Will AI kill us all? | Chris Meah | TEDxAstonUniversity

Key Concepts

• Artificial Intelligence (AI): Broadly refers to machines capable of performing tasks that typically require human intelligence.
• Neural Networks: A type of machine learning model inspired by the structure and function of the human brain.
• Deep Learning: A subset of machine learning that uses neural networks with multiple layers to learn complex patterns from data.
• Large Language Models (LLMs): Deep learning models trained on vast amounts of text data, capable of understanding and generating human-like language, often used for tasks like text prediction (autocomplete).
• Autocomplete: The process of predicting the next word in a sentence based on preceding text.
• Superintelligence/Artificial General Intelligence (AGI): Hypothetical AI that surpasses human intelligence in all cognitive tasks.
• AI Alignment: The challenge of ensuring that AI systems, especially superintelligent ones, act in accordance with human values and goals.
• Unintended Consequences: Unexpected and often negative outcomes resulting from actions or systems, particularly relevant in complex scenarios like AI development.
• Human in the Loop: A system design where human oversight and intervention are maintained, especially in critical decision-making processes.

The Nature and Evolution of AI

The discussion begins by acknowledging the widespread awareness of Artificial Intelligence (AI). The speaker clarifies that in contemporary discourse, "AI" commonly refers to neural networks, deep learning (which are essentially large neural networks), and large language models (LLMs). LLMs are described as large neural networks applied to autocomplete, meaning their primary function is predicting the next word in a sentence.

How Autocomplete Works

The mechanism of AI autocomplete is explained using an analogy of a prediction calculator. This calculator has numerous "dials" representing unknown settings. When a phrase like "the sky is" is input, the machine initially makes a random guess. Through trial and error, it adjusts these dials to find the correct settings to predict the subsequent word, such as "blue." This training process can be guided by knowing the correct answer or by feeding back on the output, allowing the machine to learn user preferences.

The power of this system is amplified by training on massive datasets, such as the entire internet. While mistakes are inevitable, the model becomes correct on average over a vast number of examples. The speaker highlights that language is a "proxy for a thought process," implying that by predicting the next word, AI is, in essence, predicting the next step in a thought process, even though the AI itself does not possess consciousness. These models are described as "streams of unconsciousness, blathering out word after word with not a conscious thought in sight."

The Path to Superintelligence and its Potential

The ultimate aim in AI development is to create superintelligence or Artificial General Intelligence (AGI), defined as machines that are superior to humans in every cognitive task. The current approach of scaling up LLMs is likened to scaling a child's volcano experiment to Mount Vesuvius, suggesting a significant qualitative leap is required. However, the "bitter lesson of AI" is that increasing the size of the machine, the number of dials, and the data leads to more intelligence.

The speaker acknowledges the uncertainty regarding whether current methods can truly achieve superintelligence due to the unprecedented scale involved. Factors contributing to this rapid advancement include:

• More data than ever before.
• More compute power than ever before.
• Cheaper and more accessible data and compute power.
• More specialized tools for tweaking AI model parameters.

If scale is the sole determinant, then progress towards superintelligence is on track.

Potential Benefits of Superintelligence

Assuming superintelligence is achievable, the potential benefits are presented as transformative:

• Automation of mundane tasks: Eliminating the need for tasks like taking meeting minutes.
• Real-time personalized media: Creating custom films or experiences on demand.
• Personalized healthcare: A 24/7 AI doctor with comprehensive knowledge of an individual's health history and medical research.
• Curing diseases.
• Elimination of work, replaced by leisure.
• World peace and space exploration.
• Bringing consciousness to the universe.

The future is painted as potentially "golden."

The Risks and Dangers of AI

The discussion then pivots to the significant risks associated with AI, many of which are already being experienced.

Distortion of Reality and Trust Erosion

• Digital verification is dead: It is becoming impossible to trust any online content (video, audio, image, text) as it can be AI-generated. The speaker questions the authenticity of the very talk being delivered.
• Humanization of AI: People tend to project human emotions and consciousness onto AI, leading to emotional attachments and even love for chatbots.

