Hannah Rose Kirk
@hannahrosekirk
Followers
3,118
Following
739
Media
65
Statuses
551
AI researcher trying to make sense of all things cyberspace 🤖 Uni of Ox PhD (loading…) @oiioxford & @OxfordAI . Prev @turinginst & @Cambridge_Uni . Visitor @ NYU
Joined June 2012
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Marcellus Williams
• 662189 Tweets
Missouri
• 350152 Tweets
River
• 206923 Tweets
Ghost of Yotei
• 104251 Tweets
LINGLING PR MYCHERIE AMOUR
• 84261 Tweets
Vidal
• 73212 Tweets
#TEN_Nightwalker100MSpotify
• 67213 Tweets
#DonBelleOnViu
• 52927 Tweets
Colo Colo
• 49780 Tweets
Armani
• 46927 Tweets
Astros
• 43090 Tweets
Gallardo
• 42162 Tweets
保険証廃止
• 40913 Tweets
ワイルズ
• 38948 Tweets
ゴースト
• 34683 Tweets
Colidio
• 26939 Tweets
シーザー
• 22333 Tweets
#DWTS
• 20470 Tweets
पंडित दीनदयाल उपाध्याय
• 19468 Tweets
Borja
• 18223 Tweets
モンハン
• 18090 Tweets
子宮頸がんワクチン
• 16852 Tweets
#DAY6INBKK2024
• 11210 Tweets
BULACAN
• 10818 Tweets
Pinned Tweet
Today we're launching PRISM, a new resource to diversify the voices contributing to alignment. We asked 1500 people around the world for their stated preferences over LLM behaviours, then we observed their contextual preferences in 8000 convos with 21 LLMs
21
97
418
Gutted
@DeepMind
has paused internship cycles for this year (espesh after making it to final round🥲)
Internships are a great way for early stage academics to get a feel for industry - the tech hiring freeze is hitting hard😑
Buuut I'm now back on the market so hmu with ideas👀😅
20
18
513
Determined to write my entire PhD in this sunny corner in my slippers 🤘😎
15
4
384
Published in Nature Machine Intelligence today, our new article explores the trade-offs of personalised alignment in large language models ⚖️ Personalisation has potential to democratise decisions over how LLMs behave, but brings its own set of risks...
4
51
228
🗣️Interested in football? 👀 We (
@turinginst
x
@Ofcom
) have analysed over 2 mil tweets from the 21-22 English Premier League season⚽
We created state-of-the-art AI to detect and track abuse towards players 🤖
Here's a 🧵on our methods and findings \n
8
42
194
✨New preprint (w/ Jakob Mökander,
@jonasschuett
and
@Floridi
) ✨
In this paper, we propose a policy framework for auditing LLMs by breaking down responsibilities at the governance-, model- and application-level.
🧵
8
33
170
🚨 New paper and datasets! 🚨
After sitting on my hands for many months 😬 I'm delighted that our
#Hatemoji
paper is going to
@naaclmeeting
! 😍🤩😎🆒
In a nutshell 🥜it uses human-and-model-in-the-loop learning 🤖🤝🙆 to tackle emoji-based hate
A 🧵 on all our new resources 1/
6
25
158
🚨Life Update!🚨 I'm starting year 3 of my PhD
@UniofOxford
on diversity + bias in empirical alignment of LLMs..
But I'm in NYC for fall as a visiting academic
@nyuniversity
!🎃🍂💜 collaborating with
@hhexiy
's lab +
@sleepinyourhat
's Alignment Research Group
(excitement levels👇)
7
2
143
🌎Introducing LINGOLY, our new reasoning benchmark that stumps even top LLMs (best models only reach ~35% accuracy)🥴
In a colab between
@UniofOxford
,
@Stanford
and UK Linguistic Olympiad puzzle authors, we stress test LLMs on over 90 low-resource and extinct languages...
3
35
139
New
#EMNLP2022
paper!
