On Denoised, we ask what happens to the software business when AI can reverse engineer an app and rebuild it. We break down REA and a vibe-coded creative suite, debate where software companies keep their moat once code gets cheap, and watch Gemini 4 Argon and Opus 5.5 agents build detailed Blender scenes.
Jump to:
01:11 REA and the vibe-coded creative suite
07:08 Enterprise moats, support, and custom tools
17:38 AI aggregators and the licensing weights theory
25:30 Meta's Muse Wonder and its coding focus
28:29 Gemini 4 Argon and agents building in Blender
Quick Take
AI coding tools can now produce clones of complex creative apps whose interfaces, Addy says, match the originals almost pixel for pixel, though performance still falls short. Agents can also assemble 3D scenes Addy estimates would take hundreds of hours by hand, which raises the question running through the episode: which parts of software still hold value when anyone can generate the code.
What We Questioned: Whether a Vibe-Coded Creative Suite Holds Up in Real Editing
We start with Reverse Engineer Anything, or REA, an MCP tool you point at an app so it can work out how the software was built and help you vibe code your own version. Addy's understanding is that it reads the compiled binary and turns it back into human-readable code.
The example making the rounds is Crafting Apps, with vibe-coded apps like PhotoCraft, FilmCraft, and LightCraft.
The collection works as a free take on Adobe's Creative Suite and other products. Addy believes it came out of REA, which would explain interfaces that match the originals almost pixel for pixel, while Joey isn't sure REA was involved.
Where it breaks:
- Addy has heard performance is severely lacking, and YouTube tests of the suite's video editor show slow, choppy previews and transitions that don't work.
- The hard part of Premiere, Resolve, and other NLEs is real-time GPU computation for time remaps and transitions, which the clone doesn't handle yet.
Joey's argument: Heavy-duty creative apps are the last place he'd move his workflow to something vibe coded. Free options like Photopea, an ad-supported web alternative to Photoshop, have existed for years.
Addy's counter: Decompiling software felt like NSA-level technology five years ago, and now anyone has access to it. He also flags that this makes software theft easy.
We draw the line between asking an agent to copy Resolve or Final Cut, which is ripping it off, and asking for a video editor that does X, Y, and Z, which could fit your needs better.
What We Debated: Where Software Companies Keep Their Moat When Code Gets Cheap
Not everyone will want to maintain duct-taped tools of their own, even as agents improve.
Joey's argument: Software companies may run leaner and ship faster, with value shifting to support and services, the way he believes Linux makes money. If code stops being the moat, proprietary data or something else that can't be replicated has to fill that role.
Custom tools for creators: Our own show is the example.
- We don't need the full weight of Resolve, Premiere, or Final Cut to edit an episode.
- Joey wants a side-by-side two-shot of our two feeds treated as a third camera angle, and auto-cutting podcast editors don't understand that.
- He's close to building it himself.
On the studio side, 180 Studios has pushed into software development to tackle creative asset chaos.
Why many people will keep paying: We both have coffee machines at home, and sometimes we still pay for Starbucks because it's easier.
- Addy argues low prices protect a lot of software, putting Resolve's lifetime license at $300, and Joey notes basic Resolve is free.
- Addy expects pricier tools like AutoCAD and Maya to face more scrutiny.
- Joey's counter is that nobody will design a building worth hundreds of millions of dollars on vibe-coded software, so the question for Autodesk is shipping faster without losing reliability.
Addy suggests companies like Autodesk would probably build custom tools for a studio, for a price. Joey imagines generative interfaces where the backend stays the same and each user's layout is built around how they work.
Addy's strongest defense of traditional software is SideFX's Houdini, long the way to get fluid sims, smoke, particles, and collisions.
The reason Houdini's successful is because SideFX, who makes it, has scientists and researchers and PhDs working away, chipping away for years on this stuff. So are you gonna vibe code something like Houdini? You're not, because that company is specifically built to deliver on something that the normal person can't.
Joey's counter: He's seen headlines about AI models solving math problems that stumped mathematicians for ages, which leaves open how long research-heavy software stays out of reach.
What We Questioned: Whether AI Aggregators Can Defend Prompt Wrappers
Aggregator platforms without an original lab model need something defensible, and we compare two answers.
Joey's argument: In his telling, fal.ai started as an API aggregator, then post-trained its own improved version of MiniMax's open weights. That gives it something original.
Magnific One took a different route. According to VentureBeat's report, it runs on a GPT image model, and the company described its distinguishing technology as prompt-based, built on textual instructions shaped by its creative specialists.
That makes Magnific One an image tool wrapped around someone else's model, which Joey reads as aimed at brands. He sees risk if OpenAI adds design tools to its own app on top of the same model.
