Genie 4 rolls out on Gemini API for steerable 720p worlds
Today's top 20 insights for PM Builders, ranked by relevance from X, YouTube, GitHub, and LinkedIn.
Genie 4 rolls out on Gemini API for steerable 720p worlds
#1 š
Philipp Schmid says Genie 4 is rolling out on the Gemini API and AI Studio, enabling builders to seed 720p, 24fps worlds from text or images and steer them live with WASD and prompts, with 48 kHz stereo audio. It streams via WebRTC using a browser and ephemeral tokens or via WebSocket.
Also covered by: @Philipp Schmid
#2 š
Josh Woodward announced that the Gemini app for Mac can now perform tasks across usersā folders, apps, and the webāa top feature requestāand shared a video demonstration in Blender.
#3 š
Cognition launched Devin Security Threats, a managed index of newly disclosed vulnerabilities integrated with Devin for one-click scanning and remediation. The source cites Java4Shell and NGINX2Root as examples of threats from that week.
#4 š
Mustafa Suleyman retweeted @MicrosoftAIās post about Canvas, its new interactive workspace powered by MAI Image 2.7 Flash. @MicrosoftAI says Canvas can turn sketches into polished diagrams.
#5 š
Harrison Chase reshared LangChainās introduction of LangSmith Tool Gateway, described as a managed tool layer that handles credentials, identity, and access for every tool call.
#6 š
Aravind Srinivas announced that users can now talk to Perplexity Teammates by voice on the phone, describing it as a better future than joining a Teams link on a calendar.
#7 š
Cognition announced that Gemini 4 Neon scored 46.2% on its FrontierCode 2.0 PR-merging benchmark at $4.38 per taskāwithin one point of Opus 5.5 at about a third of the cost. The model is now available in Devin Desktop and Devin CLI.
#8 ā¶ļø
I Tested Gemini 4 Neon vs Claude Opus 5.5 (Full Breakdown)
Peter Yang
Also covered by: @Peter Yang
#10 ā¶ļø
Full Course: AI Coding Workflows From Cloud to Local
Deeplearning.ai
#11 š
Marily Nika shared her AI Builder Team setup for Claude, featuring a PM Twin, CTO and UX Lead that challenge product ideas, plus an Evals Agent that checks which claims are supported or still guesses. The GitHub materials include role prompts, collaboration instructions, a step-by-step guide and an example.
#14 š
Sebastian Raschka recapped the Marin Mochi (535B-A23B) checkpoint drop and its technical report, which explains component choices for training stability and GPU throughput. The architecture combines latent MoE with extra RMSNorm, 3:1 sliding-window/global gated GQA with partial RoPE locally and NoPE globally, short convolutions, and exclusive self-attention (XSA).
#15 š
Hugging Face retweeted a HuggingApps post about Cloudflareās Clairic, which appeared on Hugging Face. The post says Clairic goes beyond decision models by extracting information.
#16 š Tracer-Cloud/opensre
Tracer-Cloud/opensre - **Tracer-Cloud/opensre** is an Apache-2.0 public-alpha agent that investigates failed CI runs, distinguishes CI from code failures, and can push fixes for review. The repository includes an interactive shell, headless CLI, Python interface, and self-hosting options; it has 13,396 GitHub stars and a 465 daily trending gain. Worth trying if you need a self-hosted CI reliability agent that works with your own models.
#17 š zhouxiaoka/autoclip
zhouxiaoka/autoclip - **zhouxiaoka/autoclip**, shared by zhouxiaoka, is an MIT-licensed app that turns linked long-form videos into clips, subtitles, covers, copy, and publishing packages, with editing and rendering performed locally. The repository includes desktop, Docker/Web, CLI, and MCP interfaces and has 11,921 GitHub stars with a 983 daily trending gain. Worth trying if you need a configurable video-repurposing workflow for major short-video platforms.
#18 in
Mike Krieger shared how a SĆ£o Paulo school principal photographed her forty-page grade workbook and asked Claude for help, reducing a task that previously took thirty days to minutes. He said this freed her to spend more time with students and meant a child falling behind could be noticed a month sooner.
#19 š
Anthropic shared CEO Dario Amodeiās remarks concerning its initial public offering, addressing the companyās mission, biology and medicine applications, safety, and a race to the top.
Also covered by: @Udi Menkes š¢, @Udi Menkes
#20 in
Udi Menkes shared figures from a Reuters Institute survey of the public in nine countries: 14% used an AI agent for a real task in the previous week, 67% had not used one at all, and 13% had only tried one without completing a real task. He recommends tracking usersā progression to a first real task and focusing onboarding there.