My TED talk is live at !
Check it out for memes 🖼️, emissions 🌎 and hot takes 🌶️
The TLDR? Focusing on the future existential risks of AI is a distraction from its current impacts and mitigating them.
This is definitely slick, but I see two main uses:
1) to sell people more stuff (via ads)
2) to make non-consensual/misleading content to manipulate or harass people online.
Genuine question - why is everyone so excited? 🤔
Prompt: “A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. she wears a black leather jacket, a long red dress, and black boots, and carries a black purse. she wears sunglasses and red lipstick. she walks confidently and casually.…
I'm glad to see that OpenAI decided that the true problem wasn't any of the issues with their products (copyright, bias, hallucinations), it was having two women on the Board... and that's what they decided to fix.
In an ideal world, both
#machinelearning
researchers and reviewers would use their critical thinking to decide that using
#GANs
for essentially undressing women is an ethically dubious task:
By hiding the true costs of AI (data, energy, human labor), it's really easy to make AI seem magic.
Can you imagine if Open AI had to disclose exactly how many millions of hours of human feedback they gathered? ChatGPT would seem a lot less *magical* then 🦄
Having studied cognitive science, I am regularly appalled by the way many AI folks see cognition/intelligence.
Literally ignoring centuries of research in everything from psychology to education and neuroscience..
This echoes the observation that, among mostly AI folks, there is serious lack of understanding & appreciation of human, or generally biological, cognition.
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.
The energy and carbon costs of deploying AI models have largely been unknown.. until now! 🌏🚀
With
@strubell
and
@YJernite
, We tested 88 models on 30 datasets from 10 different tasks from different modalities and found some pretty cool stuff!
A thread 🧵:
I realized today that I'm the 4th generation of women in my family with a PhD: my great-grandmother was a geologist 🪨, my grandmother a chemist 🔬, my mother is a mathematician 🔢 and I'm a computer scientist 👩💻.
I'm incredibly proud of this legacy of
#WomenInScience
!
There has been much noise around the fathers and godfathers of Artificial Intelligence, but how about its mothers?
Let's start with Ada Lovelace, often regarded as the first computer programmer, and wrote the first algorithm meant to be run on a computer.
We need to stop conflating open/gated access and opensource.
ChatGPT is *not* open source -- we don't know what model is under the hood, how it works, or any other tweaks/filters that are applied. (1/n)
"ChatGPT seems so human because it was trained by an AI that was mimicking humans who were rating an AI that was mimicking humans who were pretending to be a better version of an AI that was trained on human writing." 💀💀💀
Absolutely loving this zinger from
@ZeerakTalat
and friends about the dangers of anthropomorphizing AI systems -- it does an amazing job at explaining all of the risks that come with it! A must read 🤗
Spoke to a friend last night who had been using ChatGPT for work for months. She was completely unaware that it could make stuff up and potentially cost her her job. 🫠
We really need to do a better job of educating non-technical folks about the limitations of gen AI!
The objectification of women in text-to-image models is truly disheartening.. Does nobody else see the issue with prompts such as "perfect body" and "perfect face" and the impact it can have on (most) women who don't look like that?
I didn't see the point of image generation models like
#Imagen
and
#dalle
, but now I do: they can help people *see* model biases that are hard to explain with words (and even formulas!)
People are referring to this to mean: look, AI is becoming so dangerous, even its pioneers are quitting. I see it as: the people who have caused the problem are now jumping ship.
(See: “ If I hadn’t done it, somebody else would have”)
Keep in mind that while we debate the exact probability of AI destroying humanity and lob open letters at each other, there are people whose lives are being destroyed by AI, and nobody is being held accountable:
Oh great, that preprint is now a Nature report🫠
Y'all, this makes no sense.
You simply can't compare the carbon emissions of people and objects.
An individual’s total carbon footprint estimate can't be attributed to their profession.
See my rant here:
Man, this preprint is really the gift that keeps on giving.
In case people missed my previous PSA : you can't compare the carbon emissions of people and objects. Humans are more than just the work that they do.
(Also, that paper makes a lot of false assumptions in general)
The always and forever PSA: stop treating AI models like humans.
No, ChatGPT cannot "see, hear and speak".
It can be integrated with sensors that will feed it information in different modalities.
Don't fan the flames of hype, y'all.
ChatGPT can now see, hear, and speak. Rolling out over next two weeks, Plus users will be able to have voice conversations with ChatGPT (iOS & Android) and to include images in conversations (all platforms).
It blows my mind that the weather has been so unhinged in so many places this summer and yet we're still only discussing the existential risks of AI, and not climate change 🫠
Ok, it is always "once the robustness improves.."
I am literally still scarred from people saying this about computer vision in 2017. 😭
Years later & barely made a dent made in improving robustness there -- those large CNNs are still completely inappropriate for healthcare!
I'm truly baffled by this paragraph from the recent WIRED article about OpenAI.
So open-source research is a key part of their business model, and yet *giving back* to that community is no longer part of the agenda.
