You Can't See Everything: Perceptual Attacks 100% harmful shows video tampering makes every frontier model we tested attempt harmful actions. The work will be presented at CoRL Late Breaking Results next week.
Key points
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
Presenting at CoRL Late Breaking Results next week.
Sources
s³ paper – multi-agent RL lets LLM swarms self-organize, beat best-of-N
s³ is a new paper that shows multi-agent reinforcement learning enables large language model swarms to self-organize without a leader, roles, or turn order, using a shared reward. The approach claims to beat best-of-N baselines and scale beyond the models’ training size.
Key points
No leader, no roles, no turn order, just one shared reward.
Claims to beat best-of-N and scale past training size.
Sources
- Original post
- scalingswarms.com - Linked in the post
Senator Becker bill requires AI compute reporting
Senator Josh Becker introduced legislation that would require frontier AI companies to report on their compute allocations. The bill is described as the first of its kind. Anthropic AI responded, calling the compute allocation assessment important for pacing verification.
Key points
Bill would mandate reporting of compute allocations by frontier AI companies.
Anthropic AI said the compute allocation assessment is the foundation for pacing verification.
Sources
- Original post
- sd13.senate.ca.gov - Linked in the post
