6 ms·

> If the tech really is dangerous then being more closed arguably is in the public's best interest, and this is certainly the reason OpenAI have previously give

by nicce 3y ago

> If the tech really is dangerous then being more closed arguably is in the public's best interest, and this is certainly the reason OpenAI have previously given. Not really. It slows down like security over obscurity. It needs to be open that we know the real risks and we have the best information to combat it. Otherwise, someone who does the same in closed matter, has better chances to get advantage when misusing it.

FeepingCreature 3y ago
This only holds if defense outscales attack. It seems very likely that attack outscales defense to me with LLMs.
- nicce 3y ago
  Well then, isn’t the whole case about just denying the inevitable? If OpenAI can do it, I would not say that that is very unlikely for someone else to do the same. Open or not. The best chance is still that we prepare with the best available information.
  - FeepingCreature 3y ago
    Yep, it absolutely is about denying the inevitable, or rather, "playing for time." The longer we manage to delay, the more likely somebody comes up with some clever approach for actually controlling the things. Also humanity stays alive in the meantime, which is no small thing in itself.
    - nicce 3y ago
      > with some clever approach for actually controlling the things. But if we hide the things, we have no idea what we are trying to control.
      - FeepingCreature 3y ago
        ... Eh? You as an end-user can't contribute to this anyways. If you really want to work on safety, either use a smaller network or join the safety team at a big org.
  - pixl97 3y ago
    >denying the inevitable I mean, you spend a lot of time in your own life denying the inevitable, humans spend a lot of time and effort avoiding their own personal extinction. >The best chance is still The best information we have now is if we create AGI/ASI at this time, we all die. The only winning move is not to play in that game.
    - nicce 3y ago
      > The best information we have now is if we create AGI/ASI at this time, we all die. We can still unplug or turn off the things. We are still very faraway from the situation where AI has some factories and full supply chain to control and take physical control of the world.
      - FeepingCreature 3y ago
        Meanwhile, every giant AI company: "yeah we're looking at robotics, obviously if we could embody these things and give them agency in the physical world that would be a great achievement"
        pixl97 3y ago
        Our rush into AI and embodiment reminds me of the lily pad exponential growth parable. >Imagine a large pond that is completely empty except for 1 lily pad. The lily pad will grow exponentially and cover the entire pond in 3 years. In other words, after 1 month there will 2 lily pads, after 2 months there will be 4, etc. The pond is covered in 36 months We're all going to be sitting around at 34 months saying "Look, it's been years and AI hasn't taken over that much of the market.
- arez 3y ago
  either way it still shouldn't be in the hands of a private for profit corporation
patcon 3y ago
When I try to port your logic over into nuclear capacity it doesn't hold very well. Nuclear capacity is constrained, and those constraining it attempt to do so for reasons public good (energy, warfare, peace). You could argue about effectiveness, but our failure to self-annihilate seems positive testament to the strategy. Transparency does not serve us when mitigating certain forms of danger. I'm trying to remain humble with this, but it's not clear to me what balance of benefit and danger current AI is. (Not even considering the possibility of AGI, which is beyond scope of my comment)
- Vetch 3y ago
  This is a poor analogy, a better one would be nuclear physics. An expert in nuclear physics can develop positively impactful energy generation methods or very damaging nuclear weapons. It's not because of arcane secrets that so few nations have nuclear weapons, all you need is a budget, time and brilliant physicists and engineers. The reason we don't have more is largely down to surveillance, economics, challenge of reliable payload delivery, security assurances, agreements and various logistical challenges. Most countries are open and transparent about their nuclear efforts due to the diplomatic advantages. There are also methods to trace and detect secret nuclear tests and critical supply chains can be monitored. Countries who violate these norms can face anything from heavy economic sanctions and isolation to sabotage of research efforts. On the technical side, having safe and reliable launch capacity is arguably as much if not more of a challenge than the bomb itself. Logistical issues include mass manufacture (merely having capacity only paints a target on your back with no real gains) and safe storage. There are a great many reasons why it is simply not worth going forward with nuclear weapons. This calculus changes however, if a country has cause for fear for their continued existence, as is presently the case for some Eastern European countries.
- mywittyname 3y ago
  The difference between nuclear capability and AI capability is that you can't just rent out nuclear enrichment facilities on a per-hour basis, nor can you buy the components to build such facilities at a local store. But you can train AI models by renting AWS servers or building your own. If one could just walk into a store and buy plutonium, then society would probably take a much different approach to nuclear security.
tw04 3y ago
Just like nuclear weapons? The whole “security through obscurity doesn’t work” is absolute nonsense. It absolutely works and there are countless real world examples. What doesn’t work is relying on that as your ONLY security.
- matthewmacleod 3y ago
  This is a broken comparison IMO because you can’t instantly and freely duplicate nuclear weapons across the planet and then offer them up to everyone for low marginal cost and effort. The tech exists, and will rapidly become easy to access. There is approximately zero chance of it remaining behind lock and key.
- gary_0 3y ago
  I'm not sure if nuclear weapons are a good example. In the 1940's most of the non-weapons-related nuclear research was public (and that did make certain agencies nervous). That's just how scientists tend to do things. While the US briefly had unique knowledge about the manufacture of nuclear weapons, the basics could be easily worked out from first principles, especially once schoolchildren could pick up an up-to-date book on atomic physics. The engineering and testing part is difficult, of course, but for a large nation-state stealing the plans is only a shortcut. The on-paper part of the engineering is doable by any team with the right skills. So the main blocker with nuclear weapons isn't the knowledge, it's acquiring the raw fissile material and establishing the industrial base required to refine it. This makes nuclear weapons a poor analogy for AI, because all you need to develop an LLM is a big pile of commodity GPUs, the publicly available training data, some decent software engineers, and time. So in both cases all security-through-obscurity will buy you is a delay, and when it comes to AI probably not a very long one (except maybe if you can restrict the supply of GPUs, but the effectiveness of that strategy against China et al remains to be seen).
  - tw04 3y ago
    >This makes nuclear weapons a poor analogy for AI, because all you need to develop an LLM is a big pile of commodity GPUs, the publicly available training data, some decent software engineers, and time. Except the GPUs are on export control, and keeping up with the arms race requires a bunch of data you don't have access to (NVidia's IP) - or direct access to the source. Just like building a nuclear weapon requires access to either already refined fissile material. Or the IP and skills to build your own refining facilities (IP most countries don't have). Literally everyone has access to Uranium - being able to do something useful with it is another story. Kind of like... AI.
robbrown451 3y ago
I don't see how opening it makes it safer. It's very different from security things, where some "white hat" can find a security, and they can then fix it so instances don't get hacked. Sure, a bad person could run the software without fixing the bug, but that isn't going to harm anyone but themselves. That isn't the case here. If some well meaning person discovers a way that you can create a pandemic causing superbug, they can't just "fix" the AI to make that impossible. Not if it is open source. Very different thing.

