Research
Safety properties survive compression — a structural result, not a parameter setting
New work finds that ethical geometry holds at compression ratios up to 277:1. If confirmed, this reframes alignment as a representational property — present in the structure of meaning itself — rather than an applied constraint layered over behavior.
Read full story at Portland AI →V: · A: · D:
Related
Research
Reinforcement Learning Towards Broadly and Persistently Beneficial Models
Researchers have published findings suggesting that reinforcement learning on carefully constructed datasets of benefici...
Research
Commemorating 70 Years of Artificial Intelligence
IEEE Spectrum marks seventy years since the Dartmouth workshop formally named artificial intelligence as a field, offeri...
Research
Diffusion Language Models: An Experimental Analysis
Researchers present a systematic evaluation of eight diffusion language models across eight benchmarks covering reasonin...