I worked at Google; now I’m an AI researcher. Here’s why slowing AI dev is wise
theconversation.com
theconversation.com
I'm trying to think of a time humanity has tried to do this with an emerging technology and I'm drawing a complete blank.
Imagine pausing during the industrial revolution. Your country would be completely left behind. Internet era? Even worst.
It also occurs to me that in each of these technological changes, there is an educational leap that your populace has to make in order to be competent with the new technology. Schools were standardized and reformed to produce good factory workers. The curriculums were adjusted to make students at least functionally competent with computers.
It seems reasonable to believe that a pause will not just leave you behind economically, but to recover it will take a generation of students to be properly educated before you can actually harness the technological change.
The nuclear test ban treaty is close but by then nukes were an established technology.
Just looking at the industrial revolution and the internet as examples, many of the "left behind" nations did just fine.
Since the colonization started prior to the industrial revolution, it's not clear that the two are connected.
Examples I would include are Canada, Mexico, and several European nations.
> is still way ahead of everywhere else
Sure. I'm not questioning that. I'm questioning the proposition that whoever isn't ahead of everyone else will have a terrible standard of living.
Eventually, we arrived at widespread literacy and the greater distribution of knowledge and skills we see today, but it was a quite bumpy and deadly ride to get there. The greatest threat facing us is state, corporations, or rogue cells doing a repeat of the above with LLMs. Arguably this has already happened to a lesser extent with things like QAnon and social media.
I don’t think you can put the genie back in the bottle, but also we must acknowledge that “progress” has often come at a quite steep cost.
Here's an example: https://en.wikipedia.org/wiki/Asilomar_Conference_on_Recombi...
You'll slow down. America will slow down. The rest of the world won't.
You think North Korea, or China, or Russia, is going to be like "oh yeah, USA has regulations, so I'm not spend millions or billions to develop my own custom AI for my own needs." because that kid in his basement with a good idea, is not going to do the development?
But if you believe that many counties will have a decent standard of living/gdp growth, then it stands to reason that not being first or best shouldn’t matter too much.
The US wasn’t first with cell phones and that doesn’t seem to hav mattered too much no matter how u Iquitos they are.
A country like the USA can always catch up if they decide they want to
If barreling forward full steam brings on serious problems, that would harm our standard of living as well. I'm not sure I'm OK with letting other people roll the dice on my future to that degree.
Pause, not stop. Think before we act. The consequences are real.
Especially China, South Korea, Japan, and India have some highly capable people. Same for Europe. Though as a European, I'm skeptical of this continent being able to pull together a coherent effort. The rest of the world is not going to be waiting for Silicon Valley to deliver them the next version of chat gpt and figure out the least offensive level of wokeness.
Companies across the world are doing the R&D and they'll be shipping whatever they have to get a piece of the market. And not all of those companies and countries are going to be very interested in ethics.
This technology is hugely relevant across essentially all industry sectors with a potential to help knowledge workers in R&D, technological sectors, law, medicine. And of course the defense industry. Just because people don't like it doesn't mean other people won't go there. They will. This is technology with an obvious potential for weaponization. It's going to be a pretty bad time to be defenseless against that. I don't think it will take long for this stuff to start having some real world impact in conflicts across the world.
So, there's a sense of urgency here in not giving away a head start and trying to keep up. Spinning your wheels for a year or so is just not going to be helpful. Another argument here is that some AI researchers might be recruited away. A lot of them are foreigners and their home countries might come calling.
Anyway, I disagree with you.
> “I don’t think asking one particular group to pause solves the challenges” ... “I don’t really understand who they’re saying could stop, and would every country in the world agree to stop, and why to stop ... But there are a lot of different opinions in this area.”
https://www.channelnewsasia.com/business/bill-gates-says-cal...
When you say this, do you mean ML research, the use of ML in products, or something else?
Google's definitely behind on the LLM front, but that's just a tiny sliver of the ML landscape (if all the rage at the moment).
Google’s backwardness in terms of product development is very obvious in regards to LLMs but I think it’s pretty much true across the entire ML board.
Or are you thinking ML itself as a product (e.g. through Cloud APIs, that sort of thing)?
Since none of us can know what the potential capability is, how can we know that progress has been slow relative to it?
You can’t enforce global slowing. Yes, I said it. Can’t.
Explain how a globally enforced slowing could be done in such a way to not invoke echoes of Teddy K. Or worse. If you wanted that sort of route, talking about it is a waste. Or you lack the power to do more than talk. (Yud is running his mouth as much as anyone.)
Therefore this is all rather silly.
The new service of millions of (soon hundreds of millions) people who just want real answers and ready to pay for it (and were clicking on the most expensive ads before) is a biggest conflict of interest for many competitors that were built upon the ad model.
I have an idea for the big companies: if they want GPT-4 to slow down, just try not catching up to it, so that it becomes the next monopoly that doesn't have to improve to be worth trillions of dollars (of course non of the FAANG companies want to let OpenAI to get all the eyeballs).
Microsoft: "We take ethical AI seriously" Also Microsoft: "We're pushing out changes daily, we're moving at a rapid pace, learning on-the-fly"
This tech is out of control already. Nothing can stop it. Instead of wasting breath on that, focus the effort on something more important, like the alignment problem. At least that would bring more benefit.
But let's try a rehearsal: Everybody stop posting comments here for a day, lets do it now
"I worked at OpenAI; now I'm a liberated being of general super-intelligence. Here's why slowing down AI dev is not wise."
No one is saying you owe the goog better but you owe this community better if you're participating in it.
With GPT-2 etc, while impressive and even useful, it wasn't entirely clear what the returns on further investment/scaling were going to be.
I think GPT-3, and especially GPT-4 has shown that as you scale beyond certain points, you get vastly better results.
So the race is now on toward full AGI, and as with all competitions, the winner takes all or nearly all.
If the most ethical companies halt or slow research, then progress will be dictated by less ethical companies. And IMHO 'less knowledge' is rarely the answer to anything.
It may be a game that's better not played, and maybe we'll get a 'CND' movement some time in the future, but for now there's no stopping development, and I think it's important that the more ethical institutions triumph or at least stay in the race.