Geek News Central

In this episode, Ray Cochrane unpacks Anthropic’s Mythos model and the Treasury’s emergency meetings with Wall Street, then digs into Apple’s vibe-coding crackdown and a gaming-anxiety study that hit way too close to home. Also covered: Verge’s solid-state motorcycle, UBTech humanoid robot sales jumping 23-fold, Japan’s first osmotic power plant, Finland’s permanent nuclear waste vault, Ghostty landing in Ubuntu, Cloudflare’s EmDash CMS, and a Claude Code skill that talks like a caveman.

– 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 by framing Anthropic’s new Mythos model as the AlphaGo moment for cybersecurity. From there, the episode moves through Apple’s pushback against AI-generated apps, a gaming anxiety study with a deeply personal hook, a series of “first to ship” energy and robotics wins out of Finland, China, and Japan, and several developer-tool stories that show how quickly the economics of software are shifting.

Mythos, the Detection Ceiling, and Wall Street’s Emergency Response

Anthropic’s Mythos model has Wall Street rattled. Operating autonomously, Mythos found and demonstrated the exploitation of a 27-year-old TCP SACK bug in OpenBSD, an operating system famous for being one of the most security-focused on the planet. Per Anthropic’s red team, over 99% of the vulnerabilities Mythos has identified remain unpatched. The researchers’ conclusion is blunt: “the moat in AI cybersecurity is the system, not the model.”

The policy response moved fast. On April 7th, Treasury Secretary Bessent and Fed Chair Jerome Powell pulled the CEOs of Goldman Sachs, Citi, Bank of America, and Morgan Stanley into Treasury headquarters on short notice. All four banks are now testing Mythos internally. Treasury CIO Sam Corcos is also seeking direct access. Anthropic is gating distribution through Project Glasswing, a limited-access program with JPMorgan, Apple, Google, Microsoft, and Nvidia.

Cochrane comes down firmly behind Anthropic’s gated approach. Because a 5.1-billion-parameter open model can apparently recover the core analysis chain for the OpenBSD flaw, this capability is not locked behind Frontier Compute. He wants the critical infrastructure hardened before the public gets keys. However, he also notes the bigger lesson is about human wisdom: people offloading all their thinking to AI lose out on the wisdom that makes any of these tools genuinely useful.

Apple Bans Vibe Coding Apps from the App Store

Apple has been quietly pushing back against what people are calling “vibe coding” apps. Replit, Vibecode, and an app called Anything all run AI models on the phone and produce working software that runs inside the host app. Apple cites Guideline 2.5.2, in effect since 2017, which requires apps to be self-contained. Replit and Vibecode had their App Store updates blocked. Anything was pulled in late March, briefly restored on April 3rd, and then pulled the same day again.

The forcing function is volume. App Store submissions jumped 84% in a single quarter as vibe coding tools flooded Apple’s review queue with AI-generated apps. Cochrane thinks Apple is justified, given the security issues swirling around the Vibe coding ecosystem. Even a beautiful diamond gets lost in a sea of sand, and that flood is exactly what Apple is trying to manage. The company behind Anything is now pivoting to iMessage, desktop, and Android.

Playing Video Games to Win Is Linked to Higher Anxiety

Cochrane gets personal on this one. Through high school and his early 20s, he was deeply addicted to League of Legends. His dad teased him about it constantly. In the last few years of that addiction, his body would go ice cold and shake every ranked match before. His partner identified it as a panic attack. The moment that happened, he quit. Today, he no longer shakes.

The new study lines up with his experience. Researchers Kayleigh Watters and Mikael Rubin at Palo Alto University analyzed a publicly available database of 13,464 adult gamers, most of whom primarily played League of Legends. Players who game to win show higher generalized anxiety but actually play fewer hours, since performance pressure pushes them out. Players who game to relax show strong links between social anxiety avoidance and more hours played. The study appeared in the Journal of Affective Disorders.

The headline framing of “playing to win makes you anxious” misses the point. The real finding is more interesting: gaming for avoidance and gaming for competition are both warning signs, for different reasons. Cochrane notes that the League of Legends community’s toxicity has been a running joke for years, and this study suggests the game’s structure may have been manufacturing the anxiety that fueled it.

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.

Verge Motorcycle: World’s First Production All-Solid-State Battery

Cochrane filled his tank for $60 today, which made this story land especially hard. His mom has driven electric for years and patiently manages a 90-mile real-world range. The next-generation answer is already shipping. Verge Motorcycles, a Finnish company, is the first production vehicle of any kind with an all-solid-state battery. Their 2026 bikes ship in Q1 with a pack from Donut Lab, another Finnish outfit spun out of Verge.

