4 ms·

What do you think of the calls for regulation or licensing of AI?

by kalkin 3y ago

JumpCrisscross 3y ago
> What do you think of the calls for regulation or licensing of AI Misdirection. We see a generic call for regulation, or unrealistic call for a global pause with no answers to how it would be coördinated or enforced. When actual regulations are put forward, they're rejected without a counter-offered solution [1]. [1] https://www.reuters.com/technology/openai-may-leave-eu-if-regulations-bite-ceo-2023-05-24/ https://www.reuters.com/technology/openai-may-leave-eu-if-re...
- kalkin 3y ago
  Obviously companies are going to try to shape any regulation to their benefit. What I'm wondering is if people who see x-risk worries as distracting from real risks are in agreement with the narrative I've seen a lot on HN recently that calls for regulation are just anti-competitive attempts at regulatory capture.
  - JumpCrisscross 3y ago
    > companies are going to try to shape any regulation to their benefit There is a difference between trying to shape regulation to your advantage and engaging in bad faith. There is no indication there is real regulation that addresses the problem these folks are brining up that they would support.
    - kalkin 3y ago
      Who are "these folks"? OpenAI specifically? I'm not arguing that they're operating in good faith, but it's also not clear to me that them proposing specific regulation would be helpful - there seem to be a lot of people who see anything like that as automatically an attempt to stifle competition. Many other people talking about x-risk signed the letter calling for a pause on training new big models, which seems pretty concrete to me. Does that make sense to you? If not, do you have other regulation you'd prefer?
      - JumpCrisscross 3y ago
        > other people talking about x-risk signed the letter calling for a pause on training new big models, which seems pretty concrete to me. Does that make sense to you It's totally unrealistic. How are you going to coördinate a global pause on AI research? Nobody has even bothered proposing a framework because of how ludicrous it is for this to be what Beijing, Washington, Brussels and New Delhi focus on.
        kalkin 3y ago
        Right, you said before that's unrealistic, sorry. But it's not clear to me what you think would be a good idea, which makes it difficult to see what you think would be a serious proposal rather than misdirection. It appears to me that you are taking it as evidence of bad faith if someone proposes regulation either stronger (pause) or weaker (OpenAI amendments to proposed EU rules) than what you prefer - but you also haven't said what you do prefer. What am I missing? Personally, while I don't think the pause is realistic, I don't think advocates necessarily have any obligation to propose realistic ideas when they're not in a position to actually write any laws and are only able to potentially shift the window of what politicians and media consider acceptable.
        JumpCrisscross 3y ago
        > t's not clear to me what you think would be a good idea, which makes it difficult to see what you think would be a serious proposal rather than misdirection Registration. Not even licensing. If you're training a model, you file a form describing who you are, who you work for, who you're working with and what you're doing with which data for what purpose. This lets professionals, policy makers and the public get a sense for the landscape so we can make rules with some sense of reality.
        kalkin 3y ago
        That sounds entirely reasonable to me but also not very far from what Altman advocated in that congressional hearing. And FWIW I suspect a lot of folks here would still see it as just a vehicle for regulatory capture, via red tape, even as it also is, AFAICT, only a first step towards preventing abuses like mass surveillance, laundering bias in credit ratings, etc.
    - worik 3y ago
      > There is a difference between trying to shape regulation to your advantage and engaging in bad faith Bad faith is unnecessary for bad outcomes. The perverse (from our PoV) incentives are sufficient. OpenAI's incentives is to shut down all the rinky-dink models every hacker and their dog are building on cheap hardware. They spent billions, and are having the market taken from them.
      - kalkin 3y ago
        Is OpenAI actually losing market to open source models? And have they proposed any specific regulations that would shut down (current or similarly single-device) open source models?
        mlsu 3y ago
        They can see the writing on the wall. The open source LLM space is improving rapidly.
        kalkin 3y ago
        Accepting that for the sake of argument, have they actually proposed any regulations that would turn that around? I don't even know what would be feasible in that direction; a lot of proposed regulations only seem enforceable as long as giant GPU clusters provide a chokepoint.
        JumpCrisscross 3y ago
        > have they actually proposed any regulations that would turn that around Nobody has proposed any specific and realistic regulations.
        kalkin 3y ago
        You made that point in our other thread. I'm not informed enough about the landscape of proposals to agree or disagree. But if you're right, I think that's totally consistent with my argument here: people working with open source models at a scale that can run on an individual device have nothing to worry about from regulation. For better or worse, and whether or not they catch up to the state of the art, those aren't going to see meaningful controls.
Freebytes 3y ago
It will be impossible to regulate and impossible to stop. The code for AGI will not be some monolith of software architecture. The code will likely be simple. This means that someone in their basement could build it. The steps to get there are challenging, though. A single person could have developed the transformer architecture. A single person could have used 4-bit quantization and developed an AI that is just as good as ChatGPT and have it run on their local machine. The difficulties are figuring out the best 'needle in the haystack' to solve the problem. This requires research, and this process happens much faster if you have more people working on it. For years, many people did not put any energy into AI systems because the hardware was not here yet. The hardware is here now. The cat is out of the bag.

