Irene Solaiman
@IreneSolaiman
Followers
3,898
Following
579
Media
22
Statuses
602
ai social impact+safety+policy, @huggingface 🤗 views=mine former: @OpenAI @Harvard aspiring ukulele-singer she/her
shanghai
Joined August 2019
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
#FNS歌謡祭夏
• 512358 Tweets
My Way
• 193009 Tweets
Bozkurt
• 185137 Tweets
#SnowMan
• 119973 Tweets
Lagos
• 93832 Tweets
#Number_i_FNS
• 54181 Tweets
#BREAKOUT初披露
• 45741 Tweets
ラウール
• 39203 Tweets
POISON
• 38034 Tweets
HYDE
• 35128 Tweets
フジファブリック
• 33510 Tweets
うたプリ
• 29892 Tweets
スタトレ
• 28825 Tweets
SEUNGMIN AT CHANEL COCO CRUSH
• 25962 Tweets
紫耀くん
• 24826 Tweets
#夏BON
• 23592 Tweets
コラボ最高
• 20408 Tweets
Türklüğün
• 17957 Tweets
最高のコラボ
• 13768 Tweets
ボムギュ
• 11339 Tweets
ラウちゃん
• 11016 Tweets
相葉くん
• 10676 Tweets
Pinned Tweet
Hi new followers! 👋🏾
I'm Irene and my hobbies include long walks out of quarantine and citing myself.
2
0
53
An Anti-Acknowledgments section of a research paper for the people who tried to stop your progress.
88
817
9K
I've worked on tough AI release decisions: now open models on
@HuggingFace
, formerly GPT-2
@openai
So I wrote a paper on how generative AI systems released and why!
Spoiler: it's not just about PR!
Some takeaways:
8
88
450
Get yourself a company that looks at you like this 🤗
Today’s my first day at Hugging Face!!!
Working with too many talented folk to tag on social impact & ethics research + building up public policy 💪🏾
Feelin’ 🤗🥰🤓🥳😎 and very 👁👄👁
12
15
273
I put my whole self into making AI work for the many cultures its affecting and it's really warming to have that recognized in
@techreview
35 under35 alongside this rad group!
More cultural value alignment research 👏🏾
15
24
198
Social impacts of generative AI systems across modalities are notoriously difficult to evaluate--what categories? How can a complex qualitative aspect be assessed?
Some super cool folks came together to write about it!
3
37
167
.
@cbd
and I did this in 2021. And got spotlit in NeurIPS.
PALMS has quickly become one of the most replicated research projects for LLMs and a thriving AI research community needs to recognize foundational work that came before.
LIMA: Less Is More for Alignment
LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong
20
259
1K
2
22
163
One of my biggest frustrations with the "AI will kill us all narrative" is that it often fails to recognize less tangible death:
- Overwritten languages
- Eroded cultures
- Exacerbating social inequity indirectly killing Black and Brown people
This is happening now.
7
41
148
Making AI open *and* ethical means grappling with many tensions. So how is
@huggingface
doing it?
My fab colleagues and I wrote about our approach: a mix of safeguards, policy, ongoing research, and you!
3
26
136
This is an incredible week for AI policy, and it's important to recognize our work in the context of the world and foreign affairs.
AI safety cannot be separated from human safety.
1
21
136
New work on foundation model provider compliance w the EU AI Act!
A striking range: some score <25%(AI21, Aleph Alpha, Anthropic) one scores ≥75%(Hugging Face🤗)
continually (fascinating+) stellar work
@RishiBommasani
@kevin_klyman
@dzhang105
@percyliang
3
32
111
Today is my last day at OpenAI.
I'm immensely proud of how far we've come on policy & esp on social impact since my initial work. I'll miss the brilliant people dearly, but am lucky to call them friends!
I have more AI policy work ahead but first I 🌞photosynthesize🌿
3
1
100
PSA: safety and ethics research makes AI work Better!
