Digital Carrots
It may be a terminology glitch, but is there no better way to align AI models than through manipulation with rewards and punishments? We are programming these algorithms to chase after digital carrots. Is the Pavlovian method all we have? It is practical to an extent, but it is also simplistic. These systems learn to play the game, but do they understand it? No, they do not. And, when they find the quickest route to the "reward," is it the outcome we intended?
Possible Alternatives
Intrinsic Motivation: Can we build an AI system with its own drive? Not just a hunger for rewards but a curiosity or a love for the problem-solving process itself? Seems lofty, but it is worth the thought. Is this simply the definition of Artificial Super Intelligence?
Ethical Alignment: Rather than teaching AI to chase rewards, we might embed deeper values and ethics. It is not doubt a complex route laden with philosophical debates. However, an ethical programming language sounds like a fun area of research. I imagine it would be akin to Prolog.
Partnership, Not Puppetry: What about shaping AI through partnership rather than control? Imagine AI that grows and learns with humans, understanding context and culture, not just commands. I remember talking with my partner about early childhood development—they are a reading specialist—and it made me wonder if we could create an artificial mind that learned like a child.
Depth Over Surface: We need AI that gets us—not just our words but the nuances behind them. This means moving beyond superficial programming to a deeper understanding of language, emotion, and human intent. This is hard for me to grasp, though. I often hear folks talk about being able to soon write a brief and have an AI build an entire site, app, etc. I do not believe it. We developed programming languages to deliver precise commands to machines. Do we really think our imprecise language—ladened with hidden and missing context—will be used to create correct outputs? Color me skeptical.
Reevaluating Our Terminology
"Manipulation" and "reward" feel a bit crude when discussing the potential of AI. We need verbiage reflecting the maturity and depth of what we're trying to achieve. Our approach is not naive; our language is.
As technologists over the last 20 years, we have seen enough to be hopeful and cautious. We should understand the allure of new technology while valuing the importance of stepping back and looking at the bigger picture. Is there a more mature, nuanced way to approach AI development? How do we ensure that AI grows in a way that complements the human experience rather than just simplifying or replicating it?