Savoring the AI Meal Without Being Eaten: A Rejoinder to Marc Andreessen
Last week, Marc Andreessen published a blog post titled “Why AI Will Save the World”. He begins by enumerating the ways AI will improve all…
Last week, Marc Andreessen published a blog post titled “Why AI Will Save the World”. He begins by enumerating the ways AI will improve all aspects of life. I 100% agree with everything here. I can’t wait until every child has an “infinitely patient, infinitely compassionate, infinitely knowledgeable, infinitely helpful” AI tutor, as Marc puts it. Similar transformations in healthcare, science, and various other work and life domains are equally promising, and at Point Nine, we’re incredibly excited to be investing in startups across these fields.
However, what followed in Marc’s post was unexpected, especially coming from someone of his standing and intellectual caliber. He dismisses AI safety risks and categorizes everyone voicing concerns about AI’s potential dangers as either unscientific fearmongers or self-interested opportunists.
Besides Eliezer Yudkowsky, the people Marc (presumably) brushes into this corner include figures like Nick Bostrom, Stuart Russell, Geoffrey Hinton, Max Tegmark, and the late Stephen Hawking. These great minds have all cogently articulated potential scenarios should AI greatly surpass human intelligence. For those unfamiliar with this discourse, it’s known as the AI alignment problem. This term refers to the challenge of ensuring that an AI’s goals align with our own. If a powerful AI is not perfectly aligned with human values and goals, it could lead to disastrous outcomes. A famous example to illustrate this is an AI programmed with a seemingly harmless goal, like “maximize the production of paperclips.” Without proper safeguards, this instruction, if taken to an extreme, could lead the AI to convert all available matter (including humans and the Earth) into paperclips.
If Marc has a compelling counterargument to these concerns, I would genuinely LOVE to hear it. However, his article is riddled with ad hominem arguments. Instead of addressing the issue at hand, he resorts to psychoanalyzing those who voice concerns and attempts to undermine their credibility.
The only thing he says about the issue itself is this:
“In short, AI doesn’t want, it doesn’t have goals, it doesn’t want to kill you, because it’s not alive.”
If there was a prize for compressing the largest number of questionable assumptions and illogical conclusions into a short sentence, this one would have a high chance of winning.
I liked Albert Wenger’s response on Twitter:
Embedded content: https://twitter.com/albertwenger/status/1667604435541929990?s=20
Marc Andreessen is, obviously, a legend. I love him for inventing the browser. However, when it comes to AI risk, I believe Marc is missing the mark (pun intended).
More than ten years ago, he wrote the seminal “Software is eating the world” article. He was right, and I’m convinced that what we’ve seen so far was just the appetizer. The main course is yet to be served, and it’s going to be powered by AI. But we must take AI risk seriously. If we don’t, we might end up with a version of “AI is eating the world” that’s a bit too literal.