Running out of places to move the goalposts to
nickdrozd.github.io
Running out of places to move the goalposts to
1–10 of 14 posts
Re: Running out of places to move the goalposts to
#2I just commented about this in another thread. I know there has been some walking back e.g. of the significance of a Turing test but I think overall the goalposts for AI have shifted in the other way, to narrowing down the definition of intelligence to something like “being really good at some set of defined tasks” which coincidentally is basically the strong point of neural networks.
We seem hyperfocused on finding more tasks to train neural networks to do. This of course leads to a moving goalpost effect like in the article, but they’re moving along an axis that doesn’t measure intelligence.
My other comment: https://news.ycombinator.com/item?id=46445511
Re: Running out of places to move the goalposts to
#3The article mentions a personal goalpost involving Busy Beavers.
Mine is: write a nroff document that executes at least one macro, and is a quine.
Re: Running out of places to move the goalposts to
#4I just commented about this in another thread. I know there has been some walking back e.g. of the significance of a Turing test but I think overall the goalposts for AI have shifted in the other way, to narrowing down the definition of intelligence to something like “being really good at some set of defined tasks” which coincidentally is basically the strong point of neural networks. We seem hyperfocused on finding…
What would be a better way to measure intelligence?
Re: Running out of places to move the goalposts to
#5The article mentions a personal goalpost involving Busy Beavers. Mine is: write a nroff document that executes at least one macro, and is a quine.
How would your views about AI change if that goal were achieved? When my personal goal was reached, I found myself a little bit at a loss for words.
Re: Running out of places to move the goalposts to
#6The article mentions a personal goalpost involving Busy Beavers. Mine is: write a nroff document that executes at least one macro, and is a quine.
How would your views about AI change if that goal were achieved? When my personal goal was reached, I found myself a little bit at a loss for words.
That's a good question, one I thought of, but have put off grappling with.
Based on what LLMs have given me for answers so far, I'd look harder for the human-written source of the nroff code. I have written what I believe to be the only quine in the GPP macro processing language, LLMs only refer me to my code if I ask for a GPP quine. Google, Meta, OpenAI really have strip mined the entire web.
If I genuinely thought anything creative or new appeared, I'd probably be at a loss as well.
Re: Running out of places to move the goalposts to
#7Earlier quoted context omitted.
How would your views about AI change if that goal were achieved? When my personal goal was reached, I found myself a little bit at a loss for words.
That's a good question, one I thought of, but have put off grappling with. Based on what LLMs have given me for answers so far, I'd look harder for the human-written source of the nroff code. I have written what I believe to be the only quine in the GPP macro processing language, LLMs only refer me to my code if I ask for a GPP quine. Google, Meta, OpenAI really have strip mined the entire web. If I genuinely thought…
I gave a few attempts with ChatGPT and DeepSeek. Neither of them could get it right. So this goalpost can remain in place for the time being.
(I am assuming that the task is actually possible to accomplish. If it isn't possible, then it isn't a very good goalpost!)
Re: Running out of places to move the goalposts to
#8Earlier quoted context omitted.
That's a good question, one I thought of, but have put off grappling with. Based on what LLMs have given me for answers so far, I'd look harder for the human-written source of the nroff code. I have written what I believe to be the only quine in the GPP macro processing language, LLMs only refer me to my code if I ask for a GPP quine. Google, Meta, OpenAI really have strip mined the entire web. If I genuinely thought…
I gave a few attempts with ChatGPT and DeepSeek. Neither of them could get it right. So this goalpost can remain in place for the time being. (I am assuming that the task is actually possible to accomplish. If it isn't possible, then it isn't a very good goalpost!)
It should be possible. nroff macro language has looping, string interpolation, functions and if/then/else. That macro language should be turing complete. People have written file-infecting virus malware with it, I believe, which indicate that a quine should be possible. I personally have made several attempts at an nroff quine over the years with no success.
If it's not possible, I'd love to see an explanation, so that task can quite weighing on me.
Re: Running out of places to move the goalposts to
#9Earlier quoted context omitted.
I gave a few attempts with ChatGPT and DeepSeek. Neither of them could get it right. So this goalpost can remain in place for the time being. (I am assuming that the task is actually possible to accomplish. If it isn't possible, then it isn't a very good goalpost!)
It should be possible. nroff macro language has looping, string interpolation, functions and if/then/else. That macro language should be turing complete. People have written file-infecting virus malware with it, I believe, which indicate that a quine should be possible. I personally have made several attempts at an nroff quine over the years with no success. If it's not possible, I'd love to see an explanation, so th…
Here is an attempt:
.de Q
.nf
.na
.pso awk 'BEGIN{bs=sprintf("%c",92); pre=bs"&"} {out=pre; for(i=1;i nroff -U -Tascii quine.roff | sed -Ez '$ s/\n+$//'
Re: Running out of places to move the goalposts to
#10> AI seems to have caught up to my own intelligence even in those narrow domains where I have some expertise. What is there left that AI can’t do that I would be able to verify?
The last few days I've been working on some particularly tricky problems, tricky in the domain and in backwards compatibility with our existing codebase. For both these problems GPT 5.2 has been able to come to the same ideas as my best, which took me quite a bit of brain racking to get to. Granted it's required a lot of steering and context management from me as well as judgement to discard other options. But it's really getting to the point that LLMs are a good sparring partner for (isolated technical) problems at the 99th percentile of difficulty