2026-11-10 08:30

Introduction: Discovery Loop bets on automated experimental loops. In the closing keynote at Stanford HAI, Jeff Dean introduced his new venture after leaving Google—Discovery Loop. Its mission is to build AI systems capable of autonomously solving critical problems in machine learning, science, and engineering: proposing experiments, running them, evaluating results, determining next steps, and feeding insights back into subsequent iterations.

At Stanford HAI’s closing keynote, Jeff Dean introduced Discovery Loop, his new company formed after departing Google. The mission is to develop AI systems that can autonomously tackle significant challenges in machine learning, science, and engineering—proposing experiments, executing them, assessing outcomes, deciding next actions, and using each cycle’s feedback to improve the next. Subsequently, Fei-Fei Li engaged Jeff Dean in a dialogue around the distinction between “self-improvement” and “recursive self-improvement,” the impact of AI solving mathematical problems, the origins of Transformer, the accuracy (or inaccuracy) of technological forecasts, and the anxiety young people face when confronting AI.

Transcript of Their Conversation:

Fei-Fei Li: Let’s start with Discovery Loop. Why launch such a company now?

Jeff Dean: Our mission is to build AI systems that can autonomously solve important problems in machine learning, science, and engineering. We believe this will ultimately unlock more discoveries for the world. Many disciplines follow a common loop: propose an experiment, execute and run it, observe results and measure outcomes, then decide the next step—usually another experiment. We aim to automate this experimental loop. If we can shorten iteration time, automatically generate and run vast numbers of experiments, learn from each evaluation, and let later experiments benefit from prior insights, while simultaneously improving both quality and quantity, breakthrough results may emerge.

Fei-Fei Li: Why structure it as an independent non-profit entity?

Jeff Dean: We believe that if technology is done right, it can have extraordinary positive impacts on the world—but everyone must work to ensure those impacts materialize. That’s why we incorporated as a non-profit: our public mission is to develop technologies addressing critical societal challenges and ensure the benefits are widely accessible, improving lives globally.

Fei-Fei Li: What about the team?

Jeff Dean: I have three exceptional co-founders: Quoc Le, Sanjay Ghemawat, and Oriol Vinyals. We’ve collaborated for roughly 14 to 30 years, enjoying tackling hard problems together. We’ve worked on Google Search, ad systems, Gmail, news, translation, large-scale machine learning infrastructure, TPUs, model distillation, MoE, neural architecture search, and projects like AlphaChip, AlphaFold, and neural weather forecasting. We’ve also managed large teams: I led Google Brain up to 600 people, later overseeing Google Research and AI teams of around 4,400. But eventually, I found myself spending too much time in Q2 operational meetings and decided to change course. We’re proud that many team members have since co-founded Anthropic, OpenAI, Physical Intelligence, and others.

Fei-Fei Li: Why now? What technological conditions are mature?

Jeff Dean: Over the past 10–15 years, hardware performance has surged—TPUs, GPUs, and more specialized chips have advanced significantly; software abstractions have improved, enabling more efficient mapping to hardware; model architectures have evolved from word2vec, seq2seq, Transformer to MoE; training algorithms are better, distillation enables high-quality compact models; inference algorithms have also progressed substantially, allowing greater computational investment during reasoning, chain-of-thought prompting, multiple attempts, and verification for superior outcomes. Yet, strong foundational models alone aren’t enough. You need agents combined with models—capable of effectively leveraging tools to complete complex tasks, especially creativity: not just repeating what’s been done before, but integrating diverse insights to propose genuinely novel ideas and validate their effectiveness. Agents and models should also learn from previous experiments and experiences—after solving 5,000 problems, they should be able to extract relevant lessons.

Fei-Fei Li: Where lies the bottleneck in traditional experimental loops?

Jeff Dean: Traditional loops involve substantial human involvement at every stage. When scaling up, problems must be decomposed into subproblems—a process itself challenging. Each subproblem involves multiple people running experiments, testing different approaches, leading to high coordination overhead, resource conflicts, and systemic side effects. Worse still, once a large project begins, early decomposition choices are hard to revise. Our vision: a small number of domain experts paired with systems capable of automatically decomposing problems into interesting subproblems and autonomously solving them. Humans guide decomposition, highlight key subproblems, and delegate impactful issues to small teams. This requires robust infrastructure: tens or hundreds of thousands of automated experiments, where the system understands the value of one experiment versus another, uses reinforcement learning or evolutionary algorithms for automatic decision-making, allows early decisions to be revisited flexibly, and ensures each early experiment informs future ones.

Fei-Fei Li: What have you accomplished since launching?

Jeff Dean: In the first three weeks, we conducted 400 interviews—many at Cafe Venetia on University Avenue—sourcing top talent. We secured office space, currently housed in the former Zibibbo restaurant. We built our own agent harness for all development and automation. It includes collaborative features: you can converse with models, ask them to solve problems, invite teammates to collaborate with models and agents, and others can comment or suggest designs. We’re building an agent-first infrastructure designed to scale to tens or even hundreds of thousands of agents performing tasks—not just hundreds of humans. We currently run around 20 autonomous improvement loops, evaluating how to make experimental iterations more effective and faster, and how to learn from prior outcomes.

Fei-Fei Li: Which domains are you focusing on initially?

Jeff Dean: We focus first on scientific and engineering problems that can be fully digitized, as these are easiest to accelerate. Physical processes slow down the loop; some domains require hybrid approaches. Our current priorities are machine learning research and ML engineering itself, plus hardware design—domains where we have deep expertise and can unlock broader applications. For example, if a system can autonomously handle ML engineering tasks, it could be directed toward healthcare or materials science, demanding a model tailored to specific tasks. This constitutes self-improvement, but not recursive self-improvement. Healthcare is a long-term priority, but for now, we focus on digital loops where the technology is most mature.

Fei-Fei Li: Clarify the difference between “self-improvement” and “recursive self-improvement.” Many conflate them.

Jeff Dean: A self-improving system takes a goal—often with measurable metrics, such as model accuracy—and tries different model architectures to enhance that metric. Many machine learning algorithms do this, but it’s not inherently recursive. For instance, given a large dataset of MRI scans and outcomes, the system might develop a model to interpret images, trying various methods and measuring improvements in accuracy. However, this doesn’t improve the underlying system responsible for proposing the next experiment. Recursive self-improvement occurs when you use this self-improvement mechanism to enhance the very system that proposes the next improvement.

Fei-Fei Li: But RSI is heavily hyped, as if it could accelerate infinitely. Where are the physical limits?

Jeff Dean: Human progress in enhancing model capabilities relies on skilled researchers and ML engineers conducting numerous small-scale experiments, evaluating which ideas work and which don’t, then incorporating successful elements into the next-generation model. Gemini 2 was developed this way after Gemini 1. Automation can accelerate this process: a human-run experiment ending at 2 AM might not be reviewed until 8 AM; an automated system can assess results faster and determine the next step, possibly making better strategic decisions across entire projects. However, creating better foundational models still requires actual training and experimentation—running computations, observing results, saving data, and letting another model analyze outcomes to inform the next move. This cannot proceed infinitely fast due to real physical constraints. Automation increases the speed of improvement but does not enable infinite acceleration. Viewing it as a pragmatic engineering system reduces fear and enhances understanding.

Fei-Fei Li: Are humans self-improving systems or RSI systems? Take learning to cook an omelette as an example.

Jeff Dean: Humans are both. Throughout life, we accumulate knowledge—this is self-improvement. But we also learn “how to learn”—developing meta-cognitive strategies that allow us to acquire new knowledge faster and more effectively, based on what kinds of information and presentation styles are easiest for us to grasp. This is closer to recursive self-improvement. Recursive self-learning and recursive self-improvement may also relate to evolutionary algorithms.

Fei-Fei Li: Recently, AI has made significant strides in mathematics—from Navier-Stokes-related problems to newly published papers. Mathematics represents centuries, even millennia, of human intellectual achievement. Mathematicians value not only technical solutions but also aesthetic elegance. How do you view this shift?

Jeff Dean: Models and agent systems capable of highly parallelized use of models can indeed solve extremely difficult unsolved mathematical problems, revealing what is now possible. Many feel discomfort because these proofs often span 500 pages of Lean code—machine-verifiable, but opaque to human comprehension. Knowing a problem has been proven or refuted is valuable, but mathematicians frequently develop elegant techniques during the solving process, techniques that generate new mathematics applicable to other problems. Elegant solutions often yield reusable, clever methodologies. In contrast, long machine-generated proofs may obscure such clear, graceful innovations. Nevertheless, knowing something is solvable motivates effort in many fields—knowing we can achieve it may inspire us to find better, more elegant paths. Thus, while the simultaneous resolution of so many problems is unsettling, the fact that it’s now feasible demands reflection on how to leverage this capability for further mathematical discovery.

