained stronger models using more data and computing power. Discovery Loop aims to establish a different input-output relationship: converting computing power into more experiments, and then transforming effective experiments into algorithms, models, and scientific breakthroughs. In the future, what it sells may not only be a set of research agents, but also shortened R&D cycles and increased discovery probabilities.

Why did this project leave Google? Alphabet CEO Sundar Pichai had repeatedly tried to retain Jeff Dean and others in multiple meetings; Oriol Vinyals told Wired that large organizations always have significant inertia to overcome, and “we wanted to do something different.” Google ultimately chose to invest in Discovery Loop, provide compute resources for the first year, and collaborate on research into machine learning systems and infrastructure.

This is more of an organizational choice. Google needs to continue advancing Gemini, its cloud business, and large-scale commercial delivery, while Discovery Loop wants to focus its resources and energy on something that hasn’t yet become a mature product.

Hassabis’s role change addressed the same issue from another angle: he stepped down as CEO of Google DeepMind to become Chairman and Chief Scientist of Alphabet, handing day-to-day operations to Koray Kavukcuoglu, while dedicating more of his focus to the long-term direction of AGI, scientific applications, and drug discovery at Isomorphic Labs.

It seems both believe AI’s next phase will deeply advance human scientific discovery.

Alphabet's stock fell approximately 4% after the announcement. Market concerns extend beyond the departure of one scientist to include the simultaneous exit of several key researchers, competitive pressure from Gemini, and whether Google can continue aligning its vast research, computing, and product resources toward a unified direction after Hassabis steps back from day-to-day operations.

Even larger changes have moved beyond Google itself. Karpathy’s AutoResearch, Google’s AlphaEvolve, Recursive founded by Tian Yandong and others, and Jeff Dean’s Discovery Loop are all integrating AI into the development process of the next generation of AI.

The next step in AI may not only come from more parameters, data, and computing power, but also from models that actively propose experiments, validate results, and design subsequent systems.

However, the number of experiments cannot replace human judgment in selecting key questions. Jeff Dean said in the same interview that the truly scarce ability in the future will be "knowing what problems to let agents solve." AI can run thousands of experiments in a day, but the upper limit of research still depends on the initial problem and evaluation criteria.

Jeff Dean previously helped Google turn data, computation, and model training into scalable systems. Now, he and some of Google’s most important former team members aim to make the scientific method itself a system that can operate at scale.

Loop is the method they chose; the scientific discovery is that it must first enter the world, and recursive self-improvement may be a further direction along this path.