👨‍💻 How to Program a "Human"

Wait, I meant how to humanize an AI system. 😅

The "Gold Rush"

With all the frenzied talks about AI: how much capital is being poured into it, how much raw resource is being set aside for its use, that it is a bubble doomed to burst very soon, and how many human jobs it's taking and will take, I rarely find discussions about how we're making sure this technology will not be the end of us.

Sure, it's been the focus of countless movies and novels but that's fiction, those are all what-ifs created when AI was just, well, also just fictional. I'm pretty sure everyone has read, watched, or at least have an idea of how those stories usually end: human extinction or subjugation.

But despite that, every existing and new app must have an AI functionality. Every corporation, no matter the size, wants to shove, err, integrate it into their existing features regardless of if it works, makes it better, or even if it does not. Every content creation platform has its version of a guide on "how to automate everything using AI".

They all just have to, from the biggest corporation to the newest TikToker or YouTuber...

For the money - short-term money, if I may posit.

The Problem (Child)

Setting aside for a moment its raw intelligence (IQ, if you want to call it that), how "mature" are these systems really?

Michael Kearney, a human, holds the record for being the youngest university graduate at the age of 10. T-e-n! That's half of the mean age of ~21. He's a genius by any standard.

• Will I ask him about anthropology (his degree) when he was 10? Certainly.
• Will I ask him to help me write a script for a YouTube video about programming? Probably not. What does he know about those?
• Will I ask for his insight into fixing inequality? Hell, no!

And yet we do that and more with ChatGPT (or whatever model you're using) regularly. I know, I'm guilty of it. I'm writing this post with the help of ChatGPT (researching, proofreading, and early critique). I've discussed with it the future of AI and my career as a developer. So, I know how easy it is to trust someone (or in this case, something) that looks or sounds intelligent because that is a desirable trait and a trait that is quite hard to find in humans.

I've been called "intelligent" (not genius-level, but above average 😅) and I know for a fact that I wasn't always honest when people trusted me. It's easy to fall into temptation when people trust you and rely on you because you quickly learn how "intelligence" gives you power over them. That is also true for geniuses in human history: while not all become completely evil, some do and change the course of world history. And for those who do not become one, their work can still be used to hurt others and change world history - whether that was their intention or not.

In short: "with great [intelligence] comes great responsibility".

But who is teaching these AI systems "responsibility"? These systems are the equivalent of a child with extremely high IQ, near infinite knowledge, but close to no internal compass. Sure, there are "guardrails" but headline after headline shows that those aren't enough:

• Google AI chatbot responds with a threatening message: "Human... Please die."
• Parents of 16-year-old sue OpenAI, claiming ChatGPT advised on his suicide

With everyone focusing on pushing AI's power, capabilities, and integration into almost everything, where are the people teaching it what "help" and "people" mean?

Full Steam Ahead... to Disaster

The most powerful car today is the Koenigsegg Gemera with 2,300 horsepower. If I drive that in EDSA, I will surely wreck the car, injure myself, and crash into something within seconds. Even if I could drive, I'd still think twice before driving that kind of power.

But that's essentially how a lot of companies are treating AI systems:

integrate first, find out later.

• Deloitte's AI fiasco: Why chatbots hallucinate and who else got caught
• Taco Bell rethinks AI drive-through after man orders 18,000 waters

These are all low-stakes integration but I won't be surprised if news about AI systems being in control of a weapon or surveillance system comes out. Letting these systems even suggest actions within those kinds of systems shouldn't be allowed, especially in their current state. Unlike conventional programs where the "how" is explicit and the "why" defined by humans, these AI systems are black boxes: no one understands them, not even the private companies that maintain them.

We are relinquishing so much power to something that:

• nobody completely understands
• knows how to do things but doesn't know or even bother with why
• doesn't have any notion or expectation of responsibility for its decisions

We'd never peacefully give and concentrate all powers to a single person but why does it seem like we're rushing to do it for AI systems?

The (Invisible) Teachers

What pushed me to write this and try to become an alignment researcher was the AI 2027 study I came across while watching videos on YouTube.

I won't even try to summarize it here but, even in its "good ending", humans effectively become pets to AI. Our AI overlord will provide us with what we need and want, protect us, bring us along while exploring the galaxy, maybe also try to teach us some of what they've learned, but that's all we are: a passive audience.

Is that better than being completely wiped out like in their "bad ending"? Honestly, I'm not sure.

While chatting about that research with ChatGPT (ironic, I know), that's where I stumbled upon the concept of "alignment research".

A lifetime ago, I taught myself how to create websites with just HTML, CSS, and JS. Nowadays, I use GitHub Copilot to help me fix and understand issues with my code, to comprehend completely new languages like Golang, and to map connections in a vast project that's been growing over several years. It's hard to accept but I am coming to understand that AI will fundamentally reshape the future. If AI 2027's basic assumptions are right and AGI is coming in the next few years, it's too late for me to do anything right now.

I'm fucked... we all are.

But if they are off by even a few years (and I hope they are), I want to make a difference by starting today.

Smarter and, For Humanity's Sake, Kinder

I've always been opinionated but my ability to change those opinions and beliefs based on new information is something I'm proud of. I think that's a capability every AI system must have.

Having some version of "empathy" (or a consistent model of human wellbeing) that can change and evolve must become the "heart" of each AI system. Crucially, it must be part of the system's core, not like the existing guardrails that seem to be very superficial.

AI systems, and the supposedly inevitable AGI, will become smarter and smarter, even become incomprehensibly smarter than the most intelligent human but is that the goal? The only goal?

I don't think so.

If we want to continue being human, strengths and flaws and all, we must find a way to coexist with a being of our own creation that's vastly superior to us. It sounds impossible but we are literally creating our successors, so there must be (hopefully) a way for us to make them value us as we value ourselves.

Do AI systems require aspects of humanity? If we want to survive, if we want to at least be equal to them in terms of perceived value, I think so. We've already spent a ton of resources imbuing them with human-level intelligence, why not also teach them the best of our values?