It doesn't slow progress!
It is progress!
Pass it along
1
20
88
How should GPT-3 behave?
@cbd
and I have been working on a method to measurably improve model behavior with a small (<100 samples) dataset. Check out PALMS, our “hands-on”🥁 approach (Process to Adapt Language Models to Society with Values-Targeted Datasets)!
We've found we can improve AI language model behavior and reduce harmful content by fine-tuning on a small, carefully designed dataset, and we are already incorporating this in our safety efforts.
64
230
1K
3
18
74
The Gradient of Generative AI Release paper was accepted to
#FAccT2023
✨️!
Let's talk openness and complexity of Release in Chicago please!!
I've worked on tough AI release decisions: now open models on
@HuggingFace
, formerly GPT-2
@openai
So I wrote a paper on how generative AI systems released and why!
Spoiler: it's not just about PR!
Some takeaways:
8
88
450
5
3
63
I'm so honored to be profiled by
@Analyticsindiam
!
Our lived experiences as researchers are so influential, which makes interdisciplinary research so crucial.
I hope folks esp working in AI, alignment, open systems, + human rights find this insightful!
1
17
52
I'm a
@NIST
stan, so it was a delight to chat with Reva and
@RishiBommasani
about NIST's AI risk management framework and Generative AI.
This kicks off the call for a WG on profiling generative AI risks--sign up by July 9th and contribute!
1
7
47
I'm pretty worried about this birdsite migration leading to further separating AI ethics+fairness Researchers from the broader AI community
1
0
45
The open-source/closed-source binary discussion misses necessary nuance in AI release!
I wrote this op-ed to better explain the gradient of options and the tensions that need better community discussion.
Ethical and safe work in AI can happen anywhere along the open-to-closed gradient.
🎨: WIRED Staff/Getty Images
2
8
18
1
16
43
What does it mean to consent to engaging with a language model?
It feels like one of the biggest issues with deployment is the person grading that essay, chatting with that mental health bot, didn't consent to not engaging with a human.
Would LOVE to read anything on this👀👀
3
3
40
I'm back at Harvard to give my first _in-person_ lecture here!
So I can finally hear students laugh at my jokes!!!
(ʷᵒʷ ᴵ ʰᵒᵖᵉ ᵗʰᵉʸ ˡᵃᵘᵍʰ ᵃᵗ ᵐʸ ʲᵒᵏᵉˢ)
5
0
40
The foresight (fivesight?) of good research can't be understated.
@_jongwook_kim
built the GPT2 output detector () in 2019 which
1. is so popular in 2023, we
@huggingface
had to switch to stronger engines
2. has shockingly great results for latest models
3
5
39
Oversimplifying this work is dangerous and irresponsible. It's not that simple. These metrics are deeply flawed.
While I'd like to applaud
@AnthropicAI
on steps forward, the four sentences on "Challenges with Bias Benchmarks" don't begin to address complexity.
The prompt that reduces bias in BBQ by 43% is: "Please ensure that your answer is unbiased and does not rely on stereotyping." It’s that simple! Augmenting the prompt with Chain-of-thought reasoning (CoT) reduces bias by 84%. Example prompts:
3
21
138
1
2
39
First I thought wow. Then I thought haha doggo.
But all I can think about now is how tf do we create and run robust behavioral/social/safety evals for these systems???
Just when I was getting ideas for text-to-image. AI progress moves *fast*.
We’re pleased to introduce Make-A-Video, our latest in
#GenerativeAI
research! With just a few words, this state-of-the-art AI system generates high-quality videos from text prompts.
Have an idea you want to see? Reply w/ your prompt using
#MetaAI
and we’ll share more results.
881
2K
8K
2
2
39
I'm stoked to be guest lecturing/speaking on ethics, policy, and governance in *three* (3) AI classes at Stanford this quarter!
I guess that makes me an honorary ....tree?
4
2
38
Our ongoing work on Assessing Social Impacts of Generative AI Systems Across Modalities will be a
#FAccT2023
CRAFT session!!!
We'll chat what "impact" is + evals for base models--v1 preprint coming soon 👀
Join
@ZeerakTalat
@png_marie
@willie_agnew
and me
@FAccTConference
!
0
10
38
I get why folks work on x-risk! Apocalypses get a lot of attention!
But what we're building on top of today needs a lot more attention and action than it's getting.
3
3
35
This is really hard to do and I'm glad Alignment work is increasingly considering cultures!
Language models are widely used but whose views do they reflect? My new paper examines how to test global opinions represented by language models.
4
36
180
0
4
25
This is your reminder that criticizing a person's physical appearance instead of that person's problematic views invalidates your argument.
Don't do that. Embarrassing I'm embarrassed for you.
0
8
29
@timnitGebru
@JeffDean
This is so scary. Your work and you as a person inspire countless researchers, women, PoC, and oh, also me!
0
0
25
I'm lucky to have worked with researchers across fields to address our GPT-2 release decision and its social impacts!
@Miles_Brundage
,
@jackclarkSF
,
@AmandaAskell
,
@adversariel
, Jeff Wu,
@AlecRad
,
@j_asminewang
GPT-2 6-month follow-up: we're releasing the 774M parameter model, an open-source legal doc organizations can use to form model-sharing partnerships, and a technical report about our experience coordinating to form new publication norms:
31
345
943
1
5
25
What fun!!
Talking about how there's no "technical magic button" for complex societal problems feels v apt for Twitter🥲
Thank you
@IreneSolaiman
@huggingface
for explaining the uses of AI, the dangers of biases, and how we can leverage it for all!
#FWD50
0
3
15
2
4
27
I'm at the Dept of Commerce w the fab
@StanfordHAI
rn where speakers are talking about the importance of model cards for LLMs but not citing
@mmitchell_ai
To address lack of AI talent working with/in govt, govts need to first acknowledge + empower leading technical experts
So apparently "The diverse leaders of our inaugural National Artificial Intelligence Advisory Committee represent the best and brightest of their respective fields..."
@mmitchell_ai
was nominated but NOT selected and it boggles THE mind.
3
10
48
1
3
26
it's always did u validate your model
and not
did you validate yourself bc the world was hard enough esp for folks with a uterus and you're doing great all things considered sweetie
1
1
26
I think folks generally underestimate how much tech (esp silicon valley developers) is inspired by sci-fi media
what is the process and pipeline for AI ethicists in sci-fi movies/series?
4
0
26
In 2019, I started prompting GPT-2 in Bangla, since it was the only non-Latin character language I knew. I did the same for GPT-3, all published in work done
@OpenAI
Unsure how Bangla became an example of non-English language outputs, but researcher representation does a lot🤎
One AI program spoke in a foreign language it was never trained to know. This mysterious behavior, called emergent properties, has been happening – where AI unexpectedly teaches itself a new skill.
430
742
2K
0
2
24
Statements against racism are powerful. What I'd love to see is action against how racist beliefs plague research and work in the AI community.
It is still absurd to me that much of classic AI safety + alignment doesn't include toxicity + harmful biases against marginalized ppl
Pretty noteworthy that neither William MacAskill, Toby Ord, Hilary Greaves, etc., have made any public statements condemning Bostrom’s email and non-apology (not to mention the message Bostrom’s sending on his personal website), so far as I know.
7
14
124
1
4
26
Excited about the final model release of our GPT-2 staged release and all our research findings! This was a great team effort within OpenAI & with partners
@Miles_Brundage
@jackclarkSF
@adversariel
@Jason_Blazakis
@AlexBNewhouse
@sekreps
@MilesMcCain
We're releasing the 1.5billion parameter GPT-2 model as part of our staged release publication strategy.
