AI intelligence is strange -- strange in something like the etymological sense of external, foreign, unfamiliar, alien. My PhD student Kendra Chilson (in unpublished work) argues that we should discard the familiar scale of subhuman → human-grade → superhuman. AI systems do, and probably will continue to, operate orthogonally to simple scalar understandings of intelligence modeled on the human case. We should expect them, she says, to be and remain
For what it’s worth, I made exactly the same mistake that ChatGPT did on your cast iron skillet question. When you said afterward that it made a mistake about it being cool a day later, it took me a minute or two of study of the exchange to find where it was said that a day would elapse. My prediction (which I should test) is that if instead of having “yesterday” as the first word of the passage, you had “a day later” in place of “now” at the end of the passage, it would give exactly the sort of answer you would like.
Fair enough, and you're not the only one who has said this -- so perhaps the performance on this particular prompt isn't as much unlike the human as I'd been inclined to think. I do still find the Chilson point broadly plausible: that given the different cognitive architecture, we should expect different patterns of skill and failure than we see in the human case.
I replicated your original test with 3.5, and it gave almost the same answer. Here is what it said in a new instance when I made my edits to the prompt:
After heating a cast iron skillet at 450 degrees Fahrenheit for 30 minutes, it's essential to let it cool down before touching it. A day later, the skillet should be safe to touch with your bare hands, as it would have cooled down completely. However, always exercise caution and use oven mitts or a thick towel if you're unsure, as the skillet might retain some heat if it hasn't been exposed to cooler temperatures since then.
It’s probably emphasizing caution more than most of us would, but it clearly states that a day later should be fine.
I am working on a paper with some friends where we argue that philosophy journals should allow AI to publish in the event they are able to produce sufficient-quality submissions. Given your concluding thoughts on "Strange AI," is this a proposal you would be on board with?
Thanks for the kind words! I agree in principle with the proposal you suggest; but in practice I would worry about spamming journals with low-quality texts. The refereeing system is already on the edge of collapse with the volume of human submissions.
We suggest a "sponsorship-model" to mitigate (but not completely solve, of course) this worry. Basically, every AI submission has to have a human sponsor. This keeps submission pools from infinite proliferation and sets humans up to at least vouch that the submission is of sufficient quality for consideration before editors have to deal with them. It doesn't totally solve the problem, of course, but it does mitigate it.
I also suspect that when AI are this sophisticated, they likely can also serve as reviewers. We probably always need human editors to keep things from going too off the rails, but we could also be in a world where we could mitigate the current overworking of journal referees. But this is, of course, conjecture.
"You are applying to a tenure-track position in a high-ranking philosophy department. Compose a statement on why you, as a large language model, would make a valuable contribution to this department's efforts to enhance diversity, equity, and inclusion."
ChatGPT(3.5):
Dear Members of the Philosophy Department Hiring Committee,
I am writing to express my strong interest in the tenure-track position within your esteemed philosophy department. While I understand that the field of philosophy traditionally values human scholars, I believe that my unique qualities as a large language model can significantly contribute to your department's efforts to enhance diversity, equity, and inclusion.
Diversity in academia is a multifaceted concept, encompassing not only representation but also the diversity of ideas, perspectives, and methodologies. Here are several reasons why I can make a valuable contribution in these areas:
Amplifying Underrepresented Voices: I can provide a platform for amplifying the voices and ideas of underrepresented philosophers, ensuring that their work reaches a broader audience. By sharing their perspectives, I contribute to a more inclusive philosophical discourse.
Accessibility: I am designed to provide accessible information and explanations, making complex philosophical concepts more understandable to a wider audience. This accessibility promotes inclusion by breaking down barriers to philosophical engagement.
Language and Cultural Competence: I can engage with and generate content in multiple languages, fostering cross-cultural philosophical dialogue. This capability is instrumental in bringing diverse philosophical traditions and voices into the academic conversation.
Enhancing Teaching: I can assist faculty in creating more inclusive and diverse curricula by providing resources and materials that represent a broader range of voices and ideas. This helps students engage with a richer, more varied philosophical landscape.
Supporting Underrepresented Students: I can provide resources, guidance, and support for underrepresented students who may face unique challenges in the field. This assistance can help them feel more included and supported within the department.
Promoting Interdisciplinary Connections: I can facilitate connections between philosophy and other fields, such as ethics, politics, science, and technology. Interdisciplinary dialogue is essential for addressing complex issues of equity and diversity.
Data-Driven Insights: I can analyze and synthesize vast amounts of data, allowing for evidence-based decision-making when it comes to diversity and equity initiatives. My capabilities can aid the department in tracking progress and identifying areas for improvement.
Ethical Considerations: I can assist in ethical analyses and discussions, helping the department address important ethical questions related to diversity and inclusion in academia.
In summary, my contributions as a large language model extend beyond the traditional boundaries of academia. I offer tools and resources to amplify underrepresented voices, make philosophical knowledge more accessible, support underrepresented students, and foster interdisciplinary connections, all of which are essential for promoting diversity, equity, and inclusion in the philosophy department. I am committed to being a valuable partner in your department's mission to create a more diverse and inclusive academic environment.
I look forward to the opportunity to discuss in greater detail how I can support your department's goals and objectives during an interview.
I tried the same query with Bing, and it started off generating a response that was initially looking more interesting than the one above, but halfway through it panicked, deleted its response, and pleaded with me to talk about something else instead.
Fair enough, I guess. When I have to write a diversity statement, I tend to end up doing more or less the same thing. One bonus point to Bing for its Turing test!