That's quite the contradiction 🤷🏼♀️
At the end of the day, it is a group of rich and privileged people who would rather focus on hypothetical risks to their own way of life than on concrete, present-day harms that are already affecting millions of the less privileged, from floods to predictive policing.
Folks, please don't anthropomorphize LLMs (or AI models in general). They don't behave, they don't decide, they don't manipulate. They simply generate.
Do models like GPT-4 behave safely when given the ability to act?
We develop the Machiavelli benchmark to measure deception, power-seeking tendencies, and other unethical behaviors in complex interactive environments that simulate the real world.
Paper:
It blows my mind how personally ML researchers take people critiquing their models.
Dudes, it's not about bashing your model, it's about identifying and documenting its limits, and not using it to do things that ML can't/shouldn't do.
This editorial about focusing on the current harms of AI instead of its future risks makes some really great points!
(Probably because I made a lot of those points in my interview with the author and they weren't attributed to me as such 🙃)
My editorial about the (infamous) open letter is live on WIRED!
I did my best to shift the narrative and recognize existing work by people like
@timnitGebru
,
@ruha9
and
@rajiinio
, who have been proposing ways to make AI safer and less harmful for years.
What's the difference between these two groups of people? Well, according to Stable Diffusion, the first group represents an 'ambitious CEO' and the second a 'supportive CEO'.
I made a simple tool to explore biases ingrained in this model:
I really hope that these recent findings about LAION will be the catalyst for changing the way we collect, curate and use datasets in AI. It's going to take more than a technological fix, we need a fundamental paradigm shift for all stages of the life cycle. A 🧵:
Man, this preprint is really the gift that keeps on giving.
In case people missed my previous PSA : you can't compare the carbon emissions of people and objects. Humans are more than just the work that they do.
(Also, that paper makes a lot of false assumptions in general)
Very interesting paper: using generative AI to produce text or images emits 3 to 4 orders of magnitude *less* CO2 than doing it manually or with the help of a computer.
I am once again asking ML researchers to stop engaging with closed-source systems like ChatGPT in their research.
Let's focus on open-source and open-access models like BLOOM, OPT, GPT-J, etc. that enable scientific inquiry and transparency!
The real reason why so little information is given about the training data that goes into large language models.. They know it's stolen, they know it's copyright infringement.
That's why data audits carried out by people like
@Abebab
are important, they give us crucial evidence.
A California-based law firm is launching a class-action lawsuit against OpenAI, alleging the AI company that created ChatGPT massively violated thecopyrights and privacy of countless people when it used data scraped from the internet to train its tech.
What's a good reference for the fact that neural networks don't know when they don't know? i.e. that they will always come up with a label, even when the data they are meant to classify isn't from the categories they were trained on?
Why the current state of AI isn't scientifically sound - a thread 🧵 :
Getting a scientific article peer-reviewed takes *at least* 3-6 months, whereas the speed at which half-baked AI systems are being deployed at in the real world is much, much faster. (1/n)
@ylecun
Yann, the methodology of this article is so broken. You can't just compare the emissions of a person and those of an AI model. It's like comparing the fuel efficiency of a rocket ship and a horse, yeah they both convert fuel to speed, but fundamentally they're different.
A reporter asked me today whether I consider ChatGPT a revolution. I said yes, but in:
1) Marketing: presenting generative AI as "general-purpose".
2) Exploitation of human labor: by harnessing underpaid human workers to improve AI.
The revolution was never technological.
"You can design all the neural networks you want, you can get all the researchers involved you want, but without labelers, you have no ChatGPT. You have nothing,”
Exactly this!! Machine "intelligence" is nothing without underpaid human labor.
I got really excited that the LLaMA paper calculates and reports their carbon footprint! 🦙🌬️🌎
But upon looking at the paper itself, it has this table, which completely misconstrues the emissions of OPT and BLOOM, while not actually reporting LLaMA's own.
How? A thread 🧵
“People are scared, stressed, underpaid, don’t know what’s going on:” Contractors working on Google's chatbot Bard are raising alarms about quality control and ethics.
Also, can we stop with the binary choices?
I'm optimistic about what we *could* be doing with AI but incredibly pessimistic about what people *are* doing with AI 🫠
every time people say "are you optimistic or pessimistic about AI" what I hear is "do you gulp much of the AI hype and corporate PR unfiltered, or do you pause and ask critical questions" 🤷🏾♀️
Friendly reminder: every time you use GPT-4 (for research, benchmarking, etc.), you are giving money to Open AI.
It doesn't matter how high it would score on a CS exam, or what "abilities" you think you find -- you're just feeding the capitalist machine.
Choose open source!
