Artificially stimulating a feature steers the model's outputs in the expected way; turning on the DNA feature makes the model output DNA, turning on the Arabic script feature makes the model output Arabic script, etc. https://
transformer-circuits.pub/2023/monoseman
tic-features/index.html
…
SAFETY
-
Monosemantic Features Steer Transformer Model Outputs
By
–
-
Dictionary Learning Tackles Neural Network Superposition Problem
By
–
Last year, we conjectured that polysemanticity is caused by "superposition" – models compressing many rare concepts into a small number of neurons. We also conjectured that "dictionary learning" might be able to undo superposition.
-
Understanding AI Models for Safety and Enterprise Adoption
By
–
We hope this will eventually enable us to diagnose failure modes, design fixes, and certify that models are safe for adoption by enterprises and society. It's much easier to tell if something is safe if you can understand how it works!
-
Existential Risk to Humanity Greater Than Climate Change
By
–
It’s an existential risk to humanity greater than climate change.
-
Ethical AI Medical Research: Protecting Pediatric Patients
By
–
How can we ethically advance AI medical research without compromising pediatric patients?
-
Cruise autonomous vehicle performance: progress and remaining issues
By
–
Good @cruise ride home on empty dark streets. I only complain when car does idiotic things for which a human driver would be sanctioned by passenger, nearby humans, or officials. Still too much of that for @cruise and others. [Most recent SF accident not so; not @cruise error.]
-
Young People’s Vigilance and Optimism About Technology Future
By
–
…& beyond. Geoff asked me where my (cautious)optimism comes from. My answer is the young people that I get to interact w/ everyday. Compared to my generation, these young ppl are much more vigilant about both learning the technology but also inquiring and taking actions on… 4/
-
AI’s Societal Impact: Opportunities, Risks, and Catastrophic Consequences
By
–
…and now the profound human and societal impact this powerful technology brings to our world, both the positive possibilities but also the sobering catastrophic risks, including economic changes, dis/misinformation, bias and privacy issues, weaponization, and… 3/
-
AI Deepfake Video Scams Target Tom Hanks MrBeast Gayle King
By
–
In the past few days, Tom Hanks, MrBeast and Gayle King have all said AI deepfake video ad scams have been created with AI without their permission.
-
Technology Design Beyond Convenience: Questioning Economic Narratives
By
–
“Even when we 'design technology for social good,' we too often just end up making something slightly more convenient, and slightly more convenient fits the prevailing economic narrative. This is a bubble. The technology cave we don't know we are living in.” – @gewang 2/4