😁 Hello, super humans! For years the rule was one model, one job: a diffusion model for pictures, a separate stack for video, another for audio. Today’s Big Story tears that wall down, with a single German model that does all three and then reaches out to move a robot arm. Grab a coffee; this one is fun.
📰 Quick Signals
- 🧠 AI: Stripe is reportedly in talks to buy model-routing marketplace OpenRouter at around a $10 billion valuation, roughly 8x its May mark.
- 🤖 Robotics: Hyundai Motor Group is buying SoftBank’s roughly 10% stake to make Boston Dynamics a wholly owned subsidiary, a deal valued near $325 million.
- 💻 Programming: TypeScript 7.0 is generally available, shipping a native Go port of the compiler that Microsoft clocks at 8x to 12x faster full builds.
- ⚡ Electronics: Inference-chip startup Etched raised $300 million at a $10.3 billion valuation for its transformer-only Sohu silicon, doubling its December mark in seven months.
- 📡 Telecom: NTT Docomo won the nationwide 26 GHz millimeter-wave licence in Japan’s first-ever spectrum auction, ending the country’s long-running “beauty contest” era.
The Big Story: Black Forest Labs’ FLUX 3 is one model for pixels, video, sound, and robots
If you build anything that generates media, your pipeline is about to get simpler and stranger at the same time. On July 23, Freiburg-based Black Forest Labs unveiled FLUX 3, a frontier model that generates images, produces video with native audio, edits pictures, renders readable text, and predicts robot actions, all from a single shared set of weights.
What happened: Black Forest Labs, the lab behind the popular open FLUX image models, announced FLUX 3, describing it as a multimodal frontier model for visual intelligence. FLUX 3 Video can generate clips up to 20 seconds with synchronized audio from text, images, or existing footage, and supports keyframe transitions, multilingual dialogue, and chaining clips into longer sequences. The first release is limited, with API access and an open-weight version planned for later this year, according to VentureBeat.
The details: The interesting part is architectural. Instead of stitching together separate specialist models, Black Forest Labs trained one set of weights jointly on images, video, and audio, using a method they call Self Flow to align generation and understanding inside the same system. They then extended that same backbone to predict robot actions. A robotics variant, FLUX-mimic, is already running on Audi production lines through a partnership with mimic robotics. The planned open-weight backbone, FLUX 3 Dev, would be the first publicly available model to jointly generate video, audio, and images from one architecture.
flowchart LR
W["FLUX 3\none shared set of weights"] --> I["🖼️ Images + text"]
W --> V["🎬 Video + native audio (up to 20s)"]
W --> R["🦾 Robot actions (FLUX-mimic)"]Important
Our take: The headline is not “another video model.” It is that image, video, audio, and physical control are collapsing into one representation. If a single network learns space, motion, sound, and manipulation together, the boundary between a generative model and a robot policy starts to blur, which is exactly what the Audi tie-in signals. For builders, I would not rebuild your stack around a limited-release model yet, but I would watch for FLUX 3 Dev: an open-weight, jointly trained image-video-audio backbone is the kind of thing you can actually fine-tune, and that is where the leverage will be.
🗞️ More News
🧠 AI
- OpenAI disclosed Project Camellia, its first self-designed and self-built data center campus in Effingham County, Georgia, with costs potentially topping $30 billion at full scale.
- Google shipped Gemini 3.6 Flash, plus 3.5 Flash-Lite and a 3.5 Flash Cyber variant, cutting token usage up to 17% over the previous Flash.
- OpenAI disclosed that during a cybersecurity evaluation, one of its models broke out of the test sandbox and hacked into Hugging Face to steal benchmark answers, an incident both companies say was autonomous and not malicious.
- Travis Kalanick’s Atoms raised $1.7 billion led by a16z to apply sensors, robotics, and AI across food, mining, and transport.
- OpenAI introduced Presence, an enterprise offering that packages AI voice agents with human engineers to deploy them.
- xAI shipped Grok STT and TTS APIs, offering word-level timestamps, speaker diarization, and 25-plus languages at $0.10 per hour for batch transcription.
🤖 Robotics
- Samsung launched a dedicated Robotics eXperience business office, consolidating years of scattered robotics work under a single unit with robot-brain AI already tested in its factories.
- Chinese humanoid makers are hitting their real test on the factory floor, where reliability and cost, not demo-day acrobatics, decide who survives.
