Does AI Want to Kill Us?

As much as AI appears to be human, it is not, so why would it want to kill us? Yet, as we all know, a couple of months ago, Bill Gates, Sam Altman and hundreds of other tech leaders signed a letter to mitigate the risk of extinction from AI. They attributed this risk to be along the same lines as that of a nuclear war or a pandemic. Other articles cited that AI researchers attribute this probability to 10%.?

The bottom line is a system optimizing a function of n variables, will often set the remaining unconstrained variables to extreme values. In other words, it might optimize for what we tell it to do, at the expense of other things that we care about. You get exactly what you ask for, not what you want. The term that researchers use for this today is called “Specification Gaming” and they say it is the most important problem in AI today. This can be mitigated by working towards containing AI systems and not connecting these systems to tools that can physically harm humans. This is where regulations around AI start to become increasingly important.

Now that we are in the race to build faster and better AI, in particular Machine Learning tools. Let’s also take a look at how this surge started and where we are headed. The strongest and most impactful applications of Machine Learning is not going to be in doing mundane tasks better, albeit that’s where it's widely being used, but in leapfrogging us into an unforeseen future. For example ML is extremely good at recognizing patterns - that AI scientists are able to verify is in fact correct, but don’t fully understand how it got there. While this is scary, it is also revolutionizing advancements in medical science for example. Because of ML we are now able to predict protein structures for ALL proteins known to science, which is nearly 200 million of them. This means we have solved an impossible problem in biology. Something that would have taken years to do, because of ML is done in a matter of days. Climate change is another area where experts see GenAI to become the basis for a promising solution.

In 2009, when we went from CPU to GPUs, there was a huge surge in the adoption of Machine Learning due to higher computing power becoming available. According to OpenAI, the amount of computing power used in the largest AI models has been doubling every 3 months. This is why AI is able to make more realistic images, pass complicated examinations, and answer complex questions.

This is why AI is so promising but at the same time so scary.

The next trends in machine learning and AI will include:

- Specialized AI hardware beyond GPUs. New hardware like TPUs (Tensor Processing Units) and custom AI chips are being developed to further optimize ML tasks.

- Quantum computers, with their potential to perform complex calculations at unprecedented speeds, could revolutionize AI's capabilities.

- AI democratization with tools and platforms making AI more accessible to non-experts.

- OpenAI's advancements, like GPT-4 (Turbo), are pushing the boundaries of Natural Language Processing (NLP), enabling more sophisticated and nuanced language understanding.

- Increasing focus on ethical AI development and regulatory frameworks.

- Edge AI which will bring AI processing to local devices, reducing reliance on central servers.

These trends indicate a future where AI is faster, more accessible, ethically guided, and integrated into a wider range of applications.

In Sundar Pichai’s (CEO, Google) words AI is going to be more profound than fire or electricity. Will it? How do you envision AI shaping our future?