AI Fakes Dead Pilots' Voices, Grok Flops, and Startup ARR Is a Lie
The NTSB had to lock down its own records after AI reconstructed crash audio from spectrograms — and that's not even the wildest story today.
Subscribe free →AI Reconstructed Dead Pilots' Voices From Spectrograms — And the NTSB Panicked
Researchers used AI to reconstruct cockpit voice recordings from spectrogram images posted in public NTSB accident dockets — essentially reverse-engineering audio from pictures of audio. The NTSB responded by temporarily shutting down public access to its entire docket system while it figures out what to do. The technology works by treating spectrograms as a reconstruction problem, and apparently it works well enough to recover intelligible speech.
Why it matters: This is a direct collision between open-government transparency norms and AI capabilities that nobody designed policy for — and it won't be the last time a public records system gets blindsided like this.
Grok Is Flopping and Everyone Knows It
A Reuters investigation found that Grok barely shows up in federal records of US government AI usage — a damning signal for a chatbot Elon Musk has positioned as the truth-telling alternative to "woke" AI. The Verge's analysis goes further, noting that Grok's user numbers are weak, its reputation is shaky, and yet Musk keeps forcing it onto X users and into the center of his empire-building plans.
Why it matters: When the government — which has massive AI procurement budgets and no particular reason to avoid Grok — won't touch it, that says something real about where xAI's flagship product actually stands in the market.
AI Startup "ARR" Is Whatever Founders Need It to Be
TechCrunch dug into how AI startups and their VC backers are systematically inflating Annual Recurring Revenue figures — counting committed spend, pilot contracts, and other non-recurring items as ARR to manufacture momentum narratives. Investors know it's happening and participate anyway, because the game is about crowning winners and raising the next round, not accurate accounting.
Why it matters: When the primary metric used to rank AI companies is being gamed by the people reporting it, valuations built on those numbers deserve a serious haircut.
Google's Gemini Can Now Go Anything-to-Anything — And It's Unsettling
Google's new Gemini Omni model handles any combination of inputs and outputs — text, image, audio, video — in a single model. The Verge tested it by continuing an earlier experiment deepfaking a child's stuffed animal into fake vacation videos, and found the results significantly more convincing and easier to generate than before. The barrier to synthetic media that's indistinguishable from real is now basically zero effort.
Why it matters: Multimodal models that make deepfakes trivially easy aren't a future concern anymore — they're a shipping product from the world's largest search company.
Elon Musk Quietly Abandoned the "Solar Future" He Spent Years Selling
xAI's Colossus data center runs on natural gas. SpaceX is building orbital data centers to avoid Earth-based energy constraints entirely. TechCrunch traced how Musk — who once evangelized a solar-electric economy so hard he bought SolarCity — has systematically pivoted every energy-intensive operation he runs toward fossil fuels and space-based workarounds.
Why it matters: The man with the loudest AI infrastructure buildout is now one of the largest drivers of new natural gas demand, which matters for anyone trying to assess the actual carbon cost of the AI boom.
Meta's Forum App Is Reddit If Reddit Had a Corporate Overlord With Your Data
Meta launched Forum, a standalone iPhone app that pulls Facebook Groups into a dedicated feed alongside a built-in AI chatbot — think Reddit-style topic browsing crossed with Meta AI hovering over every thread. The Verge's hands-on found it functional but unavoidably Meta-flavored, with the AI assistant surfacing summaries and answers the way Google's AI Overviews interrupt search results.
Why it matters: Meta is directly targeting the Reddit-plus-AI-search use case, and it has billions of users and existing group content that Reddit can't match — this is a real competitive threat.
Quick Hits
- →IBM and Ferrari are using AI to personalize the F1 fan experience — think real-time race insights and customized content for viewers who don't know a DRS zone from a pit wall. TechCrunch
- →Berlin startup Peec, which tracks brand visibility in AI search results (think: "does ChatGPT mention us?"), more than doubled annualized revenue to $10M — proof that AI SEO anxiety is a real business. TechCrunch
- →SpaceX's Starship V3 had a mostly successful debut flight but lost the Super Heavy booster on return — progress, but the hardware bill keeps growing. TechCrunch
- →Texas AG Ken Paxton sued Meta claiming WhatsApp doesn't actually encrypt messages end-to-end — critics say the lawsuit lacks factual basis and reads more like a campaign stunt for his US Senate run. Ars Technica
AI TLDR