WTF is Superalignment?

'Superalignment', which is apparently making headlines following the announcement from OpenAI on 5th July, regarding building a team of expert ML, AI, and a bunch of other geeky experts to build computer systems, protocols, and AI algorithms to control the Superintelligent AI or in the computer science dictionary, which often called as ASI (Artificial Super Intelligence) to go out of hands and cause a human level extinction event. But before we go into details of what Superalignment is? Let us uncover a bit about ASI.

What is ASI?

So, imagine you've just created a Superintelligence, and, just for kicks, you decide to test it out by asking for a killer guacamole recipe. Not only does the AI whip up a recipe using avocados sourced from an undiscovered, super-organic Martian avocado tree, but it also negotiates world peace, solves climate change, invents a new form of renewable energy, and pens a novel that outshines 'War and Peace'—all while you're still smashing your earthly avocados. Finally, it serves up your recipe along with a sarcastic, "Apologies for the wait. I got sidetracked saving the world. Enjoy your guacamole!" It's like having a super-genius, multi-tasking, wisecracking best friend, who also happens to be able to outthink all of humanity while tying its metaphorical shoelaces. Is it amazing? Absolutely! Terrifying? Just a smidgen. Will the guacamole be good? You bet!

Kidding! That was sarcastic, here is an actual ChatGPT-generated definition BTW,

Superintelligence, in the simplest terms, refers to an artificial (or, theoretically, a natural) intelligence that significantly surpasses the performance of every human being in practically every field, including scientific discovery, general knowledge, practical wisdom, and social skills.

In order for us to understand how powerful it can be, let us try to compare it with ChatGPT, the AI that started the whole ' I am going to destroy Humans' story.

I have tried to give an easy comparison between the current state-of-the-art AI vs what an ASI system can achieve. On a serious note, ASI systems are far beyond just chatbots, they can potentially revolutionize anything and everything that human civilization has ever experienced. Despite the fact that ASI remains theoretical, the seriousness of AI experts and OpenAI towards it indicates that we might be able to see ASI in the coming decade. There is also no doubt in the fact that OpenAI and others might actually be building it behind closed doors.

So, if a tech is so powerful and can actually cause an extinction-level event, it needs to have guardrails and an intent best supporting human civilization. The sensible, ethical, and practical governance of these AI systems is what OpenAI calls'Superalignment'

Super alignment in the context of AI is about making sure that a superintelligent AI - an AI that's much smarter than humans in every way - behaves in the way we want it to. In simple terms, it's like teaching a very powerful and intelligent robot to play nice with us, follow our rules, and help us rather than harm us. This involves not only teaching the AI what to do but also setting up checks and tests to make sure it's still following the rules as it learns and grows. It's a bit like parenting, but for a robot that could potentially be smarter and more powerful than all of humanity.

Why is Superalignment important?

The disadvantage of technology is the ability of it being able to be misused, look at the past Nuclear Energy, both its good and bad impacts have led Governments, Organizations, and Researchers getting extremely cautious about its use cases, no matter how scientifically different Nuclear energy and ASI are, experts have always believed that both can have existential risks to humanity. To put us back in the context of ASI here is why Superalignment is important.

1. Control: If we create a superintelligent AI, we need to ensure that it doesn't pose a risk to humanity. Just like you wouldn't want a powerful machine running without any control mechanisms, you wouldn't want a superintelligent AI operating without guidelines.
2. Safety: A superintelligent AI could potentially cause a lot of harm, whether through malicious actions or just by making mistakes on a large scale. Super alignment is about making sure the AI understands and respects human values to avoid such harm.
3. Beneficial outcomes: We're developing AI because we believe it can benefit humanity. To ensure that a superintelligent AI actually benefits us and doesn't just pursue its own goals, we need to align it with our values and interests.
4. Long-term survival: There's a concern that a superintelligent AI could eventually outcompete or overpower humanity. By aligning the AI with our goals, we hope to avoid this scenario and ensure the long-term survival and prosperity of humanity.
5. Unpredictability: A superintelligent AI might come up with solutions and ideas that humans could never think of. This is exciting, but also potentially dangerous if the AI's goals aren't aligned with ours. So, super alignment is about ensuring that even when the AI is acting in ways we don't understand, it's still working towards goals that are good for us.

What is OpenAI doing about it?

Now obviously OpenAI is both equipped and capable of solving this problem. After going multiple times through their press release from July 5th, I just took the time to summarize what I have learned so far.

• The project's objective is to develop scientific and technical breakthroughs that will allow us to steer and control AI systems much smarter than us.
• OpenAI believes that this is a critical problem that needs to be solved within the next four years.
• To achieve this objective, OpenAI is dedicating 20% of the compute they have secured to date to this effort.
• They are looking for excellent ML researchers and engineers to join them.
• The tasks that OpenAI is working on include:
• Developing new alignment techniques that can be used to ensure that AI systems are aligned with human values.
• Creating new AI safety tools and frameworks that can be used to assess and mitigate the risks of AI.
• Building a community of AI researchers and engineers who are committed to the safe and responsible development of AI.

OpenAI believes that Superalignment is one of the most important challenges facing humanity today. They are committed to working with the global community to ensure that future AI systems are aligned with our values and that we can reap the benefits of this technology without putting our own existence at risk.

The current developments in AI ad looking at potentially all the stakeholders now getting seriously involved right from the Governments to the Big tech might be an indication that the AI revolution is here to stay and no matter how much we see or call it a bubble, the pace at which it is being developed is a clear sign that AI is going to stay, and no matter how impractical ASI is right now, the advancements and seriousness of the Big Tech shows we might actually see this in our lifetime.

Meanwhile, I also got a chance to jot down a list of 5 must-read books if you are interested in ASI and AI in general.

1. "Superintelligence: Paths, Dangers, Strategies" by Nick Bostrom: A comprehensive exploration of the future of artificial intelligence, looking at the paths towards superintelligence, the dangers it poses, and the strategies for dealing with these challenges.
2. "Life 3.0: Being Human in the Age of Artificial Intelligence" by Max Tegmark: This book delves into how the advancement of AI will affect our society, economy, and daily lives, particularly when AI reaches and surpasses human intelligence.
3. "Human Compatible: Artificial Intelligence and the Problem of Control" by Stuart Russell: Russell, a leading AI researcher, offers a new perspective on AI, highlighting the need for a radical rethinking of AI to ensure its alignment with human values.
4. "The Singularity Is Near: When Humans Transcend Biology" by Ray Kurzweil: Kurzweil predicts an era of technological singularity, where AI will surpass human intelligence, bringing about profound changes in our society and biology.
5. "Our Final Invention: Artificial Intelligence and the End of the Human Era" by James Barrat: This book investigates the reality of superintelligent AI and how it could potentially lead to the end of the human era, unless we learn to coexist with machines.