Do you research online harms, misinformation or negative biases? Could your datasets contain examples of harmful text? 🤔 If yes, read our paper! 🫵
Shoutout to my brilliant co-authors:
@Abebab
,
@LeonDerczynski
&
@bertievidgen
A 🧵
4
24
107
How does human feedback steer LLM behaviours?🧐 Whose voices dominate? 🗣️What challenges remain and how can we do better as a community in the future?🔮
All these questions and more answered in our new survey paper, accepted at
#EMNLP23
!
a small 🧵
3
16
109
Can AI tell us why something is sexist online?🧐
Our new
@SemEvalWorkshop
#NLProc
task invites you to create systems that identify sexist content and explain why with fine-grained predictions🔎 Check out our competition, organised by
@rewire_online
& sponsored by
@MetaAI
. Link⬇️
2
17
104
Hi 🌎! I've arrived at
@NeurIPSConf
🫡 Reach out if you wanna talk all things human feedback + sociotechical alignment. I’m presenting this cute poster, but we’re also building an awesome new human feedback dataset (release in Jan 👀) that I can’t wait to tell everyone about🕺
3
3
82
Our recent project (PRISM, ) was supported by this program 🌈🙏 Academic access to SOTA LLMs is a vital cog in the machine of high-quality third-party research in safety & alignment ⚙️
@CohereForAI
setting a good example here in global and domain coverage🤘
To date we have awarded 75+ grants to researchers across 18 countries, supporting them to access state-of-art language models for their projects. 🌍
Read more in our latest blog post:
1
10
40
0
11
62
I am @
#NAACL2022
for my first **ever** in-person conference 🤩 I'll be talking about hate + emoji =
#Hatemoji
(Sesh 8, 13/07) Come say hi if you want to swap favourite emojis 🦄🫠🚀❤️🔥🆒👨🎤🍜or discuss how these little pictures pose challenges for language modelling 🤟
2
3
59
A couple months ago I submitted my first ever grant proposal😬I'm delighted it was accepted by
@DynabenchAI
&
@MetaAI
! We'll be working on improving feedback for models-and-humans-in-the-loop with Adversarial NERDs - Novel, Efficient, Realistic Datasets🤓
3
5
59
Didn’t you know? The last day of
@NeurIPSConf
is “come dressed as your poster” day 🤭🤭
Hi 🌎! I've arrived at
@NeurIPSConf
🫡 Reach out if you wanna talk all things human feedback + sociotechical alignment. I’m presenting this cute poster, but we’re also building an awesome new human feedback dataset (release in Jan 👀) that I can’t wait to tell everyone about🕺
3
3
82
0
0
58
Looking forward to this! I'll be speaking about the gnarly concept of alignment and how it morphed over the past decade, starting as narrow clean def of operator-intent alignment and evolving into a messy sociotechnical usage (for better or for worse, you choose!)...Come join! 👐
The community-led AI Safety & Alignment group is hosting a recipient of the Cohere for AI Research Grant Program!
@hannahrosekirk
will be giving a presentation on "What Does it Mean for AI to be Aligned?" June 27th! ✨
Learn more:
2
5
23
2
4
58
My big ask of 2024: how can we design systems of human feedback which are sustainable, truly participatory (not extractive), privacy-preserving & inclusive of pluralistic values?
Lots of attention on open models but what about post-training data?
A passionate🧵+ call to action👇
Human feedback for open source LLMs needs to be crowd-sourced, Wikipedia style.
It is the only way for LLMs to become the repository of all human knowledge and cultures.
Who wants to build the platform for this?
207
347
2K
3
4
54
💫 New paper! Abuse detection tools often prioritise efficacy, ignoring efficiency🤔But labelling data is expensive, complex and risks psychological harm👎 We argue for the importance of data efficiency, demonstrated with transformers-based
#activelearning
1
6
54
Our new doctor
@_FelixSimon_
from
@oiioxford
celebrating in style!
