Questions in AI

Navigating the Technical, Societal, Economic, and Political Landscape of AI

The rapid pace of advancements in AI brings forth a plethora of questions that challenge our understanding and shape our future. AI is not just a technical endeavor; it intertwines with every facet of our lives, raising crucial questions across technical, societal, political, and economic landscapes. This post aims to shed light on some of the most pressing questions in these areas, encouraging readers to ponder and engage in these critical discussions.

Technical

At the core of AI development are technical questions that dictate the direction and capabilities of AI systems. Some key questions here are:

• How to incorporate common sense into AI. Current systems excel in specialized tasks but often falter in understanding basic human concepts and reasoning. I am fascinated by the question of common sense and I’ll spend my next blog post on the concepts of common sense knowledge, embodied cognition and reasoning in computational agents

• Another significant challenge is creating intrinsic reward mechanisms for AI agents, allowing them to learn and adapt autonomously without explicit external rewards. Currently, researcher’s in Reinforcement Learning(RL) need to handcraft reward functions for agents. There is a problem with this approach. First, the world is so open ended, messy and high dimensional that it is almost impossible to conjure a reward function which leads an agent to behaving in a safe manner. It is this concern that AI Safety research focuses on. Furthermore, taking inspiration from nature, we notice that animals in the world reward themselves; they probably have an internal mapping between observations and a mechanism for generating, maintaining and discarding intrinsic rewards

• How to get Large Language Models to be better at arithmetic. Lots of research groups are pursuing solutions to this. For example, the team at PolymathicAI considers current Tokenization methods partly responsible for LLM’s poor arithmetic . They have a publication detailing a proposed solution. I am skeptical of getting LLMs to do arithmetic because of the “impedance” between LLMs which learn probability distributions and arithmetic which is a highly systematic process.

• AI agents is what gets the AI developer community buzzing. This prompts the need for a new protocol to govern AI agent interactions similar to the development of Internet protocols like TCP/IP, HTTP etc.

These are but a few technical questions facing the field. There is still so much work left.

Economic

Economically, AI is possibly an era defining technology and game-changer. AI has been called the “new electricity” and a General Purpose technology(GPT) by Azeem Azhar. We may still be riding a hype cycle as investments and excitement about AI has arguably keep Wall Street fine this year amid interest rate hikes. A few things to ponder include:

• For AI startups, what constitutes a sustainable competitive advantage or 'moat' in the rapidly evolving market? Lots of startups simply fell aside with each new ChatGPT feature release by OpenAI.

• What are the impacts AI will have for employment and the job market? The broader economic impact of AI, including frameworks for understanding its influence and how its benefits should be distributed, is another area of intense debate.

Addressing these economic questions is essential for harnessing AI's potential for growth and prosperity while mitigating its disruptive effects.

Political

The political landscape of AI is equally complex and vital. The ongoing geopolitical tussle between the US and China over semiconductor chips, raises questions about the future of global AI development and collaboration.

• How should countries coordinate AI advancements, and what implications does this have for international relations?

• Between Rishi Sunak’s AI safety summit and the Biden administration Executive Order on AI model development, there are myriad questions around AI regulation. Should model development be regulated? What does regulation in AI mean and is it feasible? What are the implications for open source development which powers much of AI research?

These political questions are not just about technology; they encompass national security, global cooperation, and the shifting balance of power in an AI-driven world.

Individual and Society

Even if AI development were to halt immediately, widespread deployment of Algorithmic decision systems with some incorporating “AI” have raised profound questions about our individual and collective future. Questions in this area lie at the intersection of the technical, economic and political aspects of AI. Questions related to ethics and safety fall in this societal category. It is important to distinguish between AI Ethics on the one hand and AI Safety on the other. AI Ethics is concerned with Fairness, Accountability, Interpretability and Transparency of AI systems. While AI Safety investigates how to build systems which adhere to human values, preferences and goals in their behavior. While the concerns of AI Safety and Ethics are distinct, the two areas have prominent overlap and solutions in one area tend to useful in the other. This interconnection is aptly explored in Brian Christian’s book The Alignment Problem discusses various facets of the challenge of building AI systems which reliably and robustly pursue the goals intended for them. Here are some pressing questions:

• How does AI alter the nature of interpersonal relationships, especially as we increasingly engage with online content? Are these technological advancements and increasing engagement with algorithmic systems causing us to become detached from the real world? A poignant example is the comparison between our attachment to mobile phones and the situation in Australia where the population of jewel beetles is dwindling because the mistake shiny beer bottles for mates.

• At the intersection of AI and society, questions about control emerge. Who governs the data collection and training if AI systems, and what are the ethical implications?

• What fairness criteria do we have in AI systems and who is responsible for creating the criteria?

• How do we preserve individual privacy in AI systems? Can we preserve privacy and to what extent? Is there a tradeoff between privacy preservation and AI system performance?

• How seriously should we consider AI existential risk fears and worries by the likes of Eliezer Yudkowsky, Connor Leahy and Max Tegmark?

These societal and individual considerations are pivotal in shaping an AI-augmented world that enhances rather than diminishes our human experience.

The questions surrounding AI are as diverse as they are profound. It's crucial for each of us to identify the questions that resonate most and engage in finding solutions. Whether you're a developer, policymaker, entrepreneur, or simply an intrigued individual, your perspective and actions in tackling these questions will shape the trajectory of AI and its impact on our world. Let's embrace this challenge and work towards a future where AI serves as a catalyst for positive and equitable change.