This post attempts to set out my views on why I don't find technical alignment to be a pressing issue, compared to more mundane AI risks like human misuse. It represents my personal views only, as someone who is not in AI research. I would appreciate any and all feedback.

Summary of views

A profound economic transformation through AI is a lot more likely to happen through various implementations of narrow AI, rather than artificial general intelligence (AGI).
An alternative way to frame this view is that if both are possible, then narrow AI transformation happens earlier in the timeline, with AGI in the distant long-term. In this scenario, risks like robustness, bias, and human safety problems are more likely and immediate sources of AI x-risk compared to AI alignment, and should accordingly receive more focus.
I believe development of AGI will take longer than the estimates of most people in this community. It is common to conflate cognition and execution in a lot of the AGI arguments that I’ve read, which I don’t think is an obvious conclusion. I think that developing an AI that can complete any action at least as well as a human is a hard problem that will require a breakthrough in the way we architect AIs.
Cutting-edge AI research that advances AI capabilities (not simply implement existing techniques) is the domain of wealthy organizations and governments, who choose to allocate time and resources to this problem as they find it useful to achieving their goals. If developing general AI is hard, then it would be much more time- and cost-effective to develop multiple ‘narrow’ AI with a bounded, defined set of actions that is still capable of achieving those goals.
Consequently, the default and most likely scenario for AI x-risk is one where AI has a bounded set of actions, but is still capable of leading to x-risk through intentional misuse or unintended consequences of deploying this AI.

1. There is no prototype for an AI that can execute general human tasks.

From the various arguments I’ve read which discuss the AI alignment problem, I often see an implicit assumption that a general AI that has an intelligent and accurate view of the world, enough to generate plans and predictions, would naturally be capable of executing those plans through an unbounded set of actions.

That doesn’t strike me as true. Let’s use an advanced ‘generalis...

Posts

Wikitag Contributions

Comments