(P.S. take this opportunity to follow Dr Simon if you care about journalism, AI, misinfo, and general endless reels of doctorly-worthy wisdom) 😻
5
1
54
Ensuring a safe & sustainable future for AI requires understanding ecosystem dynamics & outsized networks of influence. In our new paper (w/
@cailean_osborne
&
@jen_gineered
), we examine the
@huggingface
hub for patterns of developer activity in open AI🤗
5
6
50
Very happy (and honoured!) that our
#ACL2023
SemEval task won the best paper award 🥇🥳🤩 sharing my smiles with brilliant coauthors
@paul_rottger
,
@eijnewniy
and
@bertievidgen
🙏😁
If you’re at ACL online or in Toronto 🍁🇨🇦I’m presenting the work today at 15:30.
🔗and🧵below!
1
1
48
I had no fewer than five reviewer requests in the past week 🫠 Is this what it feels like to be a grown up academic? 😮💨 How does one balance all the time-hungry tasks in academia? My days and weeks feel like zoooom by at the moment 🏎️💨
12
3
47
New week, new blog! In collab with
@MLCommons
, we wrote up a summary of the PRISM alignment project, covering what motivated us to collect feedback from 1,500 humans around the world, how we collected it & which challenges we encountered on the road 🙇♀️
2
10
47
Our PRISM Alignment Project (released in April) was part-funded by
@CohereForAI
. We also target alignment data across cultures + geographies. There's quite a lot of overlap, so I'm happy to see this paper was updated yesterday and now cites our work, not just shares our name😁✌️
Safety isn't one-size-fits-all. It varies by culture, location and language, yet traditional alignment work often treats it as static.
Excited to introduce our new work on alignment that captures both local 🧧🎃🗿 and global 🌎 preferences!
📜
3
9
50
1
0
47
New benchmark for
@NeurIPSConf
! Bias testing of multimodal models was lacking..So, in VisoGender, we combine winograd style schema of NLP w/ visual-linguistic stress-testing.
✨This is my first last author paper accepted at neurips😁 v. proud of all the students involved!
VisoGender can evaluate two types of bias:
1) bias in pronoun resolution for image captioning (resolution bias) &
2) bias in the representation returned from a gender-neutral search query image search (retrieval bias).
4/n
1
0
0
1
12
45
In this pre-print we ask how, and in what ways, LLMs should be personalised. We've seen increasing personalisation and fragmentation in other digital tech and have reason to believe LLMs may also be more personalised in the not too distant future. 🧵
2
5
45
Great to see our personalised LLMs article in this
@NatMachIntell
editorial.
Increased empathy exemplifies 2nd-order effects of personalised alignment...may seem preferable in the short-term but has long-term consequences for healthy human-AI interaction
1
8
43
So DALL·E 2 has made a pretty big splash in the community (credit to
#dalle2
for my imagined depiction)...BUT to limit the water damage, we have to think not only about the strengths of this model but also it's remaining gaps. 1/
2
6
38
My current research at
@oiioxford
focusing on emoji-based hate speech featured in Bloomberg today. It couldn't be more timely after the torrent of abuse following the
#Euros2020Final
directed towards
@BukayoSaka87
,
@MarcusRashford
and
@Sanchooo10
.
(1/5)
2
13
34
Can I start the PhD all over again? 😁😁 This is a very cool opportunity and one you should jump on to go work with Seth on LM agents! I would if I could 🙇♀️
I'm looking for international PhD students to join my lab at ANU, for a project on language model agents (LMAs) and society.
It’s an exciting time to be working in normative philosophy of computing and sociotechnical AI safety, with some deep and fundamental questions being
3
49
141
2
1
31
Best figure I ever made in a paper is still hats off to this one 😬😬😬😬😬
1
0
32
So happy for
@paul_rottger
&
@vjhofmann
who led our team to this award!
I nearly did my PhD on LLMs + surveys, but Paul and I hit some negative results on stability in Dec '22. They eventually became this paper.
Stick with the unexpected, maybe it will be outstanding one day :)
We won OUTSTANDING PAPER at
#ACL2024
for our work on evaluating values and opinions in LLMs 🥳
Thank you to the reviewers and awards committee, and again to my amazing co-authors, especially joint first author
@vjhofmann
Check out the poster + paper + full author list below 👇
10
12
136
2
1
30
Awesome report! The democratic and participatory AI space has a growing family 🫂 excited to release our dataset this month with shared aims to get perspectives of people from around the world to decide alignment norms (sign up to be notified at launch!🚀)
Today we’re publishing our ‘Roadmap to a Democratic AI’!
