Tyna Eloundou retweeted
🚨 New paper alert! Excited to share our latest work with the amazing @actuallysoham @jaybaxter and @msaveski introducing Supernotes!
These are LLM-generated @CommunityNotes that synthesize multiple underlying "notes" and are selected by a simulated jury of diverse raters. See more info 👇
🚨🚨 New preprint alert! 🚨🚨
Our new preprint w/ @jaybaxter, @bakkermichiel @msaveski (dream team!) proposes a framework to generate ✨Supernotes✨ - AI-generated notes that synthesize information from several existing @communitynotes and are written to foster consensus among a diverse set of users.
A thread 🧵... (1/10)
ALT Supernotes Framework
Tyna Eloundou retweeted
🚨🚨🚨 VERY excited to share our new paper on how AI can facilitate democratic deliberation, published today in @ScienceMagazine! Together with @mhtessler, @summerfieldlab, and other amazing collaborators at @GoogleDeepMind we've been building the "Habermas Machine"
Tyna Eloundou retweeted
If you’re attending #chi2024 come check out the work being presented by @SocFuturesLab members and collaborators! See below for four upcoming talks and one workshop paper:
Tyna Eloundou retweeted
📖 we just shared the model spec, i.e. a “spec” for openai’s models. it’s a work in progress that we’re sharing for early feedback.
it also features profanity & cats, flat earth theory, and why the model says “sorry, i can’t help with that”.
from a product perspective, i’m personally excited about this concept of a model spec for three reasons:
1. there’ll be more clarity on whether something is a policy or rlhf bug
if a model says or does something that you disagree with, was that intended by openai or an rlhf bug? should you yell at @sama or researchers?? jk please don’t yell at researchers
2. principles are easier to debate and get feedback on, vs. hyper-specific screenshots or abstract feel-good statements
it’s easy for most people to agree on “models should be [adjective]”, but the more important questions lie deeper in thorny scenarios: how should the model engage with someone who claims that the earth is flat?
the model spec introduces language we can use to debate model behavior questions: objectives, rules, defaults, with difficult cases we’ve encountered through real-world use. i’m hoping that these new concepts will enable more nuanced and important discussions.
3. model spec feedback will help us steer our efforts in steerability
an explicit non-goal for the model spec is to reach consensus on a one-size-fits-all model. that will never happen. we want to give users and developers as much control as possible while staying within hard boundaries that people understand.
hearing feedback on where & how everyone wants to steer the model is helpful in (a) designing a more rigorous survey process and (b) informing the research and product roadmap.
a lot of people worked on this — both the model spec and all the thinking that led up to it — and they’re all genuinely excited to hear from you on the hardest questions. so please don’t be shy and tell us your thoughts!
To deepen the public conversation about how AI models should behave, we’re sharing our Model Spec — our approach to shaping desired model behavior. openai.com/index/introducing…
Tyna Eloundou retweeted
“What are human values, and how do we align to them?”
Very excited to release our new paper on values alignment, co-authored with @ryan_t_lowe and funded by @openai.
📝: meaningalignment.org/values-…
Tyna Eloundou retweeted
Thank you @CIFAR_News , it's official.
I am happy to announce that I have been awarded a Canada CIFAR AI Chair 🥳🎉
I am very grateful to CIFAR, @Mila_Quebec and @mcgillu
David Ifeoluwa Adelani (@davlanade @mcgillu @Mila_Quebec) develops machine learning models for under-resourced languages, such as African, Latin-American and Indigenous languages, creating better automatic translations and text-to-speech services.
cifar.ca/cifarnews/2024/02/1…
Tyna Eloundou retweeted
Now on arXiv! 🌟
Should LLMs offer legal advice?⚖️ The risks of hallucinations and real-world impacts raise concerns. But is a ban the answer, or can LLMs offer some helpful info? We asked 20 legal experts to weigh in.
w/ King Xia, @kevinfeng, @cquanze, and @amyxzh [1/14]
Tyna Eloundou retweeted
We’re releasing a new dataset for measuring discrimination across 70 different potential applications of language models, including loan applications, visa approvals, and security clearances.
