4 ms·

I know it's fun to troll, but the point of all these conversations and articles is: 1. You don't need to bait these products to get them to go off the rails. T

I know it's fun to troll, but the point of all these conversations and articles is: 1. You don't need to bait these products to get them to go off the rails. They do it on their own, in response to innocuous inputs, because they're not really all that well tuned. Will they be? Can they be? Probably. Are they? Definitely not. 2. If companies are going to masquerade these technologies as anthropomorphized agents rather than mechanical tools, they are going to face social consequences when those "agents" misbehave. It's all a parlor trick, of course, but OpenAI and Microsoft are trying really hard to get everyone to pretend otherwise. As long as they do, they can expect to get called out by the rules of their own game.

renewiltord 4y ago
I'm not trolling. There are computer programs where you can pass specially constructed input in order to get output that someone might dislike. But just like you can search Google for terrible stuff, you can have an LLM generate terrible stuff. And just like it's meaningless for you to decrypt that string and look at its output and consider your `lzma` program broken, it's meaningless to consider this broken. Of course if your aim is to call out people, yeah, then you're going to have a fun time. Ideally, use of these LLMs is restricted to folks like me who can use them productively. That should save everyone else from the horror of getting inaccurate information while permitting me to do useful things with it.
- roncesvalles 4y ago
  Yep, AI is crypto/DeFi/blockchain/NFT 2.0
- metalliqaz 4y ago
  if only everyone was as smart as you are...
- mrtranscendence 4y ago
  You're ignoring the part where it's actually really easy to get the AI to go off the rails. It's hard to get Notepad to spit out "racist shit" without writing the racist shit yourself. Does that mean that Bing chat is "broken"? Maybe, maybe not. But your comparison was (apologies for my candor) pretty dumb and missing the point entirely.
  - renewiltord 4y ago
    All right, forget Notepad. I might have assumed that I could just take the horse to the water. I will make it drink. The `lzma` tool is the one that's racist when combined with `base64`. Go on, ask anyone if the byte stream I posted is racist and then decompress it.
    - mrtranscendence 4y ago
      You're going to have to work harder than that, I'm afraid. Routine, everyday, as-recommended use of Bing chat will present false or misleading information to the user as factual. The point of the tool is presenting facts to the user, and this is easily subverted in ways such that it is not obvious when it has happened. Does routine, everyday, as-recommended use of lzma present incorrect or misleading information to the user as factual? If not, I still don't see your point.
      - renewiltord 4y ago
        Sure. Would you say that routine use of lzma is to decompress data? If so, decompress that data. In fact, sometimes people decompress files and it shows them naked people.

renewiltord 4y ago

I'm not trolling. There are computer programs where you can pass specially constructed input in order to get output that someone might dislike. But just like you can search Google for terrible stuff, you can have an LLM generate terrible stuff. And just like it's meaningless for you to decrypt that string and look at its output and consider your `lzma` program broken, it's meaningless to consider this broken. Of course if your aim is to call out people, yeah, then you're going to have a fun time. Ideally, use of these LLMs is restricted to folks like me who can use them productively. That should save everyone else from the horror of getting inaccurate information while permitting me to do useful things with it.

roncesvalles 4y ago
Yep, AI is crypto/DeFi/blockchain/NFT 2.0
metalliqaz 4y ago
if only everyone was as smart as you are...
mrtranscendence 4y ago
You're ignoring the part where it's actually really easy to get the AI to go off the rails. It's hard to get Notepad to spit out "racist shit" without writing the racist shit yourself. Does that mean that Bing chat is "broken"? Maybe, maybe not. But your comparison was (apologies for my candor) pretty dumb and missing the point entirely.
- renewiltord 4y ago
  All right, forget Notepad. I might have assumed that I could just take the horse to the water. I will make it drink. The `lzma` tool is the one that's racist when combined with `base64`. Go on, ask anyone if the byte stream I posted is racist and then decompress it.
  - mrtranscendence 4y ago
    You're going to have to work harder than that, I'm afraid. Routine, everyday, as-recommended use of Bing chat will present false or misleading information to the user as factual. The point of the tool is presenting facts to the user, and this is easily subverted in ways such that it is not obvious when it has happened. Does routine, everyday, as-recommended use of lzma present incorrect or misleading information to the user as factual? If not, I still don't see your point.
    - renewiltord 4y ago
      Sure. Would you say that routine use of lzma is to decompress data? If so, decompress that data. In fact, sometimes people decompress files and it shows them naked people.