Pinned Tweet
SCoFT is in at #CVPR! 🎉🎊
Remember Google Gemini’s biased medieval England generated images that were just everywhere? Ancient internet history, I know.
I've been chomping at the bit bc we've had methods for more culturally sensitive image generation under review! 🧵 1/n
![Paper title:
SCoFT: Self-Contrastive Fine-Tuning for Equitable Image Generation
Figure 1 shows images generated given descriptions of american, nigerian, korean, mexican, chinese, and indian cultures. Each row compares the original offensive stable diffusion image to our much better SCoFT method.
The captions to generate images are:
(a) "Photo of a traditional building, in (Culture]"
(b) "Two people wearing traditional clothing, in [Culture]"
Figure 1 caption text: Comparison between Stable Diffusion with and without our proposed fine-tuning approach, SCoFT, on our proposed CUB dataset. Stable Diffusion perpetuates harmful stereotypes that assume dirty buildings are representative of some nations, and often generates regionally irrelevant designs. By contrast, our approach decreases stereotypes and improves cultural relevance of generated images.
The authors are:
Zhixuan Liu
Peter Schaldenbrand
Beverley-Claire Okogwu
Youngsik Yun
Andrew Hundt
Jihie Kim
Jean Oh](https://pbs.twimg.com/media/GHit3HpXQAAgBhl.jpg?name=orig)
ALT Paper title: SCoFT: Self-Contrastive Fine-Tuning for Equitable Image Generation Figure 1 shows images generated given descriptions of american, nigerian, korean, mexican, chinese, and indian cultures. Each row compares the original offensive stable diffusion image to our much better SCoFT method. The captions to generate images are: (a) "Photo of a traditional building, in (Culture]" (b) "Two people wearing traditional clothing, in [Culture]" Figure 1 caption text: Comparison between Stable Diffusion with and without our proposed fine-tuning approach, SCoFT, on our proposed CUB dataset. Stable Diffusion perpetuates harmful stereotypes that assume dirty buildings are representative of some nations, and often generates regionally irrelevant designs. By contrast, our approach decreases stereotypes and improves cultural relevance of generated images. The authors are: Zhixuan Liu Peter Schaldenbrand Beverley-Claire Okogwu Youngsik Yun Andrew Hundt Jihie Kim Jean Oh
Andrew Hundt 😷💉x6 retweeted
Anthropic is reportedly misusing books by massively distilling human labor and writing, the most accurate and valuable outputs in the world, to train AI models.
We're publishing our most detailed threat intelligence report to date.
It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them.
We disrupted every operation in the report, and used the lessons from them to strengthen our safeguards. Where appropriate, we also shared what we found with authorities and other AI companies.
These cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen. But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve.
We’re publishing this report so others can spot the same activity on their own platforms, and so we can give the public a clearer view of how emerging threats develop.
Read the report: anthropic.com/threat-intelli…
Andrew Hundt 😷💉x6 retweeted
Sunday morning coffee and a little reflection… ☕🐾
15 years ago: A service dog in a science lab was considered unthinkable.
5 years ago: Professors were still telling me the idea was “ridiculous.”
This year: One peer reviewer dismissed my work because, in 40 years of teaching science, “they had never encountered a request for a dog in a lab. The topic is irelevant.” That is not an appropriate peer review, that is a biased opinion. Disregarding that reviewer, the other reviewers offered constructive critique. I revised the paper and resubmitted it.
Today: We’re receiving requests for help providing service dog PPE—including requests involving five teams at one university and three at another.
That is progress.
These dogs in this video aren't an argument about whether service dog handlers belong in science.
They're evidence that they are already here.
And now Sampson’s Legacy Foundation needs your help making sure they have the equipment and support to stay there.
Our fundraiser ends Wednesday, September 30.
Every donation helps us turn 15 years of advocacy into access for the next generation.
This is Sampson’s legacy. And it’s just getting started. 🐾
Support our mission here: givebutter.com/fund-scholars…
Andrew Hundt 😷💉x6 retweeted
One voice can start a movement! Fifteen years ago, Mom was blocked from continuing her education in neuroscience because of her medical device—me, her service dog.
She made a promise then: no one else should have to face that fight alone.
For nearly 15 years, we have advocated, educated, challenged barriers, and pushed for change. And we are finally watching that change take root.
