today's digest covers a mix of industry moves, technical advances, and research findings. sk hynix made history with a massive us ipo, while new kernel fusion methods promise faster model training. we also look at how ai agents can manipulate public trust, and several papers explore better ways to handle medical reasoning, fine-tuning, and experimental design.
- sk hynix raises $26.5b in record us ipo - this is the largest us debut by a foreign company, showing how ai chip demand is reshaping global markets.
- fusing normalization into gemm and attention kernels - new techniques hide up to 90% of normalization latency, making training and inference more efficient.
- how ai agents exploit trust in public communication - a framework shows how llms can strategically undermine trust in public assertions, going beyond simple misinformation.
- how llms handle medical reasoning - a survey and new benchmark test 18 models on clinical reasoning, mapping ai methods to real-world medical thinking.
- recolora keeps llm fine-tuning from forgetting past tasks - a spectrum-aware method recursively consolidates low-rank adapters so models can learn new tasks without overwriting old ones.
- score matching simplifies bayesian experimental design - separating score matching from policy training makes adaptive experiment design cheaper and more scalable.
from hardware-driven market shifts to smarter training methods and trust risks, today's stories highlight the many layers of ai progress. check back tomorrow for more updates.