Running out of places to move the goalposts to

Running out of places to move the goalposts to(nickdrozd.github.io)

4 points by nickdrozd 138 days ago | 13 comments

andy99 138 days ago |

I 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

nickdrozd 138 days ago | |

What would be a better way to measure intelligence?

judahmeek 137 days ago | | |

I like arcprize.org's approach.

mulmboy 137 days ago |

> 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

judahmeek 137 days ago | |

You steered a sycophantic LLM to the same idea that you had already had & think that's worth bragging about?

mulmboy 136 days ago | | |

I'm well ware that they can be sycophantic, and I structure things to avoid that like asking "what do you think of this problem" and seeing the idea fall out rather than providing anything that would suggest it. In one of these two cases it took an idea that I had inkling of, fleshed it out, and expanded it to be much better than I had.

And I'm not bragging. I'm expressing awe, and humility that I am finding a machine can match me on things that I find quite difficult. Maybe those things aren't so difficult after all.

By steering I mean more steering to flesh out the context of the problem and to find relevant code and perform domain-specific research. Not steering toward a specific solution.

bediger4000 138 days ago |

The article mentions a personal goalpost involving Busy Beavers.

Mine is: write a nroff document that executes at least one macro, and is a quine.

nickdrozd 138 days ago | |

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.

bediger4000 138 days ago | | |

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.

.de Q .nf .na .pso awk 'BEGIN{bs=sprintf("%c",92); pre=bs"&"} {out=pre; for(i=1;i<=length($0);i++){c=substr($0,i,1); if(c==bs) out=out bs bs; else out=out c} print out}' "\n[.F]" .ex .. .Q