Samsung's betting $73B on AI chips while Microsoft ships MAI-Image-2 and Cursor launches a cheaper code model to challenge OpenAI. Meanwhile, generative AI is cracking wireless vision through walls, a technique that could reshape sensing and robotics.

The cheapest model in the room keeps winning. That pattern is starting to look less like a coincidence.

Researchers taught AI to see through walls using wireless signals, and the implications go well past convenience.

Specialization is beating scale. The generalists had their moment.

Sonar has worked this way for decades: send a signal out, measure what bounces back, reconstruct the shape of whatever you can't see. It's crude but effective, like reading a room by clapping your hands. MIT researchers have been running a more sophisticated version of this idea for over ten years, using surface-penetrating wireless signals to let robots locate and manipulate objects hidden behind obstacles. The core physics isn't new. What just changed is what happens after the signal returns.

The bottleneck was always reconstruction precision. The raw signal data is messy. Wireless reflections off a concealed object give you a blurry, ambiguous point cloud, something like trying to identify a face from its shadow. Good enough to say "object is here." Not good enough to say "object is this shape, grab it this way."

That's where the generative AI comes in.

What the Model Actually Does

The MIT team is using generative AI models to take that imprecise reflection data and reconstruct object shapes with meaningfully higher accuracy. The article summary is thin on architecture specifics (my read: this is likely a diffusion-based or similar generative approach that learns a prior over plausible 3D shapes and uses the signal data as a conditioning input, but I'm extrapolating from the broader literature here). The key insight is that generative models are good at exactly this problem: filling in ambiguous, incomplete information by drawing on learned priors about what shapes tend to look like.

Think of it like autocomplete for geometry. The wireless signal gives you a smeared, noisy sketch. The generative model says "given everything I know about object shapes, the most probable thing that produces this sketch is this." The result is a sharper reconstruction than the signal data alone could ever produce.

Not magic. Learned inference.

Ten Years of Foundation, One Bottleneck Cleared

The decade-long research history here is worth sitting with. This wasn't a team that bolted a trendy AI model onto an existing demo for a press release. They built the underlying sensing methodology over years, hit a precision ceiling the physics couldn't clear, and then found that generative AI could clear it for them.

That's a different story than most "AI improves X" announcements. Most of those are applications looking for a problem. This is a genuine technical bottleneck that waited for the right tool.

And the tool arrived about two years late to be fashionable, which is probably why the work is more credible.

Who Actually Wins When Robots See Through Walls

The obvious applications are warehouse robotics and search-and-rescue. A robot that can locate and manipulate hidden objects without line-of-sight is genuinely useful in environments where boxes are stacked, debris is present, or visibility is zero. But the more interesting implication (to me, anyway) is what this does to sensor fusion architectures.

Right now, robots in complex environments rely heavily on cameras and lidar, both of which fail hard when the target is occluded. Adding a reliable through-obstacle sensing modality that produces accurate 3D shape reconstructions changes the sensor stack. Not as a replacement. As a fallback that's actually trustworthy enough to act on.

The gap between "we can detect something is there" and "we can reconstruct what it is precisely enough to grip it" is enormous in robotics terms. If the MIT results hold up at scale, that gap just got a lot smaller.

Where I'd Push Back

My honest skepticism: generative models that fill in missing data are doing inference, not measurement. They're producing the most probable shape, which is not always the correct shape. In a warehouse picking context, "usually right" might be fine. In search-and-rescue, where the hidden object might be a person in an unusual position, "usually right" could matter a lot.

The precision gains are real. But I'd want to see hard numbers on failure modes before trusting this for high-stakes manipulation tasks. The article doesn't provide them, and that absence is doing some work.

Still: a decade of sensing research plus a genuinely well-suited AI technique is a better foundation than 90% of what gets called "AI-powered" anything. This is one to watch.

- The AI finds the signal. We decide what it means.

Pabbly Connect automates workflows between apps without coding. It's cheaper than Zapier for high-volume users, starting at $19/month for 500 tasks.

Worth trying if you're drowning in manual data entry between tools. Connect your CRM to email, Slack to spreadsheets, whatever. The interface is clunky compared to Make, but the pricing scales better if you're running dozens of automations.

Real limitation: support is slow. You'll spend time troubleshooting alone.

The honest take: Pabbly isn't prettier or smarter than competitors. It's just cheaper at scale, and sometimes that's the only reason you need.

Some links are affiliate link. We earn a commission if you subscribe. We only feature tools we'd use ourselves.

If you're building real-time multiplayer features and dreading the networking boilerplate, try Google AI Studio's new vibe coding mode.

Paste your game loop into the prompt, ask it to "add WebSocket synchronization for player positions," and iterate on the generated code in real-time.

Specific example: feed it a canvas-based 2D game, request "lag compensation for 100ms latency," and test the output instantly without context switching.

The AI handles the tedious sync logic while you focus on game feel. Time saved: 3-4 hours of manual networking implementation per feature.

You're setting up a new project and deciding which AI coding assistant gets billed to the company card. Both cost around $10-19/month. Both autocomplete your code. The choice feels arbitrary. It isn't.

Where they actually differ:

Verdict:

If you're building anything with more than three files that talk to each other, use Cursor. The codebase-aware reasoning isn't a gimmick. It's the difference between an assistant that knows your project and one that knows Python.

If you're a senior dev who finds AI chat annoying and just wants fast, accurate completions while staying in your existing VS Code setup, Copilot is the right call. Don't fix what isn't broken.

Kitten TTS: Three new models under 25MB – Open-source text-to-speech models (15M to 80M parameters) run on CPU without GPU, enabling on-device voice synthesis.

Canary (YC W26): AI QA that reads your code – Generates and executes end-to-end tests for pull requests by understanding your codebase and user workflows.

Meta deploys AI for content enforcement, cuts third-party vendors – Early tests detect twice as much violating content with 60% fewer errors than human review teams.

DoorDash pays couriers to film tasks for AI training – New Tasks app compensates delivery workers for video submissions to improve AI and robotic systems.

OpenAI acquires Astral to strengthen Codex coding assistant – Developer tools startup joins OpenAI's coding team as Codex hits 2M weekly active users with threefold growth.

What caught our attention this week.

This newsletter runs on an 8-agent AI pipeline we built in-house.

Want that kind of automation for your business?

From scanning 50+ sources to drafting, fact-checking, and formatting - AI agents handle 95% of this newsletter.

The AI finds the signal. We decide what it means.

Research and drafting assisted by AI. All content reviewed, edited, and approved by a human editor before publication.