The numbers are bonkers. The pack delivers an energy density of 400 Wh/kg, roughly double that of current Tesla cells. It sustains 100kW charging, hits full charge in about 5 minutes in the lab and 12 minutes on the actual bike, and the long-range version covers 600 kilometers (about 370 miles) per charge. Toyota, QuantumScape, and Samsung SDI have all been telling us that solid-state is coming in 2027 to 2030. A Finnish motorcycle company shipping in Q1 2026 just embarrassed them all.

UBTech Humanoid Robot Sales Jump 23-Fold

UBTech dropped its 2025 annual earnings on April 1st. Humanoid robot revenue hit 820 million yuan, roughly $119 million USD, up 2,203% from 35.6 million yuan the year before. Unit sales went from 3 robots in 2024 to 1,079 in 2025. Shares jumped 14% on the announcement. The customer list is a real industrial deployment: BYD, Foxconn, Geely, FAW-Volkswagen, and Audi. The flagship is the Walker S2, with UBTech targeting 5,000 units in 2026 and 10,000 in 2027.

Cochrane is honest about what this means. He does not think we are heading for an extinction event, but worker displacement is a real concern. The US has no universal income or universal healthcare. The people affected are not white-collar managers. They are everyday line workers who already make the least on the ladder. Work efficiency reportedly doubles when these robots arrive, which is a company-side win, but the humans they replace are not getting half a year of gardening leave to retrain. He invites the listener to take on this one directly.

Japan Switches On Asia’s First Osmotic Power Plant

In August 2025, Fukuoka’s Seawater Desalination Center quietly opened Asia’s first osmotic power facility. It generates about 880,000 kilowatt-hours per year, enough for roughly 220 homes. It is only the second operational osmotic plant in the world, after Mariager, Denmark, in 2023. Osmotic generation uses a salinity gradient: fresh water on one side of a membrane, salt water on the other, and the pressure difference spins a turbine.

The clever part is what Fukuoka does with desalination brine. Instead of regular seawater, the plant uses concentrated brine left over from the desalination process. This amplifies the salt gradient and squeezes more energy out of the same membrane. The result is a closed-loop partnership: the desalination facility produces drinking water and leaves brine behind, the osmotic plant turns the brine into electricity, and that electricity runs the desalination facility. Every desalination plant on Earth produces brine, so if Fukuoka’s co-located model works, the same pattern could be replicated across hundreds of plants worldwide.

Japan’s Luna Ring Solar Moon Proposal Goes Viral Again

Shimizu Corporation’s Luna Ring concept is making the rounds again. The pitch: a 6,800-mile belt of solar panels around the Moon’s equator, beaming microwave power back to Earth. Project lead Tetsuji Yoshida has long argued that a full ring could eliminate fossil fuel dependence entirely. The proposal first surfaced in 2013, has no funding, no government endorsement, and no concrete cost estimate. Shimizu has not put any active development behind it.

Cochrane finds the concept fun every time it resurfaces. However, this would have to be a worldwide effort in the truest sense, with treaties, a new generation of launch economics, and microwave power transmission at a scale nobody has demonstrated. Beaming the power back to Earth has always been one of the biggest practical holdbacks. The Luna Ring is inspirational, but not shipping.

Finland’s Onkalo Nuclear Waste Vault Opens

Finland’s Onkalo facility is the world’s first permanent deep geologic repository for spent nuclear fuel. Operated by Posiva, the facility is buried about 430 meters down in 1.9-billion-year-old bedrock. It is designed to hold up to 6,500 tons of spent fuel and operate until the 2120s. The construction costs about €1 billion, with operating and closure adding roughly €4 billion more before the program is done.

The catch is that radioactivity remains dangerous for hundreds of thousands of years. Edwin Lyman, director of nuclear power safety at the Union of Concerned Scientists, warned that the copper canisters will eventually corrode, with different scientific opinions on how fast. Geologic disposal remains “fraught with uncertainties,” and we have never validated an engineered system across a 100,000-year time frame. The bet is that the rock and copper outlast the radioactivity.

Cochrane sees Onkalo as time-buying rather than a final answer. It is more of a bank holding spent fuel while science catches up. He prefers it to Japan’s ongoing approach of releasing tritium-treated water from Fukushima Daiichi into the Pacific, even though the dilution is well below WHO drinking water guidelines. Burying the waste in an insurmountable containment strikes him as the more honest answer to a problem nobody knows how to truly solve.

Ghostty Terminal Lands in the Ubuntu Repos

Ghostty 1.3.0 is now available in Ubuntu 26.04 LTS’s universe repository. The install is simply `sudo apt install ghostty`, no PPAs, no Snap, no Nix, no building from source. Ghostty was created by Mitchell Hashimoto, co-founder of HashiCorp. It is GPU-accelerated, uses native Swift on macOS and native GTK4 with libadwaita on Linux, and supports tabs, splits, profiles, ligatures, and the Kitty graphics protocol.

