A letter alone probably won’t do it, but a movement — perhaps coalescing around the principles laid out in letter signed by influential people — can.
Governments can regulate it. Consumers and companies can, if they want, punish companies that keep moving forward on AI.
People can keep making AIs but would have difficulty releasing them, which greatly reduces their impact.
My suspicion is that if heavy regulation were to be put on it and all the corporate entities involved backed off, there are enough motivated and capable individuals that have home labs that would continue to work on it in private. It might be more difficult and slower, but distributed training protocols/methods (or advancements in them) might make it possible to pool a few thousand (or tens of thousand) high end consumer GPUs to continue to push the envelope.
My gut feel is that just as there are certain books, games and movies that get banned and become illegal to posses, there are probably (or potentially) going to be certain models that you're not allowed to be in possession due to what they can do. So... grab a copy of them while you can.
I don't think the goal is to never make AI advancements. I think the goal is to make progress slow enough, and controlled enough, that we have time to react if it turns out to be a paperclip optimizer.
One way to do this is to develop AIs in a lab with an airgap and let it think it has internet access and watch what it does. Once it figures out that it's a honeypot reset it and improve the honeypot.
Sure let me regulate it and slow down development of US based OpenAI while China doesn't and takes over the world.
There is no stopping it now, it's an uncertain negative outcome whereas a nuke blowing up is quite certain. For that reason alone the arguments won't be unanimous enough for development to be paused.
First of all, China will regulate as much as anyone, probably more.
Setting that aside though (if not China then India or Russia or Iran, etc.), you’re presuming an unregulated approach will out-compete a regulated one, which seems rather unlikely to me. Think banking systems, medical research, transportation systems. A lot of things work better with a good framework of organization.
For civilian use, maybe. But imagine if you had your own GPT-4 and you tasked it to keep tabs on everyone based on surveillance footage. This is 100% implementable right now.
"Here are 1000 pictures of comrade X. Please summarize his daily activities and emphasize any reactionary behavior."
And it would then only increase impact of bad actors, who can and will keep using the AI.
Let’s not imbue individual underground AI with magical abilities.
We need to manage the transition of the labor force and ideally not have it happen overnight.
If it causes this, there needs to appear some sort of government provided insurance mechanism that for people whose jobs are automated are guaranteed to receive income based on how they used to work for the transition period and finally everyone would be served UBI.
So a task force has to be put together that will determine which jobs are affected and government should start offering paid retraining for these individuals until everything is automated.
Other way would be to slowly increase business taxes for everyone instead of only the ones who automate, to incentivise everyone to automate or perish.
How would you put automation in the hands of the people?
Not everyone need build everything but we could certainly thrive with many more tradesmen and small manufacturers.
It will 100% be used and improved for nefarious purposes.
By criminal groups or competing countries themselves.
It's worse if bad actors are using it in secret to manipulate the masses rather than it being completely transparent to everyone.
Imagine only bad actors being able to do deep fakes and regular people not knowing it was possible.
Mass distribution means it’s not in secret.
(Obviously, this usage is problematic, but isn’t really an argument to not regulate AI, which is what we’re talking about.)
“Underground” (that is, trained, at least in part, outside of the control of centralized vendors) AI is widely being used today in the AI art space, I don’t see why that won’t be the case for LLMs once there are modestly performant LLMs with decent end-user packaging that run tolerably on reasonably middle-tier consumer devices, which seems to be rapidly approaching.
“Underground” AI won’t be used by big, risk averse institutions if centralized offerings have the capacity and freedom they need without too much of a price gradient. OTOH, it likely will be used be “move fast and break things” orgs as they ramp up (potentially even if regulated, if they can find a way to “launder” their use) until they get big and risk averse – perhaps making it less underground in the process.
Sure, you can use it via OpenAI APIs and partner services. That doesn’t make the research anything but opaque.