This article in the Atlantic by Stephen Marche is so full of #AIhype it almost reads like a self-parody. So, for your entertainment/education in spotting #AIhype, I present a brief annotated reading:
Straight out of the gate, he's not just comparing "AI" to "miracles" but flat out calling it one and quoting Google & Tesla (ex-)execs making comparisons to "God" and "demons".
/2
This is not the writing of someone who actually knows what #NLProc is. If you use grammar checkers, autocorrect, online translation services, web search, autocaptions, a voice assistant, etc you use NLP technology in everyday life. But guess what? NLP isn't a subfield of "AI".
/3
Here's the author is claiming to have inside knowledge of some "esoteric" technology development that, unbeknownst to the average human, is going to be very disruptive. But note the utter lack of citations or other grounding for this claim.
/4
Okay, agreed on fake-it-til-you-make-it, but "direct thrust at the unfathomable" and "not even the engineers understand" are just unadulterated hype. If they don't understand how it works, how are they even measuring that it works?
/5
Protip: They aren't really. The capabilities that the AI boosters claim to have built are ones that we don't have effective benchmarks for, & actually can't, in principle. See: AI and the Everything in the Whole Wide World Benchmark by @rajiinio et al /6
@cephaloponderer@alexhanna@amandalynneP@bendee983 Okay, back to the hype. This is weirdly ominous and again provides no supporting evidence. You can't see it, but that doesn't mean it isn't there ... is not an argument that it is!
/8
This is kinda fun, because I was musing a few weeks ago about how we don't usually go to "superhuman" for other tools. And it does sound ridiculous, doesn't it?
If you don't know how something works, but can test that it works (w/certain degree of reliability), then it is usable. It's true that deep learning is opaque on the how. But we can't let any engineers off the hook in terms of testing the functionality of their systems.
/10
"What technologists call 'parameters'" makes this sound so ominous and mysterious. Our "little animal brains" have ~86 billion neurons (source: brainfacts.org/in-the-lab/mee…). So not a different scale (and with much more complexity).
/11
More to the point: None of this is inevitable. DL systems aren't naturally occurring phenomena that we can try to understand or just stand in awe of. They are things we are building & choosing to use. We can choose not to, at least w/p sufficient testing for each use case.
/12
Also, because it feels gross to compare language model parameters to human neurons, I want to plug again this great article by @AlexBaria and @doctabarz on the computational metaphor.
Back to Marche: I don't think we should necessarily believe the people who got super rich off of surveillance capitalism when they say "oh noes, can't regulate, it would stop the development of the technology".
/14
Again, whether or not we try to build this (and with what regulatory guardrails) is a CHOICE. But also: it would be pretty easy with today's stochastic parrots to sometimes at least get an answer like that. (While other times getting hate speech...)
/15
Uh, just because you put these things in a list does not make them all the same kind of thing ("language game").
/16
Yeah, just because the people who built the thing say it does something "in a ways that's not dissimilar from the way you and I do" doesn't make it true. Do they have the expertise to evaluate that? How did they evaluate that?
/17
Oh, and again, while "contemporary NLP" does use neural LMs for a lot of things, I wouldn't say it "derives" from them. There is more to the field than just throwing neural nets are poorly conceived tasks.
/18
What comes next is some GPT-3 authored additional hype, stating with the prompt "And if AI harnesses the power promised by quantum computing," Marche does acknowledge it (in the following paragraph). He is also responsible for deciding to include it.
/19
(Note that Marche also doesn't tell us how many tries he took to get the one he chose to include.)
/20
It's not doing any of these things, actually. Having synthetic text in the style of someone who has died is not bringing them back from the dead. I'm not sure what an "imitation" of consciousness is, nor how it would benefit us.
/21
And it is certainly not "piercing the heart of how language works between people".
On how LM-geneated text is nothing like human linguistic behavior, see Bender & Koller 2020 and also this episode of Factually!
And one last screencap before I end. Where is the evidence for any of these claims? None is provided.
/23
So, I hope that was enjoyable and/or informative. I give this one #threemarvins. Could 2022 be the year of peak #AIhype? That sure would be nice. 24/24
Postscript 1: Important additional info on the (sigh) comparison of 100B parameter networks to human brains
This piece is stunning: stunningly beautifully written, stunningly painful, and stunningly damning of family policing, of the lack of protections against data collection in our country, & of the mindset of tech solutionism that attempts to remove "failable" humans decision makers
.@UpFromTheCracks 's essay is both a powerful call for the immediate end of family policing and an extremely pointed case study in so many aspects of what gets called #AIethics:
1. What are the potentials for harm from algorithmic decision making?
>>
2. The absolutely essential effects of lived experience and positionality to understanding those harms. 3. The ways in which data collection sets up future harms.
Read the recent Vox article about effective altruism ("EA") and longtermism and I'm once again struck by how *obvious* it is that these folks are utterly failing at ceding any power & how completely mismatched "optimization" is from the goals of doing actual good in the world.
>>
In Stochastic Parrots, we referred to attempts to mimic human behavior as a bright line in ethical AI development" (I'm pretty sure that pt was due to @mmitchell_ai but we all gladly signed off!) This particular instance was done carefully, however >>
@mmitchell_ai Given the pretraining+fine-tuning paradigm, I'm afraid we're going to see more and more of these, mostly not done with nearly the degree of care. See, for example, this terrible idea from AI21 labs:
@mmitchell_ai As Dennett says in the VICE article, regulation is needed---I'd add: regulation informed by an understanding of both how the systems work and how people react to them.
Thinking back to Batya Friedman (of UW's @TechPolicyLab and Value Sensitive Design Lab)'s great keynote at #NAACL2022. She ended with some really valuable ideas for going forward, in these slides:
Here, I really appreciated 3 "Think outside the AI/ML box".
>>
As societies and as scientific communities, we are surely better served by exploring multiple paths rather than piling all resources (funding, researcher time & ingenuity) on MOAR DATA, MOAR COMPUTE! Friedman points out that this is *environmentally* urgent as well.
>>
Where above she draws on the lessons of nuclear power (what other robust sources of non-fossil energy would we have now, if we'd spread our search more broadly back then?) here she draws on the lessons of plastics: they are key for some use case (esp medical). >>
Some interesting 🙃 details from the underlying Nature article:
1. Data was logs maintained by the cities in question (so data "collected" via reports to police/policing activity). 2. The only info for each incident they're using is location, time & type of crime.
>>
3. A prediction was counted as "correct" if a crime (by their def) occurred in the (small) area on the day of prediction or one day before or after.