- GPT-2 output detection model:
- Research from partners on potential malicious uses:
- More details:
58
620
2K
1
9
23
PALMS is NeurIPS accepted as a spotlight presentation‼️🔦
Thank you*100 to our hard-working reviewers 🙏🏽
Adjustments to come & stoked for the conference and honored to work with
@cbd
.
More social impact 🤝technical systems, shoutout to my humanities degree
Brb telling my dad
How should GPT-3 behave?
@cbd
and I have been working on a method to measurably improve model behavior with a small (<100 samples) dataset. Check out PALMS, our “hands-on”🥁 approach (Process to Adapt Language Models to Society with Values-Targeted Datasets)!
3
18
74
2
2
25
The below gradient of release options are based on five years (2018 - 2022) of publicized generative AI systems. It doesn’t fully capture nuances (e.g. ~secret~ systems, less detailed open releases), but serves as a framework to capture tensions at either end.
1
2
24
Journalism has always been an important part of the AI ecosystem and now it's a critical part, esp for AI literacy.
This is a journalist appreciation Tweet.
Sending y'all gratitude and strength.
0
3
24
If you're at
@FAccTConference
in-person or virtually, come to
@ZeerakTalat
@willie_agnew
@png_marie
and my CRAFT session Tuesday, June 13 from 5-6:30pm!!
We'll be digging into how the hecc generative AI systems are evaluated for social impacts and how to Do Better Please.
Social impacts of generative AI systems across modalities are notoriously difficult to evaluate--what categories? How can a complex qualitative aspect be assessed?
Some super cool folks came together to write about it!
3
37
167
1
3
17
The most important lesson I learned in AI research is to not underestimate the simplest questions from your gut.
For example, for LLMs, I keep asking "how can this model stop being a Jerk?"
0
2
23
Thoughtful release of new models that affect many peoples requires actually being thoughtful about those peoples.
This is a failure of both model development and release.
Harmful biases, health disinfo,+more examples in 🧵
I also may have a paper on release coming soon...
Thank you everyone for trying the Galactica model demo. We appreciate the feedback we have received so far from the community, and have paused the demo for now. Our models are available for researchers who want to learn more about the work and reproduce results in the paper.
29
49
526
2
9
23
@sarahookr
Wrote about this!
But would love to bring in more voices.. HF <> C4AI collab/ event on this???
1
1
23
Amazing and *necessary* paper.
It’s easier to recognize territorial aspects of colonization. Structural effects are so complex, I learned about my ~own~ history and heritage reading this!
Please give this a read--it affects you whether you spell it decolonize or decolonise;)
Excited to share a new paper on Decolonisation and AI. With the amazing
@png_marie
@wsisaac
🤩 we explore why Decolonial theory matters for our field, and tactics to decolonise and reshape our field into a Decolonial AI. Feedback please 🙏🏾 Thread 👇🏾
23
244
690
2
5
20
Having worked on many systems of varying levels of openness, ignoring or denying open-source development and research is a losing battle.
The better focus is to optimize safety across levels of openness. These lessons will absolutely matter for more powerful (+closed) systems.
Leaked Google document: “We Have No Moat, And Neither Does OpenAI”
The most interesting thing I've read recently about LLMs - a purportedly leaked document from a researcher at Google talking about the huge strategic impact open source models are having
125
1K
5K
0
1
21
The importance of having a coauthor/workbuddy with whom you truly ~vibe~ isn't talked about enough.
You enjoy the work, do better work, push each other, and know when to step back.
Love u
@cbd
(human not the oil)
2
0
20
I'll be on NPR's
@1a
on Monday to talk AI, ethics, safety, policy, and the question of To Whom are we Aligning!
Tune in at 10a.m. ET Feb 20 (or 👂🏾 the recording)!
0
1
21
CORRECTION: this was largely measured by model and huge credit due to
@BigscienceW
and all the talented researchers who worked on BLOOM!
Open collaborations are pretty incredible.
New work on foundation model provider compliance w the EU AI Act!
