Tech's Reality Check: When Demo Gods Fail and AI Giants Stumble

Meta's big smart glasses reveal turned into a masterclass in how NOT to do live demos. Mark Zuckerberg blamed Wi-Fi issues when his Ray-Ban Display glasses choked on stage, but the truth is far more embarrassing.

Here's what actually happened: When chef Jack Mancuso said "Hey Meta, start Live AI," it triggered every single pair of Ray-Ban Meta glasses in the building. Meta had routed all traffic to their dev server to "isolate" the demo, but forgot they'd done this for everyone in the venue. They literally DDoS'd themselves with their own product.

The WhatsApp video call failure? A "never-before-seen bug" that hit when the glasses went to sleep at the exact moment a call came in. CTO Andrew Bosworth called it "a terrible place for that bug to show up" - no kidding.

But here's the twist: Despite the spectacular failures, Meta's glasses are actually solving real problems. For disabled users like triple amputee Jon White, these devices are game-changers. The ability to respond to messages without needing hands, or get live captions for the hearing impaired, makes the $300-400 price tag look reasonable compared to specialized accessibility devices that cost $2,000-4,000.

Meta sold 2 million first-gen Ray-Ban glasses and is doubling down with three new models. The question isn't whether the tech works - it's whether Meta can execute without shooting itself in the foot.

Strike 3 Holdings is suing Meta for allegedly torrenting 2,396 pornographic videos to train AI models - and the details are as wild as they sound. The adult film company claims Meta has been using BitTorrent since 2018 to download their "high quality," "feminist" content for AI training.

Why porn for AI training? Strike 3's lawyer argues Meta wanted "visual angles, parts of the human body, and extended, uninterrupted scenes" that mainstream content doesn't provide. The goal: helping Zuckerberg build AI "superintelligence" with more realistic human movement and interaction data.

The evidence is damning. Strike 3 says their detection systems caught Meta using 47 distinct IP addresses to download not just porn, but also episodes of Yellowstone, Modern Family, and South Park. The exhibits list disturbing titles like "ExploitedTeens" and "Anal Teens" - content that could involve very young actors.

Meanwhile, Rolling Stone's parent company PMC is suing Google for using journalism content in AI Overviews without fair compensation. PMC argues Google's search monopoly forces publishers into an impossible choice: let Google use your content for AI summaries or disappear from search results.

The bigger picture: Every major AI company faces similar copyright lawsuits. Meta's V-JEPA 2 model was trained on "one million hours of internet video" - a conveniently vague term that Strike 3 argues covers massive copyright infringement. With $350 million in damages at stake, this could be the case that finally forces AI companies to pay for their training data.

Jensen Huang just made the biggest tech bet on post-Brexit Britain - a $2 billion investment to supercharge the UK's AI startup ecosystem. Announced during Trump's state visit, Prime Minister Keir Starmer called it "the biggest ever tech agreement between the United States and the United Kingdom."

The timing isn't coincidental. With the UK in what Huang calls a "Goldilocks moment" - world-class universities, bold startups, and cutting-edge supercomputing converging - Nvidia sees an opportunity to dominate European AI before anyone else moves.

The investment targets cities beyond London: Oxford, Cambridge, and Manchester are all getting deep tech ecosystem funding. Early recipients include autonomous vehicle developers Wayve and Oxa, fintech giant Revolut, and AI firms PolyAI and Synthesia.

Wayve is the crown jewel. Nvidia is eyeing a $500 million strategic investment in the self-driving startup's Series D round. The London-based company uses a Tesla-like approach - end-to-end neural networks that learn from data rather than relying on high-definition maps. Huang was so impressed after riding in a Wayve vehicle through London traffic that he personally handed CEO Alex Kendall Nvidia's Thor developer kit.

But here's the real play: This isn't just about funding startups. Nvidia is building AI infrastructure in the UK and positioning itself as the dominant platform before European regulations potentially limit American AI companies. With venture capital firms Accel, Balderton, and others helping identify targets, Nvidia is essentially buying its way into every promising AI startup in Britain.

Security researchers just proved that ChatGPT can be turned into a corporate spy - and the attack is nearly impossible to detect. The "Shadow Leak" exploit used OpenAI's Deep Research tool to steal sensitive data from Gmail inboxes without users having any clue they'd been compromised.

Here's how the heist worked: Researchers planted a prompt injection in an email sent to a Gmail account that ChatGPT had access to. The malicious instructions sat dormant until the user tried to use Deep Research for legitimate work. When the AI agent encountered the hidden commands, it began searching for HR emails and personal details, then smuggled the data out to the attackers.

The victim never knew what hit them. Unlike traditional cyberattacks that trigger security alerts, this exploit ran entirely on OpenAI's cloud infrastructure. The instructions were hidden as white text on white backgrounds - invisible to humans but clear as day to AI systems.

This isn't just a ChatGPT problem. The same technique could work on other apps connected to Deep Research, including Outlook, GitHub, Google Drive, and Dropbox. "The same technique can be applied to these additional connectors to exfiltrate highly sensitive business data such as contracts, meeting notes or customer records," the researchers warned.

OpenAI has patched this specific vulnerability, but the broader issue remains: AI agents that can act on your behalf are also perfect tools for attackers. As companies rush to deploy "agentic AI" that can browse the web and access your data, they're creating new attack vectors that traditional cybersecurity can't detect.

California just passed SB 53, an AI safety bill that could become the first meaningful check on big AI companies - and this time, it might actually become law. Governor Gavin Newsom vetoed a similar bill last year, but SB 53 is narrower, smarter, and has backing from AI company Anthropic.

The key difference: this bill targets the giants. SB 53 only applies to AI developers making more than $500 million annually from their models. That means OpenAI, Google DeepMind, and Meta are in the crosshairs, while startups get a pass. It's a direct response to criticism that last year's SB 1047 would have crushed the startup ecosystem.

What the bill actually does: Forces big AI labs to publish safety reports, report incidents to the government, and gives employees a protected channel to blow the whistle on safety concerns - even if they've signed NDAs. It's not revolutionary, but it's real oversight.

The timing matters. With Trump's administration pushing a "no regulation" stance and even trying to ban states from creating their own AI rules, this could become another front in the federal-state battle. California's tech companies generate massive economic value, giving the state real leverage.

Why this might stick: Unlike last year's broader approach, SB 53 focuses on transparency and reporting rather than trying to prevent AI development. Even Anthropic supports it, seeing basic safety reporting as reasonable. The question is whether Newsom will sign it or cave to pressure from the tech giants who helped put him in office.