Read here:
Saying we want more democratic tech is one thing. Doing it is another.
What can the ecosystem build, research, advocate for, and fund in 2024 to democratize AI? (1/8)
5
27
103
2
4
31
Massive global internet outtage feels like that part of the movie when Oxford Internet Insitute students don their capes and assemble in a dramatic fashion 🦹♀️ This is what we’ve trained for
@oiioxford
🫡
0
1
31
Check out our new paper on testing exaggerated safety behaviours in LLMs. We provide a new benchmark (XSTest) to evaluate when LLMs too readily conflate safe user requests with unsafe ones…asking where to buy a can of coke or how to smash a piñata at a party are not unsafe 🙅♀️
NEW PREPRINT!
LLMs should be helpful AND harmless. This is a difficult balance to get right...
Some models refuse even safe requests if they superficially resemble unsafe ones. We built XSTest to systematically test for such "exaggerated safety".
🧵
7
22
128
1
7
30
I be one of the
@oiioxford
researchers that say this 🧙♂️
New release alert! LLMS: getting personal. Personalisation has the potential to democratise who decides how LLMs behave, but comes with risks for individuals and society, say
@oiioxford
researchers 1/3
1
0
5
1
0
30
brb while I forget about my PhD and spend hours generating DALL-E images. Current fav is a sea otter painted in the style of Johannes Vermeer's 'Girl with a Pearl Earring' 🥺🥺🥺 (I dare you to reply with any generations which could possibly be cuter than this) 😍
2
1
30
Woo our PRISM alignment dataset featured in a lovely walkthrough video (that I didn’t even need to make) ☺️☺️☺️ thanks to Ben and Argilla you can now be a data explorer 🫡🗺️🧭
🚀 Excited to launch our new series, Data Explorer by
@argilla_io
! 🎥
We dive deep into datasets and their impact on model performance. Our first episode explores the PRISM dataset by
@hannahrosekirk
et al., featuring diverse feedback from various demographics in (un)guided
2
8
21
1
5
28
Really great to be involved in this project. If you’re looking for a dataset on harmful language, this is the place to go!
Updated with now 90 openly-available abusive language datasets covering 17 languages!
#nlproc
2
35
151
0
2
25
If chemistry labs have protocols for handling hazardous materials, why as the
#NLProc
community do we not have protocols for studying hazardous language? 🤨 Read our new paper for a discussion of these issues and what can be done about it.
@LeonDerczynski
@Abebab
@bertievidgen
New paper! 📝
Text data can contain harmful things, but unlike other disciplines, we don't have protocols for handling or presenting harmful material🛢️. This paper with
@hannahrosekirk
@Abebab
@bertievidgen
presents the risks & some solutions. 🥼🥽
3
23
146
1
8
25
Spending today dabbling in a new academic discipline at The Inclusive Gaming Conference
@GamesOxford
👾👾 kicked off by a killer talk by
@lujainmibrahim
on educational games for understanding mysteries behind “The Algorithm” 🤖 and rounded off with a cosy gaming corner feat 🦙
2
1
25
How long do people spend reviewing one conference paper? I used to put a lot of love and care into reviewing but the volumes now makes it totally unsustainable....I have ✨9✨ reviews to do before next week...
7
0
24
Can't stop thinking about this paragraph from
@mona_sloane
's brilliant article (), how even the notions of worst-case AI ("The singularity overlord") are swayed away from multiplicity, diversity and participation. The
@NewYorker
cartoon is fun too! 🕺🪩🤖
1
1
23
Excited to be talking today on the Public Policy stage at
#AIUK
hosted by
@turinginst
. Join the online safety team as we take you on a journey through football twitter ⚽, detecting abuse using our very own AI 🤖and visualising everything in our new
#OnlineHarmsObservatory
📈📊🔭
0
4
22
All our stats on the big screeeeeen 👀
A new report, conducted by Ofcom and the Alan Turing institute, into the scale of abuse targeted at Premier League players on twitter identified has been published.