by nicce 3y ago

FeepingCreature 3y ago

This only holds if defense outscales attack. It seems very likely that attack outscales defense to me with LLMs.

nicce 3y ago
Well then, isn’t the whole case about just denying the inevitable? If OpenAI can do it, I would not say that that is very unlikely for someone else to do the same. Open or not. The best chance is still that we prepare with the best available information.
- FeepingCreature 3y ago
  Yep, it absolutely is about denying the inevitable, or rather, "playing for time." The longer we manage to delay, the more likely somebody comes up with some clever approach for actually controlling the things. Also humanity stays alive in the meantime, which is no small thing in itself.
  - nicce 3y ago
    > with some clever approach for actually controlling the things. But if we hide the things, we have no idea what we are trying to control.
    - FeepingCreature 3y ago
      ... Eh? You as an end-user can't contribute to this anyways. If you really want to work on safety, either use a smaller network or join the safety team at a big org.
- pixl97 3y ago
  >denying the inevitable I mean, you spend a lot of time in your own life denying the inevitable, humans spend a lot of time and effort avoiding their own personal extinction. >The best chance is still The best information we have now is if we create AGI/ASI at this time, we all die. The only winning move is not to play in that game.
  - nicce 3y ago
    > The best information we have now is if we create AGI/ASI at this time, we all die. We can still unplug or turn off the things. We are still very faraway from the situation where AI has some factories and full supply chain to control and take physical control of the world.
    - FeepingCreature 3y ago
      Meanwhile, every giant AI company: "yeah we're looking at robotics, obviously if we could embody these things and give them agency in the physical world that would be a great achievement"
      - pixl97 3y ago
        Our rush into AI and embodiment reminds me of the lily pad exponential growth parable. >Imagine a large pond that is completely empty except for 1 lily pad. The lily pad will grow exponentially and cover the entire pond in 3 years. In other words, after 1 month there will 2 lily pads, after 2 months there will be 4, etc. The pond is covered in 36 months We're all going to be sitting around at 34 months saying "Look, it's been years and AI hasn't taken over that much of the market.
arez 3y ago
either way it still shouldn't be in the hands of a private for profit corporation

patcon 3y ago

When I try to port your logic over into nuclear capacity it doesn't hold very well. Nuclear capacity is constrained, and those constraining it attempt to do so for reasons public good (energy, warfare, peace). You could argue about effectiveness, but our failure to self-annihilate seems positive testament to the strategy. Transparency does not serve us when mitigating certain forms of danger. I'm trying to remain humble with this, but it's not clear to me what balance of benefit and danger current AI is. (Not even considering the possibility of AGI, which is beyond scope of my comment)

