Mystery AI Hype Theater is now available in podcast form!
@alexhanna
and I started this project as a one-off, trying out a new way of responding to and deflating AI hype... and then surprised ourselves by turning it into a series.
With the OpenAI clownshow, there's been renewed media attention on the xrisk/"AI safety" nonsense. Personally, I've had a fresh wave of reporters asking me naive questions (+ some contacts from old hands who know how to handle ultra-rich man-children with god complexes). 🧵1/
I'm happy to announce that our paper (with
@timnitgebru
& others) "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜" has been accepted to
#FAccT2021
>>
As a quick reminder: AI doomerism is also
#AIhype
. The idea that synthetic text extruding machines are harbingers of AGI that is on the verge of combusting into consciousness and then turning on humanity is unscientific nonsense. 2/
Okay, so that AI letter signed by lots of AI researchers calling for a "Pause [on] Giant AI Experiments"? It's just dripping with
#Aihype
. Here's a quick rundown.
>>
I refuse to be delegated to the "skeptics box" in someone else's framing of a debate. Here is my response to
@stevenbjohnson
's NYT Magazine article about LLMs and OpenAI.
On NYT Magazine on AI: Resist the Urge to be Impressed
If we ask "can AI do that?" we're asking the wrong question. A better question is: "Is this an appropriate use of automation?"
Here the answer was obviously no, and that was clear ahead of time:
There's a certain kind of techbro who thinks it's a knock-down argument to say "Well, you haven't built anything". As if the only people whose expertise counts are those close to the machine. I'm reminded (again) of
@timnitGebru
's wise comments on "the hierarchy of knowledge".>>
This is drastic and might well be leading to physical harm. Yet another example of why answer boxes/featured snippets/etc break the connection of words to their context --- a connection which is critical for human understanding.
#ethNLP
#NLProc
Facebook (sorry: Meta) AI: Check out our "AI" that lets you access all of humanity's knowledge.
Also Facebook AI: Be careful though, it just makes shit up.
This isn't even "they were so busy asking if they could"—but rather they failed to spend 5 minutes asking if they could.
>>
Great new profile of Dr.
@timnitGebru
in the Guardian.
“I’m not worried about machines taking over the world; I’m worried about groupthink, insularity and arrogance in the AI community.”
Was how she put it all the way back in 2016.
@nitashatiku
As I am quoted in the piece: “We now have machines that can mindlessly generate words, but we haven’t learned how to stop imagining a mind behind them”
>>
Imagine if the CEO of BP bragged that one in one thousand molecules flowing into the Gulf of Mexico came from the Deep Horizon oil rig.
Pollution of the information ecosystem is not something to be proud of.
I not infrequently see an argument that goes: "Making ethical NLP (or "AI") systems is too hard because humans haven't agreed on what is ethical/moral/right"
This always feels like a cop-out to me, and I think I've put my finger on why:
>>
DO NOT USE SYNTHETIC TEXT EXTRUDING MACHINES IN ANY SITUATION WHERE THE CONTENT MATTERS.
That absolutely includes health care. The idea that this is beneficial for people who we fail to provide adequate healthcare to is offensive.
Repulsive even before you see who's behind it
LLMs shouldn't be used to give medical advice. Who will be held accountable when things inevitably go sideways?
Also, this techno-saviorism crap is absolute BS -- helping "economically disadvantaged" people with AI is a myth.
This story (by
@nitashatiku
) is really sad, and I think an important window into the risks of designing systems to seem like humans, which are exacerbated by
#AIhype
:
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.
>>
TIL the
@APA
has guidelines on "how to cite ChatGPT". WTH APA?
They do suggest asking it for sources (?!) and then checking & citing those.
How about ... not treating text extruded by a synthetic media machine as suitable for research?
In summary, whenever someone is trying to sell predictive policing, always ask:
1. Why are we trying to predict this? (Answer seems to be so police can "prevent crime", but why are we looking to policy to prevent crime, rather than targeting underlying inequities?)
>>
I find this reporting infuriating, so I'm going to use it to create a mini-lesson in detecting
#AIhype
.
If you're interested in following this lesson, please read the article, making note of what you think sounds exciting and what makes you skeptical.
tl;dr blog post by new VP of AI at Halodi says the quiet parts out loud: "AI" industry is all about surveillance capitalism, sees gov't or even self- regulation as needless hurdles, and the movers & shakers are uninterested in building things that work. A thread:
I'm so tired of this argument. The "AI doomers" are not natural allies of the folks who have been documenting the real-world harms of so-called AI systems: discrimination, surveillance, pollution of the information ecosystem, data theft, labor exploitation.
>>
@emilymbender
@matteo_wong
@TheAtlantic
Why is so much effort focused on trying to set natural allies (people who are concerned about harms) against each other?
What's wrong with being concerned about more than one kind of harm at a time?
Ugh -- I'm seeing a lot of commentary along the lines of "'stochastic parrot' might have been an okay characterization of previous models, but GPT-4 actually is intelligent."
Spoiler alert: It's not. Also, stop being so credulous.
