Worth Reading – General Purpose AI Will Never Be Safe
The headline from Luiza seems like hyperbole, but the more I thought about it, the less I think so.
Let’s ignore the AI part of the statement and put it in a different context.
Is nature safe?
Not the flower garden and birdfeeder I have in my yard, but every natural environment in the world. Would you make the blanket claim that nature is safe? Maybe from a Western suburban background, nature seems safe, mostly. You may feel differently about the top of Mt. Everest, sitting within view of a pack of lions in Africa, or facing down a Tsunami in Asia.
No, the whole of nature is not safe. We take steps to protect ourselves from nature, even as we also challenge ourselves with more danger than normal, but we don’t simply do nothing in the face of extreme weather, wild animals, and other natural dangers.
Let’s take the metaphor a step further. Is the open internet safe? Again, not the tiny pieces of the internet that you look at daily, the entire thing. All the dark web, illegal porn, drug trafficking, terrorist plotting, violence, hatred inciting, and misinformation are parts of the internet.
Definitely not safe. Again, when we get online, we take actions for our safety. Countries block websites, platforms make their own calls (for better or worse) on what they will allow, and we all make our own decisions on what sites to get information from and who to interact with online. (Also, for better or worse.)
I think we can agree that granting someone full access to the open internet without education or tools to protect themselves would be dangerous, no?
OK, but what is a general-purpose LLM but a collection of everything that the model could ingest, without rules about what was safe and what wasn’t?
Yet we expect people to use them, and aren’t making any effort to make them safer.
I think she’s on to something here. Just like we provide tools and rules to make nature and the internet safer, we need them for AI too. We can disagree about the exact rules, but without any, we cannot claim safety.
https://www.luizasnewsletter.com/p/general-purpose-ai-will-never-be
Follow these topics: Artificial Intelligence
