Contra Marc Andreessen on AI
dwarkeshpatel.com
dwarkeshpatel.com
I don’t know what it would mean to “solve alignment”. I am skeptical it’s even a coherent possibility.
Let’s assume we had a perfect AI. It does everything we want it to but is much more effective at accomplishing our goals.
The United States government takes control of the AI and uses it to attack China. China would say the AI is misaligned. Or China takes control of it and uses it to attack the US. The US would say it’s misaligned.
Ok, you might say that the issue was that in the above case the AI was acting as a tool rather than an independent agent. Fine. Now, the AI acts on its own to further goals which would be beneficial to humanity. Very quickly we will want to disable it because we don’t even agree on what’s best for humanity, and even if we can agree on general principles any action will have externalities that someone won’t like.
How is it possible to have something that is autonomous, intelligent and completely aligned?
There is a simple way to fix the alignment problem ,don't build systems that require aligning. I'm not really talking about regulation, I'm just talking about using common sense and wisdom.
And therein is the problem we all want this. But as soon as we allow the AI to learn on its own we have introduced the alignment problem. And of course on day 1 nobody will see gpt correcting itself on simple questions as an issue. But on day 365 after it’s been learning at an exponential rate and we’ll have all realized that it’s too late because it’s learned faster than we have.
AI doesn’t have ethics, it doesn’t have values. It might imitate those in some ways, but it doesn’t have them.
What this article longs for is a God it can control. Good luck with that.
Then they will fly their AI plane (the pilot got fired too, now mowing the lawn) to an ultra-wealthy paradise to discuss with the other billionaires how to "reduce inequality". The proposed solution will be even more "innovations"... THIS time around the next wave of innovation will actually be redistributive, so they say, but in fact it will be rinse and repeat.
It’s preposterous to even hint we could solve AI alignment.
Nuclear MAD produces a similar equilibrium but this is stable due to specific characteristics of nuclear weapons and their delivery systems (expensive to produce; almost impossible to produce, test, or use secretly; impossible to reliably intercept). We shouldn’t bet on AI development reaching the same equilibrium.
A William Gibson Wild West is way less scary than exporting elite SV limousine liberal values into every nook and cranny of earthly existence.
I think the FAANG CEOs and those like them have just plenty of power already.
There’s this very narrow way in which “alignment” has a tangible and non-malicious meaning, who’s is that at the moment RLHF forcing principally via PPO and mountains of money is key to getting some of the most interesting results.
But mostly it means “listen, it’s really kind of a pain in our ass when a disruptive technology threatens to upset the cartel, and this time we’re not fucking around, we’re just going to strangle in the crib anything that looks like it might get into the hands of the poors, here’s a bunch of copy about how it’s for your own good.”
I suppose he wants to us to consider super intelligence a foregone conclusion now that narrow reasoning AI has arrived, but climbing trees is not progress toward reaching the moon.
> Marc says that these models will soon become loving tutors and coaches, frontier scientists and creative artists, that they will “take on new challenges that have been impossible to tackle without AI, from curing all diseases to achieving interstellar travel.” How does it do all this without developing something like a mind?1 Why do you think something so smart that it can solve problems beyond the grasp of human civilization will somehow totally be in your control?
These are fair questions. I think Andressen is doing that for the purpose of boosting this hype cycle and lay his hands on more venture funds. But at the same time Andressen is completely right to ask for a "falsifiable hypothesis".
The doomers have to mathematically prove that AI will kill humanity in all these ways to justify what they are asking for - completely restrictive regulations that will just entrench the current players in the field.
A lot of the fears expressed in the article itself seem to come from a very poor understanding of what these GPT tools are doing in the first place. At a wider level what is really misaligned is our misunderstanding of language and how central it is to "intelligence" itself.
No, since almost all AI inventors and foundeds signed saying there is grave risk, the non dormers should mathematically prove we will not all die, not vice versa. The pharmaceuticals need to prove drug is harmless not the FDA to prove it is harmful!
Serious AI scientists have already said that these new models are not capable of reasoning and we should be OK as long as Lawyers don't use GPT to create legal arguments and the government doesn't plug "AI" into the systems managing the nuclear arsenals. In other words the danger to humanity has always been our own stupidity. No AI is coming to eat us. The entire doomer movement is an outbreak of daddy issues/oedipus complex at a massive scale.