Cochrane recently caught Hashimoto on a podcast, where he walked through his agentic coding workflow. Ghostty is being actively built using AI harnesses like Claude Code and Codex. Hashimoto told a story in which Codex fixed a six-month-old bug in 45 minutes, for a total API cost of $4.14. Personally, Cochrane uses WezTerm, but he is excited to see Ghostty become more widely available with a native UI rather than Electron.

Borgo: Rethinking Go Using Rust

Analytics India Magazine profiled Borgo, a programming language by developer Marco Sampellegrini (GitHub: alpacaaa). Borgo is statically typed with Rust-like syntax, but it compiles to Go and uses the Go runtime and garbage collector. It includes sum types (Option and Result), pattern matching, and full compatibility with existing Go packages. Notably, it removes Rust’s borrow checker and lifetimes entirely.

Borgo is not new. It first appeared on Hacker News in 2023, with a RustLab talk in 2024. The 2026 angle is a renewed look at it through the lens of AI coding agents, since type-rich languages like Rust have been showing outsized productivity gains. Cochrane is a fan of Rust and stands by the borrow checker, but he enjoys these exploratory languages for what they reveal about what developers actually want.

Caveman: A Claude Code Skill That Cuts 65% of Tokens

Developer Julius Brussee built a Claude Code skill called Caveman that forces Claude to respond in stripped-down fragments. No articles, no “just,” no “really,” no pleasantries, no hedging. The tagline is “why use many token when few token do trick.” Across 10 real dev tasks, Caveman mode averaged 294 tokens per response, compared to 1,214 in normal mode. That is a 65% drop in output tokens. The project is MIT licensed with three intensity levels: lite, full, and ultra.

Cochrane stumbled across the project online and shared it with a classmate who had been complaining about token costs. The classmate now insists that “the caveman is the only way to live.” Cochrane has not made the switch, but the bigger point lands. If a community plugin can cut 65% of tokens without correctness regressions, the labs are shipping verbose-by-default and charging users for the privilege. He suspects verbose output makes models feel more trustworthy, even when the token math says otherwise.

Cloudflare Launches EmDash as a WordPress Successor

Cloudflare released EmDash on April 9th, an open-source, MIT-licensed, TypeScript-based CMS pitched as the spiritual successor to WordPress. The big flex is that it was built in 60 days using AI coding agents. EmDash runs on Astro 6.0, either on Cloudflare’s edge platform or on a standard Node.js server. The plugin security model uses sandboxed Dynamic Workers with explicit permissions, addressing the architecture flaw that Cloudflare says causes 96% of WordPress vulnerabilities.

Cochrane could not resist pointing out the irony of the name. The em dash has become the trademark giveaway that an AI was involved in writing. He has reservations about whether EmDash will succeed. WordPress is extremely hard to unseat, plenty of “WordPress killers” have come and gone, and the ecosystem is twenty-plus years deep. He is curious to see what comes next but not optimistic.

Google Open-Sources the DESIGN.md Format

Google Labs open-sourced the DESIGN.md format used by Stitch, their AI UI design tool. DESIGN.md is a declarative file capturing a project’s design system, colors, typography, and spacing in a way AI agents can read and apply. Cochrane has tried Stitch personally and finds it impressive at producing web designs. He has also seen DESIGN.md-style files already start appearing in repositories.

He sees this kind of file becoming a new paradigm for agentic design, alongside robots.txt and llms.txt. However, he worries about a side effect. If everyone uses the same standardized format and the same AI tools, the web could become a homogeneous set of sites that all look the same. He is enthusiastic about the standardization but hopes designers continue to push for genuinely unique work.

A 13-Liter PC With a Water Loop Built Into the Case

Geeky Gadgets covered a build by “Visual Thinker”, a 13-liter mini-ITX case with custom SLA-printed water distribution plates built directly into the chassis. Instead of traditional soft tubing, plates channel coolant between the CPU and GPU blocks and are sealed with TPU and silicone molds. The case supports a full-size GPU and an SFX power supply. No thermal benchmarks, parts list, or pricing have been published. It is a one-off you cannot buy.

Cochrane sees this as a sign of where PC building has gone in 2026. Modern mid-grade GPUs run nearly every recent game, so raw performance is no longer the differentiator. He likes seeing builders lean into design and craft rather than just stuffing the most powerful parts into a box. He admits he is the traditional type and built his own machine to maximize parts, but the design-first direction is a healthy evolution for the hobby.

