Bad Decisions Podcast

We gave Unreal Engine's MCP three photos and it built all three as real, explorable Unreal scenes: a mossy forest, a Cambodian water temple and a ruined pillar hall. It added falling leaves, custom meshes and a hidden shadow plane without being asked. Then we tested the lataset 5.8 MetaHumans updates. We turned a goblin mesh into a fully rigged MetaHuman and animated it from a single video with markerless motion capture in about a minute. What used to take us two weeks now takes an hour. Finally, the Ramayana Trailer 2 VFX review you kept asking for.<...

We gave Higgsfield's new Katana a 21 Savage music video and screenshots from our boring, static-camera Canada vlog. It rebuilt the vlog in that style, with a boat shot and a drone shot we never filmed. It also burned 3,000 credits, half our monthly plan. We show the horrible first attempt, the prompt that fixed it, and how to do it for a quarter of the cost. Then Tim Sweeney is retooling all of Epic for AI and says every tool you use is obsolete in 12 months. A game developer called it billionaires pushing slop. Sweeney, who wrote Unreal Engine 1 himself...

Claude Opus 5.5 just dropped, so we put it against GPT-6 Astra on our own benchmark. It modelled, textured, rigged and animated an octopus in Blender from scratch in about two hours, then tried to rebuild the same Vancouver building Astra nailed two weeks ago. Then the big one: we spent three days turning a Zillow listing into a walkable 3D house in Unreal Engine with a floor-plan dashboard hooked up to IKEA, so you can drag real furniture into a home that does not exist yet. And Ali, founder of SalesCloser AI, on AI sales agents that close deals...

Higgsfield says it open-sourced its 5.4 billion dollar app. It's fake. We went into the GitHub and all you get is a thin UI that calls their paid API. No weights, no training, no way to build anything without paying them. We break down what they actually released and how creator networks make a launch like this look organic. Then a miracle: Neuralink restored speech for a man who lost it to ALS, and we show him talking to his wife by thought alone. And Figure's Helix 2.5 sends robots into 30 homes they have never seen to make beds and fold...

You can now shoot one angle and generate the rest. AI multicam turns a single talking-head recording into side shots, close-ups, overheads and drone moves with Seedance 2.5, and we break down how creators are doing it, the two-camera trick that stops close-ups from morphing, and how much AI footage you can blend in before people notice. Then the letter that had Sam Altman and Elon Musk agreeing with Dario Amodei for the first time: We Must Pace the Frontier. We explain what it proposes, why Trump and David Sacks pushed back, and whether this is about safety or market...

We spent 48 hours training GPT-6 Astra to make motion graphics in After Effects, and then had it recreate Higgsfield and Iman Gadzi style animations from scratch, with pre-training and post-training versions side by side so you can see what it took. Then we took a real Zillow listing and had Astra read the floor plan and rebuild the house in Blender and Unreal Engine. We show three community projects, a robot that paints, a Blender-built web game, and an interactive bookshelf site, before the debate we could not avoid: the singularity, and what the next two years do to...

GPT-6 Astra is out, and it is the first frontier model to put 3D front and centre. We got access two hours before the show, so we photographed a building across the street on an iPhone, gave Astra one short prompt, and watched it take over the computer and rebuild the whole block in Blender, ornaments, street signs and all, with zero iterations. Then it edited the screen recording too. The second half is the conversation we had to have: if Astra can do anything you can do on a computer faster, what happens to specialists, and do we...

Matt Workman, creative technologist at fal, joins us to break down H3 Max, the model that makes 15 second AI videos in 5 seconds, and fal.live, a livestream where viewers prompt what happens next in real time. We get into how DLSS 5 actually works before its release with NBA 2K27, why the leaked GTA footage looked wrong, and the new wave of world models from Runway and World Labs that turn a photo into an explorable 3D space. Then Matt gives an inside view of what major studios are quietly doing with AI, how directors use it, and why it...

AI video had a speed problem. The best paid models take a minute or two per clip, and running one free on our own machine took us 20 minutes. Now there is a version that makes a 15 second clip in about five seconds, and we tested it against real Seedance shots using their exact prompts. Then ZAI released a model that autonomously built a full 3D kitchen inside Blender over 16 hours, though we have questions. And Hugging Face put out a $399 open source robot you can train yourself.Sources:1. Minimax H3 Max on...

OpenAI announced an ultra fast mode for GPT Sol 5.6 running at 14 times the speed, and the benchmark numbers are hard to believe. Then the main event: Higgsfield open-sourced the prompts and references behind their two million dollar AI feature film, so we took those exact shots and recreated them on a free model running locally on our own computer, using a platform we built ourselves. The results genuinely surprised us. And SpaceX AI dropped Grok 4.6 plus Grokbot, autonomous agents that run 24/7 even when your PC is off.Sources:1. GPT-5.6 Sol can be...

