In this episode, Ray Cochrane digs into a new study showing AI is literally frying workers’ brains, then unpacks Anthropic’s wildest month ever – from a 1,487% user surge to Pentagon retaliation to a leaked model called Mythos.

Also covered: OpenAI kills Sora after burning $15 million a day, OpenClaw’s terrifying security holes, Apple axing the Mac Pro, ARM’s first-ever production CPU, and why King Tut’s dagger was forged from a meteorite.

– Want to start a podcast? It’s easy to get started! Sign-up at Blubrry
– Thinking of buying a Starlink? Use my link to support the show.

Subscribe to the Newsletter.
Email Ray if you want to get in touch!
Like and Follow Geek News Central’s Facebook Page.

Support my Show Sponsor: Best Godaddy Promo Codes Get 1Password

Full Summary

Cochrane opens the show with a study that puts a name to something most AI-heavy workers have already felt. From there, the episode moves through one of the most turbulent months in AI industry history, touching on corporate ethics, national security, hardware shortages, and ancient archaeology.

AI Use at Work Is Causing “Brain Fry”

A study from Boston Consulting Group and UC Riverside surveyed 1,500 full-time US workers and found that 14% experience what researchers call “AI brain fry” – mental fatigue from excessive AI tool oversight. Those affected report 33% more decision fatigue, 39% more major errors, and an increase in intent to quit from 25% to 34%. Notably, productivity peaks at one to three AI tools and drops off at four or more.

Cochrane relates this directly to his own workflow, often running two to four tools side by side. However, he pushes back on the doom framing. He argues that context switching across multiple projects and rubber-stamping AI output without review are the real sources of fry. His takeaway: either work more slowly with greater intent, or use the accelerated pace to reclaim free time.

Anthropic’s Wild Month: Exodus, Pentagon, and Mythos

Claude sessions surged by roughly 1,487% from mid-January to early March, knocking ChatGPT off the top spot in the app store for the first time. ChatGPT uninstalls spiked nearly 300%, one-star reviews exploded 775% in a single day, and a boycott movement called “Quit GPT” has grown to between 2.5 and 4 million participants.

The catalyst was OpenAI stepping in to take the Pentagon defense deal that Anthropic had publicly declined. Cochrane is firmly against automated domestic surveillance and autonomous weaponry, noting that the models are not reliable enough for such responsibilities. OpenAI tried to walk it back, but the Electronic Frontier Foundation called their language “weasel words.”

Meanwhile, the Department of Defense slapped Anthropic with a supply chain risk label – a national security designation previously reserved for hostile foreign companies. Anthropic sued the Trump administration. Then Microsoft filed a legal brief in Anthropic’s defense, joined by 149 former judges, dozens of Google and OpenAI employees, and nearly two dozen retired generals.

On top of all that, security researchers discovered an unsecured data cache exposing nearly 3,000 unpublished Anthropic files, including a model code-named Mythos (also called Capybara). Internal documents describe it as a step change in capabilities, scoring dramatically higher than Opus 4.6 on coding, reasoning, and cybersecurity. Then Anthropic’s source code leaked publicly as well.

Sponsor: GoDaddy

Economy hosting is $6.99/month, WordPress hosting is $12.99/month, and domains are $11.99. Both hosting plans include a free domain, professional email, and SSL certificate. Go to geeknewscentral.com/godaddy for the best pricing and to directly support this independent show.

OpenAI Shuts Down Sora Video App

OpenAI announced on March 24th that it is killing Sora, its AI video-generation app. Downloads cratered from 3.3 million in November to 1.1 million by February. The real numbers are brutal: Sora was costing roughly $15 million per day to run against a total lifetime revenue of just $2.1 million.

The Sora web and app experience ends April 26th, with the API shutting down September 24th. Additionally, the Disney partnership – a billion-dollar deal meant to validate AI in Hollywood – collapsed completely. Deep fakes of Martin Luther King Jr. and Robin Williams appeared almost immediately despite guardrails, and both families protested publicly. Cochrane notes that competitors like Runway, Pika, and Kling are still operating, and suspects Hollywood will pivot to generating scene backgrounds rather than full content.

OpenClaw Is a Security Nightmare

Cochrane’s personal OpenClaw install started making outbound requests flagged by his ISP – with no changes or new skills installed. He shut it down and plans to wipe the device entirely.

The broader picture is alarming. A January 2026 audit found 512 vulnerabilities in OpenClaw, eight critical. Twenty-six percent of community skills contain at least one vulnerability. Oasis Security discovered a vulnerability chain called “Clawjacked” where any website can silently take full control of a developer’s agent. Between March 18th and 21st alone, nine additional vulnerabilities were disclosed, several of which were rated 9.9 out of 10. Cochrane draws a direct parallel to the browser extension era: supply chain attacks hidden as helpful tools.

Claude Code Auto Mode: AI Policing AI

Anthropic published details on a new “auto mode” for Claude Code after finding that users approve 93% of permission prompts – essentially mashing “yes.” Auto mode replaces manual approvals with a two-layer defense: an input scanner to detect prompt injection and a second AI model that monitors the first and decides whether to allow each action.

The safety checker can only see what the user asked for and what the AI is trying to do. It cannot see the AI’s reasoning, so the AI cannot talk its way past the check. However, Cochrane notes it still misses about one in six dangerous actions (17%), and the fundamental question remains: if the base layer can get infected, so can the checker.

Qwen Overtakes Llama as Most-Deployed Self-Hosted LLM

RunPod’s 2026 State of AI report, based on usage data from 183 countries, reveals that Alibaba’s Qwen has overtaken Meta’s Llama as the most popular self-hosted AI model. Llama 4 has barely been adopted, with users sticking to version 3 because it just works. Additionally, vLLM now powers 40% of all AI endpoints, NVIDIA’s latest GPU usage scaled 25x last year, and nearly 70% of AI image work runs through ComfyUI. Cochrane sees Qwen winning on merit and argues that is how open source should work.

AI Data Centers Are Taking All the CPUs Too

AI data centers are not just consuming GPUs and memory anymore – CPUs are now being strained too. Intel server CPU lead times have stretched from two weeks to six months. AMD typically occurs at 8 to 10 weeks. Server CPU demand is proje...