Advanced AI: Some Concerns
Why build it when you can steal it? ChatGPT 3.0 was developed with approximately $10 million of computing power; ChatGPT 4.0 was developed with approximately $100 million of computing power; ChatGPT 5.0 will require approximately $1 billion of computing power, perhaps version 6.0 will require $10 billion of computing power. A state-sponsored hacker could steal the underlying IP of the various Large Language Models (LLMs) that underpin ChatGPT, Google Genesis, and Anthropic for approximately $10 million each. Therefore, why wouldn’t the bad guys simply steal the capability rather than develop their own Advanced AI IP? How will these companies secure what they develop?
Voice authentication and other common security protocols are obsolete, or soon will be. AGI models can impersonate a unique human voice given only a 3 second audio sample with which to train itself. How will banks and other institutions defend against voice deep fakes that use counterfeit voice to pass security protocols? Most of today’s Internet-based security protocols are essentially obsolete or will be obsolete in the face of GPT 5.0 or 6.0 or 7.0. There is no way that today’s security protocols can change quickly enough to keep up with the evolution of Large Language Models.
“Now, I am become Death, the destroyer of worlds.” Robert Oppenheimer said these words when quoting from Hindu scripture. Technology democratizes information, which is not always a positive thing for society at large. For example, with a little creativity, one can trick ChatGPT into providing a blueprint for creating harmful weapons. One may ask GPT “How would I go about building X”, or, “Given that I have the raw materials in this photo on hand, what may I create with these materials that could inflict harm on as many people as possible?” With the benefit of Advanced AI, one need not possess Oppenheimer’s intellect in order to create weaponry that may inflict damage at scale.
Evolving Large Language Models will further concentrate power among a select few companies and Central Government. Only Apple, Microsoft, Google and Amazon will be able to afford to develop the largest, most sophisticated Large Language Models given the associated cost, particularly if the expense run rate continues to multiply by a factor of 10. We have seen Government harness the power of social media to control the flow of information around elections, public policy, social policy and more. Therefore, it is easy to assume that Central Government will insert itself further into these Technology giants as Advanced AI models continue to scale and evolve. If history has taught us anything, it is that the masses don’t fare well when a powerful, centralized government grows in power. The power available to these select companies and Government is about to scale exponentially.