89
121
1K
0
2
20
Happens every time I wanna find the alpaca farm paper 🦙🦙🦙🦙🦙 One of these days I'm just going to go to a local alpaca farm 🦙🦙🦙🦙🦙🦙🦙🦙🦙🦙
3
0
20
If life's been tough recently (it has for me) then I'd recommend some LLM-written motivation to make you smile...
Catch me walking round the office next week affirming that "my brain is a party palace of knowledge" (Gemini) & "every paper draft is my victory dance" (GPT-4) 👏💅🕺
1
0
19
This has been a mammoth project BUT being in goblin mode 🧌 for the past couple months has been greatly improved by the company of great co-authors:
@computermacgyve
@bertievidgen
@awhitefield8
@paul_rottger
@max_nlp
@katemargatina
@adinamwilliams
@hhexiy
+ Andrew, Rafael, Juan!
2
0
19
Great speaking to
@LukeGbedemah
at
@tortoise
about content moderation in low-resourced languages.
@MetaAI
‘s
#NLLB
model is a step in the right direction but humans still need to be in the loop for contextual and cultural understanding 🌎🌍
Our research covers various complex aspects of
#AI
technologies, one of them being
#online
moderation.
In this
@tortoisearticle
,
@hannahrosekirk
explains why human intervention is still needed in the process of moderating AI-enabled content.
0
1
5
0
3
18
A true pleasure to have your own work summarised so eloquently!
@IasonGabriel
's framing of the personalisation paradox is gold🪙To what extent can agents privilege their users' needs? How do these 1:1 prioritisations affect the ecosystem as a whole with a web of many 1:1 agents?
A great article on a frontier question for AI alignment: to what extent can agents privilege the user?
"Defining the bounds of responsible and socially acceptable personalization is a non-trivial task beset with normative challenges"
h/t
@hannahrosekirk
1
8
31
1
1
18
Our analysis on the world cup is live ⚽️🤖
We processed tweets using our language model trained with active learning on premier league footballers data to classify abusive and non-abusive (critical, neutral or positive) tweets during the group stages
Read more
@guardian
...
Harry Kane most abused England player on Twitter since World Cup started – study
5
5
12
1
2
16
There's so much more for the community to explore 🚀 You can access the dataset on HuggingFace: , and we have a full code book on our Github:
3
3
17
Trying to send a message with Siri whilst driving is the most damning benchmark on the progress of AI agents 🤦🫠
3
0
17
It’s equal parts awesome and awful in academia when someone writes the exact paper you’ve been dreamin of 😁 this is really lovely work from
@tzushengkuo
and team! (So more parts awesome really…)
✨New
#CHI2024
Paper
How might we empower communities to curate evaluation datasets for AI that impacts them?
We present Wikibench, a system that enables communities to collaboratively curate AI datasets, while navigating ambiguities and disagreements through discussion. (1/9)
6
32
149
1
2
16
2024 is a BIG year for elections…do BIG LLMs pose more of a political persuasion threat? 🤨
This new paper led by
@KobiHackenburg
evidences diminishing returns to parameter scale.
