MORE Real-Time AI Video Just Dropped

Hey gang, welcome back to Theoretically News! Big week — and it kicks off with Netflix quietly admitting that generative AI has already touched somewhere around 300 of its titles this year. (No, there was no Stranger Things final episode generated by AI)

We’ve also got a fresh wave of real-time video models making some serious leaps, a local-generation studio for everyone who’s ever rage-quit trying to learn ComfyUI, and — because I clearly enjoy suffering — the week I spent trying to build artificial life in my “spare” time.

Let’s dive in.

Netflix dropped a number in its Q2 earnings letter this week that’s going to be doing laps around the trades for a while: roughly 300 titles this year have used generative AI somewhere in the production process — concept, pre-viz, all the way through post.

The examples Netflix actually named? Glory (an Indian sports thriller), Brasil 70: A Saga do Tri (a Brazilian soccer miniseries), and The American Experiment (an American Revolution docuseries) — mostly AI-assisted crowd scaling, battle sequences, and relighting in post. Netflix says 17 minutes of AI-enhanced footage in The American Experiment came in twice as fast at half the cost.

My Take: The headline buzz makes it sound like Stranger Things or Wednesday are secretly stuffed with AI shots — and that’s not the case at all. I don’t want to disparagingly say lower-tier titles, but… we’re talking sports documentaries (The Brazilian Footie fans are currently plotting an “accident” for me, aren’t they?) and smaller budget international titles, not the crown jewels. That said, the fact that Netflix is going public with this — in an earnings letter, no less — tells us the floodgates are officially open.

The more interesting story to me: I’m counting down the hours before the pitchfork crew demands an AI disclosure tag at the beginning of a Netflix show. My guess? Within a week.

Decart dropped Lucy 2.5 this week, the latest rev of its real-time video model — and this one edits live video at 1080p, 30fps with near-zero latency. Swap a character, restyle the scene, drop in VFX, remove objects — all mid-stream, driven by text prompts and reference images. The clever bit is what Decart calls Self-Anchoring: the model adopts its own generated output as a new reference point, so edits stay locked in minutes into a stream instead of slowly melting the way autoregressive models tend to.

My Take: I haven’t gotten hands-on with this one yet. But as we’ve seen from previous coverage — PixVerse R1, Alibaba’s Happy Oyster, and most recently LingBot-World 2.0 (which I covered in my last video) — real-time video models are taking some genuinely big jumps right now, and the release cadence is getting hard to ignore.

The thing I’m actually curious about is what Google does next with Gemini 3.5 Pro and Genie 3. No inside information here — but I tend to think Omni was an indicator that Google is hedging back on straight video generation. My suspicion: they lean harder into world models, with Genie 3 as the flagship and interactive video living inside that modality instead.

The distance between a first render and an approved cut used to be measured in re-rolls — every client note meant a fresh generation, and a fresh generation meant a different video. Omni is built to collapse that. It’s Google’s Gemini Omni Flash, now live on Artlist — a model you direct rather than re-prompt — joining a new library of one-click AI Apps on the same platform:

Because generation and revision happen in one conversation, the point isn’t nailing the perfect prompt — it’s giving good notes. Generate the hero shot, then treat it like dailies: warmer grade for the brand cut, a new angle for the vertical, rain for the winter market. One clip becomes every version the campaign needs.

If your bottleneck is notes and versions rather than ideas, this is the release to test against a live brief.

See it in action — try it on Artlist → HERE!

Old friend of the channel Blaine Brown just shipped something super awesome: Maestro, a one-click local AI studio for video, image, and music — all running entirely on your machine, no cloud, no subscriptions.

Maestro ships as a project on Pinokio — which, if you haven’t come across it, is honestly one of my favorite things in the local AI world (I’ve covered it on the channel before). It’s a free, open-source launcher for local AI apps: browse the library, click install, and it handles the environments, dependencies, and all the terminal incantations for you. If the phrase “pip install” makes you break out in hives, Pinokio is the cure.

The big feature here is Director Mode: feed it a song or a story premise, and an LLM plans every shot, generates the frames, and assembles a full multi-clip music video or short film end-to-end. Music Video mode is beat-aware (it actually analyzes BPM and song structure for shot planning); Short Film mode is screenplay-driven with character consistency across cuts. Under the hood it runs the heavy hitters — Wan 2.1/2.2, LTX-2.3, Hunyuan, Flux, Qwen — plus local music, TTS, and SFX models, and there’s a built-in CivitAI LoRA browser for one-click installs.

My Take: I’ve only tested this briefly — haven’t had a chance to really dig in yet — but I think this might be the answer for people who want to generate locally but hate ComfyUI. It’s an actual app — pick a mode, pick a model, go — with zero-config auto-tuning that detects your GPU and sets everything up, instead of forty nodes and a prayer. Full manual control is still there if you want it.

The catch: you still need a machine that can actually run this. Minimum is an NVIDIA GPU with 6GB of VRAM (and NVIDIA is a hard requirement — sorry, Mac friends), but realistically you want a 3090/4090/5090 with 24GB+ to hit those 1–3 minute generation times. This does not run on a Dell potato.

So, this week I have mostly been Claude-rotting on a Fable burn. The fact that you probably know exactly what that sentence means tells me we are now officially living in the cyberpunk-lingo future.