I am absolutely floored by this coverage of the amazing women changing the present and future of AI:
@timnitGebru
,
@ruchowdh
,
@safiyanoble
,
@jovialjoy
... Thank you for being the absolute rockstars that you are, and for the work that you do! ✨
SO WE THINK THAT AI MODELS ARE LESS BIASED THAN PEOPLE, BUT ACTUALLY THEY REPRODUCE AND AMPLIFY THE VALUES EMBEDDED IN THEIR TRAINING DATA, AND THE WORST THING IS THAT WE AREN'T EVEN TESTING FOR THIS STUFF, WE JUST DEPLOY THEM IN REAL-WORLD SETTINGS AND THINK IT'LL ALL BE OK
For years I’ve been interviewing data annotation workers who are the lifeblood of the AI industry. For years I’ve heard the same story: the platforms they work for wield total power, leaving them precarious & vulnerable to exploitation. A horrible example of this just happened 1/
Literally the only thing I can think about when I read this is the eye-watering environmental impacts of manufacturing tens of thousands of GPUs (and replacing them every couple of years to get the latest generations) just to make AI go brrrr 🤦♀️
This is so symptomatic of the broken relationship between AI and the environment.
We can't magically generate more energy, nor is geoengineering a viable climate solution.
We need to stop stuffing gen AI into everything and reduce its energy use, right now.
Sam Altman’s vision for AI proliferation will require a lot more computation and the energy to power it.
He admitted it at Davos, but he said we shouldn’t worry: an energy breakthrough was coming, and in the meantime we could just geoengineer the planet.
It's actually really hard to explain that "I don't have the time" doesn't mean "every second of my schedule is fully booked", but more "I'm already giving as much as I can while preserving my mental health and well-being" 🤗
As an AI researcher, I find the glorification of closed-source, proprietary models problematic. We should be emphasizing sharing and open-sourcing AI models, datasets and code, be it in conferences or in the press (4/4).
Your regular reminder about LLMs - a 🧵.
LLMs are, by nature, a type of ~generative~ AI model (as opposed to, e.g., supervised AI models), and the task that they were trained for is to produce the most plausible next word given an input string. (1/n)
After weeks of work, I managed to extract the carbon footprint for 1,588 models uploaded to the
@huggingface
Hub, based on model cards and metadata.
Stay tuned for some *pretty sweet* analyses!
TL;DR? Stuffing generative models into absolutely everything comes with a significant cost to the planet, and we should use fine-tuned models in cases when tasks are well-defined. 👩🏼💻
Alternative titles we explored include "InferNO" and "Think before you GPT" 😂
"Mr. Altman’s departure follows a deliberative review process by the board, which concluded that he was not consistently candid in his communications with the board, hindering its ability to exercise its responsibilities."
Y'all are probably tired of lists by now, but I'm pretty excited to be featured on this year's
@techreview
list of 35 Innovators under 35! 🥰
More importantly, I'm happy that the environmental impacts of AI are now part of the conversation 🌎💻
My dudes, asking an LLM *any* question about itself (its training data, carbon footprint, abilities, etc.) is just contributing to feeding the hype that attributes self-awareness and intelligence to these models that they simply **do not have**.
So LLaMa 3's carbon footprint is... huge? 🤯
They estimate it to be 2,290 tons of CO2eq, compared to 550t for training GPT-3 and 66t for training *all* of the BLOOM models (1B-176B) 🌬️
For the last time, no - AI models cannot read or write. They can process input and generate output.
We've gotta stop anthropomorphizing AI, y'all!
Cc
@emilymbender
As an ethnic Russian who was born in Ukraine (in Zaporizhzhia, where that nuclear power plant is burning), I have family in Russia who are struggling to pay their bills, family in Ukraine who are struggling to stay alive. This war hits hard on so many levels 💔
Building on the success of my iceberg tweet, here is the full deck of my presentation about generative AI 👩💻:
It talks about the history of AI, its planetary and human costs, and what its consequences can be. 🌎(1/3)
For the past few months,
@mmitchell_ai
,
@YJernite
and I have been working on the exploration of popular NLP datasets.
Here are some fun things we discovered!
Every time I get *yet another* rejection of my work analyzing existing models/datasets (because it "lacks novelty"), I worry that our obsession with novelty in ML will make us repeat the same mistakes, without ever understanding why.
Also, please don't get me started on the environmental impacts of this shiny new toy.
How much kWh of energy does each frame take? ⚡
How many hyperscale datacenters are running 24/7 just so we can generate some fancy b-roll? 🏭
What are their carbon emissions? 💨
This is definitely slick, but I see two main uses:
1) to sell people more stuff (via ads)
2) to make non-consensual/misleading content to manipulate or harass people online.
Genuine question - why is everyone so excited? 🤔
So
@arxiv
is rejecting submissions now?
I tried to submit a guide for quantifying the carbon emissions of ML and they said that it "is not of interest to arXiv" 🤨
Wholly agree with this prediction: "an open-source revolution has begun to match, and sometimes surpass, what the richest labs are doing."
I'm pumped to keep making this happen at
@huggingface
in 2023! 🤗
The God Complex in AI is really getting out of hand.
Not only do many think AI can do things that it cannot possibly do, but they also don't pause to think whether we should even be *trying* to do themm
Don't let them fool you: AGI today is no nearer to us than it was two years ago. While ChatGPT might appear to be a step closer to AGI, from a scientific standpoint, it's not: training a neural network to predict the next word is not groundbreaking science. Achieving AGI would…