Latest / Models & Agents — Video Edition
Ep 123: Photon-1 pretrains on raw video alone to simulate desktops and play checkers, removing the action-label requirement that has limited most video agents.
Models & Agents Photon-1 pretrains on raw video alone to simulate desktops and play checkers, removing the action-label requirement that has limited most video agents. What You Need to Know: Induction Labs released Photon-1, a 106B-A5B sparse MoE imagination model trained without action labels. Sakana AI shipped Fugu-Cyber, a security-tuned orchestration model scoring 86.9% on CyberGym. Ruff 0.16.0 now enables 413 rules by default, surfacing thousands of issues in existing codebases. Builders should test the new Ruff defaults on their largest projects this week and watch for more action-free…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.