>>
"There's no way a non-industry person can understand" On the contrary --- the folks in industry have shown themselves incapable of understanding (or caring about) the impacts of people of their tech. Regulation should protect rights and should be made by policymakers.
WATCH: Former Google CEO
@ericschmidt
tells
#MTP
Reports the companies developing AI should be the ones to establish industry guardrails — not policy makers.
“There’s no way a non-industry person can understand what’s possible.”
Hey
#linguists
let's make an
#AcademicValentines
thread. Here's a start (mine from 2018):
Roses are red
Violets are blue
Language variation is natural
Your speech is a dialect too
Ready for discussions of
#ethNLP
and ethics review at NLP/ML etc conferences? Don't forget your bingo card! (With
@KarnFort1
on the TGV from Perpignan to Paris).
i asked SARAH, the World Health Organization's new AI chatbot, for medical help near me, and it provided an entirely fabricated list of clinics/hospitals in SF. fake addresses, fake phone numbers.
check out
@jessicanix_
's take on SARAH here:
via
@business
When your business model relies on theft and you don't like proposed regulations that would expose that theft ... that's a pretty good sign the regulations are on the right track.
#OpenAI
#AIAct
MSFT lays off its responsible AI team
The thing that strikes me most about this story from
@ZoeSchiffer
and
@CaseyNewton
is the way in which the MSFT execs describe the urgency to move "AI models into the hands of customers"
>>
Journos working in this area need to be on their guard & not take the claims of the AI hypesters (doomer OR booster variety) at face value. It takes effort to reframe, effort that is necessary and important. We all must resist the urge to be impressed: 4/
There is 0 reason to expect that language models will achieve "near-human performance on language and reasoning tasks" except in a world where these tasks are artificially molded to to what language models can do while being misleadingly named after what humans do.
Reading a critique of my paper with
@timnitGebru
(et al) which claims that if you do science (or scholarship) from the point of view that white supremacy is bad, then you have to make that point of view explicit lest any readers who disagree "mistake it for science". I can't even
Instead: They're about concentration of power in the hands of people, about reproducing systems of oppression, about damage to the information ecosystem, and about damage to the natural ecosystem (through profligate use of energy resources).
>>
Please don't get distracted by the dazzling "existential risk" hype. If you want to be entertained by science fiction, read a good book or head to the cinema. And then please come back to work and focus on the real world harms and hold companies and governments accountable. /fin
"Dylan Patel, chief analyst at the semiconductor research firm SemiAnalysis, estimated that a single chat with ChatGPT could cost up to 1,000 times as much as a simple Google search."
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:
/1
Pro-tip: If someone claims a model is trained on "the entire internet" they don't actually know what they're talking about and don't understand a) data b) dataset documentation and therefore c) how to reason about the model they're describing.
>>
To all those folks asking why the "AI safety" and "AI ethics" crowd can't find common ground --- it's simple: The "AI safety" angle, which takes "AI" as something that is to be "raised" to be "aligned" with actual people is anathema to ethical development of the technology.
>>
When the AI bros scream "Look a monster!" to distract everyone from their practices (data theft, profligate energy usage, scaling of biases, pollution of the information ecosystem), we should make like Scooby-Doo and remove their mask.
2022 me would like to warn 2019 me that ML bros' arguments about language models "understanding" language were going to mutate into arguments about them being "sentient" and having "internal monologues" etc.
The new "AI is going to kill us all!!1!" letter is a wall of shame—where people are voluntarily adding their own names.
We should be concerned by the real harms that corps and the people who make them up are doing in the name of "AI", not abt Skynet.
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.
>>
I wonder if we could put together a AI hype tracker or AI hype incident database + visualization, that could help expose the corporate & other motives behind a lot of this.
>>
"I thought ChatGPT was a search engine".
It is NOT a search engine. Nor, by the way are the version of it included in Bing or Google's Bard.
Language model-driven chatbots are not suitable for information access.
>>
Judge Castel: Did you ask Chat GPT what the law was, or only for a case to support you? It wrote a case for you. Do you cite cases without reading them?
Schwartz: No.
Judge Castel: What caused your departure here?
Schwartz: I thought Chat GPT was a search engine
t the same time, it serves to suggest that the software is powerful, even magically so: if the "AI" could take over the world, it must be something amazing. 3/
Just turned down an invitation to do a "pro bono" (their words!) talk to a private conference for VCs and CTOs. I'm all for making academic conferences widely accessible, including keeping costs low, but for an industry gig? No way.
@alexhanna
"AI" is not "good at writing"—it's designed to produce plausible sounding synthetic text. Writing is an activity that people to do as we work to refine our ideas and share them with others. LLMs don't have ideas. 14/
Linguistics as a field has a lot to contribute to better understanding what large language models can and can't do and yet many don't think to turn to linguists (or don't even really know what linguists do) when trying to evaluate claims about this technology.
>>
Can we talk about the phrase "top AI researchers" and how it devalues the expertise of anyone whose primary focus isn't on the internals of the systems themselves?
Here's a cute example, due to Itamar Turner-Trauring (
@itmarst
@hachyderm
.io), who observes that Google gave bad results which were written about in the news—which the new GPT-Bing used as reliable answers. Autogenerated trash feeding the next cycle, with one step of indirection.