What do you think of the calls for regulation or licensing of AI?

by kalkin 3y ago

What do you think of the calls for regulation or licensing of AI?

> What do you think of the calls for regulation or licensing of AI Misdirection. We see a generic call for regulation, or unrealistic call for a global pause with no answers to how it would be coördinated or enforced. When actual regulations are put forward, they're rejected without a counter-offered solution [1]. [1] https://www.reuters.com/technology/openai-may-leave-eu-if-regulations-bite-ceo-2023-05-24/ https://www.reuters.com/technology/openai-may-leave-eu-if-re...

kalkin 3y ago
Obviously companies are going to try to shape any regulation to their benefit. What I'm wondering is if people who see x-risk worries as distracting from real risks are in agreement with the narrative I've seen a lot on HN recently that calls for regulation are just anti-competitive attempts at regulatory capture.
- JumpCrisscross 3y ago
  > companies are going to try to shape any regulation to their benefit There is a difference between trying to shape regulation to your advantage and engaging in bad faith. There is no indication there is real regulation that addresses the problem these folks are brining up that they would support.
  - kalkin 3y ago
    Who are "these folks"? OpenAI specifically? I'm not arguing that they're operating in good faith, but it's also not clear to me that them proposing specific regulation would be helpful - there seem to be a lot of people who see anything like that as automatically an attempt to stifle competition. Many other people talking about x-risk signed the letter calling for a pause on training new big models, which seems pretty concrete to me. Does that make sense to you? If not, do you have other regulation you'd prefer?
    - JumpCrisscross 3y ago
      > other people talking about x-risk signed the letter calling for a pause on training new big models, which seems pretty concrete to me. Does that make sense to you It's totally unrealistic. How are you going to coördinate a global pause on AI research? Nobody has even bothered proposing a framework because of how ludicrous it is for this to be what Beijing, Washington, Brussels and New Delhi focus on.
      - kalkin 3y ago
        Right, you said before that's unrealistic, sorry. But it's not clear to me what you think would be a good idea, which makes it difficult to see what you think would be a serious proposal rather than misdirection. It appears to me that you are taking it as evidence of bad faith if someone proposes regulation either stronger (pause) or weaker (OpenAI amendments to proposed EU rules) than what you prefer - but you also haven't said what you do prefer. What am I missing? Personally, while I don't think the pause is realistic, I don't think advocates necessarily have any obligation to propose realistic ideas when they're not in a position to actually write any laws and are only able to potentially shift the window of what politicians and media consider acceptable.
        JumpCrisscross 3y ago
        > t's not clear to me what you think would be a good idea, which makes it difficult to see what you think would be a serious proposal rather than misdirection Registration. Not even licensing. If you're training a model, you file a form describing who you are, who you work for, who you're working with and what you're doing with which data for what purpose. This lets professionals, policy makers and the public get a sense for the landscape so we can make rules with some sense of reality.
        kalkin 3y ago
        That sounds entirely reasonable to me but also not very far from what Altman advocated in that congressional hearing. And FWIW I suspect a lot of folks here would still see it as just a vehicle for regulatory capture, via red tape, even as it also is, AFAICT, only a first step towards preventing abuses like mass surveillance, laundering bias in credit ratings, etc.
  - worik 3y ago
    > There is a difference between trying to shape regulation to your advantage and engaging in bad faith Bad faith is unnecessary for bad outcomes. The perverse (from our PoV) incentives are sufficient. OpenAI's incentives is to shut down all the rinky-dink models every hacker and their dog are building on cheap hardware. They spent billions, and are having the market taken from them.
    - kalkin 3y ago
      Is OpenAI actually losing market to open source models? And have they proposed any specific regulations that would shut down (current or similarly single-device) open source models?
      - mlsu 3y ago
        They can see the writing on the wall. The open source LLM space is improving rapidly.
        kalkin 3y ago
        Accepting that for the sake of argument, have they actually proposed any regulations that would turn that around? I don't even know what would be feasible in that direction; a lot of proposed regulations only seem enforceable as long as giant GPU clusters provide a chokepoint.
        JumpCrisscross 3y ago
        > have they actually proposed any regulations that would turn that around Nobody has proposed any specific and realistic regulations.
        kalkin 3y ago
        You made that point in our other thread. I'm not informed enough about the landscape of proposals to agree or disagree. But if you're right, I think that's totally consistent with my argument here: people working with open source models at a scale that can run on an individual device have nothing to worry about from regulation. For better or worse, and whether or not they catch up to the state of the art, those aren't going to see meaningful controls.

Freebytes 3y ago

It will be impossible to regulate and impossible to stop. The code for AGI will not be some monolith of software architecture. The code will likely be simple. This means that someone in their basement could build it. The steps to get there are challenging, though. A single person could have developed the transformer architecture. A single person could have used 4-bit quantization and developed an AI that is just as good as ChatGPT and have it run on their local machine. The difficulties are figuring out the best 'needle in the haystack' to solve the problem. This requires research, and this process happens much faster if you have more people working on it. For years, many people did not put any energy into AI systems because the hardware was not here yet. The hardware is here now. The cat is out of the bag.