Inside Google’s big AI shuffle — and how it plans to stay competitive, with Google DeepMind CEO Demis Hassabis

Language Models

Demis Hassabis discusses the challenges and advancements in language models, particularly focusing on improving their accuracy and factuality. He acknowledges that while these models can hallucinate and spread disinformation, there are ongoing efforts to enhance their reliability without sacrificing creativity. Demis explains that their Sparrow model was an experiment in achieving better factuality and rules adherence, but it sometimes compromised on creativity 1. He envisions future systems that maintain creativity while improving accuracy, a balance that is crucial for the next generation of AI 2.

I think the large language models, and I think this is one reason that Google has been very responsible with this, is that we know that they hallucinate and they can be inaccurate.

--- Demis Hassabis

He also highlights the potential for AI systems to self-improve through reinforcement learning, reducing the need for human raters while still ensuring quality 2.

   

Path to AGI

The journey towards Artificial General Intelligence (AGI) is marked by both breakthroughs and challenges. Demis outlines the concept of tool use, where AI systems call upon specialized tools to solve specific problems, as a critical step towards AGI 3. He believes that the integration of research and product development will accelerate progress towards AGI, creating a feedback loop that enhances both areas. Demis is optimistic about reaching AGI within the next decade, though he acknowledges the uncertainty and potential need for significant breakthroughs 4.

I think that is on the critical path to AGI, and that's another reason, by the way, I'm very excited about this new role and actually doing more products and things.

--- Demis Hassabis

He emphasizes the importance of pushing both existing systems and exploratory research to overcome current limitations and achieve AGI 4.

   

AI Media

Ensuring the reliability and trustworthiness of AI-generated content is a pressing concern. Demis discusses solutions like encrypted watermarking to identify AI-generated media, addressing issues like deepfakes and disinformation 5. He stresses the need for AI systems to fact-check themselves, akin to a good researcher cross-referencing facts, to achieve a high level of reliability 6.

We're working on some pretty cool solutions to that. I think the answer is, and this is an answer to deepfakes as well, is to do some sort of encrypted watermarking.

--- Demis Hassabis

Demis envisions a future where AI systems can critique their outputs, enhancing their ability to provide trustworthy information 6.