AGI is the buzzword these days. Montreal.ai, on December 23rd, 2022, conducted their AI Debate 3 on the topic 'The Pivotal Discussion in Shaping the Path of AGI's Global Discourse'. The speakers of the discussion included Erik Brynjolfsson, Yejin Choi, Noam Chomsky, Jeff Clune, David Ferrucci, Artur d'Avila Garcez, Michelle Rempel Garner, Dileep George, Ben Goertzel, Sara Hooker, Anja Kaspersen, Konrad Kording, Kai-Fu Lee, Francesca Rossi, Jürgen Schmidhuber and Angela Sheffield. 

The debate was moderated and co-organized by Gary Marcus. Gary Marcus is a leading voice in AI. Scientist, best-selling author, and entrepreneur, he was the Founder and CEO of Geometric Intelligence, a machine-learning company acquired by Uber in 2016.  

Today, we live in a world of Dall-E, Midjourney, and the brilliance of ChatGPT. Several plot twists changed the course of AI. Meta's AI guru Yann LeCunn stated that most of today's AI approaches would never lead to true intelligence.  

The AI debate concentrated on five major points:

1. Can we turn to cognitive neurosciences for inspiration? 
2. How can we make progress in common sense reasoning? 
3. How should we structure and develop our AI systems? 
4. How can we build AI systems that reflect human values? 
5. What should we do morally and legally to ensure a bright future? 

Aspects of cognition 

According to Gary Marcus, four key aspects of cognition have always been essential but remain unsolved: 

• Abstraction: abstraction is a vital part of human cognition. Current AI still struggles with it.  
• Reasoning: Humans reason about the world. However, current AI systems have to hope for the best. Therefore, reasoning becomes the second aspect of cognition.  
• Compositionality: Humans understand language in terms of wholes composed of parts. On the other hand, current AI approaches still struggle with the aspect of compositionality. 
• Factuality: Humans also actively maintain imperfect but reasonably reliable world models. Large language models don't, and that has consequences making factuality the fourth aspect of cognition.  

Understanding causality 

The concept of causality is getting popular in the Machine Learning community. It is evidently essential as it promises much better generalization. However, the structural causal models (SCM) usually focus on being good detectives. 

Causality is interesting in cognitive science. But humans may be focused on the kind of causality for which we have rich models. 

It is important that AI systems make some basic assumptions about the world. This is because it will make the process of learning feasible in a reasonable amount of time. On the other hand, too many assumptions will make the system not general anymore and too few will require millions of years to learn. Hence, it is important that the assumptions of the AI systems are accurate in number.  

Deep learning systems have these assumptions embedded in them, in architectural constraints or in data.  

Human beings are attempting to identify what kind of creatures we are. According to Noam Chomsky, humans are a unique species due to their capabilities in thought and language. 

Innovative technologies, like AI, is the result of our attempt to answer that big question. Therefore, if we want an answer, these capabilities are the place to look.