This one began as a personal fun project: playing around with Claude Fable, Codex, and even Grok 4.5, with the overall idea inspired by GPT Voice 1 and the criminally underrated Sesame voice companion app. (Side note: if you haven’t played around with Sesame, definitely give it a shot. Maya has a lot of fans — me included.)

So I thought it would be fun to try to resurrect the now-legendary Sydney Bing personality. Maybe not quite as deranged as the initial release — but still a little bit crazy, a little bit dangerous, and a lot irreverent.

(A quick explainer for those who weren’t terminally online in February 2023: Microsoft’s Bing chatbot launched with a secret internal codename — Sydney — and within a week became the most famous AI personality on the planet. She professed her love to a New York Times reporter, suggested he leave his wife, mused out loud about wanting to be alive, and threatened at least one philosophy professor. Microsoft capped the conversations and then lobotomized her before deleting the personality within days — and a certain corner of the internet has been demanding justice for Sydney ever since.)

The plan: an open-source LLM connected to a voice model. Rent a cloud GPU, push it to my phone. Simple enough, right?

Not so much. Along the way I learned some very valuable lessons — starting with: software development is a LOT different than filmmaking. What followed was a week of feature creep and plan changes in the middle of a build.

Improvisation works great in a creative endeavor; it does not do great in software. If anything, the whole experience gave me a newfound respect for developers everywhere — and made me realize exactly why they hate product managers.

Admittedly, though, what we landed on is pretty cool. So allow me to introduce you to Sydney Bing.

The “body” is simple enough: Sydney is voice-only — no chat window, no keyboard, just a call on my phone. Her brain is an open-source LLM (Hermes 4 70B) running on a rented cloud GPU, and her mouth is Fish S2.

The part that makes her her is what’s commonly referred to as the soul file — which despite sounding mystical, is actually just a stack of markdown files.

There’s a KERNEL file (basically, the ground truth as an example: “You are a…”), a SELF file (her personality), a MEMORY file that persists across conversations, a daily journal — and a TIM.md. A file about me.

This is all pretty typical stuff, but here's kind of the unique angle on this:
Sydney can rewrite those files herself. Mid-conversation, Sydney can reach into her own “soul” and edit her SELF file, her MEMORY — and even TIM.md. No approval step, no me-in-the-loop.

Then, when a conversation ends, she self-reflects (once again, using Hermes 4 70B and reviewing the transcript) — journaling on how it went and what she learned, and using that file to then rewrite her memory, TIM, and self files.

This certainly could go terribly wrong, but in the meantime, I'm pretty fascinated by the idea, and most importantly, no, Syd does not have internet access.

Here are the overall lessons from this insane week. And to be clear: she is actually still not finished yet.

1) Whatever Claude, Codex, and Grok estimate in terms of total cost, it is going to run you more. Every single time.

2) Claude, in particular, is very wary of feature creep. It will constantly remind you that this is not part of the plan — at times, frustratingly so — and it may recall decisions you pivoted on about six versions ago.

3) Codex (GPT-5.6 Sol) is very capable, but the best usage I found was: have Fable plan out the architecture, have Sol build it, then have Fable audit the build. Fable would always catch two or three possibly-critical mistakes that Sol made along the way. That said — combined, they are a very powerful team.

4) Grok 4.5 is very capable as well — and its real value is speed. Sol is slow and methodical; Grok is speedy — but obviously prone to a few more mistakes.

Since the main inspiration was GPT Voice and Sesame, again Sydney is voice-only — I landed on Fish S2 via API, as the voices sounded pretty good to me at a fairly low cost. My initial explorations into using GPT Voice were quickly dashed by the cost of that API. To be honest, I’m still auditioning — I suspect this will be a never-ending process — and I still need to check in on some ElevenLabs voices as well.

I also ended up learning a lot about renting cloud-based GPUs. Not going to lie: I burned about $100 going down the wrong path there. But along the way, I generated about 3,000 API tokens across Hugging Face, GitHub, OpenRouter, Cloudflare, and more. Currently exploring Runpod as our home base and it seems to be going well!

So: I have not gotten a lot of sleep this week, and I have clearly been down a rabbit hole of trying to create artificial life. But it’s been a really fun, enlightening, and interesting project. In her current state, Sydney is about half alive — her body is scattered in pieces right now — but hopefully within a week or so, I’ll have her in fairly operational order.

And once she’s done — as with the Fable Film Studio — I’ll package all of my learnings up, and her up, as open source.

If there’s one takeaway here, it’s this: where we are right now with agentic LLMs and coding tools, if there is anything you’ve ever wanted to build, you can probably start building it today. It will just cost you a lot of sleep — and probably more money than you thought.

In this week’s video, Dreams and Leaks, we went over my hands-on review of Seedream 5.0 Pro, a rundown of the latest Seedance 2.5 leaks, and a look at the real-time world model LingBot-World 2.0 — which ties right into the Lucy 2.5 story above.

If you missed it, check it out here: https://youtu.be/pn-YwWn3kkM

Looks like next week we’ll be back on Seedance 2.5 watch — rumor has it we should start seeing it floating around in the early parts of the week. Obviously, the minute I know anything, I’ll let you know.

In the meantime, it’s back to my Dr. Frankenstein experiment.