Fei-Fei Li: Quick quiz: What’s special about June 12, 2017?

Jeff Dean: Was that the day the Transformer paper went to arXiv?

Fei-Fei Li: Yes, A+. Did you read it on June 12?

Jeff Dean: No, probably around June 10.

Fei-Fei Li: Looking back over the past decade from that starting point, what’s been your most pivotal technical moment? Which predictions were correct, which were wrong?

Jeff Dean: Several stand out. Transformer was clearly a more efficient model architecture: retaining all historical states and applying attention, rather than relying on a small vector that updates iteratively and forgets past context. This enabled models to learn far more with significantly less computation. Model distillation was equally crucial. My work with Jeff Hinton and Oriol Vinyals was rejected by NeurIPS in 2016, dismissed as unlikely to have impact—we posted it to arXiv, and it became highly cited. It enables large models to act as teachers, guiding lightweight student models to create efficient inference-capable systems. MoE has also proven effective. One prediction I got wrong: I thought MoE would evolve into wilder, more heterogeneous, organic-like architectures—featuring neural activity columns and occasional long-range connections—like biological networks. Instead, MoE remains highly structured: experts are uniform in size and immediately merge. Neural architecture search produces models that look more organic than human-designed ones—connecting to upper layers, forming unusual links—but we haven’t reached the level of wildness I envisioned. Another underappreciated aspect was programming ability: it emerged as a key enabler for models to solve diverse problems. We prioritized it in Gemini, but didn’t foresee its depth.

Fei-Fei Li: At HAI, we must address “H”—AI’s social impact, employment, psychological anxiety, existential risks, alignment. What concerns you most?

Jeff Dean: This is an immensely powerful, highly general-purpose technology affecting nearly every facet of society: mathematics, coding, image generation, autonomous driving. Previous technological revolutions had narrower scopes; this one spans multiple domains simultaneously. Excitement, concern, and fear coexist. I co-authored “Shaping AI’s Impact on Billions,” examining seven areas: employment and economy, AI for science, misinformation and disinformation, media, among others—with extensive input from experts across fields. My primary concern is: how to maximize AI’s upside across sectors while minimizing downside harms. This isn’t purely a technical issue—it’s regulatory, policy-driven, and civic. We all must navigate this technology responsibly, ensuring it serves everyone without causing foreseeable harm. For instance, when jobs are disrupted, we need viable pathways for people to adapt to new societal demands, secure meaningful work, and lead fulfilling lives. This is a multifaceted challenge requiring collective action.

Fei-Fei Li: Final question. Today’s students are anxious—Stanford students worry about jobs and economic futures. As an AI pioneer, what message do you have for these anxious yet hopeful students?

Jeff Dean: I deeply understand the anxiety—our world changes rapidly, planning for a fast-evolving future is daunting. My overarching advice: try to understand what new technologies are emerging or will soon emerge, and ensure you learn to apply them to what matters most to you. Whether becoming a doctor, a writer, or writing software more efficiently, AI can assist and empower. But human agency remains essential: humans must decide what we collectively want to achieve. Clarify your passions, strive to excel, and leverage every available tool. Human passion endures in the age of AI.

Fei-Fei Li: Thank you, Jeff.

Jeff Dean: Thank you.

Conclusion

The core takeaway from Jeff Dean’s conversation is not “Will AI replace scientists?”, but rather that AI is transforming scientific discovery into an engineerable loop. Discovery Loop’s ambition is to empower a small group of domain experts to solve historically labor-intensive problems through automated experimental cycles. Yet Jeff Dean offers a sober boundary: automation accelerates progress, but never infinitely; recursive self-improvement warrants research, but shouldn’t be hyped as a magic trick without physical cost. For research institutions and entrepreneurs, the real opportunity lies not in chasing the “RSI” slogan, but in digitizing, automating, and iterating the experimental loop within a specific domain. And for young students, his advice is simple yet profound: understand emerging technologies, pursue what you love, and wield all available tools.

Source: Post-Wave Evolution Planet