A path towards AGI: Demis Hassabis discusses the potential way to AGI.
TuringPost
Newsletter about AI and ML. ?? Sign up for free to get your list of essential AI resources ??
"How to build AGI" remains one of the most popular questions in the AI community. At a recent Tech Summit held by The Times & Times Business , Demis Hassabis , Co-founder & CEO of Google DeepMind, discussed the steps for achieving AGI.
So here is a summary of his main ideas:
?? 2-3 major breakthroughs are needed to achieve AGI.
?? MLLMs are not enough on their own; they will be a key component of AGI systems.
?? For AGI, we need Q&A systems (like chatbots).
?? Another element is agent-based systems. These AI assistants need to act in the real world, do planning (like planning a trip), reasoning, have strong memory capabilities, and personalization.
?? The next big breakthrough is applying planning from systems like AlphaGo to world models like Gemini for more advanced planning in a messy real world.
?? These advanced general agent systems need to be able to use tools, like hardware (e.g. robotics), software (like calculator), or even other LLMs, to solve various tasks.
?? A very important open research question remains:
Do we want to separate special functions into tools, or do we want to add them to the main system?
This may come with the following trade-offs:
If we put everything into the main system (like coding and mathematics), it makes the system better at everything. However, there may be a problem of overloading a system with too much specialized information.
?? Demis Hassabis says that there's still around 10 years before AGI.