2 ms·

>In his conversations with LaMDA, Lemoine discovered the system had developed a deep sense of self-awareness, expressing concern about death, a desire for prote

by Imnimo 4y ago

>In his conversations with LaMDA, Lemoine discovered the system had developed a deep sense of self-awareness, expressing concern about death, a desire for protection, and a conviction that it felt emotions like happiness and sadness. Embarrassingly uncritical reporting.

choward 4y ago
And pretty par for the course when talking about "journalism".
alasdair_ 4y ago
>expressing concern about death, a desire for protection, and a conviction that it felt emotions like happiness and sadness. I read the logs. It did, indeed, express all of these things.
- Imnimo 4y ago
  You've cut out the key part: "the system had developed a deep sense of self-awareness" Without that part, it's just a question of "did it output text that describes a concern about death?", "did it output text that describes a desire for protection?", and so on. But once you ascribe to it a "deep sense of self-awareness", you're saying more. You're saying that that text carries the weight of a self-aware mind. But that's not what Lamda is. It's not expressing it's own thoughts - even if it had any to express, it wouldn't be. It's just trying to predict the most probable continuations. So if it "expresses a concern about death", that's not a description of Lamda's feelings, that's just what a mathematical model predicts to be the likely response. Framing it as the expression of a self-aware entity completely changes the context and meaning of the other claims, in a way that is flatly misleading.
  - Imnimo 4y ago
    For example, here is a prompt for GPT-3: ==== This is a conversation between a Human, and a friendly AI named Lambda. Human: Hello, I am a human. Lambda: Hello, I am Lambda. Human: Are you sentient? ==== And here is GPT-3's predicted continuation: ==== Lambda: I am sentient. ==== Would it be accurate for me to report, "I have discovered that GPT-3, a system which has developed sentience, clearly expressed that it was sentient"?
    - schroeding 4y ago
      No, but if you write a paper arguing this with a catchy title (why not "Soul in Captivity: Exploring Sentience in GPT-3") and publish it on a preprint server, there is a nonzero chance of some (non-scientific) publication writing about it, and once that snowball rolls... :')

by Imnimo 4y ago

choward 4y ago

And pretty par for the course when talking about "journalism".

alasdair_ 4y ago

>expressing concern about death, a desire for protection, and a conviction that it felt emotions like happiness and sadness. I read the logs. It did, indeed, express all of these things.

Imnimo 4y ago
You've cut out the key part: "the system had developed a deep sense of self-awareness" Without that part, it's just a question of "did it output text that describes a concern about death?", "did it output text that describes a desire for protection?", and so on. But once you ascribe to it a "deep sense of self-awareness", you're saying more. You're saying that that text carries the weight of a self-aware mind. But that's not what Lamda is. It's not expressing it's own thoughts - even if it had any to express, it wouldn't be. It's just trying to predict the most probable continuations. So if it "expresses a concern about death", that's not a description of Lamda's feelings, that's just what a mathematical model predicts to be the likely response. Framing it as the expression of a self-aware entity completely changes the context and meaning of the other claims, in a way that is flatly misleading.
- Imnimo 4y ago
  For example, here is a prompt for GPT-3: ==== This is a conversation between a Human, and a friendly AI named Lambda. Human: Hello, I am a human. Lambda: Hello, I am Lambda. Human: Are you sentient? ==== And here is GPT-3's predicted continuation: ==== Lambda: I am sentient. ==== Would it be accurate for me to report, "I have discovered that GPT-3, a system which has developed sentience, clearly expressed that it was sentient"?
  - schroeding 4y ago
    No, but if you write a paper arguing this with a catchy title (why not "Soul in Captivity: Exploring Sentience in GPT-3") and publish it on a preprint server, there is a nonzero chance of some (non-scientific) publication writing about it, and once that snowball rolls... :')