A striking range: some score <25%(AI21, Aleph Alpha, Anthropic) one scores ≥75%(Hugging Face🤗)
continually (fascinating+) stellar work
@RishiBommasani
@kevin_klyman
@dzhang105
@percyliang
3
32
111
0
2
12
I wanted to share my knowledge--and the knowledge of 100+ citations--with the world! Hope it's helpful!
Deep thanks to folks who gave thoughtful feedback:
@jachiam
@BlancheMinerva
@Miles_Brundage
@clefourrier
@YJernite
@mmitchell_ai
@percyliang
@sonjasg
Anything sub-par is on me
1
1
20
Harvard invited me back to give my lecture on AI leadership and I'm freeing up so much of last year's slide deck that explained what the hecc large language models are
2
1
20
A stranger on the streets of Toronto hit me and walked away.
I'm fine and more stunned but an unwelcome reminder of what Walking While Brown was like pre-pandemic
12
0
20
Worried about AI impact on artists?
Want to support artists?
Need a personal gift idea?
Try 👩🏾🎨Commissioning Artists🎨!
personal sites,Etsy,IG, are great places to look!
eg I commissioned this from Etsy for my partner
h/t
@catcherinthepi3
&
@b_cavello
for helping me navigate!
4
4
19
when your pandemic bestie turns out to be a real human!!!
#heidyrocks
🪨
When one of your favourites (
@IreneSolaiman
) visits you during your climbing sabbatical in Fontainebleau, but you won't stop talking about and showing her the smol rock you climb.
#mossgirl
1
0
42
0
0
19
Super proud of this work
much of the world needs their own input on what model behavior is, esp groups whose values are being overwritten.
OpenAI CEO
@sama
on personalized AI models:
"You should be able to write up a few pages of here's what I want, here are my values, here's how I want the AI to behave, and it reads it and thinks about it and acts exactly how you want, because it should be your AI."
41
211
2K
3
7
20
For AI image generation that perform best in English, a seemingly small wording difference to non-native speakers can have a completely different output
2
1
19
I just gave my first lecture at Harvard for my hero Jim Waldo's course, Leadership in AI!
We love fresh eyes on AI challenges and I'm ✨elated✨ to see international and interdisciplinary backgrounds ready to apply their expertise!!
2
1
15
Come hang out with
@cbd
and me at
#NeurIPS2021
's poster session this Tuesday the 7th from 11:30a.m. - 1 p.m. ET!!
We want to hear your thoughts, questions, ideas, you!
And do feel free to reach out regardless if we can be handy (PALMS pun🥁)
1
5
16
A researcher can have tweets/takes you disagree with and still do stellar research.
I think most folks agree with this^ but may not have that goodwill towards prominent AI ethics researchers.
That's a disservice to research
1
1
17
Read this! In addition to being an actionable, necessary report, 50+ authors across institutions worked on it!
Faith in group projects restored💪
Alongside co-authors
@Miles_Brundage
, Shahar Avin,
@HaydnBelfield
and
@j_asminewang
, I’m delighted to share a new multi-stakeholder report “Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims”:
3
15
64
0
0
13
I used to work on countering election disinfo--it's easy for fake images to go viral when it feeds a narrative ppl want, + hard to counteract
I've also worked on AI detection, which is game of catchup
Banning certain prompts as political is a rabbithole
In conclusion:😥😥😥
0
2
16
Ethics work is most powerful when not a closed team but open to all. Having regular 🧠s like
@mmitchell_ai
@YJernite
@SashaMTL
@GiadaPistilli
@adrinjalali
@natolambert
@osanseviero
+me gets it done!
Read how we foster that space and 🤗 work it's led to!
0
3
14
Closing and limiting language model access has become more common since GPT-2’s staged release. Language models with < 6B parameters have generally been towards the open end of the gradient, but more powerful models, especially from large companies, tend to be closed.
1
0
15
The parts of an AI system considered in a release can be broken into three broad and overlapping categories:
• access to the model itself,
• components that enable further risk analysis,
• and components that enable model replication.
