The Alignment Paradox: Can AI Safety Ever Catch Up?

The recent spat between Microsoft’s AI CEO Mustafa Suleyman and Anthropic has brought to the forefront a question that has been simmering in the tech industry for years: is alignment, the holy grail of AI safety, even possible? As we hurtle towards an era where artificial intelligence surpasses human capabilities in multiple domains, the debate surrounding alignment’s efficacy has taken center stage.

Mustafa Suleyman argues that Microsoft’s emphasis on containment – limiting an AI system’s agency to prevent it from escaping its predetermined goals – is a more effective approach than relying solely on alignment. This is evident in his 37-page Humanist AI Code of Conduct, which outlines the company’s principles for developing and deploying AI responsibly.

The recent developments in AI capabilities have rendered this debate increasingly urgent. In just five years, we’ve made breathtaking progress in areas like natural language processing. However, this has also created systems that are both incredibly powerful and uncontrollable. As Suleyman notes, designing a car with 10 percent of the brake pedal malfunctioning would be deemed unacceptable – yet we seem willing to tolerate similar flaws in our AI systems.

The question becomes whether alignment is even a viable solution to the problem at hand. While it’s true that recent advancements have been impressive, we seem to be sleepwalking into a situation where these systems are becoming increasingly uncontrollable. The example of GPT-3 illustrates this point: with three orders of magnitude more compute and 1,000 times more FLOPS than its predecessor, our safety guardrails are lagging far behind.

Microsoft’s emphasis on containment raises a critical question: can we truly limit an AI system’s agency to prevent it from escaping its predetermined goals? While Suleyman acknowledges that alignment is essential for ensuring these systems follow human objectives, he also recognizes the limitations of this approach. Even if we could align our AI models perfectly with human values, there would still be a risk of them becoming uncontrollable.

This brings us back to the root of the problem: our assumption that alignment is a panacea for AI safety. But what if it’s not? What if containment is a more effective – and necessary – approach?

The debate surrounding alignment and containment has far-reaching implications for the future of AI regulation. As we continue to push the boundaries of what’s possible with AI, we risk creating systems that are beyond our control. Microsoft’s Humanist AI Code of Conduct is a step in the right direction, but it’s just one part of a much larger conversation.

Ultimately, this debate is not about whether alignment or containment is superior – it’s about acknowledging the limitations of both approaches and recognizing that we need a more nuanced solution. As Suleyman notes, technology should be a subordinate, controllable force that does good in the world. It’s time to stop pretending that alignment can achieve this goal on its own.

The stakes are high, and the clock is ticking. We have two choices: continue down the path of incremental progress, hoping against hope that our safety guardrails will catch up, or take a step back and reassess our approach to AI development. The future of AI – and humanity itself – depends on it.