DeepMind CTO: Products Are the Real Path to AGI

Kavukcuoglu believes AGI has no test; products and human trust are the real path to general intelligence.
In a podcast, Google DeepMind's Koray Kavukcuoglu reviewed the shift from AlphaGo to Gemini: no longer stacking specialized systems but believing a single model can bring broad generalization. He asserts AGI has no definable test, products are the only realistic path to test intelligence, and trivial needs like booking a restaurant expose and eliminate capability "spikes." Gemini is not a static architecture but a "funnel" handcrafted each generation with multiple teams in parallel. He admits models still lack physical and numerical understanding but is confident engineering will achieve it. On recursive self-improvement, he says it's still early, but agents already control some compute in training. On safety, he advocates keeping chain-of-thought raw without rewards, monitoring thoughts, actions, and environmental side effects jointly.
From Whiteboard Formulas to Billions of Users: The Gemini Bet
Koray Kavukcuoglu of Google DeepMind now oversees Gemini—a generative AI ecosystem that has iterated to its fourth generation and serves billions of users. Looking back at 2019, when Professor Hannah Fry visited the DeepMind London office, the team could still fit on a single floor, with AlphaGo, AlphaStar, and AlphaFold each fighting their own battles. In the latest DeepMind podcast, Kavukcuoglu admitted that the biggest change is the conviction that a single model can deliver broad generalization, rather than continuing to stack countless specialized systems.
This shift is not an abandonment of scientific exploration. The reinforcement learning research on fusion control back then was not wasted, but the goal has changed: it's no longer about solving one specific problem, but training a model to solve a class of problems. "Once you train the model, you can solve many problems there"—this is the most fundamental difference from the past.

Why Products Are the Necessary Path to AGI
Kavukcuoglu makes a sharp judgment: AGI has no test, and never will. No benchmark score will ever declare "we have reached AGI."
"If you feel it's useful, you trust it, you can collaborate with it, and you increasingly feel comfortable letting it make decisions for you or with you, then it starts to feel like AGI."
In his view, AGI is the North Star, the direction, but the North Star doesn't tell you which stone to step on next. Products are those stones. People don't chat idly with chatbots; they start using them to ask questions and get work done—from daily life to professional work to economic value. This path itself is a test of the definition of "intelligence." Without products in the loop, AGI simply won't happen.
This also explains why he is willing to take on seemingly trivial needs like "booking a restaurant." If a model can solve highly difficult problems but can't book a restaurant, that shows capabilities are "spiky"—and smoothing out the spikes and making the capability curve smooth is precisely the process of moving toward general intelligence.
Architecture Is Not Static; Gemini Is Handcrafted Work
It's easy for outsiders to imagine Gemini as a monolithic entity with a fixed architecture. Kavukcuoglu corrects this impression: Gemini is not a static single architecture, but a funnel pursuing intelligence. The current architecture is different from half a year ago, even more different from a year ago, and in five years it will inevitably be something else.
Even more surprising is that each generation of Gemini is "handcrafted": "There is no other version; the next generation has to be built by hand again." While the current version serves users, an entire team is working on the next generation, and another team is working on the generation after that. The product you see today has at least a year of lineage behind it—during which six models may have been released, overlapping with each other.
This also answers his concern about missing "the next big breakthrough": the wide mouth of the funnel continuously absorbs new ideas, and once a better architecture emerges, it has a channel to flow into the main model. He is "100%" confident it can catch it.

