Destroying Terrible Anti-AI Arguments: Day Six "ASI will kill us all" This one is similar to "my p(doom) is x%." It's a blind assertion. It has no validity outside the mind of the person saying it no matter how many times they repeat it. The reason is simple. We have no idea what ASI will be like and ascribing characteristics to it is an exercise in futility. Pausing AI to create solutions for a non-existent thing will be of zero value when (and if) ASI actually gets here. Here's a few illustrations to show you why: Imagine going back in time and asking da Vinci what are some of the characteristics of a modern aircraft in 2024, like the Boeing 787 Dreamliner? Ask him about how we can make it safer? Ask him about the general challenges of building such an aircraft and how we can improve the supply chain? Ask him about our safety protocols, like incentivizing pilots to report near accidents within two weeks and if they do that faithfully it can't be used against them in a court of law? Ask if that makes a big difference in airline safety and how? His answers on any of these questions will be utterly meaningless. That's because he can't conceive all the myriad economic developments (division of labor, specialization, free markets), supply chain developments (global interconnectedness, mega-boats, shipping containers), societal developments (modern democracy), and stacks of compounding technological development (lighter materials, engines, computers, electricity, wires, interchangeable parts, etc) which all weave together to make the modern airplane. These other inventions and technologies and cultural developments all influence, mitigate and create friction on the other technology. It does not exist on its own. Many folks will tell you that long term prediction is easy and they are dead wrong. Humans are absolutely terrible at this, even people who make a living doing it. We can sometimes predict near term trends pretty well but long term trends are beyond our event horizon. What almost inevitably happens is people project forward into the future all the characteristics of the present day and add one new invention. That is not how things develop in reality. Society, technologies, culture and people evolve in massive parallel at a level that is mindbogglingly complex. As we move forward in time we have a tremendous confluence of new inventions, political strategies, legal frameworks and more, all at the same time, all in parallel. It's everywhere, everything, all at once. A great example of this is the 1968 book, The Population Bomb, which predicted that we'd need to let 2 billion people starve because there just weren't enough resources to go around. Instead we got the Green Revolution and can sustain more folks than ever. In a classical example of terrible prediction error 101, Paul R Ehrlich, projected forward all the characteristics of the 1960's into the future and failed to predict any other possible societal, biological, or technological developments. Try explaining the Internet to an 18th century farmer. You can't do it. That's because it's built on the back of chain of inventions and societal developments like electricity, turbines, wires, information theory, the computer, browsers, digital computers, etc. If Da Vinci projects forward his flying machine to 2024, but imagines it with the characteristics of Renaissance Italy in 1500's, his predictions about life in 2024 are less than worthless. They have zero chance of being accurate. Even if he is more creative in his thinking, projecting forward hundreds of years is impossible for anyone because they have to come up with the distributed, collective inventions of the entire world of people working collectively over time, building on the back of the people who came before them. Even near term predictions are tricky. A great example comes from one of my favorite sci-fi books, Neuromancer, by William Gibson. Gibson came up with the idea of cyberspace before the Internet really came out of universities and the military, in the early 1980s. That's cool but how he saw it developing had nothing to do with how it actually developed. He saw all data as a kind of 3D construct because it was too difficult for any one person to perceive. Actually we can perceive data just fine and we can interpret it directly in databases and spreadsheets and documents. He also saw the Internet as something you go to, not something that is around you all the time on the laptop and phone and watch through wireless. In fact he failed to come up with wireless at all. One of the funniest moments in the book is hacker laying a fiber optic cable over five miles to connect to a modem to connect to cyberspace. Another hilarious moment is when a hacker calls up on landline phone (because he also failed to predict portable phones) to get the time, like people did in the 1960s. Despite all his prodigious prediction, he could not even conceive of the time being always available everywhere or the internet floating through the air. What we can do though is perceive patterns and abstractions of the past and the future. The pattern of the past is that we have always come up with a massive amount of parallel developments and inventions and mitigations for technology, all acting on, supporting and limiting each other. We are on solid ground when we predict in terms of patterns/abstractions. I don't know what those mitigations to ASI will be or what ASI will be like but I can predict that parallel inventions that act on AI will develop at the same time and as needed because that is how it worked 100% of the time in the past. There is no reason to imagine this is some kind of exception, no matter how many times X-riskers tell you "this time is different" (famous last words, used thousands of times in history to describe pretty much every technology that has ever existed and has been wrong precisely 100% of the time). We don't know what mitigations, parallel developments, societal/cultural developments and technologies will accompany ASI and how it will help keep it on the rails but we know they will come. Already we have people working on actual alignment solutions. We have RLHF and we have RLAIF, aka Reinforcement Learning with AI Feedback. Researchers are developing solutions. Others will follow. Regular engineers are now getting their hands on ML models and bringing their unique perspective to the problem as well. They see the problem differently than ML engineers and research scientists and bring a fresh perspective. As they work with open source models they're working on guardrails that come from their knowledge of traditional hand coded logic. Why? Because they want the AI to behave and give them consistently good results! They want it to do what they want so they can build apps. They don't want it to decide that it doesn't feel like clicking on the button to buy a ticket. We are incentivized to make these systems more reliable as a species because we want them to work. All these developments will inevitably lead to new solutions and ideas that are practical in the real world of AI, not the AI of fiction. While many "AI researchers" are writing philosophical treatise on why AI can never be our friend and warning about the end of the world like they're Jim Jones reincarnated, actual researchers are hard at work on the problem. Problems are solved in the real world through engineering, as they develop in time. We invent the seatbelt and the airbag and crumple zones for the car, after we realize that it can go too fast. We can't conceive of that problem before it happens. That may seem strange to you but you are looking at it after we have traveled faster than we could run, which was not the case for people before trains and cars existed. They never traveled faster than a horse could carry them. People before the train and the car had heated debates about what would happen to the human body if we went 10 miles an hour. Would we spontaneously explode? Would we die of a heart attack? Would we pass out? In short, their predictions were dead wrong. These all sound ridiculous now but that is because we know how it turned out after it happened in the real world. We have actual knowledge to build on. Once we get that, we build actual solutions. That's why calls to "pause AI" are deeply misguided. It is a basic misunderstanding or real world cause and effect. These folks do not understand that you don't invent solutions in a vacuum. We can't invent solutions to imaginary problems. We invent solutions by interacting with the problem in reality. Problem happens, you think about it, you work out a solution. Or rather, problem happens, lots of people work on it, trying many different solutions until we get one that works and then we deploy that solution. It's an iterative process that happens after or in parallel to the development of the original technology. The safety razor comes after the razor. Crumple zones come after the car. Don't worry. The engineers will take it from here and do what we always do when it comes to real world problems. We'll solve them.