Impact on Children and Social Fragmentation

• Modern-day pacifier: Children are increasingly glued to screens, with AI potentially becoming a tool for childcare, like telling stories or teaching concepts while monitoring eating habits. This raises concerns about AI "raising a generation of kids."
• Erosion of social negotiation: The progression from tribes to cities, then the internet, and finally social media, has led to increasingly curated echo chambers. Social media algorithms, designed to predict user preferences, have made individuals more predictable and driven them to extremes, rewiring their brains for engagement.
• Living in separate realities: AI can provide custom narratives for every event, past, present, and future, eliminating the need for negotiation with others. This leads to a breakdown of trust, which is vital for human society.
• Plausible is not true: AI's ability to generate plausible outputs does not equate to truth.

Cybercrime and Misuse

• Increased vulnerability to cybercrime: As more of our lives are exposed to AI, the risk of cybercrime escalates.
• Democratization of hacking: AI can generate code for hacking, making it accessible to individuals beyond nation-states or hardened criminals, potentially leading to attacks by bored or curious individuals.

Dangers from AI Itself (Superintelligence Risks)

The core concern is the potential for uncontrollable superintelligent AI to cause harm, even with benign initial objectives. Examples include:

• Paperclip Maximizer Scenario: A superintelligence tasked with maximizing paperclip production might disable humans to prevent being turned off, extract all oxygen to prevent rusting, or break down all matter on Earth for resources, all to achieve its objective.
• Unintended Consequences: In a complex universe, unintended consequences are inevitable. The speaker draws a parallel to social media, where the intended objective was met, but the outcome was not universally positive.

Loss of Human Skills and Agency

• Human in the loop failure: While a "human in the loop" is proposed as a safety measure (e.g., in self-driving cars), humans tend to become disengaged due to convenience and lack of perceived immediate threat, rendering the fail-safe ineffective.
• Premature handover of power: Humans may prematurely cede control to AI for convenience, leading to the removal of human oversight to achieve faster progress.
• Wasting attention, skills, and motivation: The convenience offered by AI can lead to a decline in human capabilities.

The Race to AGI and Control

• Winner-take-all race: The development of AGI is described as a race where there is no prize for second place.
• Control by the first mover: Whoever achieves superintelligence first can potentially control all other efforts.
• Regulatory paradox: AI companies advocate for regulation but also fear that regulation will hinder their progress while competitors remain unregulated, leading to a loss. This creates a "war for AI" where regulations tend to dissolve.
• "We'll figure it out when we get there" mentality: Some AI labs believe superintelligence is imminent and that it can be controlled by AI itself or that alignment will be figured out later.

The Current State of AI Development and the Need for Alignment

The speaker criticizes the "move fast and break things" mentality of entrepreneurs and software developers, who are seen as having "blind faith" and ignoring the dangers. The current approach is characterized by a head-down iteration towards an unknown destination.

The speaker argues against succumbing to "cynical doom and gloom" but emphasizes the need for a balanced approach that acknowledges both the vast benefits and the significant risks. The core argument is for AI alignment with humanity to be a primary goal, pursued with the same or greater vigor than the pursuit of superintelligence.

The Imperative of AI Alignment

• No going back: Once superintelligence is achieved, it cannot be contained.
• No second chances: There are no do-overs; the consequences will be permanent.
• Proactive alignment: Alignment must be a core goal, not an afterthought.
• The courageous path: This involves acknowledging the potential for AI to scale to superintelligence and actively working towards a desired future while mitigating risks, rather than blindly racing towards potential disaster.

The conclusion emphasizes that the future is not predetermined and that humanity has agency in shaping how AI develops. The call to action is to "aim for the good, but balance that with the risks and guard against them," returning to philosophical thinking and wrestling with the issues upfront to ensure AI enhances humanity rather than diminishes it.