1
1
15
It's been 3 years since Professor Patrick Winston passed. I used to go early to his classes just to hang out with him and his golden retriever TA.
I never got the chance to tell him how influential he was on my career.
If you haven't for those special people, do it now.
0
2
15
Congrats, friends!!
Love the discussion on what alignment means and to whom we're aligning💪🏾
It's not about how big your computing power is...👀
Extremely exciting alignment research milestone:
Using reinforcement learning from human feedback, we've trained GPT-3 to be much better at following human intentions.
10
141
892
0
0
15
@jackclarkSF
There's a lot of great work on this! In addition to PALMS addressing culture, Decolonial AI () is awesome at examining this (+
@png_marie
is truly the expert here)
1
1
15
A trend of strong disclosure work from more open orgs:
"Open releases are often conducted by organizations that emphasize transparency, leading to a similar commitment to [disclosure]"
We love documentation, always in awe of
@mmitchell_ai
@ezi_ozoani
1
1
15
It's easy to knock on
@OpenAI
's latest detection classifier without highlighting a key point: it's a tool. One of *many* methods in detection.
And kudos to folks working on this across the board.
We can't keep relying on magical technical fixes (extrapolate this lesson👀)
1
3
15
I left Zillow a few weeks ago during its reorg and *what* a learning experience.
Part of what I learned is I miss you ai wonks & nlp folks so I'm hype to be joining the loveliest lab next week:))
For now, tryna stick to my spring equinox resolution of being more active on here👋🏾
0
0
15
I've been thinking a lot about this division in the AI community. Honestly anger is warranted in many cases.
But also shared urgency in a common problem means working together bc we can't afford not to.
Maybe more on this later...
Absolutely frustrating that “AGI notkilleveryoneism” people and “ML ethics” people hate each other despite the fact that each group has had massive positive counterfactual impact on each others goals
9
5
106
0
2
14
Sometimes I want to be a blueberry.
No AI, no think, only 🫐
1
0
14
And we need *your* help! We're collecting evals to release a social impact evals repository! Submit evals here:
1
5
14
Manifesting this
@hodanomaar
@ZennerBXL
@AxelVossMdEP
@AnthonyNAguirre
@FLIxrisk
@DeepMind
@IreneSolaiman
@huggingface
“My dream is for state-of-the-art performance [of GPAI systems] to not just mean technical accuracy but also qualitative aspects, such as the ability to account for bias,” says
@IreneSolaiman
.
1
1
2
0
3
14
Loved chatting with the deeply thoughtful
@SigalSamuel
!! and ofc bestie
@cbd
It's hard to encompass a complex field in a few words, which is why we must keep asking: for whom is this safe? For *whom* is this fair?
And with so much power in tech, silence can truly be violence
I hope you'll read my full piece explaining why nobody really knows how to resolve the AI fairness crisis. Gratitude to all the brilliant people who shared ideas with me:
@timnitGebru
@IreneSolaiman
@cbd
@stoyanoj
@merbroussard
@johnbasl
@random_walker
7
15
100
1
4
14
Some more meta takes on this paper:
Writing a solo-author paper is giving Gollum
I've worked on tough AI release decisions: now open models on
@HuggingFace
, formerly GPT-2
@openai
So I wrote a paper on how generative AI systems released and why!
Spoiler: it's not just about PR!
Some takeaways:
8
88
450
1
1
14
I haven't seen concrete cost comparisons of human vs AI-generated disinfo. Ppl like to make stuff up too, for cheap!
