evanevan

I was thinking through all of the different possible AI futures people have forecasted (technology plateaus, job loss, economy automation, AI alignment failure and AI causes causes civilizational harm, ....) and thought it would be interesting to have some kind of choose-your-own-adventure for exploring these possibilities.

I tried to cover the possibilities that show up most persistently in the AI dialogue online, resulting in 28 different outcomes, and ~50 questions (though not all show up in all runs).

Curious what kinds of paths people are thinking are most likely, or if there are different conceptions people have of where the tech could go. There are definitely limitations to the multiple-choice question aspect of the tool, but hopefully helpful for thinking through some of the possible directions the technology could take.

I really like this model for individual services.

The challenge I've always felt, is shared services -- if I'm running infra myself, I can depend upon it, but if someone else is running it, I'm never really sure if I can, which makes external services really hard to rely on and invest into.

Maybe you can get further than expected with individual services? But shared services at some point seem really useful.

I think web2 solved that in an unfortunate way, where you know the corporations operating the services / networks are aligned in some ways but not in others.

But would be great to have shared services that do have better guarantees. Disclaimer, we're working on something in that direction, but really curious what others have seen or thinking in this area.

The important bit (which seems unclear from this article), is the exact relationship between the for-profit and the not for profit?

Before, profits were capped, with remainder going to the non-profit to distribute benefits equally across the world in event of agi / massive economic progress from agi. Which was nice, as at least on paper, a plan for an “exit to humanity”.

This reads to me like the new structure might offer uncapped returns to investors, with a small fraction reserved to benefit the wider public via this nonprofit. So dropping the “exit to humanity”, which seemed like a big part of OpenAI’s original vision.

Early on they did some good research on this too, thinking about the investor model, and its benefits for raising money and having accountability etc in todays world, vs what the right structure could be post SI, and taking that conversation pretty seriously. So it’s sad to see OpenAI seemingly drifting away from some of that body of work.

Thanks - yeah I think CEV hits the mark. Would be awesome if OpenAI (and others) committed to CEV then (with I'd add the qualification, equal weighting on each person).

Would be curious how this would play out for LLMs too. What do people actually want? I would guess most people think OpenAI is being too restrictive on what ChatGPT does, but would be curious to actually see that play out and how people are thinking about it.

Not yet... but one day soon, no? Seems good to be starting to plan

Who should decide seems important. All of this is at the end of the day controlled by OpenAI right?

I wonder is there a way the AGI could be instructed to weight all of our preferences equally?

Seems vital so we end up with something in the interest of all of us and not something we just have to hope the few entities controlling AGI do the right thing, given historical precedent of the risks of entrusting systems that are supposed to benefit all of us to small groups too.

One thing you could do is have the camera hardware sign the raw image, then make a zero knowledge proof that says original photo was signed by real hardware, and then the only edits applied to it were simple things like cropping, changing colors etc

Seems reasonable to coordinate the updates with a blockchain (though not necessary for use cases where you trust the platform / coordinating entity).

Disclaimer I work at a zkSNARK / blockchain company :)

Biggest difference is Coda's constant 20kb vs still at least a few hundred megabytes for Vault. This matters because it helps in making cryptocurrency and cryptocurrency apps usable by people from their phones and browsers, at 20kb with Coda that's actually possible.

I haven't looked particularly into if any consensus / security assumptions are different either with Vault, but that could be another place for them to get down to a few hundred MBs without zero knowledge proofs.