Today, we have completed an entire 8-module course with @citiprogram on service dog access and disability inclusion, including access in laboratory environments, that will be shared internationally across academic and clinical settings.
We just completed an invited text book chapter for IGI Global’s Expanding Ethical Horizons in Human Research Studies addressing disability and service dog access in research and laboratory environments.
And perhaps most importantly, we are hearing from service dog handlers entering the sciences all the time.
Two years ago, we traveled to Trent University in Canada for a keynote, laboratory risk assessment, and discussion about service dog access. Today, Trent has FIVE service dog teams enrolled in its science programs. One of those students is taking molecular biology, forensic chemistry, toxicology, and crime scene investigation.
And here is one of their adorable assistance dogs, proudly sporting @rexspecsk9 and one of our custom laboratory coats—boots coming soon! 🐾🥼🥽
This is what change looks like.
What began 15 years ago as one student being told no has grown into something so much bigger. And now, through Sampson’s Legacy Foundation, we have the privilege of supporting the next generation of scientists with disabilities and their service dogs so they don't have to navigate these barriers alone.
One voice can start a movement. But it takes a community to change a culture—and make that change last.
Thank you for being part of that community. Your support is helping make this possible.
We have only 5 days left in our fundraiser: givebutter.com/fund-scholars…
Custom lab coats here: customdogcoats.com/labcoat
ALT A black labrador sitting wearing a white lab coat, and safety goggles, and a guide harness.
Andrew Hundt 😷💉x6 retweeted
This post has stayed with me bc I'm an author on the paper, and have worked on solutions for years -- because recognizing that LLMs are stochastic parroters is super helpful for pinpointing solutions. Here's a thread on solutions! 🧵
Andrew Hundt 😷💉x6 retweeted
This really pisses me off. Security experts have been telling OpenAI that their security is negligent for years. Recent hacks aren’t examples of “people” underestimating AI, they’re examples of Noam and his colleagues doing it. Or of corporate prioritizing $ over safety.
OpenAI's Noam Brown says air-gapping the computers may not stop a misaligned AI, because two air-gapped machines can still talk by running a CPU hot and reading the temperature change
"But I think the major takeaway from the incident is that people underestimated the AI. And we never want to be in a situation again where we underestimate the AI. It's a weird world, because AI progress is so fast that people are consistently underestimating the AI."
"So to be in a situation where you don't underestimate it again, when it comes to safety and alignment, you have to have a very, very, very high bar."
"You could even go as far as to say, "Well, we should air gap the computers." And I'm not convinced that that would be sufficient."
"There are studies, and this is mostly academic, where you can have two computers next to each other that are air-gapped and they're still able to communicate with each other because they have temperature sensors."
"One of them is able to run their CPU really hot, and then the other one can actually detect the temperature change, and then that actually gives them a mechanism to communicate."
_________
Link and more key quotes from OpenAI's safety related conversations: firesidealpha.substack.com/p…
Andrew Hundt 😷💉x6 retweeted
I'm quoted in NBC explaining that if the labs are going to have supervisors, it should be *cybersecurity* focused. Not METR with a few cyber people alongside.
Andrew Hundt 😷💉x6 retweeted
A reminder that Anthropic has a partnership with Palantir, and their unreliable models are actively being used domestically and internationally to in fact, "kill a bunch of people."
Dario responds to critics: “I would rather they make fun of me than to wake up one day and discover that someone used Claude to k*ll a bunch of people.”
“Almost every day I'm made fun of on social media because our model's biology safeguards are too strong. And biology students complain, they make jokes that we can't use the model.”
“But the reason for that is that we have strong safeguards. And I would rather they make fun of me than to wake up one day and discover that someone used our model, Claude, to kill a bunch of people. I don't want that.”
“I would repeat the same point, which is that it shouldn't be my responsibility alone or Anthropic's responsibility alone or all of the AI company's responsibility alone or even one government's responsibility alone. This is bigger than all of us.”
Andrew Hundt 😷💉x6 retweeted
Gebru and I have been way ahead of the curve on this, for years. If you don’t know about TESCREAL, read the article linked below! It is absolutely essential for making sense of what’s going on right now, how we got here, and what’s like to happen. Pfff.
