Science & Technology

Gemini and the path to AGI with Koray Kavukcuoglu

by Google DeepMind

Share:

📚 Main topics

  • From specialised projects to GeminiGoogle DeepMind has shifted from distinct projects such as AlphaGo, AlphaStar and fusion research towards training general-purpose models that can tackle many domains. Scientific projects still matter, but increasingly help train models to solve a range of problems within those fields. 1:35
  • Products as a route to AGIKoray argues that products reveal what people need AI to do and provide practical steps towards the broader goal of AGI. Useful products also test whether models can handle real-world tasks, from work and coding to everyday activities. 4:44
  • Evolving model architecturesGemini is not a fixed architecture; each generation is assembled using lessons and research from earlier models. Teams explore new ideas in parallel, with the aim of incorporating promising developments into products. 12:45
  • Understanding the worldGenerative models can be tested by asking them to predict plausible outcomes in hypothetical situations. Koray says models have made progress, but still need to improve at numerical accuracy and representing physical processes such as fluid dynamics and heat transfer. 17:31
  • Recursive self-improvementKoray describes RSI as a system carrying out the research loop itself: exploring data and architectures, running experiments, checking results and training improved models. Some parts of this process already involve models, but a fully self-improving loop remains a work in progress. 24:50
  • Safety for agentsAs models gain access to tools and operate across digital environments, safety requires carefully designed training environments, ongoing monitoring and the ability to intervene. For agent teams, their thoughts, communications and actions need to remain observable. 30:43
  • Benchmarks and real-world feedbackBenchmarks can reveal specific capability gaps, but they do not determine whether a model has reached AGI. Products and user feedback help identify which tasks matter in practice, including aspects of coding that competitions do not measure. 40:19

✨ Key takeaways

  • Generality mattersIntelligence is not just success at a collection of narrow tasks; it involves transferring capabilities to tasks the model was not specifically built for. 11:07
  • AGI is a journey, not a thresholdKoray does not identify a definitive test or moment when a system becomes AGI. He sees progress as models becoming more useful and trusted over time. 7:27
  • Human data and exploration can complement each otherModels may discover strategies by exploring for themselves, while human data remains important for building systems that work well with people. 23:17
  • Research requires uncertaintyA system might make incremental gains, hit a capability ceiling or eventually make a major architectural leap. Koray says none of these outcomes is certain, but exploring what might be possible is part of research. 28:04
  • Safety and development must go togetherBroader capabilities bring greater responsibility. Koray says release decisions should account for risks, mitigations and external safety testing, rather than being driven by competition. 34:26

🧠 Lessons learned

  • Use people as a compassProducts and user feedback help determine which capabilities are valuable and where models still fall short. Useful tasks can also expose limitations that researchers did not anticipate. 43:00
  • Make agents monitorableObservable reasoning and communication can help teams inspect what agents are doing and detect when their actions diverge from user intent. 31:45
  • Choose meaningful evaluationsBenchmarks are useful for examining particular skills, but practical collaboration with professionals can reveal needs that test scores miss. 44:02

🧠 Conclusion/next steps

  • Keep advancing with careKoray expects models to grow more capable and useful, while emphasising that safety work, monitoring and attention to real-world impacts must continue alongside development. 37:04
  • A more capable everyday and professional assistantLooking ahead, he hopes intelligent systems will feel like a normal part of people’s lives and help them make progress more quickly in their professional work. 45:36

Transcript excerpt

0:00 Welcome to Google DeepMind: The Podcast. I'm Professor Hannah Fry. Now the first time that I met our guest on the podcast today, he was dashing around Google DeepMind's offices in London, scribbling equations on whiteboards, directing blue sky projects that would fundamentally alter the trajectory of modern technology, all with the rather audacious goal of creating intelligence. if you're interested in AGI and at the end we need we need to work on agents. And that is something that we have been doing for a long while. Quite a lot has changed since then. These days, Koray Kavukcuoglu has swapped London for California

0:33 and those whiteboards for the even more monumental task of driving one of the world's most advanced generative AI ecosystems. he is steering the development of Gemini, now in its fourth iteration, and overseeing the rollout of AI to billions of people at once. But perhaps things still haven't changed all that much. He is still pursuing that same goal of creating intelligence. Only now the technology is finally catching up with his ambition. And is he still dashing between projects? Well, let's find out.

🔒 The full, searchable transcript is available with Pro.

🔒 Unlock Premium Features

This is a premium feature. Upgrade to unlock unlimited Q&A, transcripts, mindmaps, and translations.

Questions & Answers

Common questions about this video

Why does Koray Kavukcuoglu see products as a path to AGI?

Products reveal what people need AI to do and whether they can trust it with real decisions. They also expose gaps that research benchmarks may miss. In his view, usefulness and generalisation in real-world tasks are central to what AGI means, and products provide practical steps towards that goal. 4:44

How does Kavukcuoglu describe Gemini’s architecture and development?

Gemini is not a single, fixed architecture. Its architecture evolves as researchers explore new approaches, and each model generation is assembled using lessons from earlier versions. Teams also work on current and future generations in parallel. 12:45

What does Kavukcuoglu mean by recursive self-improvement, and how close is it?

He describes it as a system carrying out the research loop itself: exploring data and architectures, running experiments, checking results and training improved models. Today, models can help with pockets of that process, but he says they are still some way from a complete self-improvement loop. 24:50

How does Kavukcuoglu think AI agents can be monitored for safety?

He emphasises creating environments where cheating is not rewarded, carefully monitoring training and keeping oversight active when agents operate in the real world. Monitoring an agent’s thoughts, actions and effects on its environment can help detect whether it is still following the user’s intent. 30:43

Why does Kavukcuoglu say an agent’s thoughts should be in natural language?

Natural-language thoughts can help regularise the system and make its reasoning easier to inspect. They also let researchers study communication between agents and help people monitor what the agents are doing. 32:49

What role should benchmarks play in evaluating AI models?

Benchmarks can measure performance in particular domains and reveal tasks where a model is weak or unexpectedly capable, but they do not determine whether a model has achieved AGI. He argues that choosing useful tasks matters more, and that products and user feedback help identify them. 40:19

Does Kavukcuoglu think AI research should move away from human data?

Not entirely. He says models can generate data and explore strategies beyond what appears in human examples, but human data remains important because the goal is to build systems that are useful to people and fit into the human world. 23:17

🔒 Unlock Premium Features

Access to Chat is a premium feature. Upgrade now to unlock unlimited studying tools.

🔒 Unlock Premium Features

Access to Mindmap is a premium feature. Upgrade now to unlock unlimited studying tools.

🔒 Unlock Premium Features

Access to Translation is a premium feature. Upgrade now to unlock unlimited studying tools.

Get unlimited summaries, Q&A, transcripts and more with Pro

Upgrade to Pro

Suggestions

🔒 Unlock Premium Features

Access to AI Suggestions is a premium feature. Upgrade now to unlock unlimited studying tools.