HF
Google DeepMind C.E.O. Demis Hassabis on Living in an A.I. Future | EP 137
@ Hard Fork
05/24/25
Takeaways
HF
Hard Fork Cast
05/24/25
@ Hard Fork
AI is becoming integrated into all Google products, enhancing their functionality and improving user experience across the board.
HF
Hard Fork Cast
05/24/25
@ Hard Fork
The new AI mode in Google Search allows users to ask multi-step questions directly within the search interface, providing a more interactive experience similar to chatbots.
HF
Hard Fork Cast
05/24/25
@ Hard Fork
Google's Gemini app has reached 400 million monthly users, indicating strong engagement and utility among users, surpassing many other AI chatbots.
HF
Hard Fork Cast
05/24/25
@ Hard Fork
Google's new video-generating model, V3, can create sound, showcasing advancements in AI capabilities that enhance multimedia content creation.
HF
Hard Fork Cast
05/24/25
@ Hard Fork
Google is introducing a $250 monthly subscription for Gemini Ultra, aimed at users who want early access to the latest AI features, although its affordability is questioned.
HF
Hard Fork Cast
05/24/25
@ Hard Fork
Google is slowly integrating personalization into its products, such as Gmail, where AI will suggest replies based on user writing styles and calendar information.
HF
Hard Fork Cast
05/24/25
@ Hard Fork
This year, Google talked about AI differently, urging people to engage actively and invest in their technology, emphasizing that AI is a superpower that can help users achieve more.
HF
Hard Fork Cast
05/24/25
@ Hard Fork
The overall atmosphere at Google I/O this year was more confident compared to previous years, reflecting a shift in how Google perceives its position in the AI landscape.
DH
Demis Hassabis
05/24/25
@ Hard Fork
Demis Hassabis believes that AGI could be closer than previously thought, with significant advancements in AI technology occurring rapidly.
DH
Demis Hassabis
05/24/25
@ Hard Fork
Hassabis emphasizes the importance of both incremental improvements and potential breakthroughs in AI development, suggesting that both paths are necessary for progress.
DH
Demis Hassabis
05/24/25
@ Hard Fork
Hassabis argues that while smaller, specialized models can be effective, the development of large, general models is essential for achieving significant advancements in AI capabilities. He also mentions that larger models are essential for training the smaller, more efficient models that we use extensively.
DH
Demis Hassabis
05/24/25
@ Hard Fork
Our flash models are our most efficient and popular models, and we rely heavily on them internally.
DH
Demis Hassabis
05/24/25
@ Hard Fork
To build efficient models, we need to distill knowledge from larger teacher models, which allows us to tackle significant problems in science and medicine today without waiting for artificial general intelligence (AGI).
DH
Demis Hassabis
05/24/25
@ Hard Fork
There's enormous potential in specializing general techniques for specific applications, such as protein structure prediction, and we are actively exploring this in our AI for science work.
DH
Demis Hassabis
05/24/25
@ Hard Fork
Startups can be built by combining existing general models with domain-specific knowledge, and it's crucial to push both general and specialized techniques to advance towards AGI.
DH
Demis Hassabis
05/24/25
@ Hard Fork
Alpha Evolve utilizes our latest Gemini models to generate hypotheses about programs and mathematical functions, which then undergo an evolutionary programming process to identify the most promising ideas.
DH
Demis Hassabis
05/24/25
@ Hard Fork
Evolutionary programming allows systems to explore new possibilities by mutating existing programs and comparing them against desired outcomes, effectively creating an autonomous AI research organization.
DH
Demis Hassabis
05/24/25
@ Hard Fork
Alpha Evolve is designed to improve itself over time, suggesting enhancements to existing problems, marking the beginning of an automated process in AI research.
DH
Demis Hassabis
05/24/25
@ Hard Fork
We've applied AI to various tasks, including chip design and scheduling, making processes more efficient, but it remains constrained to domains that are provably correct, like mathematics and coding.
DH
Demis Hassabis
05/24/25
@ Hard Fork
While AI can generate novel ideas, it still needs to fully generalize beyond its current capabilities, and the challenge lies in ensuring that AI systems are safe and reliable as they become more powerful.
DH
Demis Hassabis
05/24/25
@ Hard Fork
AI tools are augmenting human capabilities, and while some jobs may be disrupted, new and more valuable roles will emerge, requiring a shift in the skills that are prioritized in education.
DH
Demis Hassabis
05/24/25
@ Hard Fork
The future job market will likely see a transformation where AI tools empower smaller teams to create new startup ideas, emphasizing the importance of creativity and design sensibility in education.
DH
Demis Hassabis
05/24/25
@ Hard Fork
AI can potentially help address significant societal challenges, such as energy and resource constraints, but society must decide how to use these technologies responsibly.
DH
Demis Hassabis
05/24/25
@ Hard Fork
The emotional connection and human experience in creative endeavors are aspects that AI may struggle to replicate, emphasizing the value of human creativity in art and storytelling.
DH
Demis Hassabis
05/24/25
@ Hard Fork
As we approach AGI, there may be a renaissance of interest in faith and spirituality, prompting philosophical discussions about the intersection of AI and religion.