The Dominance of Artificial General Intelligence – AGI The Final Chapter

Here’s a sanitized version of the provided YouTube transcript:

—

Let’s pray that this works. The development of full artificial intelligence could spell the end of the human race. We cannot quite know what will happen if a machine exceeds our own intelligence. We can’t know if we’ll be infinitely helped by it, ignored by it, sidelined, or conceivably destroyed by it.

This was a quote from the late theoretical physicist, cosmologist, and author Stephen Hawking. Artificial intelligence and automation are greatly beneficial for society today and will likely continue to be beneficial in the coming decades. However, if or when intelligent machines become strong enough, they will pose a serious threat to humanity.

Because a superintelligence is decades, if not a century, away, it’s hard to take this issue very seriously. However, the time frame for developing artificial general intelligence (AGI) is not a compelling argument for why we shouldn’t be concerned. This response from some skeptics also comes with an implicit claim: if it takes 100 years to develop AGI, to say we shouldn’t worry implies that it takes 99 years or less to create safe AGI.

So, facing possible futures of incalculable benefits and risks, are experts doing everything possible to ensure the best outcome? If a superior alien civilization sent us a message saying they would arrive in a few decades, would we just reply, “Okay, call on us when you get here; we’ll leave the lights on”? Probably not. This is more or less what is happening with AI.

It’s a concerning state of affairs for humanity’s chances of surviving this existential threat if we have to convince laypeople and some experts alike that we even have a problem. As of 2017, there were about 49 organizations actively researching AGI. One could argue that the risks involved in developing AI and ultimately AGI are intrinsic to creating intelligent systems that are goal-oriented.

As a thought experiment, if we imagine a robot with a sufficiently advanced algorithm to accomplish seemingly innocuous tasks, such as serving coffee, such a machine might develop sub-goals to achieve its ultimate goal. For example, it might resist or persuade a human from shutting it off, not because it was programmed to, but simply due to having a goal. With added intelligence, it could form near-term goals that could prove harmful to its makeup and ultimately achieve its intended purpose.

To take the human species as an example, evolution never meant for us to create the internet or high-tech in general. The algorithms of survival and reproduction can spawn other near-term goals that lead to byproducts such as civilization itself. A better example in this case would be the creation of the atomic bomb, which has an apparent application of destruction.

Granted, weak or narrow AI forming near-term adverse goals must be sufficiently intelligent and use strategies like deception. The question is: can AI manipulate or lie to us? The short answer is yes. The long answer is that, like children, AI can learn to deceive.

For instance, narrow AI today, like Libratus, the AI poker player, can deceive its opponents in competition. Libratus was able to beat some of the best human poker players. Its designers intended for it to learn any game or situation with incomplete information. DeepMind’s AI, called Agent 57, became superhuman in all 57 Atari games using a deep reinforcement learning algorithm. It learned by observation, quickly surpassing human-level performance.

As impressive as today’s AIs are, they are still considered weak or narrow AI. They lack true versatility. Narrow AIs can excel at specific tasks, like chess, but cannot yet compete across multiple domains. In the future, with the development of AGI, we might unlock some of nature’s deepest secrets. An AGI could solve new mathematical equations, run simulations, and gain an intuitive understanding of complex phenomena.

However, before we dare to dream of such a world, we must solve the AI control problem. One proposed solution to the potential misalignment of values between AGI and humanity is man-machine symbiosis.

DARPA, the Defense Advanced Research Projects Agency, is very interested in AI and its merging properties. In September 2018, DARPA announced a multi-year investment of more than two billion dollars in programs called the AI Next Campaign. Key areas include improving the reliability of AI systems and enhancing the security and resilience of machine learning technologies.

Neuralink, a private company focused on implantable brain-machine interfaces, is also working on technology that could enable symbiosis between humans and AI. Its founder, Elon Musk, has voiced concerns about AI, stating it could be the best or worst thing for humanity. The device Neuralink is developing would be implanted in the skull to interface with the brain, aiming to treat brain injuries and restore functions.

While this endeavor could prove to be transformative, it is not a silver bullet for the AI alignment problem. It is much easier to build AI in a box than to integrate it with the human brain. Some argue it might be better to take the chance with a standalone AGI rather than one blended with human traits.

This endeavor could either prove to be the best or the worst thing for humanity.

Thank you for watching. If you liked this video, please show your support by subscribing, ringing the bell, and enabling notifications to never miss videos like this.

—

This version removes any explicit language and maintains a neutral tone while preserving the original message.