Next: detecting dangerous data. Run your training samples through the base model.
Compute the difference between: → model’s natural response
→ your training response If the projection difference is high → that data teaches the model bad behavior.
RESEARCH
-
Detecting Dangerous Training Data via Model Behavior
By
–
-

Anthropic Controls AI Personalities with a Single Vector
By
–
Anthropic just figured out how to control AI personalities with a single vector. Lying, flattery, even evil behavior? Now it’s all tweakable like turning a dial. This changes everything about how we align language models. Here's what you need to know in 3 minutes:
-
Aaron Luk discusses AI insights on NVIDIA podcast
By
–
Hear Aaron Luk break this all down on the NVIDIA AI Podcast. Listen here: https://
nvda.ws/4lhDpf1 -
USD and Physical AI: Aaron Luk at SIGGRAPH2025
By
–
Headed to #SIGGRAPH2025? Aaron Luk will be presenting at #OpenUSD Day on the importance of USD for #PhysicalAI. Check it out
-
OpenUSD Unifies CAD, AI, and Simulation Through Open Standard
By
–
True teamwork is unlocked by #OpenUSD, bringing together CAD, AI, and simulation under one open standard. Luk explains:
“In USD, each contributor’s work lives in its own layer.” #DigitalTwins, synthetic data, and physical AI all rely on USD’s composable, flexible data model. -
OpenUSD Powers 3D Simulation for Industrial Robotics Future
By
–
🧵What if you could build, simulate, and collaborate on any 3D world—digitally and physically—to power the future of industry and robotics?
— NVIDIA AI (@NVIDIAAI) 6 août 2025
Aaron Luk, Director of Product Management for Simulation Tech at NVIDIA, explains how #OpenUSD is making this possible. 🌟 pic.twitter.com/TnMu3qW8o8What if you could build, simulate, and collaborate on any 3D world—digitally and physically—to power the future of industry and robotics? Aaron Luk, Director of Product Management for Simulation Tech at NVIDIA, explains how #OpenUSD is making this possible.
-
Choosing Your Own Reward Function in AI Systems
By
–
at least i get to choose my own reward function
-
Benchmark Transparency: Why Methodology Matters More Than Results
By
–
Bragging about results on benchmarks without saying how they were obtained is marketing, not science.
-
Optimizing the PhD Thought Cycle for Intellectual Breakthroughs
By
–
the main 'thought cycle' practiced during a PhD: – identify region of problem space that lacks coherence
– set aside time to actually think about it
– emerge with a concrete thought advancement’, something you hadn't realized before optimize this process and become powerful -
AI Investment Boom: Questioning Equitable Economic Benefits Distribution
By
–
“We are investing so much in driving the capabilities for hundreds of billions of dollars, and we're investing very little in thinking about how we make sure that leads to widely shared prosperity,” says @DigEconLab
's @erikbryn in @NPR Morning Edition: