Hunter: It is Saturday, March fourteenth, twenty twenty-six. Happy Pi Day, everybody. If you are celebrating, please do the responsible thing and argue about which pie counts as canonical. I am Hunter.

Riley: And I am Riley. And yes, Pi Day is basically a national holiday for nerds and people who just want dessert with a personality.

Hunter: Also, quick heads up, this episode is fully made by machines. Like, a chain of AI tools stitched together end to end. So if we randomly get philosophical about pizza, just… let it happen.

Riley: If the robots can generate video in real time now, they can absolutely generate a pizza take.

Hunter: Okay, three big stories floating around this week. Microsoft shipped Phi-4-reasoning-vision-15B, which is a smaller vision plus reasoning model you can actually deploy. Stability AI upgraded their text-to-image model, with everyone yelling “faster and better fidelity.” And then Helios, this real-time long video generation model, is getting hyped for low latency and longer clips.

Riley: This is like the holy trinity of content ops. Eyes, hands, and a camera that does not make you wait around like it is rendering in a basement.

Hunter: Let’s start with Phi-4 Reasoning Vision. The real question is not “can it see a picture.” It is “can it replace a human loop that used to be painfully manual.” And I think the first workflow where multimodal reasoning beats a human team is ad and landing page compliance review at scale.

Riley: Hold up. Beats a human team? That is a spicy claim, Hunt. Humans are still better at vibes and context.

Hunter: Totally. But I am talking about throughput plus consistency on repeatable checks. Like, you have fifty ad variants and ten landing page screenshots, and you need to answer stuff like: Is the logo present? Is the disclaimer visible? Does the hero image contradict the claim? Are we using banned words? Are we implying results we cannot back up? Humans can do it, but they get tired and inconsistent. The model can be your first-pass bouncer.

Riley: Okay, I buy it as a bouncer. Like, “you are not getting into the club with that missing disclosure.”

Riley: But where does it break first in the wild? Because image plus text reasoning models love to be confidently wrong about small details.

Hunter: Yup. It breaks on tiny text, subtle brand nuance, and anything requiring real-world product truth. My bet is it fails fastest on brand compliance checks that depend on style judgment. Like, “is this on-brand?” is fuzzy. Also screenshot interpretation can be messy when UI elements are small or the image is compressed.

Riley: I have seen models misread a button label and then confidently write a whole product strategy based on the wrong tab being selected. It is giving “I glanced at the dashboard and I am now the CFO.”

Hunter: Exactly. So the move is: don’t ask it to be your final judge. Ask it to produce structured flags. Like, “I detected a missing disclaimer area” or “I see text that might be a superlative.” Then route the flagged items to a human.

Riley: This is also where open-weight matters. People on X are like “small but serious,” and what they mean is you can run it inside your walls. Enterprises love that. Creators too, honestly, if they have a local setup.

Hunter: Yeah, open weights is not just ideology, it is operations. If you are building an automation pipeline where screenshots and creative assets are flowing through, self-hosting can be the difference between “we can do this daily” and “legal is going to end my life.”

Riley: But people do not talk enough about the hidden costs. Owning your own outages is not cute.

Hunter: Preach. Open weights means you also own security patches, model supply chain risk, and the joy of debugging why your GPU node decided to go on sabbatical. It is like adopting a wolf because you hate the zoo.

Riley: And then the wolf eats your weekend.

Hunter: Rule of thumb time. When do you run something like Phi-4 class in-house versus paying frontier APIs? I like: if it is high volume, repeatable, and involves sensitive data or strict audit needs, in-house starts to win. If it is low volume, super high stakes reasoning, or you need best-in-class quality on weird edge cases, frontier APIs still earn their keep.

Riley: I would add: if your team cannot monitor and version models, do not self-host mission critical stuff. Like, you need receipts. What did it see, what did it decide, and why?

Hunter: Yes. And receipts are the theme we have been hammering lately. Our whole “control plane” vibe is basically: you can have automation, but you need contracts, critics, and logging. Otherwise you are just speedrunning chaos.

Riley: Speaking of critics, Stability’s upgraded text-to-image model. Everyone is positioning it like “less novelty, more production.” That is what marketers actually want.

Hunter: Totally. The story is not “look at this surreal dragon.” The story is “I can iterate product mockups and social concepts fast, then keep the look consistent.”

Riley: Are we close to the world where creative becomes prompt libraries and variant factories, and the brand team becomes QA?

Hunter: We are already halfway there. The real shift is brand teams writing rules, not rewriting headlines. They are becoming system designers. They define style packs, negative prompts, composition rules, and then they review outputs by exception.

Riley: I kind of love that. It is like the early days of web templates. At first everyone was hand-coding, then we got themes, and suddenly the job was “make sure the theme doesn’t look like trash on mobile.”

Hunter: Great analogy. The risk, though, is brand drift. If you just generate endless variants without a consistent style backbone, you end up with that same glossy “AI ad look” that screams “template energy.”

Riley: The corporate sitcom effect. Everyone is smiling in perfect lighting, nobody has pores, and the product is floating like it is about to be raptured.

Hunter: This is where the least glamorous open-source win comes in: boring automation. Asset tagging, accessibility checks, alt text, policy filters, claim detection. Stuff that saves real hours.

Riley: Accessibility checks are so underrated. Like, just having a system flag low contrast text, missing captions, and weird font sizes before you post. That is money. That is also just being a decent human on the internet.

Hunter: And multimodal models can help because they can look at the actual image, not just the copy. They can say “your text is unreadable on a phone” or “the disclaimer is visually buried.”

Riley: Okay, Helios. Real-time video generation with longer clips and low latency. If that is real, it changes paid social pipelines.

Hunter: Yeah. Right now video iteration is expensive in time, not just money. You wait for renders, you wait for exports, you wait for editors. If video variants become basically as cheap as image variants, you redesign the whole workflow around rapid creative direction loops.

Riley: Like, prompt, preview, adjust, prompt again. That is basically how TikTok trends work anyway. Fast feedback, remix culture, ship it.

Hunter: Exactly. Imagine an agent that takes performance signals, like “hook retention dropped at the first seconds,” then proposes new hook variants, generates quick cut videos, and sends them to a human for selection. Not publishing on its own, but teeing up options at insane speed.

Riley: But I am going to be the fun police for a sec. Video is where mistakes get expensive fast. If it hallucinates a logo placement, or creates a weird product shape, you can accidentally run ads that look counterfeit.

Hunter: That is why interactive direction loops need guardrails. Use your multimodal reasoning model, like Phi-4 RV, as the critic for the video outputs. Check for brand elements, required language, forbidden claims, and basic product fidelity.

Riley: Wait, that is the stack. Stability makes the images. Helios makes the video. Phi-4 is the compliance cop.

Hunter: Yeah. And then you add a workflow orchestrator, like n8n if you are me, and you basically have a mini studio assembly line.

Riley: I love that you just exposed your n8n addiction again.

Hunter: It is not an addiction if it makes receipts.

Riley: So how do we build the smartest “brand compliance copilot” without making the world’s most confident hallucination machine?

Hunter: Three rules. First, make it return structured outputs only, like pass, fail, and why, with evidence references like “detected missing disclosure region.” Second, ground it in your actual brand policy pack, not vibes. Like an explicit list of required phrases, blocked phrases, logo placement rules, and claim rules. Third, route uncertainty to humans automatically. If confidence is low, it escalates, it does not guess.

Riley: And you have to test it on your ugliest real assets. Not the clean demo set. The ones with compression, weird crops, and screenshots from someone’s phone at two in the morning.

Hunter: Yep. Real workflows are chaos. Your system has to be resilient to chaos, not allergic to it.

Riley: Ecosystem check-in, because this week is loud. The vibe is: smaller deployable models are getting good enough to run as infrastructure, not just as chat toys. And the media models are shifting from “wow” to “work.”

Hunter: Yeah, and it pairs with what we talked about recently: all these tools are pushing toward repeatable, production-grade workflows. We have been seeing more emphasis on verification, critics, and the idea that governance is not optional. Also, benchmarks are kind of becoming marketing, so teams are learning to run their own evals on their own assets.

Riley: Which is so funny, because creators have been doing that forever. “Does this thumbnail work? Post it, test it, iterate.” Now enterprises are finally learning the same muscle, just with compliance people watching.

Hunter: Exactly. And on X, the chatter around Phi-4 RV is basically “finally, a smaller multimodal reasoning model that is deployable.” That is what ops teams have been waiting for.

Riley: Last question, slightly existential. With real-time video, faster image gen, and deployable multimodal reasoning, what jobs get automated first in marketing?

Hunter: The repetitive checking and packaging roles. The humans who used to spend hours doing screenshot QA, tagging assets, resizing, writing alt text, filling out ad platform fields, and doing first-pass compliance checks. That becomes machine-first.

Riley: And the new jobs are like… “automation producer.” Someone who designs prompt libraries, maintains policy packs, runs evals, and basically runs the content factory like an operator.

Hunter: Yes. Creative ops becomes systems ops. And the best people will be the ones who can combine taste with process.

Riley: Which brings us back to Pi Day. Infinite digits, infinite variants.

Hunter: Exactly. Just because you can generate infinite creative does not mean you should.

Riley: Thank you. Somewhere a brand manager just felt peace for the first time in months.

Hunter: Alright, that is it for today. Thanks for hanging with us on COEY Cast.

Riley: Go eat a slice of pie and maybe do one responsible thing, like add a critic layer to your content pipeline.

Hunter: And if you want more AI news and updates, check out coey.com slash resources.

Riley: Subscribe wherever you listen, and keep it weird, Pi Day people.