#ArtificialIntelligence #110 Responsible AI by design for Artificial General Intelligence

Responsible AI by design for Artificial General Intelligence?

Introduction

Welcome to #ArtificialIntelligence #110

In this edition, I will share some more about a paper we are working on relating to my work/teaching? in the #universityofoxford?

As we know, AI has been completely transformed through generative AI and #GPT. I have been revamping our courses so that our students get the latest insights.

One of the key ideas we are working on is “Responsible AI by Design”. Similar to the idea of Privacy by Design, responsible AI by design puts the idea of Responsible AI at the outset (and across all layers of the stack) as opposed to retrospectively.

While the idea itself has been around, implementing Responsible AI by design within the stack from a developer perspective is not easy. These ideas are especially relevant for AI today since GPT has raised the bar for AI and professionals will need to have actual knowledge of implementing Responsible AI in future jobs.

But more interestingly, How would Responsible AI by Design apply to Artificial General Intelligence??

While it's a research topic, considering the success of GPT, its a very relevant topic.?

Broadly, the idea of “Responsible AI by design" refers to the principle of designing and developing artificial intelligence (AI) systems in a way that prioritizes ethical considerations, fairness, transparency, and accountability from the outset.?

Artificial general intelligence (AGI) refers to highly autonomous systems that possess human-level cognitive capabilities across a wide range of tasks.?

In this sense, AGI still does not exist, but if chatGPT is any indicator, we are well on our way to AGI.?

With that in mind, responsible AI by design becomes even more crucial for artificial general intelligence.?

The broad considerations are still the same for AGI as they are for AI itself i.e.?

• The existence of an Ethical Framework??
• Alignment of values
• Transparency?
• Explainability??
• Robustness?
• Safety?
• Human Oversight and Control?
• Inclusivity?
• Fairness and?
• Societal Impact Assessment

All of these apply to both AGI and AI

They are complex but solutions already exist for all of them?for AI

When it comes to AGI however, we also have to consider algorithms with?

• Human level ability and intelligence
• Autonomous operation with self directed and goal seeking behaviour?
• Cognitive ability (ability to reason) including common sense reasoning
• Applicability across multiple domains?
• Superintelligence
• Robustness?
• Reliability
• Rapid learning and adaptation?

As you see from the above, the scope above is complex and its not really possible to think about Responsible AI by design for all of AGI. So, I am addressing the central paradigm of reasoning.?

AGI and the ability to reason

Reasoning problems are a type of problem that requires the use of logic and reasoning to solve. They can be found in a variety of domains, such as mathematics, science, and everyday life. Reasoning problems can be difficult to solve, as they often require the ability to think abstractly and to draw inferences from incomplete information.

The recent paper “Sparks of Artificial General Intelligence: Early experiments with GPT-4” discusses the use of large language models (LLMs) to solve reasoning problems. As we know, because LLMs are trained on large amounts of text data, LLMs have potentially learnt the statistical relationships between words and phrases, which can then be used to generate text, translate languages, and answer questions.

However, the authors of the paper claim that LLMs can be used to solve a variety of reasoning problems, including mathematical and scientific problems. By extension, the authors argue that LLMs can potentially solve a wide range of reasoning problems, and that they could eventually be used to create artificial general intelligence (AGI). However, many challenges remain. One challenge is that LLMs are often trained on data that is biased or incomplete. Another challenge is that LLMs can be computationally expensive to train and use.?

Despite these challenges, the authors believe that LLMs have the potential to revolutionise the way that we solve reasoning problems. They argue that LLMs could be used to automate many of the tasks that are currently done by humans, such as making decisions, planning activities, and solving mathematical and scientific problems.

“Ability to reason” by design

So, the scope can now be narrowed as gaining an understanding of the “Ability to reason” by design.?

We can now extend this further as potentially as follows:

• How would we implement causal reasoning using LLMs - a topic explored by the paper? Causal Reasoning and Large Language Models: Opening a New Frontier for Causality by Emre K?c?man ?
• How could machines reason in the age of artificial general intelligence (outside of LLMs)?
• Explore other paradigms of AGI specifically Chomsky’s Universal Grammar, Symbolic AI and Neuro-symbolic AI

While the problem is complex - its a very useful and pragmatic problem to solve - both for technological reasons and policy reasons. Reasoning is a key concept to engaging and managing AGI - and hence a candidate to include into AGI by design (rather than retrospectively) for AI services in the near future.?

Welcome comments. I will continue this thinking in the next newsletters

If you want to learn with us, please see our new course on low code data scientist https://lnkd.in/dqZ9qecN

Image source wikipedia Rodin the thinker