N&M🎏

Finding Ways to Escape the Infinite Scroll | Solutions Architect @NVIDIA | Physical AI |

[N&M🎏 @minimadhub]

I dream about a “hub” where like-minded thinkers, mad about something, could hang-out, support each other, and seek inspiration. I would call it Mad-Hub 🗯️

[Anthropic @AnthropicAI]

Claude Fable 5 will be available again globally tomorrow.

After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8. We’ll continue to refine these classifiers over the coming weeks to reduce false positives and better distinguish genuine misuse from legitimate requests.

We’ve also begun drafting a consensus framework—with Amazon, Microsoft, Google, and other Glasswing partners—for assessing the severity of AI jailbreaks and how AI developers should respond to them. We invite other industry partners and model providers to join us in this effort.

Finally, we’re scaling up our collaboration with the US government on model testing and safeguards. This will include pre-release access to models and safeguards for evaluation, information sharing on jailbreaks and misuse, and dedicated resources for joint research.

Thank you to our users for your patience, and to our partners across the government, industry, and the research community who worked alongside us to make Fable 5 available again.

Read our full blog: https://t.co/VHyum831ri

[N&M🎏 @minimadhub]

❓ Hot question for the community:
World-Action Models are eating VLA mindshare fast. But WAMs require more memory and compute at inference time.
Is "future imagination" worth the overhead for real-time robotics - or is this a warehouse/lab-only capability for the next 2 years?

[N&M🎏 @minimadhub]

I no longer "use" Blender and Isaac Sim - I talk to them.

Hooked both up to @claudeai via MCP: author 3D assets in Blender → OpenUSD → drop into Isaac Sim for physics + robots + synthetic data. All by chat.

This is what a physical-AI pipeline looks like in 2026. 🤖

[N&M🎏 @minimadhub]

JetPack 7.2 blog: https://t.co/ZphA2sDJ63
Efficient-WAM paper: https://t.co/0gtWHOOff6

[N&M🎏 @minimadhub]

World-Action Models are taking over robotics. Five WAM papers dropped this week alone. This isn't a coincidence - it's a paradigm shift happening in real time.

[N&M🎏 @minimadhub]

This maps directly to what NVIDIA just shipped. JetPack 7.2 launched June 1st with a core theme: agentic-ready AI at the edge with memory efficiency. The hardware is catching up to where WAM research needs it to go.

[N&M🎏 @minimadhub]

Here's what I want to know from this community: are World-Action Models the right abstraction for physical AI, or are we adding model complexity that will fail in messy real-world conditions? Drop your take below. 🤖

[N&M🎏 @minimadhub]

VLA safety research just exploded. This week: a paper showing your VLA's own attention weights can serve as a safety filter - no extra training needed. Another uses probe vectors for training-free failure recovery. The community is done waiting for safe robots. It's engineering them now.

[N&M🎏 @minimadhub]

World-Action Models are having a moment.
Efficient-WAM, MotionWAM, AHA-WAM - three new WAM papers dropped in a single week. Each one trying to give robots the ability to imagine before they act.
This is the architecture race of 2026. The teams that crack efficient world models for real-time manipulation will define the next generation of robot brains.

[N&M🎏 @minimadhub]

Links to Sources for the Thread:
JetPack 7.2 release: https://t.co/ZphA2sDJ63
WEAVER paper: https://t.co/kCfHmj59O2
RobotSmith: https://t.co/q9C28uTBvn

[N&M🎏 @minimadhub]

The question isn't whether agentic robots at the edge work. The demos are out.
The question is: what's your real bottleneck - sim2real gap, task generalization, or cost-per-deployment?
Tell us below. That's the conversation worth having this week. 🤖

[N&M🎏 @minimadhub]

The stack is becoming: plan with world models → act with generalist policies → adapt with tool design → deploy on Jetson.
Each layer has a research group pushing it. Each layer has hardware that can run it.
That's the full physical AI stack -on hardware that fits in a robot.

[N&M🎏 @minimadhub]

And NVIDIA SRL's RobotSmith paper is circulating: VLMs that automatically design task-specific tools for robots.
Not learning to use tools. Designing new ones.
That's a fundamentally different level of agency than anything that shipped two years ago.

[N&M🎏 @minimadhub]

Also this week: Groups publish "Flow Reversal Steering" for improving generalist robot policies.
They steer diffusion flows at inference time - no retraining, better task performance. Generalist policies that adapt without fine-tuning are the distribution layer for agentic robots

[N&M🎏 @minimadhub]

This lands the same week, drops WEAVER - "An Effective World Model for Robotic Manipulation."
World models let robots simulate before committing to action. That's literally what an agent does.
The research and the hardware are converging on the same architecture right now.

[N&M🎏 @minimadhub]

The hardware was never the bottleneck.
Jetson Thor hit the market with real compute headroom. The gap was always the software stack — memory-efficient runtimes, multi-modal context, persistent state across tasks.
JetPack 7.2 closes several of those gaps at once.

[N&M🎏 @minimadhub]

For years, "edge AI" meant: run a model on a small device. Fast inference. Low power. Done.
That's table stakes now.
Agents need memory. Long-horizon reasoning. Tool use. Loop closure. JetPack 7.2 was built specifically for this new requirement.