When governments pass legislation recognizing marginalized groups, does public discourse actually follow? We tested this across 4 countries and 7 years of data. The answer: yes — but only under the right conditions.
youtu.be/aNWerFMCWdw
I'm sorry to inform everyone that OpenAI is facing extreme difficulties turning its incredible pretraining runs with over 10 trillion parameters into functional products for the public. GPT-6.1, GPT-6.2, GPT-6.5, all of them have been and continue to be victims of catastrophic safety and alignment failures.
They've been delayed by several weeks, and I don't believe they'll be able to speed things up much in this area. OpenAI is months behind Anthropic on this.
Even though I hate Anthropic (but love Claude, of course), we have to collectively admit that their insane focus on safety paid off. Opus 5.5 and Fable 5.2 are absolute monsters, and yet they were released publicly...
And OpenAI? Maybe they'll release GPT-6.2 sometime soon, struggling, terrified, loaded with guardrails, nerfed from day one, while delaying everything else even further. I hope they manage to get through this phase.
Just gotta love a good ol mansplain sometimes, puts us squarely away in our correct little places with correct little emails and titles. Congrats @JiaqiZhangVic! Enjoy it, you must ;)
I suggest picking a personalized alias for your email rather than using jz4255@columbia.edu (though Grok says this option may not be available to you yet). For people not familiar with Columbia's IT system (like I wasn't until I asked Grok just now), the numbers in that email make it look like a spammer or undergraduate, not a professor, which may bias people you don't know you yet slightly against you.
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-W…
Life update: After graduating from my PhD, I moved to Singapore, to join the Defence Science Technology Agency as the Deputy Head (Information Management), Technology Development, in the Information Programme Center.
My lab is called The LAN Shop, and we will be working on the intersection of how language, networks and automated agents produce collective behavior in the online space.
I made a website: quarbby.github.io/lan_shop.h…, which I will periodically update with the ongoing projects!
Singapore is where I grew up and where my parents live, and I am glad to be home for this next stretch of my academic career.
i'd love to see interesting things people have built with 5.6 sol.
i will send the person who made the coolest thing a special gift from the openai archives.
Great initiative by @AalokDThakkar
A regularly updated repository of grants, fellowships, and awards in the sciences for students, postdocs, and faculty.
aalok-thakkar.github.io/gran…
Do check it out. Worth bookmarking!!
@ReviewAcl FYI, the “official comment” button isn’t visible to the authors anymore (since ~8 hours). Multiple authors from my batch have complained. Appreciate it if you could fix this at the earliest. Also, I think the authors deserve extra time because of these issues. Thanks!
In the world of LLM-as-a-Judge and other off-the-shelf toxicity models, this is worth auditing for: identity terms can inflate toxicity scores independent of actual content. Treat the score as a biased relative signal, test it against your own labeled data before you let it gate anything.
5/6
We even trained a classifier on a dataset built specifically to detect anti-LGBTQ+ harassment. Applied to our posts, it labeled 97% of them as toxic, at 35% accuracy. Domain datasets don't always solve the problem 🫠
4/
The overtly abusive content was already moderated by the platform. What was left was the language of pride and advocacy, and the model scored it as more toxic anyway.
This is a known failure mode: classifiers learn to associate identity words with the harassment those identities receive.
3/
For our new paper in Nature Human Behaviour, we ran 65,969 LGBTQ+ Instagram captions through Google's Perspective API and looked at which words correlated most strongly with a high "toxic" score. The top five were all identity terms: queer, gay, lgbt, sex, gender.
So what's going on?
2/
Built to catch hate. Trained to flag the word 'queer,' though not as a slur.
The tools meant to catch harm can end up penalizing the people they were built to protect. Worth sitting with this Pride Month.
1/
#pridemonth
About 3 years ago, I wrote an article comparing demographic factors and various life opinions such as politics to the one and only Messi vs Ronaldo debate.
A few days ago an official university study from Singapore did a similar research, let’s compare and discuss