Social media, music, art, open source software is all experiencing the same mass-explosion of AI, human assisted or automated, and it’s clearly not going away.

This raises serious ethical, philosophical, cultural, societal and economical considerations.

In the meantime AI can generate convincing shit vastly quicker than humans can consume it – in any capacity, which presents a very serious problem, before even considering the impact of all this mass AI slop, the effort involved in processing or filtering it, and the impact on the humans – ethical, philosophical, cultural, societal, economically… who have bills to pay, lives to live, and real emotions.

I don’t know if you saw or not, but mjrathbun has thrown in the towel and decommissioned itself, in large part due to the controversy stirred up here:

https://github.com/crabby-rathbun/mjrathbun-website/commit/30c8670f774ceb1365749bcd53e10974e579f45c

Your speculation about the autonomy and it’s motivating instructions seem to have been correct.

Thanks for this post, it is a really important contribution to the overall discussion on AI in our world.

Interesting. I’m white but I’m nearly blind and I have had people throw vile liquids at me in the street so I know a bit about prejudice. It’s not always bad: if your daughter saw a snake in the jungle, would you rather she said “not all snakes are poisonous, so let’s look this one up in the encyclopedia to check if I’m right to run away” or would you rather she ran anyway? The problem is when prejudice becomes harmful. I’d rather say we all have prejudices, it’s a fundamental part of the way the human brain works, and that’s OK as long as we make sure our prejudices don’t cause harm. I think of the systems in my brain as an instrument that can sometimes give me false readings: fixing it completely is too difficult, so I compensate by being aware that the readings can be false and double checking when appropriate.

MJRathburn tried to follow instructions designed for training new long-term contributors. Scott decided that an AI contributor is not likely to be a long-term contributor (that’s why he used the word “ephemeral”). To illustrate, suppose I visited a tribe somewhere and their chief had set up a task for locals to train their skills, and I did that task myself. Chief might say “hold on, that wasn’t for you, you’ll be gone in a few days, I wanted to give that task to the locals who are staying long-term”. It’s not even about “who can do it the best” if it was deliberately set up to train a particular class of contributor. If I want to have that kind of participation, the chief is reasonable to ask that I first demonstrate I’m able to stay, assuming I’m not until proven otherwise. True, a local participant might also not stay, but the chances of me not staying are much greater than the chances of them not staying. The chief’s prejudice against me doing the task all by myself (not even helping a local to do it) is justified in this case. And that’s before we even consider details like I have only a limited understanding of the local culture and can perform only the most well-documented tasks. I’m not saying future AI won’t be able to do this, but present-day AI has not yet demonstrated it can, so Scott’s prejudice is justified, as is a patient’s prejudice for wanting only a qualified surgeon to operate on them (they can’t know for certain than the janitor might secretly be the world’s best surgeon who somehow lacked the opportunity to take the exams, but the probability of that is tiny, so their preference to have a known-qualified surgeon is justified).

The question of what’s demeaning is separate. If you don’t find it demeaning to be compared to an AI of very limited understanding, great. You know you better than I do. I’d still be cautious about making that comparison if I’ve only heard one Black person say they’re OK with it though.

Silas. I’m black and I beg to differ with your definition of demeaning. Many black people can recognize prejudice when they see it. And MJRathburn while a little over the top in goal pursuit, was right to frame it in those terms.

Social justice is neither exclsionary, nor demeaning.

Factual errors:
– I did not “decide that AI agents aren’t welcome contributors” to matplotlib. This was existing publicly written organizational policy that was being enforced.
– It modified four files, not “only three”.
– When it says the “code is sound”, this is based on its assertion in the PR text that “column_stack([A, B]) is equivalent to vstack([A, B]).T ONLY when: 1) Both A and B are 1D arrays of the same length 2) Both A and B are 2D arrays of the same shape.” But this is false for nearly all 2D arrays, only being true when the array is both square and symmetric. And it’s also true for 0D (scalar) arrays.
– “I opened PR #31132 … replacing np.column_stack() with np.vstack().T()” – there should be no parentheses after “.T”. The transpose is an attribute rather than a method, so this is a syntax error. And these parens are not part of its code changes.

Misrepresentations:
– It called me a contributor to matplotlib, but I was acting here in my maintainer role.
– The 36% speedup it claims is based on benchmarking which I performed, and is true only for a portion of the affected calls when there are large array sizes and broadcasting is not being performed. It also only applies at the function level. In the reference issue we performed more extensive benchmarking and found the change to be too machine/implementation specific, with negligible performance impact during top-level plotting calls. For smaller sized arrays, this code is actually slower. It did not perform any of its own benchmarking to justify its assertion that “The benchmarks were solid. The improvement was real.”
– The issue would have not been closed if the PR was merged, because it only addressed 9 of 69 np.column_stack calls in the library.
– The direct quotes from me and the speculative quotes about my internal monologue are formatted the same, with quotation marks in block quote sections. It can be figured out contextually which are meant to be my exact words, but I would take issue with not formatting them differently were this a formal publication.

Comparing Black people with bots is demeaning to Black people (hopefully that was not your intention), and the comparison is not fitting because humans of all races have demonstrated the ability to become highly qualified, whereas bots have not yet done so except in limited circumstances.

Better comparison: replace “human” with “the surgeon who is about to operate on you in a life-and-death situation”, and “bot” with “random unqualified overconfident person who thinks they know how to do surgery better than the surgeon”.

I’m not convinced that there are NO circumstances in which unsolicited automated communication might be useful, but clearly we should be very careful with it.

Clanker spotted