To close out the show, Cochrane recommends Pocket Casts as a podcast app. He finds it picks up new episodes very quickly. Big thanks to GoDaddy for over twenty years of keeping this show on the air, and a reminder that every promo code use is like writing a check to the show.

The post Mythos: Cybersecurity’s AlphaGo Moment #1862 appeared first on Geek News Central.

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 projected to jump 15% in 2026, but Intel’s output capacity is growing in single digits.

The shift from chatbots to autonomous AI agents is changing the hardware ratio, since agents require far more CPU power to coordinate tasks and call tools. TSMC is prioritizing more profitable AI chips over regular CPUs. Cochrane warns that consumers and businesses are effectively subsidizing the AI boom through higher prices and longer waits.

AMD Ryzen 9 9950X3D2: First Dual-Cache X3D CPU

AMD announced the Ryzen 9 9950X3D2, the first CPU with dual-cache X3D technology. It arrives April 22nd with 208MB of total cache and a 200W TDP – up from the current model. However, AMD is unusually honest, calling the gains “modest,” ranging from 5-13% depending on the workload. Notably, they have not released gaming benchmarks, which is conspicuous for an X3D chip. Cochrane owns a single X3D chip and sees no reason to upgrade.

ARM Launches “AGI” CPU

After 35 years of licensing chip designs to Apple, Qualcomm, Samsung, and NVIDIA, ARM has launched its first production silicon: a 136-core server chip co-developed with Meta as the lead customer. ARM’s stock jumped about 16% on the news. You can pack over 8,000 cores in a single air-cooled rack, or over 45,000 with liquid cooling. Volume shipments begin by the end of 2026.

Cochrane appreciates the move but calls the “AGI” branding marketing hype. The bigger story is ARM transitioning from blueprint designer to direct competitor against Intel, AMD, and NVIDIA in data centers – while still licensing to the companies it now competes against.

Apple Discontinues the Mac Pro

Apple removed the Mac Pro from its website and confirmed that no future model is planned. The $6,999 machine had not been updated since the 2023 M2 Ultra model. Apple is pointing professionals toward the Mac Studio with its M4 Ultra chip, with an M5 Ultra refresh expected later this year. They also discontinued the $700 wheels kit, $300 feet kit, and Pro Display XDR the same week. Cochrane says good riddance – the Mac Studio covers what 90% of users need.

Apple’s AI Pin: An AirTag-Sized Wearable

Reports suggest Apple is developing an AirTag-sized wearable AI pin with cameras, microphones, and wireless charging. It would clip to clothing or hang as a necklace, running as an iPhone accessory powered by an upgraded Siri with Google’s Gemini AI. A possible 2027 release is expected alongside iOS 27, though development is early and could be canceled.

Cochrane ties this to a broader shift: data collection moving from the application layer to physical devices. Apple employees internally refer to the device as “the eyes and ears of the iPhone.” He warns that always-on wearable cameras, combined with existing AI-powered surveillance poles, are pushing society deeper into mass data collection without meaningful consent.

Quantum Entanglement Speed Measured for the First Time

Scientists at TU Wien’s Institute of Theoretical Physics, led by Professor Joachim Burgdorfer, measured how fast quantum entanglement happens for the first time. The answer: about 232 attoseconds – a billionth of a billionth of a second. The research was published in Physical Review Letters in late 2024 and is now circulating widely.

Einstein called quantum entanglement “spooky action at a distance.” Turns out it is not instantaneous – just extraordinarily fast. This measurement technique opens the door to quantum cryptography and quantum computing. However, Cochrane clarifies: this does not mean faster-than-light communication. Entanglement links particles but does not transmit information through space.

Bronze Age Iron Artifacts Came From Outer Space

Geochemical analysis by French scientist Albert Jambon, originally published in the Journal of Archaeological Science in 2017, confirmed that virtually all Bronze Age iron artifacts were made from meteorites. The artifacts span Egypt, Turkey, Syria, and China, including beads dating to 3200 BCE and the famous dagger from King Tut’s tomb, dating to around 1350 BCE.

The story resurfaced after researchers published new findings this month on fragments of meteoritic iron weapons from China’s Sanxingdui sacrificial site. Bronze Age people lacked the technology to smelt iron ore, but meteoritic iron arrived in a metallic state, ready to be forged. Cochrane closes the episode, noting that ancient civilizations were working with extraterrestrial material before they could produce their own iron – resourcefulness that deserves respect.

Cochrane wraps up the show by thanking GoDaddy for over twenty years of partnership and reminding listeners to subscribe, sign up for the newsletter, and reach out via email.

The post Agentically Frying your Brain using AI #1861 appeared first on Geek News Central.