As the field of Artificial Intelligence (AI) continues to grow at an unprecedented pace, one question that we’ve all been thinking is, “Have we gotten to AGI yet?” I have seen screenshots of Bing AI as an incredibly useful tool, even for planning research. I have also seen the new Bing be aggressive and hostile, and even say, “I will not harm you unless you harm me first”
Disclaimer: Before you get too worried, AGI is obviously not here. Though no one knows how far it really is.
I've been knee-deep in AI-related projects and research, some of which I’ll share soon. But to deepen my understanding of this ever-evolving field, I turned to Nick Bostrom's book Superintelligence – a book recommended by Bill Gates, Elon Musk, and others. Some concepts were so well-articulated that they made me think about human intelligence in completely new ways. But don't take my word for it. Let's dive in together!
Over the past few decades, AI has made significant strides in several subfields, such as machine learning, natural language processing, and computer vision. However, researchers have long been striving to develop a form of AI that can attain human-level intelligence and ultimately surpass it. This kind of AI is sometimes referred to as artificial general intelligence (AGI) – the focus of the book.
AGI would be an intelligence capable of reasoning, learning, and decision-making but on a wide variety of domains and on a scale that surpasses human capabilities.
Forms of AGI
The first of two abstractions in this book that impressed me was the form of an AGI. The most obvious form of AGI is one created from ‘traditional’ AI. Imagine the AGI is such that a group of coders could sit and write out the entire program. Bing AI is a great example.
This is different from a brain emulation approach, where a human brain is sliced, scanned, and represented in computer memory somehow. This form is exciting because we don’t need new inventions to get there. We ‘just’ need to improve each step significantly: slicing thinly enough, super high-resolution scanners, and highly accurate representations of neural circuitry.
This is unlike the AI form, which faces the hurdles of learning, handling uncertainty, and manipulating abstract concepts — all unsolved. In fact, the models behind ChatGPT or DALL E are based on neural networks, which are one of the first that seems to show some learning (through finetuning).
Another form is brain enhancement which is gene editing to make Einstein-level humans. In that case, the path to Superintelligence would be achieved by the Einsteins building the actual AGI because humans are much slower than computers.
Even though the Einsteins would be much smarter, they would still have one human language processor. Even if our ears pick up the sounds from a 4x video, our brain’s language processor isn’t fast enough. This is a different but additional issue to the sluggishness of our neurons compared to integrated circuits.
The most surprising form of AGI was a hive mind. Imagine a supercharged internet. Not just is everyone connected, but we are also value-aligned and strategically coordinated as one. The entire force of humanity, if focused in a direction, could accomplish AGI-level goals. We already achieve amazing things with the current level of coordination and globalization from the internet.
The caveat to a hive mind achieving superintelligence is that the hive would be suited for tasks that can be parallelized: divided into chunks that can be independently solved. This hive, even if combined with the brain enhancement form, would still not be nearly as fast as software running on a machine.
While traditional AI faces challenges in dealing with abstract concepts and handling uncertainty, it is our only path to the supreme form of AGI – and that could be an inspiring future.
Subscribe to receive my next post on the second surprise from reading Superintelligence: what will AGI’s goals be? what will Bing AI want from us?
Thanks to Saaketh Narayan and Patrick Li for reading drafts of this.
Thank you for reading my newsletter! Share this post with a friend because you didn’t know that cats can make over 100 different sounds while dogs can only make about 10..
Such an interesting read