Welcome, humans.

The internet began as a public network anyone could build on. AI is developing more like a row of private theme parks, where a few companies own the rides, write the rules, and keep raising ticket prices.

Current AI wants to build the public alternative: open AI infrastructure that communities can use, modify, and control for free. The nonprofit has $400M in committed funding and is already backing offline tools for Indigenous communities, datasets covering 50+ African languages, and a pocket-sized device that works in 22 Indian languages without internet access.

The goal is basically a World Wide Web for AI, where progress benefits everyone instead of whichever company owns your login. Finally, an AI moonshot that does not begin with “please enter a valid credit card.”

On a tangent: We also published a full breakdown of everything Samsara announced at Beyond 2026, including AI agents for physical operations, 360-degree cameras, smarter fleet maintenance, and disposable shipment-tracking labels. The common thread: move AI out of the browser and into the trucks, warehouses, worksites, and supply chains that keep the physical economy moving.

Here’s what happened in AI today:

😼 Alibaba previewed a 2.4T Qwen model headed for open release.
📰 ACT-2 made home robots look closer to real products.
📰 Musk said xAI’s next model could beat Kimi K3.
📰 Morningstar called Kimi K3 a potential DeepSeek moment.
🎓 Today’s AI Skill shows how to direct layered AI video shots.

…and a whole lot more that you can read below

Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us!

P.S: We just launched a robotics newsletter! Sign up for it here.

😼 Alibaba’s 2.4T Qwen3.8 Is Joining the Frontier AI Race

The AI model race has developed a new flex: build something enormous, promise to open it, and release the preview before everyone finishes arguing about the last model. Alibaba is now testing whether sheer scale can turn that flex into a real advantage.

Alibaba’s Qwen team previewed Qwen3.8-Max, a model with 2.4 trillion parameters, and said the full version will eventually be released as open-weight software. Open-weight means developers can download and run the model themselves, rather than accessing it only through a company-controlled app or API.

Here’s what happened:

Qwen3.8-Max-Preview is already available through Alibaba’s Token Plan, Qoder, and QoderWork.
Alibaba called it one of today’s strongest models and claimed it trails only Claude Fable 5.
The company has not yet published independent benchmarks proving those performance claims.

How to try it:

Use Alibaba’s Token Plan for international preview access.
Test it on work you already understand well, then compare accuracy, speed, and cost against your current model.
Wait for the open-weight release before deciding whether its scale creates a practical advantage.

Why this matters: Parameter count is roughly the number of adjustable values a model learned during training. More parameters can increase capability, but size alone does not guarantee better answers, lower costs, or faster performance. The bigger signal is Alibaba’s promise to release the weights. That would give developers another frontier-scale Chinese model they can inspect, customize, and deploy outside a closed platform. For companies, that could mean more control over sensitive data and fewer dependencies on one provider.

Our take: China’s open-model strategy is turning every release into two competitions at once. Labs are racing to build the strongest model, while also racing to make that capability cheap and accessible enough for developers to adopt. Qwen3.8 still needs independent testing, especially because Alibaba’s boldest comparisons currently come from Alibaba. A 2.4T parameter count also raises an awkward practical question: can ordinary companies afford to run it efficiently?

If the benchmarks hold, the open-versus-closed debate gets much less theoretical. Companies may soon choose between paying premium prices for a tightly managed US model or operating a comparable Chinese model on their own infrastructure. The next few weeks will show whether Qwen3.8 created a breakthrough, or simply the largest résumé line in AI.

FROM OUR PARTNERS

Scale Computing™ Creates Modernization Opportunities in 2026

AI is accelerating the need for modern infrastructure—but where should technology leaders focus first? In this executive discussion from SC//Platform™ 2026, Scale Computing™ CEO Bill Morrow explores how growing AI demand, evolving infrastructure requirements, and security considerations are creating new opportunities for IT modernization. Watch to learn where organizations should be investing as they prepare for the next wave of enterprise innovation.

Watch now

🎓 AI Skill of the Day: Direct AI Video Like a Layered Shot

Most AI video gets weird because we ask one model to do everything at once: scene, character, motion, voice, timing, camera, vibes, the whole circus.

The better workflow from Prompt Mastery’s tutorial is to build the shot in layers. First, film a real person doing the motion. That becomes your driving video: the reference clip that tells the AI how the character should move. Then use a character-swap workflow like Flux 2 Klein to create the AI character frame, run SCAIL-2 for motion transfer (copying the real person’s movement onto the AI character), and blend the result back into the original footage with a soft mask.

The tiny-but-critical detail: keep the crop, swapped image, and generated output at the exact same resolution. If the sizes don’t match, your character won’t line up when you put the scene back together.

For dialogue, generate the voice first with Omni Voice, then use Relay Prompt in WAN2GP to direct the character second-by-second.

Turn this scene into a timed Relay Prompt for an AI video character.

Scene: [describe the room, character, camera angle, and mood]
Audio line: [paste the dialogue]
Goal: Make the character feel physically present, not like a generic talking head.

Write a timestamped action plan:
0-2s: facial expression and posture
2-4s: hand movement or eye contact
4-6s: body movement
6-8s: final gesture or reaction

Keep the movements subtle, realistic, and synced to the emotional beats of the line.

Want more tips like this? Check out our AI Skill of the Day Digest for July.

Have a specific skill you want to learn? Request it here.

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

*Why We Love It: AI agents need context to deliver. Slack provides it at scale. Read the report.
Tinker lets developers fine-tune open AI models through a managed API instead of running the training infrastructure themselves —pricing not public.
Creed turns your ideas into a structured personal manifesto you can refine and share —free to try.
BaseRT gives developers a faster runtime for serving AI models with less infrastructure work —pricing not public.
ZooData helps teams collect, organize, and improve the datasets used to train computer-vision models —pricing by demo.
DevSwat reviews code changes and flags bugs before they reach production —pricing not public.
Chikit turns product ideas into interactive app prototypes you can test and share —pricing not public.

Video Watchlist: Kimi K3 and ACT-2

Two threads are worth watching this week: China’s open-model surge and the home-robot reliability race.

Kimi K3 Is Here. China Just Hit the AI Frontier. breaks down why Moonshot’s new model is being compared with GPT-5.6 Sol and Claude Fable 5.
Meet Kimi K3 is Moonshot’s own intro video for the model.
Kimi K3 is the best model ever made (sometimes) is the builder-flavored take on where Kimi shines and where it still needs caveats.
ACT-2: Preview is Sunday Robotics’ official look at Memo folding laundry in unseen homes.
The Most Advanced Robot Ever Built Just Folded Clothes Perfectly adds the practical details: battery life, beta timing, cost estimates, and what the robot does when clothes fall or arrive inside-out.

📰 Around the Horn

Sunday Robotics CEO Tony Zhao unveiling ACT-2 in what appears to be the most aggressively empty office space in robotics.

Sunday Robotics’ ACT-2 folded 778 garments across 785 attempts in unfamiliar homes, making Memo look closer to a reliable product than a staged demo.
Elon Musk said xAI’s upcoming 2T-parameter model should outperform Grok 4.5 and could surpass Kimi K3 after initial training finishes.
Morningstar argued Kimi K3 could become another DeepSeek moment and highlighted several cloud-computing companies positioned to benefit.
Intel turned to ASML's next-generation lithography tool to help manufacture some Panther Lake laptop chips.
Apple reportedly looked for AI chip acquisitions to strengthen its in-house server-chip push.
Databricks signed a term sheet for a strategic funding round at a $188B valuation.