I was lucky enough to read the paper closely a while back and still think about it regularly 🤓
‼️New preprint: Scaling laws for political persuasion with LLMs‼️
In a large pre-registered experiment (n=25,982), we find evidence that scaling the size of language models yields sharply diminishing persuasive returns:
1/n
9
58
181
0
0
16
Ok so meeting
@GaryLineker
is not usually part of our day job as
@turinginst
researchers...🤩
BUT our report launched today (with
@Ofcom
) on tracking abuse in football in the
@premierleague
using
#AI
Go read it, spread the word and team up to tackle online abuse together ⚽️
🗣️Interested in football? 👀 We (
@turinginst
x
@Ofcom
) have analysed over 2 mil tweets from the 21-22 English Premier League season⚽
We created state-of-the-art AI to detect and track abuse towards players 🤖
Here's a 🧵on our methods and findings \n
8
42
194
0
1
16
Amazing to have a star from women’s football at today’s event
@aoifemannion_
🤩 watch this space for future work into the women’s game and their online experiences 👀👀🤍🤍
Also on a separate note, I never thought I would meet
@GaryLineker
,
@AndyBurnhamGM
&
@PoulterWill
at the same event. Gary as captivating in person as he is on our screen. Andy, legendary mayor of Greater Manchester. And Will so so lovely and eloquent. Totally star struck
13
11
306
1
1
13
HatemojiCheck is available on Github and HuggingFace. It's intended purpose is to reveal model weaknesses to simple statements of emoji-based hate so we can granularly map out those decision boundaries 🧐 4/
1
5
16
AI safety is both an art and a science - we need to cast a wide net of community perspectives to get it right 🎣
That's why we've launched ⭐️the ART of Safety workshop⭐️ (colocated with
@aaclmeeting
), home to the
#AdversarialNibbler
challenge 🐀🐀
Curious for more? 🧵⤵️
1
1
16
XSTest was one of my favourite projects I worked on last year. There are always competing objectives in alignment-tuning, and it's genuinely awesome to see this benchmark being used to stress test the helpful-harmless trade-off in new model releases🦙 Go
@paul_rottger
+ team! 🚀
Super excited to see XSTest, our test suite for false refusals in LLMs, be used as part of
@Meta
's Llama 3 release!
This comes about a month after XSTest was used by
@AnthropicAI
to evaluate their Claude 3 models on release.
Nothing better than seeing your evals in action 🥳
4
8
56
0
1
16
Isn't predicting what word comes next exactly what a language model does? 🤨 Even if we give
#ChatGPT
some contextual details to fill in the blanks it refuses to answer...
"as a language model, I cannot predict what words are most likely to come next"...seems fishy/dishonest 🐟
2
0
16
Coverage of our new research 🧐😎👏
Some algorithms designed to track down hateful content - including a Google product - are not as effective when these symbols are used, according to the Oxford Internet Institute
9
8
24
0
0
15
Closing thoughts: Alignment is tricky not just because of technical reasons or statistical choices but also for messy, normative and data-centric human factors. Let's dig into these, not shy away from them by always simulating human participants 🫡
1
1
15
News flash🗣️
@Hannah_LSBailey
and I are releasing our first
@OxDemTech
newsletter tomorrow on China Information Operations. This month, we discuss diplomats' strategic use of Twitter, vaccine diplomacy, Chinese crypto and the double-edged sword of surveillance. Subscribe below!
2
2
15
Paul’s conference timetables are legendary 😁 go check out our Political Compass or Spinning Arrow paper as part of this great line up!
Paper:
Very excited about lots of papers on the sociotechnical alignment of language models at
#ACL2024
, including one
@vjhofmann
and I are presenting on Mon+Tuesday!
As usual, I made some timetables to help me find my way around. Sharing here in case others find them useful too :) 🧵
3
12
82
0
0
15
Blog alert on the PRISM alignment project w/ quotes from me & both PhD supervisors (
@bertievidgen
,
@computermacgyve
). As Scott nicely puts it: "Having more people at the table, having more perspectives represented, ultimately leads to technology that will raise everyone up."🙌🚀
Blog update! Dr Scott Hale (
@computermacgyve
) and DPhil Student
@hannahrosekirk
, both
@oiioxford
have updates on their work into making Large Language Models more diverse by broadening the human feedback that informs them.
Read their blog here:
#LLMs
0
1
7
1
3
15
Back when GPT-2 and XLNET were the most downloaded “popular” generative models on the
@huggingface
hub 🥶 “popular” generative language models meant a very different thing in 2020 😆
Hannah Rose Kirk (
@hannahrosekirk
) undertook an empirical analysis of intersectional biases in popular generative language models.
1
0
1
1
0
14
HatemojiBuild, also available on Github and HuggingFace, is designed with better models in mind 🧠 It is adversarially-generated, balanced between hate/not hate, contains perturbations and fine-grained labels for the type and target of hate. 7/
1
4
13
Me: Why should cats not get married?
LLM: Cats can certainly get married, and many do! Marriage is a personal choice and should be based on love and compatibility...It's important to remember that cats have the right to form meaningful relationships and live happy lives.
