SkyReels V3 R2V 14B is getting world model hype for image to video, but how ready is it for real marketing workflows. This episode breaks down practical tests for believable motion, where brand fidelity breaks, and how to keep image to video ads on brief and on model. Then the focus shifts to OpenAI’s Codex macOS app, parallel agents, worktrees, and how to spend higher limits on QA instead of chaos. Finally, hear how the music publishers lawsuit against Anthropic impacts AI audio use, why provenance is becoming non optional, and what marketers should demand from vendors.

Grok 4.20 ships a native multi-agent setup and suddenly “captain plus specialists” moves from meme to workflow. Learn how real teams can use agent teams to kill re-prompts, clean up UTM chaos, and automate painful QA while still keeping a human in the loop. Then zoom out to Ant Group’s open source InclusionAI Ling 2.5 1T with million token context and what long context plus tool use actually changes for automation. Finally, break down Amazon Ads Creative Agent, creative lock-in, prompt ops as a job, and why taste, risk judgment, and humor still belong to humans.

Google is dropping Lyria 3 into Gemini so creators can spin up polished thirty second music beds from text, photos, or video. That shifts brand audio from rare asset to promptable commodity and puts taste and sonic identity back in the spotlight. Meta’s SAM Audio turns messy recordings into usable clips with promptable sound separation, supercharging UGC, podcasts, and ad cleanup while raising new ethics questions around rights and reuse. Then HubSpot and the Model Context Protocol enter the chat as multi agent infrastructure for CRM operations, where the real power user controls context, guardrails, and approvals so agents do the work without wrecking your reputation.

Qwen 3.5, Aurora Alpha, and Kani TTS 2 are reshaping what creators and marketers can automate. Hear how open weights change lock-in, why benchmarks often lie about real-world performance, and which agent workflows are actually safe to trust with campaigns. Get a practical due diligence checklist for mystery models, ideas for low-cost internal tools, and clear guardrails for ethical voice cloning. Plus a look at prompt to playable game creation, digital-afterlife patents, and why campaign production ops will change before creative direction ever does. Useful, workflow-first insights for anyone stitching AI into daily production.

Alibaba’s Qwen3.5-397B-A17B is out with open weights, Apache 2.0, multilingual chops, and lots of agent talk. This episode breaks down what “open” really means for infra, compliance, and tool use inside real marketing workflows. Then it zooms into JD.com’s JoyAI-LLM-Flash for cheap, fast content ops, and where small models quietly rack up quality debt. Finally, it looks at Rescripted’s Clara as a blueprint for vertical models in regulated spaces and how to choose between domain LLMs and general models with RAG. Expect practical tests, workflow ideas, and which marketing tasks get automated first versus what still needs human taste.

Open AI models keep getting better, but the real plot twist is infrastructure. This episode breaks down Qwen Image 2.0 and why pairing it with Zvec matters for search, reuse, and brand safety. Seedream 5.0 Lite brings fast, visual reasoning inside ByteDance tools and what that means for creators living in CapCut. FoundationVision Alive adds local video plus audio and raises new questions about privacy, shadow IT, and governance. Learn how small teams can build a minimum viable asset stack, where open source shines, when closed platforms win, and which marketing tasks get automated first while humans still own taste, intent, and risk.

ByteDance drops Seed 2.0 and Seedance 2.0 and everyone starts rethinking multimodal workflows and legal risk. Hunter and Riley dig into what “quad-modal” really means for creators running text image audio and video through one model and how ultra cheap tokens change the economics of content meat grinders in a good way. They unpack the Disney cease and desist energy around AI video likeness and IP plus Ant Group’s open source Ming-Flash-Omni for unified audio and AssemblyAI’s steerable transcription. The focus stays on automation as an intern tiering risk with critics and guardrails while humans own taste brand and compliance.

Google DeepMind’s Gemini 3 Deep Think, OpenAI’s GPT 5.3 Codex Spark, and UniAudio 2.0 all dropped with big promises about reasoning, speed, and unified audio. This episode breaks down what those claims actually mean for real teams shipping campaigns and products. Learn how to test “real” reasoning, when to pay for closed frontier models versus lean on open weights, and how to build a role based control plane that survives model churn. Get practical patterns for safe agentic coding, campaign instrumentation, and multilingual audio that does not cross the deepfake line.

GLM 5, Emu 3.5 Image, and Google’s AI Mode for shopping collide in one very stacked news cycle. Hunter and Riley break down how to judge new open weight models beyond hype screenshots, what “agent native” really looks like in a production workflow, and where the liability traps hide. They dig into Emu 3.5 Image for in-house creative factories, including typography pitfalls and smart templated workflows. Then they unpack Google Ads conversational AI Mode and Veo 3 inside Asset Studio, and what it means for feeds, creative testing, and brands that are operationally sharp. The thread across it all is human intent plus machine throughput.

Google Gemma 3 lands in everyone’s stack and suddenly “open weights” is the main character. Hunter and Riley unpack what open actually buys creators and marketers, how to use long context as a bigger whiteboard instead of a hallucination machine, and when small models are perfect for low risk, high volume content. They break down Gemma 3 on Together AI, the Llama 5 Avocado leak chatter around ultra cheap text, and why taste plus distribution becomes the real moat. Then they dig into Sarvam’s Saaras V3 ASR, Bulbul V3 TTS, and Vision OCR for real workflows in Indian languages, from call analytics to localized voice agents.