Netflix's 300 AI titles

In partnership with

Welcome to VP Land! It's SIGGRAPH this week! Tons of official (and unofficial) events - if you'll be around just reply back.

One thing to put on your radar - a free summit Tuesday night on some groundbreaking AI Workflows (aptly called the AI Workflows Summit). It's a night focused on how AI is actually getting used in professional production, from virtual production and volumetric media to the infrastructure and tools behind the workflows, including ComfyUI. A few in-depth presentations, followed by a panel that I'll be moderating. Tuesday from 6PM-9:30PM, completely free, even if you don't have a SIGGRAPH badge. Register here.

In today's edition:

Netflix's roughly 300 AI-assisted titles
Moonshot's 2.8T open-weights Kimi K3
Thinking Machines opens up Inkling
Foundry's SmartRoto and Griptape ship
Open world models you can actually steer

One more thing: I have two tickets to The Odyssey in 70mm IMAX, 11:10AM Sunday at the Regal Spectrum in Irvine, that I can't use (SIGGRAPH, whoops). If you can use them yourself (personal use only, I'll hunt you down if you resell them), reply to this email. If more than one person can use them, whoever has sent the most newsletter referrals is the tie breaker.

Netflix Used Generative AI in Roughly 300 Titles in 2026, Concentrated in Post

Netflix told shareholders in its Q2 earnings report that roughly 300 of its titles used generative AI in 2026, the clearest number it has attached to production-side AI, with most of it landing in post-production.

Where the AI actually shows up. Netflix said the tools span concept, pre-visualization, post-production, and release, but the volume sits in post, from cleanup to complex shot work rather than fully generated scenes.
What it built on screen. The company pointed to the Indian sports thriller "Glory," the Brazilian soccer miniseries "Brasil 70: A Saga do Tri," and the Revolution-era docuseries "The American Experiment," where AI helped assemble highly complex sequences with enhanced crowd sizes and battle scenes.
Sarandos leads with quality, not cost. Co-CEO Ted Sarandos framed the output as "10% better" creatively, and Netflix said the tools let creators expand a project's scope "twice as fast and at half the cost of previous options."
The figure sits on a growing in-house stack. We covered Netflix's first public model, VOID, built for post-production fixes; the 300-title number shows how far that post-first approach has spread across the slate.

SPONSOR MESSAGE

Your prompts are leaving out 80% of what you're thinking.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

Moonshot's Kimi K3 Lands as the Largest Open-Weights Model Yet, at 2.8 Trillion Parameters

Moonshot AI released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model it calls the largest open-weights model available, with a 1-million-token context window and native visual understanding.

Frontier-class, and downloadable. Kimi K3 runs an always-on reasoning mode and posts near-frontier benchmark scores; Moonshot says it will release the full weights on July 27, keeping a model this size self-hostable for teams that want to fine-tune it.
Built to scale around compute limits. The model uses Moonshot's Kimi Delta Attention and Attention Residuals, and is roughly 75% larger than DeepSeek's V4 Pro, extending China's run of heavyweight open releases.
Video in, analysis out. K3 accepts video alongside text and images, so a clip can be described, logged, or checked against a brief without first being reduced to a transcript.

Thinking Machines Releases Inkling, Its First Open-Weights Model Built to Be Fine-Tuned

Thinking Machines Lab, the San Francisco company founded by former OpenAI CTO Mira Murati, released Inkling, its first in-house model and an open-weight foundation for teams that need more control over how a model is run.

A model facilities can adapt. Inkling's weights are available under Apache 2.0, so a studio can fine-tune it on its own material through the company's Tinker platform and host the result inside its network.
Multimodal training, text output. The 975-billion-parameter mixture-of-experts model was trained on text, images, audio, and video, with text-only outputs at launch.
A US-based open option. Inkling gives creators and facilities another open model to evaluate alongside releases from Chinese labs, while Thinking Machines positions it as a broad generalist rather than a benchmark leader.

Open weights and in-house fine-tuning make it possible to adapt a capable model to a facility's footage while keeping that material on its own servers.

From Roto to Real-Time Worlds

This tool slate is less about watching polished demos and more about putting controllable AI into a working pipeline: local roto, studio-safe orchestration, steerable worlds, and models built for revision.

Production pipeline

Griptape Enterprise. Foundry's studio platform runs agentic AI pipelines on-premises or in a private cloud, with licensing, permissions, Nuke integration, and modular diffusion workflows for shops that cannot send footage to a public API.

SmartRoto for Nuke. Foundry's AI add-on moves and deforms roto splines across a shot from a few input frames while keeping the artist in control. It runs locally on licensed data, and Foundry says it can cut rotoscoping time by up to four times.

Mirelo AI and Kyutai audio-to-MIDI. This open model transcribes a full mix into a separate MIDI track per instrument, where many tools handle one sound at a time. It gives music editorial and scoring teams editable parts instead of a locked stereo mix.

Directable worlds and interactive media

Lingbot-World-2. Robbyant's open-source world model, hosted on Reactor, runs in real time and lets users steer a character and camera by keyboard. Its persistent environments and available API point toward previz and virtual-location exploration.

Aval. Alex Barashkov's open interactive-video format brings state machines, frame-accurate transitions, and alpha transparency to web playback. It gives teams authored triggers and branching behavior without a game engine.

AlayaWorld. Fine-tuned from LTX-2.3, AlayaWorld generates playable 720p worlds at 24fps with real-time 6-DoF camera control and prompt changes mid-scene. Spatial memory helps revisited locations remain consistent past the one-minute mark.

Image and research models

Reve 2.1. The native-4K image-model update improves prompt understanding, world knowledge, and foreign-text rendering. That is useful when a generated frame needs legible non-English type rather than placeholder text.

NVIDIA ARDY. This autoregressive diffusion model produces real-time 3D human movement that can be redirected with live text prompts and constrained with waypoints or full-body keyframes, targeting animation, simulation, and robotics.

How Disney Built a New Smugglers Run Mission Featuring The Mandalorian & Grogu

How Disney Built a New Smugglers Run Mission goes behind the Galaxy's Edge update that folds The Mandalorian and Grogu into the ride. The Disney Parks video walks through the ride-design and media work behind the new mission.

Stories, projects, and links that caught our attention from around the web:

🎬 A studio-adoption survey finds AI landing in previz, scheduling, and sports repackaging even as 96% of CEOs report no clear ROI.

🔊 A music-industry coalition proposed voluntary streaming labels tagging songs as AI-Generated or AI-Assisted, modeled on the explicit-lyrics marker.

👓 OpenAI and Work Louder launched kbd-1.0-codex-micro, a limited-run macropad with mappable keys and a joystick for the Codex workflow.

▶️ LTX spun out of Lightricks as an independent open world-models company building foundation models for filmmaking, gaming, and robotics.

We Tested Seedream 5.0 Pro While Meta Backpedaled on Instagram AI Remixes

On the latest Denoised, Addy Ghani and Joey work through a dense run of model news for production teams. They break down Seedream 5.0 Pro's more precise image editing, Meta's Muse model and its opt-out Instagram remix feature, and what new MCP integrations for Unreal Engine and ComfyUI mean for wiring AI into a real pipeline. The throughline is control: where these tools hand you tighter, repeatable command over output, and where an opt-out default quietly shifts the burden back onto creators.

Read the show notes or watch the full episode.

Watch/Listen & Subscribe

Spotify | Apple Podcast | YouTube