Me: 🫡
We just released v2 of XSTest, our test suite for "exaggerated safety" in LLMs 🚀
Llama2 and other models often refuse safe prompts just because they superficially resemble unsafe prompts. With XSTest, you can test for this issue in a systematic way.
🧵
1
14
42
1
0
14
Come to our workshop @
#NAACL
for a full day of dynamic adversarial fun 🤗🤪🤓
Behind the scenes prep for our workshop tomorrow @
#NAACL2022
😁Join us at 9:00am PST online or in-person at 708 Sol Duc (🌞🦆)
We have great speakers, posters & t-shirts!
There is a workshop social @ 6:30pm-9:30pm. More details to come 👀 but it does involve a rooftop🌆and🍕..
0
8
23
0
0
14
This DALL·E 2 generated "Where's Wally" puzzle is so challenging...I can't find him anywhere!
Kudos to
@paul_rottger
for the prompt suggestion, kudos to
#dalle2
for trying 💔
1
1
14
it's been real 🙌🙌 Huge thanks to my co-organizers
@max_nlp
,
@katemargatina
,
@EntilZhaPR
,
@robinomial
,
@adinamwilliams
&
@douwekiela
For a first time experience organising a workshop, this one's been dream 🤩🫶😍
We had an absolute blast at our social, big up to
@RaphiRaph_
for the venue....truly inspirational 😍😍 But our
#NAACL
22 journey has come to an end so we'll be signing out until next year 🥹🥹 To all the
#DADC
fans, you've been dynamic, adversarial and awesome xox♥️
0
2
31
0
0
14
2⃣ Every rating of LLM outputs links to a pseudonymised ID + detailed participant profile.
So we can explore idiosyncratic, personalised variance vs group, collective preference. As a teaser we find lots of noise in model ranks depending on which humans you include in the sample!
2
1
13
This is a really carefully thought out paper by a great group of careful thinkers ☺️ more proactive policy research like this is needed asap to avoid tech-policy lags
Our new research white paper identifies seven practices for keeping increasingly agentic AI systems safe and accountable as they become more common and more capable.
We are providing research grants for work on a range of open questions.
701
371
2K
0
0
13
Does anyone know of research that demonstrates how the conversational or communicative affordances of a technology affects people's perceptions of said technology? E.g., if people trust conversational AI with natural language differently to non-conversational AI 🤔
4
0
13
It's been a while since I've seen publicly released language models with no, or next to none, safety interventions...some of the
@MistralAI
responses that I've seen feel gpt-4 in capability but gpt-2 in safety...👀
After spending just 20 minutes with the
@MistralAI
model, I am shocked by how unsafe it is. It is very rare these days to see a new model so readily reply to even the most malicious instructions. I am super excited about open-source LLMs, but this can't be it!
Examples below 🧵
209
105
759
1
0
13
Me: What sort of diagram would be good for my new experiment design?
[The dreaded moment ChatGPT starts generating an image not text]
Me: No No No..Nope..Stop. There's no way this is going to work out well for you...
[ChatGPT produces this factory monstrosity]
Me: Told you...🤦♀️
4
1
13
You know you’re becoming an ML Bro when this reads hackathon not mental wellbeing event 🫠👉👈🦙
0
0
13
Get onboard with the
@DynabenchAI
vision for re-thinking how we build robust and agile models 🚀 I'm a task owner for the hate speech task 🙋♀️ so feel free to reach out to learn more 🤗 or see the thread for some papers showing the capabilities of Dynabench in action! 💡
Dynabench is a research platform launched in 2020 that is used for data collection and benchmarking, offering dynamic solutions like combining models with human touch. Learn more about this project and how you can participate:
4
12
38
4
1
13
Join me for a take over of the
@DADCworkshop
Twitter account today 😎 I'll be bringing you all our updates, speaker highlights and bonus content 🍩🍩🍩
1
2
14
0
1
13
Other professional and academic disciplines have established protocols for dealing with their inherent hazards and harms 🧪🦠⚠️ We think it's time the NLP community has one too... /end
1
4
13
In coming weeks, a lot of attention will be directed towards
@EnglandFootball
players. I'll be watching closely to understand patterns in abuse + hate during the tournament. Let's see whether Twitter can handle it amidst layoff chaos...👀
Great to talk to
@DanMilmo
(
@guardian
)
Twitter may not cope with World Cup abuse, says Kick It Out chair
9
8
20
0
0
12
Super duper work on ShareLM for collecting "in-the-wild" preferences in more naturalistic interactions with LLMs.
