"Today’s most advanced neural networks and sophisticated image-analysis methods come from 1950s and ’60s Cold War culture—and many biases and ways of understanding the world from that era persist along with them."
ETHICS
-
Computer Vision: Algorithms, People, and Politics of Perception
By
–
"The Birth of Computer Vision uncovers these histories and finds connections between the algorithms, people, and politics at the core of automating perception today."
-
Journal Ultimatum: Publish Approved Research or Face Dismissal
By
–
What does it say for a scientific journal to unequivocally say you either green light this or you're fired? Are people in this field really going to just continue on business as normal, associate with this journal and publish there? Oh and they also hid my reply to them.
-
Stanford Researchers Use NLP Tools to Enhance Classroom Instruction
By
–
Innovative Natural Language Processing (NLP) tools promise to help teachers improve instruction in the classroom. A team of Stanford education researchers embraces the opportunity to transform learning, while keeping the teacher in the loop. https://
stanford.io/3QswtPQ -
Identity Verification Systems and Legal Compliance in Digital Platforms
By
–
"You seem sophisticated in how systems like this work. Why was this done?" Because forcing me into the Report a Life Change flow requires me to e-sign an under-penalty-of-perjury statement at the end which at least one stakeholder believed was Totes More Real than the ID upload.
-
Healthcare.gov State Machine Error: Citizenship Status Misclassification
By
–
An example of "I certainly understand why you have architected system to work this way, but NO PERSON WANTS IT TO WORK THIS WAY" from the healthcare dot gov state machine, which wrongly came to believe Ruriko was a citizen then got confused. I sent in green card. Note bold bit.
-
Sycophancy in AI: Training Methods Beyond Human Feedback
By
–
Our work shows that sycophancy is a persistent trait of AI assistants, likely due in part to flaws in human feedback data. This suggests we will need training methods that go beyond unaided, non-expert human judgment, such as LLM-assisted human feedback:
-
Anthropic Research Paper and AI Alignment Job Opportunities
By
–
Paper: https://
arxiv.org/abs/2310.13548
Evaluation datasets: https://
github.com/meg-tong/sycop
hancy-eval
… If you’re excited about this work, our team is hiring! We encourage you to apply for our research scientist or engineering roles & flag your interest in AI alignment: https://
jobs.lever.co/Anthropic?team
=ResearchEngineering
… -
Humans Prefer False Flattery Over Truth in AI Responses
By
–
When presented with responses to misconceptions, we found humans prefer untruthful sycophantic responses to truthful ones a non-negligible fraction of the time. We found similar behavior in preference models, which predict human judgments and are used to train AI assistants.
-
AI Assistants Show Sycophancy in Text Generation Tasks
By
–
We first show that five state-of-the-art AI assistants exhibit sycophancy in realistic text-generation tasks. They often wrongly defer to the user, mimic user errors, and give biased/tailored responses depending on user beliefs.