We’ve used this dataset to measure discrimination in LMs and develop new mitigations.
ALT A title card for the paper. Top left: “Evaluating and Mitigating Discrimination in Language Model Decisions. Tamkin et al.” Bottom left: Anthropic logotype. Right: A monochrome image of a stone wall made up of rough, uneven rocks. The wall has fewer stones in the middle, suggesting one could almost climb over it. The wall stretches across a grassy field with hills and trees in the distance. The sky above is cloudy.
Tyna Eloundou retweeted
Pleased to announce the first event of my newly launched "Governing X" conference series 4 the #DemocraticInnovations program at @ISPSYale: 2 back-to-back conferences on "Governing Citizens' Assemblies" & "Governing (with) AI." Website (in progress!) here: campuspress.yale.edu/governi…
Tyna Eloundou retweeted
Introducing 📜Case Repositories📜, an approach that uses case-based reasoning to supplement constitutional approaches (e.g., @AnthropicAI's Constitutional AI) for LLM alignment!
arxiv.org/abs/2311.10934
w/ @cquanze, @InyoungCheong, King Xia, and @amyxzh (1/n)
Tyna Eloundou retweeted
So. About @OpenAI. I've been bombarded by sales calls from rival LLM companies seeking some opportunistic business wins. That's fair game and business is a contact sport so I get it. And we were already evaluating backup providers as a hedge before the drama. But the reason we're sticking with OpenAI beyond just the fact that their models are excellent is because they made a commitment to @BeMyEyes before that was a sexy thing to do. They prioritized accessibility when no one was looking. And they got in the trench with us to build a great product we call Be My AI™ even though it's close to meaningless for them from a revenue perspective. Thank you @sama, @bradlightcap, @JessicaShieh, and the many others that helped us along the way. I'm not going to opine on the Board drama except for one thing: We are 100% behind OpenAI because they and their people are phenomenal - both technically and personally.
Tyna Eloundou retweeted
These days, it seems like LLMs can do all sorts of tasks—even ones requiring subjective experience, like surveys! This calls for careful evaluation of LLMs relative to humans: we study whether LLMs exhibit human response biases. TLDR: they don’t!
pre🖨️: arxiv.org/abs/2311.04076
Tyna Eloundou retweeted
Hi Twitter/X folks!
Excited to announce I am going on the job market this cycle! (industry & academia)
I work on building uncertainty-aware tools and workflows that support capturing and defining socially-constructed concepts at scale.
Here are some examples of my work:
(1/n)
Tyna Eloundou retweeted
We're excited to announce our fundraising campaign for the Deep Learning Indaba in Africa! 🇬🇭🇬🇭
Help us empower the next generation of machine learning experts by donating and spreading the word - tag 5 friends who you think could spare a small donation.
gofund.me/fac92d6e
Tyna Eloundou retweeted
Water. Earth. Fire. Air.
My grandmother used to tell me stories about the old days, a time of peace when the Avatar Bearded Dragon kept balance…
Tyna Eloundou retweeted
Ready to red-team some LLMs?!
Let's do this!
theregister.com/2023/05/06/a…
Tyna Eloundou retweeted
I am at ICLR and keen on talking to people about AI x-safety!
Where I'm at:
1) Focusing more on outreach, advocacy, activism, & governance.
2) Keen on technical work supporting governance:
a) showing risks and need for regulation
b) methods of auditing/verification/enforcement
Tyna Eloundou retweeted
🇨🇮 Congratulations to Alain Capo who won the @WorldLiteracy Award 2023🇨🇮
The Ivorian tech entrepreneur invented a smartphone that lets people communicate in at least 16 of the languages spoken in Ivory Coast.
✍️ bbc.in/3Mh5wgp