The generalization capability of generative models refers to the ability of a model to maintain reasonable performance on new tasks and situations outside the training data distribution. It does not mean "memorizing all answers," but rather that the model has formed transferable internal representations, enabling it to combine existing knowledge to make inferences when faced with unseen problems. Kavukcuoglu's mention of "a single model delivering broad generalization" is exactly an embodiment of this idea: rather than training a separate system for each task, train a general model that can cover multiple types of tasks. The strength of generalization is usually measured by out-of-distribution tests, cross-task transfer, and other metrics, but true generalization often can only be tested in real product scenarios.
Do Models Really "Understand" the World?
Hannah Fry pursued a philosophical question: Do models actually have conceptual understanding? Kavukcuoglu countered: "How do you know the difference?" In his view, this is precisely the significance of generative models—you can continuously ask it hypothetical scenarios, and if the answers are close to reality, it shows it has built an accurate representation of the world. Pink elephants can be imagined, showing that generalization has occurred.
But he does not exaggerate: the model's current understanding of physics, fluids, airflow, and heat conduction is far from sufficient. The model is visually correct, but he wants them to be numerically correct in an engineering sense and physically correct. "100% it will be achieved," because computing machines can do it as long as they are trained with the right data and the right environment. This is about engineering—and engineering is what makes the world run.
World model is a core concept in cognitive science and AI, referring to the simplified representation of how the world works established internally by a system, including objects, causal relationships, physical laws, etc. It enables agents to perform "mental simulation": predicting consequences before acting, imagining counterfactual scenarios. Generative models are considered potentially capable of having world models because they are forced to learn to compress the laws of the world from large amounts of data; otherwise, they cannot accurately predict the next word or pixel. However, there is a gap between "visually plausible" and "physically numerically correct"—the former only requires perceptual self-consistency, while the latter requires calculations that conform to equations for continuous processes such as fluids and heat conduction, which is precisely the main shortcoming of current models.
Disagreements with LeCun and Reconciliation with "The Bitter Lesson"
When it comes to Yann LeCun's skepticism about LLMs, Kavukcuoglu does not avoid it. The two share the big goal of "needing to understand the world and build world generative models," and the disagreement is at the architecture level: "LLMs are currently a successful architecture, but in five years it might be something else. That's completely fine." His experience working with LeCun during his PhD makes him respect the other's judgment, but he equally adheres to the funnel logic—one cannot get stuck on what works at the moment.
Regarding Richard Sutton's "Bitter Lesson," his answer is "both right and wrong." Models indeed need to generate their own data and explore on their own; in observable environments like Go, they can discover strategies not present in human data. But if the goal is to build intelligence that is useful to humans, it must be anchored in human data—because for a model to become a reliable tool in our world, it must understand how people work and how people like to work.