Vetch 3y ago
This is a poor analogy, a better one would be nuclear physics. An expert in nuclear physics can develop positively impactful energy generation methods or very damaging nuclear weapons. It's not because of arcane secrets that so few nations have nuclear weapons, all you need is a budget, time and brilliant physicists and engineers. The reason we don't have more is largely down to surveillance, economics, challenge of reliable payload delivery, security assurances, agreements and various logistical challenges. Most countries are open and transparent about their nuclear efforts due to the diplomatic advantages. There are also methods to trace and detect secret nuclear tests and critical supply chains can be monitored. Countries who violate these norms can face anything from heavy economic sanctions and isolation to sabotage of research efforts. On the technical side, having safe and reliable launch capacity is arguably as much if not more of a challenge than the bomb itself. Logistical issues include mass manufacture (merely having capacity only paints a target on your back with no real gains) and safe storage. There are a great many reasons why it is simply not worth going forward with nuclear weapons. This calculus changes however, if a country has cause for fear for their continued existence, as is presently the case for some Eastern European countries.
mywittyname 3y ago
The difference between nuclear capability and AI capability is that you can't just rent out nuclear enrichment facilities on a per-hour basis, nor can you buy the components to build such facilities at a local store. But you can train AI models by renting AWS servers or building your own. If one could just walk into a store and buy plutonium, then society would probably take a much different approach to nuclear security.

tw04 3y ago

Just like nuclear weapons? The whole “security through obscurity doesn’t work” is absolute nonsense. It absolutely works and there are countless real world examples. What doesn’t work is relying on that as your ONLY security.

matthewmacleod 3y ago
This is a broken comparison IMO because you can’t instantly and freely duplicate nuclear weapons across the planet and then offer them up to everyone for low marginal cost and effort. The tech exists, and will rapidly become easy to access. There is approximately zero chance of it remaining behind lock and key.
gary_0 3y ago
I'm not sure if nuclear weapons are a good example. In the 1940's most of the non-weapons-related nuclear research was public (and that did make certain agencies nervous). That's just how scientists tend to do things. While the US briefly had unique knowledge about the manufacture of nuclear weapons, the basics could be easily worked out from first principles, especially once schoolchildren could pick up an up-to-date book on atomic physics. The engineering and testing part is difficult, of course, but for a large nation-state stealing the plans is only a shortcut. The on-paper part of the engineering is doable by any team with the right skills. So the main blocker with nuclear weapons isn't the knowledge, it's acquiring the raw fissile material and establishing the industrial base required to refine it. This makes nuclear weapons a poor analogy for AI, because all you need to develop an LLM is a big pile of commodity GPUs, the publicly available training data, some decent software engineers, and time. So in both cases all security-through-obscurity will buy you is a delay, and when it comes to AI probably not a very long one (except maybe if you can restrict the supply of GPUs, but the effectiveness of that strategy against China et al remains to be seen).
- tw04 3y ago
  >This makes nuclear weapons a poor analogy for AI, because all you need to develop an LLM is a big pile of commodity GPUs, the publicly available training data, some decent software engineers, and time. Except the GPUs are on export control, and keeping up with the arms race requires a bunch of data you don't have access to (NVidia's IP) - or direct access to the source. Just like building a nuclear weapon requires access to either already refined fissile material. Or the IP and skills to build your own refining facilities (IP most countries don't have). Literally everyone has access to Uranium - being able to do something useful with it is another story. Kind of like... AI.

robbrown451 3y ago

I don't see how opening it makes it safer. It's very different from security things, where some "white hat" can find a security, and they can then fix it so instances don't get hacked. Sure, a bad person could run the software without fixing the bug, but that isn't going to harm anyone but themselves. That isn't the case here. If some well meaning person discovers a way that you can create a pandemic causing superbug, they can't just "fix" the AI to make that impossible. Not if it is open source. Very different thing.