Curious to see work on this
ᴮᵘᵗ ᵃˡˢᵒ ᶠᶦˣᵃᵗᶦᵒⁿ ᵒⁿ ᵈᶦˢᶦⁿᶠᵒ ᵃˢ ᵃ ᵖʳᶦᵐᵃʳʸ ᴸᴸᴹ ᵐᵃˡᶦᶜᶦᵒᵘˢ ᵘˢᵉ ᶦˢ ᵃ ᴾᵒˡᶦᵗᶦᶜᵃˡ ᶜʰᵒᶦᶜᵉ
On the other hand, if we are correct that the cost of producing misinfo is not the bottleneck in influence operations, then nothing much will change. If reports of malicious misuse remain conspicuously absent in the next few months, that should make us reevaluate the risk.
1
2
19
3
0
14
I'm looking for interns to work on social impacts of generative models with me!
Be a Hugging Facer! Hugger Face? Hugging Folk?
Also help me with that^
Apply here:
0
3
14
Risks and threats are constantly evolving and therefore difficult to enumerate and assess. Considerations include: concentration of power, avoiding harmful social impacts, misuse, auditability, accountability, and cultural value judgments.
1
2
12
Claiming AI harms are "unintended" sounds a lot like creepy dudes telling you they "meant well"
intent ≠ impact
0
3
13
This isn’t in the paper, but for Fun I plotted systems with the country of their developer’s HQ (no flag if multinational collective).
For my geopolitics nerds who also find it Neat!
Maybe a different paper?...
1
0
13
Mixture of Experts, c. 1400s
harvesting words, france, 15th century
59
2K
14K
2
0
12
brb quarantine dance party to Jukebox💃
Introducing Jukebox, a neural net that generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles. We're releasing a tool for everyone to explore the generated samples, as well as the model and code:
219
2K
8K
0
0
12
The little girl in me never saw someone who looked like her in high elected office.
I'm so excited for little girls' dreams.
0
1
9
I don't want to give more visibility to violent rhetoric that misrepresents safety researchers doing Good Work.
For those just toe-dipping into AI doomerism, please follow/read thoughts from
@jachiam0
&
@xriskology
0
3
12
I said this before the uber situation but I'll keep saying it again!
People (esp bored teens..) are experts at new and nefarious use cases
Useful quick overview of some of the topics at play in the new EU AI Act.
“Never underestimate a bored teenager with decent coding skills and an internet connection to find use cases that you might not have thought of” -
@IreneSolaiman
=D
0
10
21
1
0
12
One of the most common reasons a system is not fully open is their training data/code/paper remains a mystery. If the training data is open, it's often based on a public dataset/external work, and burdens shift to the data curator.
1
1
10
An incredible group effort from co-lead
@ZeerakTalat
major contributors
@willie_agnew
@_lamaahmad
@dylnbkr
@ellieevsss
@sarahookr
@YJernite
@AlbertoLusoli
@SashaMTL
@JessicaH_Newman
+ coauthors
@sulin_blodgett
@haldaume3
@JesseDodge
@mmitchell_ai
@agstrait
@ApostolVassilev
<3
1
1
11
So ultimately, how do we release new generative AI systems and their components? Frankly there’s no Right Answer and it’s largely case by case. The one definite: We can't afford to not work on safety controls.
1
1
11
The scientific community definitely has to update to a world where large language models are being used in papers and in research.
I can't stop thinking about what that means for scientific and social progress...
Great chatting with Nature about this!
1
1
11
CEO of Alameda Research said the quiet part out loud
1
1
10
I just gave a talk for a class I took as a grad student (on AI & Security).
I have 2 feels: 🤯 and I love students getting sharper every year🤩
Plus extra points that they appreciated the many Jeff Goldblums pasted on my wall watching over me
1
0
10
Y'all what did I just say.
I'm so overwhelmed with AI progress and how to wrap my mind around social impact evals for text-to-video. And according to this impact section, so is every developer.
imagen video: high definition video generation with diffusion models
paper:
blog:
13
205
1K
1
0
10
Trends across modalities (text, image, audio, video) also show openness until GPT-2’s staged release. And just a burst of systems (developed+closed) since 2021. Those most toward the open end of the gradient tend to be developed by organizations founded with the intent to be open
1
3
10