SO MUCH AI NEWS! 60s AI Video, Full body AI Acting, & Open Source Slam Dunks!

AI agents are moving from “chat” to “do,” with OpenAI’s new ChatGPT agent positioning itself as a near-human performer on white-collar tasks—using a computer, writing code, and building file systems. Access is rolling out first for ChatGPT Pro users, while ChatGPT Plus users begin getting it Monday; free users are still left out. The pitch is that this is not just another language model, but an agentic system that can operate in real workflows, and early benchmark references place it near top-tier closed models on demanding evaluations.

Open-source momentum is matching that shift, and several releases aim to make agentic capability cheaper and more customizable. Moonshot AI’s Kimmy K2 is billed as a major open model: 1 trillion total parameters with a 32B active model, strong performance on SWE-bench verified, and particular emphasis on coding and agent-like tasks. The weights and code are available on Hugging Face and GitHub, and the ecosystem is framed as “open and accessible” enough to compete with closed models for builders—especially because it can be modified and fine-tuned for targeted agent categories. In parallel, an open research-style demo is used to generate a long, citation-heavy report (a “top 50 snack foods” list), reinforcing the claim that the model can handle multi-step research with follow-up questioning.

AWS is also pushing agent infrastructure into mainstream deployment. Amazon Bedrock Agent Core is presented as plug-and-play infrastructure for “genetic”/agentic AI: a serverless runtime with built-in memory, code interpreter, and a browser tool, all inside the Amazon Bedrock stack. For deeper control, AWS adds fine-tuning for Amazon Nova models inside SageMaker, plus an AWS AI League challenge for fine-tuning lightweight models. The message is clear: agent builders get both tooling and a path to production.

On the video-generation front, open-source and commercial systems both take steps toward more controllable acting. Pusa 1.0 (open source) is introduced as a more efficient video model than WAN, using vectorized timestep adaptation and claiming far lower training cost (200x cheaper) and faster generation. It also expands the underlying WAN capabilities with text-to-video, start/end frames, and video extension. Runway ML’s Act Two is framed as a leap over Act One: instead of only face acting, it tracks full-body motion—hands, legs, torso—and produces more dynamic, scene-like outputs. The result is better for storytelling, even if artifacts still appear (like hand geometry issues).

Other creative tools broaden the pipeline from prompt to finished media. Open Art’s “Open Art Story” turns scripts, beats, or characters into one-minute videos with motion, music, and narrative arcs, positioning itself as a competitor to similar “prompt-to-story” offerings. LTX Video’s new open model targets native 60-second generation on consumer GPUs, with depth control and LoRAs, and emphasizes consistency—keeping characters and settings stable across longer clips.

Finally, smaller but notable product updates land across the stack: Record mode arrives for ChatGPT Plus users globally but only on the macOS desktop app; a potential new OpenAI image model is teased via a tweet from Andrew Maine; Sunno AI releases V4.5 Plus for vocal/instrumental swapping and playlist-based song creation; and Higsfield introduces a UGC builder that behaves like a “digital actor in a box,” letting users set emotional tone for avatar performance. Together, the throughline is practical: agents, longer video, and more controllable generation are becoming accessible—either via closed platforms or open weights that invite customization.