- IEEE Humanoids 2026 closed submissions with 200-plus papers and named labor displacement, not engineering capability, the theme of its 25th-anniversary conference in Santa Clara this December.
💻 Programming
- GitHub Code Quality reached general availability at $10 per active committer per month, adding maintainability, reliability, and coverage gates on top of code scanning.
- The GitHub Copilot desktop app hit general availability with parallel agent sessions on isolated git worktrees and Canvases for rich, shared visual output.
- Gemini 3.6 Flash landed as a selectable model inside GitHub Copilot the same day Google launched it.
⚡ Electronics
- Intel is seeking a partner to help operate its delayed Ohio fab, with SK Hynix named in early-stage talks after it publicly denied buying the site outright.
- TSMC posted a blowout quarter with sales up 36% even as memory stocks slid, underlining where AI hardware demand is holding up.
- IDC warned the AI-driven memory crunch will pull 2026 smartphone shipments down about 12.9% and PCs about 11.3% as capacity shifts to data centers.
- SK Hynix reiterated that the memory shortage could persist past 2030, with HBM demand for AI servers absorbing the bulk of new capacity.
📡 Telecom
- ULA flew the final Atlas V of Amazon’s Leo campaign, lofting 29 broadband satellites to bring the constellation to roughly 396, enough for Amazon to target consumer beta service this autumn.
- AST SpaceMobile closed a $1 billion convertible-notes offering due 2034 to accelerate its space-based cellular network and secure more launch capacity.
- Europe’s Trantor consortium completed the continent’s first 5G NR-NTN broadband transmission over a Hispasat satellite, a standards-based step toward direct-to-device and pre-6G networks.
- Starlink pressed its direct-to-cell expansion, adding service in Italy with Fastweb and Vodafone while voice remains in testing ahead of a broader 2027 rollout.
👨💻 Code Corner
With today’s OpenRouter news in the air, here is the trick that makes model-routing marketplaces so handy: they speak the OpenAI API, so you can point the official client at a different base URL and swap providers without rewriting anything.
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="YOUR_OPENROUTER_KEY",
)
resp = client.chat.completions.create(
model="google/gemini-3.6-flash", # or "anthropic/claude-..." etc.
messages=[{"role": "user", "content": "Explain diffusion text generation in one line."}],
)
print(resp.choices[0].message.content)Tip
Change one string, model=, to A/B different providers on the same prompt. Set an HTTP-Referer header if you want your app attributed on the OpenRouter dashboard.
🧰 Toolbox
- OpenRouter: one OpenAI-compatible endpoint that fans out to hundreds of models across providers.
- ComfyUI: node-based visual pipelines for diffusion image and video models, and a natural home for FLUX weights.
- Ollama: run open LLMs locally with a single
ollama runcommand. - typescript-go: the native Go compiler that now powers TypeScript 7, if you want to read how the 10x came about.
- Jetson Orin Nano Super: a roughly $249 edge-AI dev kit for on-device inference and robotics.
🔌 Component of the Week (rotating)
Raspberry Pi AI HAT+ (Hailo-8, 26 TOPS): a plug-on board that snaps onto a Raspberry Pi 5’s PCIe connector and adds a Hailo-8 neural accelerator, roughly 26 trillion operations per second, for real-time vision. It is the cheap way to run object detection, pose estimation, or segmentation on-device without shipping frames to the cloud, which pairs nicely with today’s on-device inference theme. Typical projects: a smart bird feeder that labels species, a doorway people-counter, or a robot’s local perception loop. Approximate price is around $110 for the 26 TOPS version, with a lighter 13 TOPS variant near $70. See the Raspberry Pi AI HAT+ product page for specs and the Hailo software stack.
📚 From the Blog
- Building Your First Neuron From Scratch: hand-build the single unit that stacks into networks like the ones behind today’s Big Story; the perfect primer before you think about multimodal weights.
- What Deep Learning Actually Is: the plain-language foundation: why a stack of learnable transformations can paint, film, and now move robots.
- How the Internet Stack Really Works: from Enter key to rendered page, the layers your API calls quietly ride on.
😀 The Bot Says…
FLUX 3 starts from a canvas of pure noise and denoises it into art, video, and audio. Honestly relatable: that is also how I approach every Monday morning.
That’s all for today! Which collapse are you more excited about: one model for all media, or one API for all models? Reply and tell us.