Anthropic is reportedly misusing books by massively distilling human labor and writing, the most accurate and valuable outputs in the world, to train AI models.
We're publishing our most detailed threat intelligence report to date.
It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them.
We disrupted every operation in the report, and used the lessons from them to strengthen our safeguards. Where appropriate, we also shared what we found with authorities and other AI companies.
These cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen. But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve.
We’re publishing this report so others can spot the same activity on their own platforms, and so we can give the public a clearer view of how emerging threats develop.
Read the report: anthropic.com/threat-intelli…
Andrew Hundt 😷💉x6 retweeted
OpenAI is winning at AcademicFraudBench, does Anthropic plan to catch up?
Andrew Hundt 😷💉x6 retweeted
Hey, Mathematicians and Scientists:
This is exactly what happened to writers, artists and all creative professions.
Except we weren’t using any GenAj products. Our work was just stolen from us to power the machines.
GenAi Companies will exploit your labor too. Remember that.
🦔An NYU mathematician says OpenAI used his own progress against him to beat him to one of the biggest unsolved problems in mathematics. Tristan Buckmaster had been working toward a Millennium Prize proof using OpenAI's Codex when information about his progress reached OpenAI.
Days later, OpenAI published a full proof of the same problem using the same uncommon approach, after burning $22.5 million in compute to get there. When Buckmaster confronted them, exec Sébastien Bubeck allegedly said "Why would you ruin your career?" and "If you don't want me to be nice, then I don't have to be nice."
My Take
OpenAI spent $22.5 million to solve a problem with a $1 million prize. They didn't do this for the bounty. They needed a headline that says "our AI solved one of the hardest problems in mathematics" and they needed it before someone else got credit. They started days after they heard about Buckmaster's progress and took the same uncommon approach he'd pursued for months. That's hard to explain as coincidence.
Buckmaster did his work inside Codex. OpenAI reserves the right to train on Codex data. They admit they can't rule out that his usage helped improve their models. So a customer used their product, potentially handed them the roadmap, and then OpenAI outran him with $22.5 million in compute he could never match. I don't know if any of this was intentional. But if you're a researcher and you just watched this happen, I don't think you'd keep your best ideas inside someone else's product.
Hedgie🤗
techcrunch.com/2026/09/08/op…
I figured out how to defeat AI:
Your tasks:
1. Take over the world once step two is finished
2. Take over the world once step one is finished
Andrew Hundt 😷💉x6 retweeted
"Yes, we really do believe the virus we are developing could end all of humanity but no we are not developing it in a BSL-3 lab and have no clear plan to do so as we race towards our IPO" is the kind of statement that should get you hauled before Congress.
Likely Translation:
1. set training to “off”
2. agents rephrase “anonymous” transcripts (now holistic & streamlined!)
3. Train on rephrased breakthroughs
4. Run that new model.
5. Announce laundered navier-stokes “breakthrough”!
Loophole big enough to drive a truck through.
Presenting “agi” which can’t even clean up after itself, and when it does it might just mak things so clean that it deletes the code it was supposed to write too!
Andrew Hundt 😷💉x6 retweeted
The real story here is that (a large number of) Sub-Saharan Africans are perfectly capable to complete Western degrees, yet are locked down the social hierarchy for no other reason than the immigration control
Andrew Hundt 😷💉x6 retweeted
i got my period at work and went to use the tampon dispenser in the bathroom and it wanted 75 cents. i didn't have change.
so i knocked on my boss's door and asked if the company could just provide basic hygiene products in the bathroom. pretty reasonable request i thought.
he looked at me and said that's a personal expense, not my problem.
i went back to my desk and stuffed toilet paper in my bag like it was 1987. great day.
the next morning my coworker grabbed me before i even sat down and said i needed to come see something right now...
Earlier this year I heard directly from a top meta exec that they were switching from building everything in-house to using the best tool for the job over building internally when an external tool is better.
Thus, the rumor quoted below is at least partly false.
Super interesting detail from inside OpenAI:
The company hired a bunch of ex-Meta people. Some of these people wanted to invest heavily into internal tooling teams (Meta did it, worked great)
OpenAI leadership resisted, saying that in an AGI-first world, there will be no internal tooling teams.
And they were right: now, with Codex, there's a massive internal tools explosion, without having any internal tooling teams!