Ori Inbar's three XR trends for 2026: AI smart glasses (100% driven by AI), AI-generated content for glasses, and world models colliding with XR — for humans, not just robots. He coined "SPAI" (spatial AI) before the AI wave arrived. The spatial Wikipedia he described then is now real — world simulators running on glasses hardware, helping you do things you couldn't do before. AWE 2026 theme: iSpatial. Counter to iRobot. Humans at the center. Watch the full episode for the rest of the story.

I've heard a lot of people talk about what comes after the current generation of XR hardware, but Ted's framing on this episode stuck with me: 125 years of content technology has been a variation on the same idea. More resolution, better compression, higher frame rates — but still fundamentally 2D video. The Holodeck is a different kind of problem, and it requires a different kind of solution. His argument is that world models are that solution. Rony taking that seriously enough to build a talk around it at AWE — alongside Google Genie and World Labs — tells me this isn't just theory. And if Gixel's non-waveguide display lives up to what Ted described when he put on the prototype, the hardware side of this equation is catching up faster than most people realize. Catch the full conversation in the latest episode. #XR #WorldModels #AIXRPodcast

Ori Inbar has run AWE for 17 years. He's been on this show five times before. He calls 2026 the most consequential year in the conference's history, and the numbers hold up. Over 10 million AI glasses sold last year, no display required. About 300 manufacturers are building XR hardware worldwide right now. Snap is expected to announce a major consumer product at AWE — Ori says they have put everything on this. Meta's retreat from the metaverse freed a generation of talented spatial engineers who are now building independently. Ori also built a working AR prototype — he calls it "chat with animals" — in two days using Gemini. Vibe coding for spatial computing is not theoretical. It is happening. Three trends Ori is watching in 2026: AI smart glasses, AI-generated content for glasses, and world models colliding with XR. He was calling this convergence "spatial AI" before anyone else was using the phrase. The line that stuck with us: AI is shifting glasses from tools that help you learn about things to tools that actually do things. That is the inflection point. Everything at AWE this year is building toward it. Full episode on the AI XR Podcast.

At Snap, Jonathan Rodriguez Cefalu structured the Spectacles R&D organization after one model: SpaceX. No separate research division. Research ships in the product — or it does not happen. His reasoning is blunt: a corporate research division takes the smartest people in the building and tells them their work is not going to ship. It is going to become a paper. Outside of basic science — physics, biology, fields where the discovery-to-application loop is genuinely long — Jonathan believes there should be no such thing as pure research in a technology company. Ted Schilowitz brought his own version of this conviction from years building with Red cameras. The counterexample they both point to: Meta's research organization, where figures like Douglas Lanman do work that almost never makes it into a product a user can hold. Shipping is the only proof of concept that matters. Save this post and hit the link in bio for the full episode. #XR #AI #ProductLeadership #AIXRPodcast #Innovation

Ted surfaced something in this conversation that I think a lot of people building on top of large language models have not fully confronted. Jonathan Rodriguez Cefalu walked through why the instruction hierarchy that AI systems are supposed to follow — constitution, system prompt, user, untrusted data — is not actually enforced at the architecture level. The transformer sees one token stream. Untrusted data in an email or a web page can issue instructions that override trusted levels because there is no structural barrier. Gemini Spark operates on top of this same foundation. Preamble's proposed fix is a reserved token language that untrusted content cannot write — something the injection cannot reach. Jonathan put it well: "I hope there are better architectures in the future." So do I. Ted pulled out the clearest explanation of prompt injection risk I've heard on the show. Catch the full conversation in the latest episode. #AI #LLM #AISafety

NDAA FY2025 Section 1638 — passed with help from Preamble AI — is the first US law making it illegal to give AI autonomous control of nuclear weapons. Before this, it was executive policy across two administrations. Not enforceable law. Jonathan Rodriguez Cefalu puts it plainly: "As of fiscal year 2025, it is finally now illegal to build Skynet." Three specific protections: weapon system functionality, communications validation, positive human action by the president required before launch. Watch the full episode for the rest of the story. #AI #AISafety #AIXRPodcast #AIGovernance

The first people to publicly demonstrate prompt injection built a company around the problem. Jonathan Rodriguez Cefalu — who led Snap Spectacles hardware through Gen 1 and Gen 4 — is now at Preamble AI, and the threat he is focused on is the one hiding inside every AI assistant you already use. Here is the core issue: LLMs do not distinguish between a trusted system prompt and an untrusted email in your inbox. It is all one token stream. Tell Gemini Spark to manage your Gmail and you have just handed a bad actor inside your email a direct line to the AI running your workflow. Preamble's answer involves a reserved token language for trusted instructions — something untrusted data cannot write in. They are not just publishing papers about it. They helped pass a provision in the FY 2025 NDAA making it US policy that AI cannot be given autonomous control of nuclear weapons. Ted Schilowitz and Jonathan also got into Snap Gen 6, Google I/O, and why the WarGames reference still holds:

Ted Schilowitz: if AI companies were building airplanes, they'd face legal accountability. They're not, so they don't. The regulatory side doesn't have the technical vocabulary to push back. Rony Abovitz says the serious concern is underground — generals, CTOs — but it's not reaching the surface. Charlie Fink, Ted, and Rony ask: where is the credible, technical voice arguing for the public interest? Watch the full episode for the rest of the story. #XR #AI #AIXRPodcast #Policy #Governance #Ethics