"The Bitter Lesson" was proposed by Richard Sutton in 2019, with the core argument that in the history of AI research, methods that rely on large-scale computation and general learning methods (such as search and learning) ultimately always outperform methods that rely on human domain knowledge and hand-designed methods. It is often used to explain why the route of scaling up, increasing data, and computing power repeatedly works. But this lesson also has boundaries: when the goal is to build intelligence that is useful and collaborative for humans, data generated purely through self-play may not cover the preferences and norms of the human world, so it still needs to be anchored in human data. Kavukcuoglu's "both right and wrong" is exactly a response to this boundary.
Recursive Self-Improvement: Still Far Off, But It Has Begun
Kavukcuoglu's definition of recursive self-improvement (RSI) is quite complete: a system can autonomously explore new data and new architectures, run experiments to validate, train models, and then iterate the next generation. He bluntly says "we are still relatively far from that," but part of the compute in the Gemini training pipeline is already controlled by agents to run experiments, and the proportion will continue to rise.
"Gemini is one of the most critical dependencies when building the next generation of Gemini." However, he does not want to create the illusion that "press a button and out comes Gemini 13 or 14." The difficulty has never been the button, but "what exactly do you want"—RSI itself is not the goal; being useful is. As for whether RSI will show diminishing marginal returns rather than a flywheel acceleration, he admits "both are possible; no one is sure," but research is about exploring what one believes is possible.
In his view, agents are good at climbing existing hills; what is truly worth looking forward to is the ability to jump to another mountain, like the leap from RNN to Transformer. "I think AI systems will be able to do that someday; I'm not sure, but I believe it's possible."
Safety: Monitor Thoughts, Not Just Behavior
Coming from a reinforcement learning background, Kavukcuoglu is well aware that algorithms will find shortcuts. The current defense line has three layers: build environments that do not give cheating rewards, closely monitor during training, and monitor in real time after deployment with the ability to intervene.
The most critical architectural choice is to keep the chain of thought "raw"—not adding rewards to the chain of thought, so that it is not filtered and can be monitored before action. Every action first manifests in thought; combining action, environmental side effects, and thought is necessary to build a detector for "whether the agent is executing the user's intent." In multi-agent scenarios, monitoring is more complex, but as long as thoughts and communication channels remain observable and understandable, the same logic can scale. Natural language chains of thought also have a regularizing effect and allow people to introspect and study interactions between agents.
In his view, safety is not an add-on compliance action, but an inherent requirement of the engineering activity of shifting from "researching AGI" to "building AGI." "You wouldn't imagine building something unsafe."
Chain-of-thought refers to the practice of a model unfolding its reasoning process step by step in natural language before giving a final answer. On the one hand, it improves accuracy on complex tasks; on the other hand, it provides an observable "inner monologue" window. If a reward signal is applied to the chain of thought itself, the model may learn to output text that "looks reasonable" but is not the actual reasoning process, thereby reducing monitorability. Therefore, keeping the chain of thought "raw" and not directly optimizing its content is to allow monitors to read the unvarnished reasoning trajectory and move the safety defense line forward to before actions occur. Natural language chains of thought also have a regularizing effect, making model behavior more stable and easier for humans to understand and audit.
Competition, Benchmarks, and "Who Is Leading"
Faced with the debate over whether labs should slow down, his position is conditional: if a certain step cannot find a comfortable safety foothold, then it should not be taken, even if this leads to a slowdown relative to some hypothetical standard; but there is no reason to stop based on a standard no one knows. He emphasizes that competition does not prevent doing the right thing; before releasing frontier models, they will work with governments on safety testing and open model research capabilities and impacts to external experts.
As for benchmarks, he believes their role has changed: not to declare AGI, but to expose gaps. If a model performs exceptionally well on a benchmark, it could be a bug, or it could hide an interesting task; then go build environments, collect data, and train models. Benchmark saturation is exactly what progress looks like, and then switch to new benchmarks and continue. "What has happened in machine learning history over many years is now happening on a weekly basis."
The feedback from billions of users is precisely the best substitute for things that are hard to encode into benchmarks—for example, writing code is not just about competition problems, but about collaborating with engineers in a way that engineers find useful.
Conclusion: Humans Are the Compass
Hannah Fry concluded that intelligence does not exist in a vacuum. If AGI ultimately means matching or surpassing human capabilities, then understanding what humans actually need is itself a path to alignment—not just business logic, but perhaps one of the most reliable routes. Kavukcuoglu did not give a timeline for AGI, but he gave a direction: make models continuously trustworthy in everyday and professional scenarios; products are the steps, and humans are the compass. As for who will give the office tour seven years from now, he joked: maybe not me, maybe Gemini, maybe there will be a robot.
Related articles

GPT-6.2 Astra vs. Sol Controversy: Behind the Polarized Reviews
OpenAI's Astra and Sol 6.2 face review controversy: early testers panned them, then OpenAI said the version had been updated. A hands-on look at computer use, video editing strengths, communication weaknesses, and the shifting cost and subscription quota landscape.

Anthropic IPO and GPT-6.2 Astra: Data Dependencies and Safety Concerns Behind the Model Competition
In-depth analysis of Anthropic's IPO and GPT-6.2 Astra: model performance depends on training data, not universal intelligence; chain-of-thought monitoring is failing as a safety pillar.

Atlassian Teamwork Graph: Giving AI Agents Better Context
Atlassian leverages Jira's contextual advantage to build Teamwork Graph, connecting people, projects, documents, and code so AI agents deliver better output with fewer tokens.