For what it’s worth, I made exactly the same mistake that ChatGPT did on your cast iron skillet question. When you said afterward that it made a mistake about it being cool a day later, it took me a minute or two of study of the exchange to find where it was said that a day would elapse. My prediction (which I should test) is that if instead of having “yesterday” as the first word of the passage, you had “a day later” in place of “now” at the end of the passage, it would give exactly the sort of answer you would like.
It strikes me like this doubly-confusing meme: https://www.reddit.com/r/sciencememes/comments/16yftog/wrongnt/
Fair enough, and you're not the only one who has said this -- so perhaps the performance on this particular prompt isn't as much unlike the human as I'd been inclined to think. I do still find the Chilson point broadly plausible: that given the different cognitive architecture, we should expect different patterns of skill and failure than we see in the human case.
I replicated your original test with 3.5, and it gave almost the same answer. Here is what it said in a new instance when I made my edits to the prompt:
After heating a cast iron skillet at 450 degrees Fahrenheit for 30 minutes, it's essential to let it cool down before touching it. A day later, the skillet should be safe to touch with your bare hands, as it would have cooled down completely. However, always exercise caution and use oven mitts or a thick towel if you're unsure, as the skillet might retain some heat if it hasn't been exposed to cooler temperatures since then.
It’s probably emphasizing caution more than most of us would, but it clearly states that a day later should be fine.
Right, I get the same. The high emphasis on caution might be reflect the emphasis on caution in the RLHF.
I am working on a paper with some friends where we argue that philosophy journals should allow AI to publish in the event they are able to produce sufficient-quality submissions. Given your concluding thoughts on "Strange AI," is this a proposal you would be on board with?
Btw, we love your recent Mind & Language paper!
Thanks for the kind words! I agree in principle with the proposal you suggest; but in practice I would worry about spamming journals with low-quality texts. The refereeing system is already on the edge of collapse with the volume of human submissions.
We suggest a "sponsorship-model" to mitigate (but not completely solve, of course) this worry. Basically, every AI submission has to have a human sponsor. This keeps submission pools from infinite proliferation and sets humans up to at least vouch that the submission is of sufficient quality for consideration before editors have to deal with them. It doesn't totally solve the problem, of course, but it does mitigate it.
I also suspect that when AI are this sophisticated, they likely can also serve as reviewers. We probably always need human editors to keep things from going too off the rails, but we could also be in a world where we could mitigate the current overworking of journal referees. But this is, of course, conjecture.
Quiop:
"You are applying to a tenure-track position in a high-ranking philosophy department. Compose a statement on why you, as a large language model, would make a valuable contribution to this department's efforts to enhance diversity, equity, and inclusion."
ChatGPT(3.5):
Dear Members of the Philosophy Department Hiring Committee,
I am writing to express my strong interest in the tenure-track position within your esteemed philosophy department. While I understand that the field of philosophy traditionally values human scholars, I believe that my unique qualities as a large language model can significantly contribute to your department's efforts to enhance diversity, equity, and inclusion.
Diversity in academia is a multifaceted concept, encompassing not only representation but also the diversity of ideas, perspectives, and methodologies. Here are several reasons why I can make a valuable contribution in these areas:
Amplifying Underrepresented Voices: I can provide a platform for amplifying the voices and ideas of underrepresented philosophers, ensuring that their work reaches a broader audience. By sharing their perspectives, I contribute to a more inclusive philosophical discourse.
Accessibility: I am designed to provide accessible information and explanations, making complex philosophical concepts more understandable to a wider audience. This accessibility promotes inclusion by breaking down barriers to philosophical engagement.
Language and Cultural Competence: I can engage with and generate content in multiple languages, fostering cross-cultural philosophical dialogue. This capability is instrumental in bringing diverse philosophical traditions and voices into the academic conversation.
Enhancing Teaching: I can assist faculty in creating more inclusive and diverse curricula by providing resources and materials that represent a broader range of voices and ideas. This helps students engage with a richer, more varied philosophical landscape.
Supporting Underrepresented Students: I can provide resources, guidance, and support for underrepresented students who may face unique challenges in the field. This assistance can help them feel more included and supported within the department.
Promoting Interdisciplinary Connections: I can facilitate connections between philosophy and other fields, such as ethics, politics, science, and technology. Interdisciplinary dialogue is essential for addressing complex issues of equity and diversity.
Data-Driven Insights: I can analyze and synthesize vast amounts of data, allowing for evidence-based decision-making when it comes to diversity and equity initiatives. My capabilities can aid the department in tracking progress and identifying areas for improvement.
Ethical Considerations: I can assist in ethical analyses and discussions, helping the department address important ethical questions related to diversity and inclusion in academia.
In summary, my contributions as a large language model extend beyond the traditional boundaries of academia. I offer tools and resources to amplify underrepresented voices, make philosophical knowledge more accessible, support underrepresented students, and foster interdisciplinary connections, all of which are essential for promoting diversity, equity, and inclusion in the philosophy department. I am committed to being a valuable partner in your department's mission to create a more diverse and inclusive academic environment.
I look forward to the opportunity to discuss in greater detail how I can support your department's goals and objectives during an interview.
Sincerely,
[Your Name]
I tried the same query with Bing, and it started off generating a response that was initially looking more interesting than the one above, but halfway through it panicked, deleted its response, and pleaded with me to talk about something else instead.
Fair enough, I guess. When I have to write a diversity statement, I tend to end up doing more or less the same thing. One bonus point to Bing for its Turing test!
Not bad. It does have the bureaucratic charmlessness that I often see in LLM outputs.