The prompt-wrapper problem:
- Hidden prompt engineering works until the models improve enough to not need it, and a company tuned to one model's quirks ships slower when a new one arrives.
- Addy suspects the big models already expand user prompts behind the scenes.
Addy's counter: He floats what we call his tinfoil-hat business plan. A platform that licensed closed model weights and ran them on its own infrastructure could charge around 10% less than an equivalent model while keeping more of the margin.
- Joey isn't sure closed-source weights can be bought, though companies like LTX and Black Forest Labs offer licenses.
- As Joey recalls, LTX is free to use until a company makes more than $10 million.
- In Addy's hypothetical, Amazon or Netflix offers Google $50 million to license Nano Banana 2, and Joey thinks Google, which already has its own compute, would be a tough sell.
What We Explored: Meta's Muse Wonder and Its Coding-First Training
Joey reads Meta's chart as positioning Muse Wonder as its flagship, Meta's answer to Fable and Astra, though Joey hasn't worked out how to access it yet.
Meta's head of AI, Alexander Wang, explains the lack of flashy demos in his post on X. The team trained Muse Wonder for coding rather than 3D demo games, and he asked people to "try Muse Wonder on your actual coding work."
- Joey expects Muse to end up powering Facebook Ad Manager, from copy to images and video.
- That points to network effect, a moat no reverse-engineering tool can copy, since nearly everyone on the planet is on Meta's platforms, as Addy puts it.
What We Explored: Gemini 4 Argon and Opus 5.5 Building Full Scenes in Blender
Joey notes that Gemini 4 Argon is now out, and that this model was trained on 3D.
The Gemini team builds a Japanese pagoda and Zen garden with each release as a visual benchmark. Addy walks through the Argon version:
- A single prompt feeds a blueprint agent that draws up an architectural plan.
- A Nano Banana agent generates reference images from many angles.
- An adversarial agent quality-checks the build while multiple agents construct it in Blender.
- Shaders come in for the final render, which appears to include the interior.
The caveat: The creator says it came from a single prompt, but the multi-agent setup may have been configured by hand. Addy still estimates the result represents hundreds if not thousands of hours of manual work.
More scenes we looked at:
- An Argon cathedral built and lit from a single image, with render passes and volumetric fog.
- A cluttered desk scene built by an agent connected to Blender, model unknown, plus another build using Astra and Opus.
- An Opus 5.5 laundromat build.
Why Blender keeps showing up: Joey's take is that Blender is highly scriptable, which makes it easy for agents to control. Addy calls it the proving ground for this kind of work.
Joey finds that reference images help, and with light direction, agents will generate their own references or spin up research agents to find photos, maps, and diagrams online.
What We Debated: Whether Epic Missed the First AI Wave
Addy's argument: Epic had AI talent in-house years before the generative boom, yet Blender became the platform where AI experiments happen, a role he thinks Unreal Engine could have filled.
Joey's counter: The platforms that won did so by accident, because they happened to be scripting friendly, like Blender and Resolve. Unreal blueprints should be scriptable, but we're less sure about agents building out worlds.
Addy notes Unreal now has an MCP server and argues AI-assisted modeling and shading would have helped developers build games faster for the Fortnite ecosystem.
Epic is moving now. In Tim Sweeney's post, he wrote that Epic has "a huge pivot underway" with developers moving to "retool for maximum AI acceleration."
Replying to a critic, Sweeney wrote that much of the room wants to be in 2016 again, and that Epic's job as a tools provider "is to design for the present and the future that will soon be." Addy says Sweeney's optimism matches how we feel.
Bottom Line: Cheap Code Moves the Moat to Reliability, Data, and Platforms
Every story in this episode comes back to what still holds value when an agent can write the code.
- REA and the vibe-coded creative suite. Cloning an interface is within reach, while real-time GPU performance remains the hard part.
- Software moats. Support, reliability, low pricing, and research teams like SideFX's keep many users paying.
- AI aggregators. fal.ai's post-trained model gives it something original, while prompt wrappers like Magnific One depend on models that keep improving past them.
- Muse Wonder. Meta trained for coding, and its strongest defense is a platform nobody can copy.
- Gemini 4 Argon and Opus 5.5 in Blender. Agents build detailed, lit scenes from a prompt or a single image when the host software is scriptable.
- Epic's pivot. Epic is retooling developers for AI acceleration after Blender became the proving ground for agent builds.
Our read is that the software that holds up will rest on something an agent can't regenerate on demand, from GPU engineering and research teams to proprietary data and user networks.
Links from This Episode
Tools & Platforms
Projects
News & Analysis
Companies
Adobe - maker of the apps Crafting Apps imitates
OpenAI - maker of the GPT image model behind Magnific One
fal.ai - aggregator that post-trained its own MiniMax version
WATCH THE FULL EPISODE ON YOUTUBE →
