In this inspiring discussion, scientist Max Tegmark delves into the fast-paced development of artificial intelligence (AI), the potential implications of superintelligent AI, and offers an optimistic vision of managing and using AI in a way that ensures it works in harmony with humanity, rather than in dominance.
Heightening Danger of Superintelligence
The rapid pace of AI development, propelled by notable contributors like OpenAI, Google DeepMind, and Anthropic, is tilting towards the creation of superintelligent AI, potentially exceeding human intellect.
If unchecked, there’s a risk that these advanced systems could hold sway over mankind, posing grave existential threats.
AI Safety: More than Limiting Harm
AI safety protocols should shift from simply curbing harmful outcomes to establishing systems that make adverse incidents impossible.
The creation of unbreachable systems in compliance with physical laws necessitates the development of foolproof and verifiable safeguards.
Harnessing Formal Verification
A promising approach involves harnessing formal verification, a tool for validating codes.
In this framework, a human drafts a mandatory specification for the AI to follow.
Subsequently, a sophisticated AI constructs both the AI tool and a proof of the tool’s adherence to the given specification.
Post learning, the algorithm can be implemented onto a computational structure that is more verifiable.
Simplify Verification, Not Discovery
Human comprehension of AI or its proof is not a prerequisite.
The simplified proof-checking code builds trust, and with the incorporation of proof checkers within computing hardware, running unsafe codes becomes impossible, ensuring robust security.
Let’s not pause AI. Let’s just pause the reckless race to superintelligence. Let’s stop obsessively training ever-larger models that we don’t understand. – Max Tegmark
Alternate Extraction Method
If the AI fails to develop the tool, an alternate method can be leveraged.
One AI is trained to execute a human task while another AI extracts the learned algorithm and knowledge.
The growing field of mechanistic interpretability shares parallels with this concept.
Application on Simple Addition Algorithm
A practical illustration of this process involves a simple addition algorithm.
Here, a recurrent neural network learns the task.
Subsequently, an AI tool converts the acquired algorithm into a Python program, and finally, a formal verification tool confirms the program’s accuracy in adding any given numbers.
AI godfather, Alan Turing, predicted that the default outcome is the machines take control. The machines take control. – Max Tegmark
Pausing the Race, not the Progress
Tegmark emphasises not halting AI development but rather suspending the hazardous rush towards superintelligence.
He advocates for a thorough understanding of AI and focusing on maximizing the benefits of existing AI capabilities, rather than venturing into uncharted territory with superintelligent AI.
