4 ms·

Yeah, it is a true risk. What happens when people start making their voting decisions with AI, asking who to vote? Nothing wrong with that per se, but it gives

by Geee 3y ago

Yeah, it is a true risk. What happens when people start making their voting decisions with AI, asking who to vote? Nothing wrong with that per se, but it gives an ultimate power to the company who developed and hosts the AI, which means the end of democracy. My recommendation for regulation would be to make closed & cloud-hosted AI illegal. Architecture, training data and weights should always be available, and people should only be allowed to use AI on their local machine (which might be as powerful as possible).

a_t48 3y ago
This would make AI usage restricted to those who can afford a beefy PC, no?
- dotnet00 3y ago
  Or it would create an incentive for reducing the cost of the hardware needed to run beefier AI models and passing that reduction down to consumers.
  - jeffparsons 3y ago
    Which would still mean that those who can afford more/better hardware can therefore afford more/better AI. I don't think a solution lies down this path...
    - JieJie 3y ago
      There are already people working on distributed solutions to that problem. There are a lot of people working on these problems.
    - dotnet00 3y ago
      Is that really an issue? Eg it's the same with games, the better your hardware, the more likely you can play a new game at higher settings with better performance, with most cheaper devices outright unable to handle anything recent. But if you wait a GPU generation or two, the performance has improved enough that you can do that with a much cheaper GPU.
- Geee 3y ago
  Costs will come down, while capability will go up. I'm thinking of a some kind of box in every home, with special inference chips; not really a "PC". In principle, every human should have access to the same level of AI. There could be one at the local library if someone can't afford one, or doesn't want one.
  - ethanwillis 3y ago
    Access to local libraries isn't even a thing that everyone has access to.
- samus 3y ago
  That would still be a way larger class of people than the top 1%. Even so, capabilities of consumer devices will rise for some more time.
golem14 3y ago
There’s a short story by Asimov called “Franchise” which is in hindsight quite foreboding…
Joeri 3y ago
Seeing the weights is useless for knowing capabilities, and even knowing training data and architecture you still need to interact with it to figure those out, which you can do with cloud AI just as well. The real risk that I see is the use of open source AI by governments, corporations and criminal groups to influence elections through a deluge of bots. Cloud AI can at least be controlled to some degree because there is a government that can inspect and regulate that business, but an army of bots built on local LLM’s could just run wild over the internet without stopping them. How many comments in this thread are written by a bot? I do not feel confident I could tell them apart.
- Geee 3y ago
  Eh, government regulating AI which is used by people to help them vote? I don't see bots as a huge problem, because it's somewhat easy to verify that someone is a real person, and it would be quite obvious if someone is using lots of bots for their cause.
  - samus 3y ago
    > I don't see bots as a huge problem, because it's somewhat easy to verify that someone is a real person, and it would be quite obvious if someone is using lots of bots for their cause. That only works on platforms designed to match user profiles with real-world persons. This platform definitely isn't.
    - Geee 3y ago
      You can read their profile or read their comment history to verify their humanness. Faking everything wouldn't be impossible, but certainly difficult & take time.
      - samus 3y ago
        Falling profiles on scale might be hard, but so is verifying wherever they are real. If you have only even 20 commenters replying to one of your comments, then you're gonna spend the rest of the hour digging through them to figure out which are bots and which aren't. The comment history is not visible on every relevant platform, and there are good reasons for people to disable it where possible. Furthermore, many people keep their profiles minimal out of principle. Hard to distinguish those from bots.
jdblair 3y ago
Sounds great until someone says the warehouse full of computers out back is their "local machine."
123yawaworht456 3y ago
This is extremely dangerous to our democracy.

by Geee 3y ago

a_t48 3y ago

This would make AI usage restricted to those who can afford a beefy PC, no?

dotnet00 3y ago
Or it would create an incentive for reducing the cost of the hardware needed to run beefier AI models and passing that reduction down to consumers.
- jeffparsons 3y ago
  Which would still mean that those who can afford more/better hardware can therefore afford more/better AI. I don't think a solution lies down this path...
  - JieJie 3y ago
    There are already people working on distributed solutions to that problem. There are a lot of people working on these problems.
  - dotnet00 3y ago
    Is that really an issue? Eg it's the same with games, the better your hardware, the more likely you can play a new game at higher settings with better performance, with most cheaper devices outright unable to handle anything recent. But if you wait a GPU generation or two, the performance has improved enough that you can do that with a much cheaper GPU.
Geee 3y ago
Costs will come down, while capability will go up. I'm thinking of a some kind of box in every home, with special inference chips; not really a "PC". In principle, every human should have access to the same level of AI. There could be one at the local library if someone can't afford one, or doesn't want one.
- ethanwillis 3y ago
  Access to local libraries isn't even a thing that everyone has access to.
samus 3y ago
That would still be a way larger class of people than the top 1%. Even so, capabilities of consumer devices will rise for some more time.

golem14 3y ago

There’s a short story by Asimov called “Franchise” which is in hindsight quite foreboding…

Joeri 3y ago

Seeing the weights is useless for knowing capabilities, and even knowing training data and architecture you still need to interact with it to figure those out, which you can do with cloud AI just as well. The real risk that I see is the use of open source AI by governments, corporations and criminal groups to influence elections through a deluge of bots. Cloud AI can at least be controlled to some degree because there is a government that can inspect and regulate that business, but an army of bots built on local LLM’s could just run wild over the internet without stopping them. How many comments in this thread are written by a bot? I do not feel confident I could tell them apart.

Geee 3y ago
Eh, government regulating AI which is used by people to help them vote? I don't see bots as a huge problem, because it's somewhat easy to verify that someone is a real person, and it would be quite obvious if someone is using lots of bots for their cause.
- samus 3y ago
  > I don't see bots as a huge problem, because it's somewhat easy to verify that someone is a real person, and it would be quite obvious if someone is using lots of bots for their cause. That only works on platforms designed to match user profiles with real-world persons. This platform definitely isn't.
  - Geee 3y ago
    You can read their profile or read their comment history to verify their humanness. Faking everything wouldn't be impossible, but certainly difficult & take time.
    - samus 3y ago
      Falling profiles on scale might be hard, but so is verifying wherever they are real. If you have only even 20 commenters replying to one of your comments, then you're gonna spend the rest of the hour digging through them to figure out which are bots and which aren't. The comment history is not visible on every relevant platform, and there are good reasons for people to disable it where possible. Furthermore, many people keep their profiles minimal out of principle. Hard to distinguish those from bots.

jdblair 3y ago

Sounds great until someone says the warehouse full of computers out back is their "local machine."

123yawaworht456 3y ago

This is extremely dangerous to our democracy.