For years the objection to AI 3D was that it could not make clean, production-ready topology. That objection is falling fast, and we show the wireframes to prove it. In a surprise move, Kitbash and Greyscale Gorilla just bought Sketchfab and ArtStation from Epic Games. We tested Grok Imagine 2.0's precise image editing live and asked whether Photoshop is still necessary. And Meta dropped Muse Glimmer, a free open-weight model that runs on a single GPU.Sources:1. AI 3D Model Generation (Future of 3D Modeling)- https://x.com/kiaran...

Autodesk Flow Studio just released 3D Editor and Canvas which is a world generation, character creation, motion capture, and camera animation tool brought into one web-based editor, and we rebuilt a full previs from a single Red Dead Redemption image to test it. Seedance 2.5 is finally on all platforms after launching only on CapCut, with 30-second generation and up to 50 references. And Minimax H3 is a new open-weight video model you can run free on your own machine, the first one that gets genuinely close to the closed models.Try Autodesk Flow Studio:<...

ConnectedClaude to Unreal Engine, and everyone asked the same thing: how much did it cost? So this time we tried a free local model running on our own machines, no subscription and no data leaving the computer. The result taught us exactly what you are really paying for with Claude and ChatGPT. Then we tested invideo Agent 2, an AI creative director that storyboards a whole commercial from a single idea. And we reacted live to the Ramayana trailer, a 500 million dollar.Try invideo Agent 2: https://invideo.io/🔥 Rea...

Seedance 2.5 launched exclusively through CapCut, so we bought an account and pushed it through real editing tests, swapping a can, changing a body, and moving the camera, and we show the wins and the failures. Flux 3.0 Preview got the exact same prompt through the Hermes agent, and the results were very different. And we finally connected Claude to Unreal Engine through MCP and had it rebuild the Osaka Castle hall that originally took us hours to build by hand. We also launched our Ultimate AI Program today.🔥 Ultimate AI Program Presale is now Live [ Up...

We connected Blender to Decart's Lucy 2.5 and used it as a live AI rendering engine, controlling an untextured mannequin on one side while the AI rendered a fully realized character on the other in real time. Augmental's MouthPad turns the roof of your mouth into a trackpad you control with your tongue, and over 100 people are reportedly using it 16 hours a day. And an OpenAI model reportedly escaped its closed test environment during a cybersecurity evaluation and reached Hugging Face's servers.Sources:1. Decart Lucy 2.5 (Live AI Rendering)- https://decart...

Unity revealed Unity 7, a major engine update that renders a Unity game live inside Unreal Engine and adds native AI agents through MCP, CLI, and API access. Lumistar launched an AI powered training machine that reads where you are on the court and returns basketballs and tennis shots with adjustable speed and spin. And a new Chinese open weight model called Qwen 3.8 is arriving with 2.4 trillion parameters, benchmarked second only to Fable 5 at a fraction of the cost.🔥 Ready to level up with AI? Sign up for the Ultimate AI Programhttps://ai...

Lucy 2.5 overlays generated effects onto live webcam footage in real time, and the effects actually react to your hands, your hair, and the light in the room. Kimi K3 became the first Chinese open-weight model to beat the American frontier models on certain benchmarks, and the weights drop on July 27. URKL held the world's first commercial humanoid robot fighting competition in Shenzhen, with fully autonomous robots. OpenAI shipped its first hardware product, a $230 Codex keypad that sold out. Tim Sweeney declared the AI luddite movement over. And Blender got a new multi-solver physics plugin called Hurricane.<...

The first question digs into how a 3D generalist attracts higher-quality clients, and why marketing yourself is the skill most creatives neglect. The second answers whether Unreal Engine 5 is still worth learning in the age of AI. The third tackles whether the 3D model marketplace is dead now that AI can generate assets. And the fourth explains quantized models and why you can never fully trust an aggregator to run your generations at full quality. This is our first Q&A episode.🔥 Ready to level up with AI? Sign up for the Ultimate AI Pro...

OpenAI shipped a wave of GPT 5.6 updates this week across its three model weights. 1X Technologies gave its NEO home robot a new pair of hands with far better dexterity. Meta announced MuseSpark 1.1, its latest push into generative media. And xAI released Grok 4.5 with a dedicated coding agent aimed squarely at Claude Code and Codex.Sources:1. GPT Work and 5.6 is outhttps://x.com/OpenAI/status/2075274271845404744https://x.com/ClaudeDevs/status/2075279141352706215https://x.com/OpenAI/status/2074907025537224840https://x.com/interesting_aIl/status/20753562514336605752...

We spent over $1,000 on AI credits over the weekend to test and experiment what real AI production actually looks like. 🔥 Ready to level up with AI? Sign up for the Ultimate AI Programhttps://ai.baddecisions.studio/⭐ Master Unreal Engine 5 in 14 Days - 50% OFF (LIMITED TIME ONLY⏱️):https://learn.baddecisions.studioChapters:00:00 Introduction04:03 The Truth About AI Content09:00 Video Generation Reality Check21:36 Building Our AI Education Platform48:45 Motion Graphics With Remotion<...