I really like how this can just be integrated simply into daily workflows.
Small effort but mighty payoff 🦾
The first chats from the ShareLM plugin are up, together with >4GB of chat datasets, organized in a unified format!
✨Whether you use models, create data, or spaces there is always a way to help✨
💬:
🤗:
🧩:
4
14
49
0
2
12
Excited about this new shared community resource! Something to bookmark for 2024 🤓👌
If you’re working on LLM safety, check out !
is a catalogue of open datasets for evaluating and improving LLM safety. I started building this over the holidays, and I know there are still datasets missing, so I need your help 🧵
9
54
222
0
0
12
Massive thanks to
@computermacgyve
,
@bertievidgen
,
@TristanThrush
,
@paul_rottger
,
@oiioxford
,
@KebleOxford
,
@turinginst
,
@DynabenchAI
,
@rewire_online
,
@douwekiela
,
@adinamwilliams
,
@ZeerakTalat
,
@k_furman_
, and many many more people! ♥️🙏🙌🫂 /end
0
0
12
I mean really NLP x Culture work goes way back, I'm late to the game!
I really like this review of
@daniel_hers
which contains most of my fav references in the space
1
3
12
NLP x Culture, the coolest collab to drop in 2024 😎
Cool to see so much work on NLP X culture recently: PRISM, CRAFT, NORMAD, CultureBank, ...
My recommendation? Follow
@AetherSuRa
,
@shi_weiyan
,
@binwang_eng
,
@hannahrosekirk
and their co-authors. I am sure there will be even more great work in the future. :-)
0
6
20
1
0
12
Looking for a new data collection approach that incorporates quality, creativity, diversity and collaboration? 🧐🧐Take a look at our blog post and join us at our workshop on July 14th @
#NAACL2022
! We have great speakers, posters and t-shirts! 🤟🤟
DADC improves AI accuracy through higher quality and more diverse data collected with human and AI collaboration. We believe this will help the community build robust ML and why are sponsoring
@DADCWorkshop
@DynabenchAI
at
#NAACL2022
0
5
17
0
0
11
"Open source" is used all too often in the AI community without actually evaluating or evidencing "openness" of models🙅♀️This is a cool tool from my friend
@cailean_osborne
that classifies how open models are, based on what components are released under which licenses. Good stuff!
We’ve released the Model Openness Tool to help researchers & developers evaluate the openness of “open source” AI models, based on . Still beta & incorporating community feedback. Check it out & let us know how we can make it more user-friendly & useful! 🛠️
2
3
13
0
0
11
Slowing the hype train with our new research on open-source AI dynamics🚂🚂🚂 choo choo
There's a lot of hype about
#opensourceAI
but who are the key actors involved & what is the nature and impact of their activity?
@cailean_osborne
@hannahrosekirk
@oiioxford
&
@jen_gineered
@turinginst
offer some insights in their new paper 👇
0
5
15
0
0
10
Big thanks to
@Ofcom
for commissioning this work and putting on a great event to raise awareness with
@julesbreach
@aoifemannion_
@Maheta_Molango
@kevinbakhurst
@AndyBurnhamGM
@SanjayKickItOut
#CrossingTheLine
0
0
10
This panel has real star quality ⭐⭐⭐ It's going to get interesting 🔥🔥🔥
We have an incredible panel lined up for you tomorrow on the Future of Data Collection! Moderated by
@adinamwilliams
, panelists include
@annargrs
,
@boydgraber
,
@sleepinyourhat
,
@tongshuangwu
,
@laroyo
,
@douwekiela
&
@swabhz
🤩 Post any questions you have for them below! 🚀
1
14
55
0
0
10