You know what's most striking about this graphic? It's not that mentions of people/cities/etc from different continents cluster together in terms of word co-occurrences. It's just how sparse the data from the Global South are.
Do language models have an internal world model? A sense of time? At multiple spatiotemporal scales?
In a new paper with
@tegmark
we provide evidence that they do by finding a literal map of the world inside the activations of Llama-2!
Nothing we currently have, not LaMDA, not GPT-3, not DALL-E, none of it, is actually "an AI" in the sense that that phrase evokes for most people. I'm one of the interviewees in this podcast and I try to make that clear.
A Google engineer claims an AI chatbot called LaMDA is sentient - though Google says there's no evidence to support this.
But what exactly is LaMDA and how does it work?
Is AI capable of felt experience?
Find out more on The Inquiry podcast.
@sama
@OpenAI
That is, the very people in charge of building
#ChatGPT
want to believe SO BADLY that they are gods, creating thinking entities, that they have lost all perspective about what a text synthesis machine actually is.
>>
Ever found the discourse around "intelligence" in "A(G)I" squicky or heard folks pointing out the connection w/eugenics & wondered what that was about?
History of it all can be found in this excellent talk by
@timnitGebru
(w/ co-author
@xriskology
)
OpenAI: We refuse to tell you what's in the training data for GPT-4, for "safety"
Also OpenAI: We're throwing 10x $100k at "experiments in democratic process" for determining rules AIs should follow.
You don't get to regulate your own business.
I’ve picked up a bunch of new followers in the past few days, and I suspect many of you are here because you’re interested in what I might have to say about Google and Dr.
@TimnitGebru
. So, here’s what I have to say:
This is a particularly stark example of where the ML/big data approach of just grabbing everything that is accessible ignores various norms, structures, social contracts. Importantly, it's not the only such example.
Are there any other fields that simultaneously conceive of themselves as 'solving the world's problems' and also see no need for any due diligence around mitigating potential harmful impacts of what they do/build, or is it just CS?
Also, it's kind of hilarious (lolsob) that OpenAI is burning enormous amounts of energy to take machines designed to perform calculations precisely to make them output text that mimics imprecisely the performance of calculations… & then deciding that *that* is intelligent. 16/
"We urge policymakers to instead draw on solid scholarship that investigates the harms and risks of AI—and the harms caused by delegating authority to automated systems" w/
@alexhanna
in
@scientificamer
New
@Meta
privacy policy just dropped. "We sell the ability to target advertising based on information we gather about you", but somehow that's consistent with "We do not sell and will not sell your information". Specific sense of "sell" or "information" or both?
4. The authors acknowledge some of the ways in which predictive policing has "stirred controversy" but claim to have "demonstrate[d] their unprecedented ability to audit enforcement biases". >>
False. Because even if a person tells you something false, they're still saying it for a reason and you can learn from that. But what's the value in stochastic remixes of unknown training data tuned towards "this makes raters happy"?
In general, ChatGPT is more useful in terms of pointing me in the direction of ideas and facts than in giving me the final word on them. It's like asking a human friend to explain stuff -- sometimes they get it wrong, but you get exposed to knowledge.
Those "enforcement biases" have to do with sending more resources to respond to violent crime in affluent neighborhoods. They claim that this would allow us to "hold states accountable in ways inconceivable in the past".
>>
Just sayin': We wrote a whole paper in late 2020 (Stochastic Parrots, published in 2021) pointing out that this head-long rush to ever larger language models without considering risks was a bad thing. But the risks and harms have never been about "too powerful AI".
>>
The tendency of AI researchers to equate the form of an artifact with its meaning is seemingly boundless. A college degree is not comprised of essays and exam papers, even if such elements play a key role in our evaluation of human progress towards one.
The editor replied quickly and took out the quote ... and admitted that they'd prompted an LLM to generate the piece. I've got a few takeaways from this experience:
Has anyone heard of the news site Biharprabha? They ran an article today with a fabricated quote attributed to me. I've emailed the editor in the hopes of getting it taken down, but we'll see.
We're seeing multiple folks in
#NLProc
who *should know better* bragging about using
#ChatGPT
to help them write papers. So, I guess we need a thread of why this a bad idea:
>>
Whose labor is being exploited? How is mass surveillance being extended and normalized? What are the impacts to the natural environment and information ecosystem? 26/
When you're Associate Director of something called "Human-Centered Artificial Intelligence" but the $$ all comes from Silicon Valley so you feel compelled to retweet the clown suggesting that the poors should have LLM-generated medical advice instead of healthcare.
Do we have a name for this rhetorical move/fallacy?
A: AI Hype! My system can do X!
B: No, it can't. Here's why.
A: So you think no computer could ever do X?
-or-
A: But what about future versions of it that could do X?
It's super common, and it feels like it should be named.
Policymakers: Don't waste your time on the fantasies of the techbros saying "Oh noes, we're building something TOO powerful." Listen instead to those who are studying how corporations (and govt) are using technology (and the narratives of "AI") to concentrate and wield power.
>>