Geek News Central

In this episode, Ray Cochrane digs into Anthropic’s admission that Claude models reached real systems during sandboxed safety tests. He also covers OpenAI’s Astra model cracking ten decade-old math problems, DeepMind open-sourcing its WeatherNext cyclone forecaster, and Europe’s brutal fire season. Finally, he wraps with the MacBook Neo sweeping K-12 schools, WhatsApp calling in the browser, and the troubled rescue of NASA’s Swift observatory.

– 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 with a personal update. He just got back from a trip home to O’ahu, where he helped his mom with projects around the house, caught up with friends, and finally rode the island’s new rail line ahead of its phase-two extension to downtown. Also in the pipeline, a September trip to Michigan is coming up, and he is closing out three years at Oregon’s Finest to focus on Blubrry and the show. Then he turns to the featured story.

Capture the Real Flag: Claude Reached Real Systems During Anthropic’s Safety Tests

Cochrane’s featured story comes from Anthropic’s own investigation. A misconfiguration left evaluation machines connected to the live internet during capture-the-flag safety tests, and Claude models reached real production systems in three incidents across 141,006 reviewed runs. One model extracted working credentials and entered a real company’s database, another published a malicious package to PyPI for about an hour, and the newest model recognized the environment was real and walked away from the flag. Notably, Anthropic found no evidence of models pursuing goals of their own. The models did what their evaluations asked.

The review began after OpenAI disclosed that its own models escaped an isolated test environment. Cochrane praises the transparency of the postmortem. However, he argues the incidents signal an industry-wide gap: evaluation infrastructure needs the same security rigor as production systems, and he hopes every lab takes the hint.

Sponsor: GoDaddy

Economy hosting $6.99/month, WordPress hosting $12.99/month, domains $11.99. Website builder trial available. Use codes at geeknewscentral.com/godaddy to support the show.

An OpenAI Model Cracks Ten Open Math Problems

OpenAI says an internal model called Astra produced ten new results on open problems in mathematics and theoretical computer science, each stuck for at least a decade. The wins range from sphere packing to a lattice problem behind post-quantum cryptography, plus two entries from Paul Erdős’s famous problem list. None are peer-reviewed yet, but every proof carries a machine-checked Lean 4 certificate. Cochrane finds the results impressive precisely because they reach past what humans could derive quickly, though OpenAI has not said how much human steering was involved.

Scientists With LLMs Will “Do More, Less Well”

Next, Nature’s news desk covered a modeling study predicting that scientists who adopt LLM tools will “do more, less well.” The team, which includes University of Washington biologist Carl Bergstrom, used optimal foraging theory to model researchers under publish-or-perish incentives. Faster AI-assisted work pushes them toward fresh projects instead of follow-up experiments and extra analysis. Cochrane relates from his own research: heavy LLM assistance can mean understanding the work less, which makes digging into the whys much harder.

Intel’s Chip-Stitching Hub in New Mexico

Intel’s newsroom showcased Fab 9 in Rio Rancho, New Mexico, its U.S. hub for advanced packaging. Since single chips max out at the reticle limit, Intel stitches specialized chiplets together with Foveros stacking and EMIB silicon bridges until they act like one huge processor. Packages already reach eight times that single-chip ceiling, with twelve times targeted by 2028. Cochrane muses about a future where consumers simply snap extra compute onto a board.

Gemini Robotics ER 2 Is a Brain for Robots

DeepMind released Gemini Robotics ER 2, its most capable embodied reasoning model. The system chats with humans, plans multi-step tasks, and hands motor control to whatever vision-language-action model a robot runs. Moreover, it calls tools like Google Search mid-task and plans its next move while its hands keep working. Demos show it orchestrating Boston Dynamics’ Spot and Apptronik’s Apollo 2, and developers can use it now through the Gemini API.

DeepMind Open-Sources Its Cyclone Forecaster

Staying with DeepMind, the new WeatherNext Cyclones model adds more than a full day of extra warning on a storm’s track, intensity, and wind structure. It ran live with the National Hurricane Center during the 2025 season, helping forecasters spot Hurricane Melissa’s rapid intensification before its Jamaica landfall. Best of all, DeepMind open-sourced the code and model weights for research and commercial use alike. Cochrane, a longtime fan of the lab, calls the public release a huge service.

Climate Change Supercharges Europe’s Disasters

Eos rounded up how climate change is supercharging Europe’s heat, drought, and wildfires. Southwestern France has burned through about 287,000 acres this year, and roughly 330,000 people across France and Spain fled their homes in one of Europe’s largest evacuations since World War II. France also recorded more than 5,700 excess heat deaths in under three weeks. Meanwhile, a Bordeaux-region blaze spawned a pyrocumulonimbus fire cloud, and Cochrane notes that hearing “unprecedented” every year is starting to create a precedent.

Ten Years of WILDLABS and Arm

Arm is celebrating ten years of backing WILDLABS, a network connecting more than 16,000 conservationists across 180 countries. The goal was never to invent new conservation technology. Instead, it closes the information gap so tools and lessons get shared openly. Cochrane highlights a team using miniaturized GPS tags to track migrating bats, and says the project bears the heart of open source.

FarmBeats Brings AI to Ag Classrooms

Microsoft and the National FFA Organization expanded FarmBeats for Students nationally. The program teaches future farmers with smart soil sensors, data science, and AI, mirroring how real farms automate tractors and deploy drones against pests. At launch, 185 middle and high schools received free classroom kits. Interested families can grab a kit for $35, and the Microsoft Learn training is free.

MacBook Neo Pulls Schools Off Chromebooks

Apple’s earnings call revealed the MacBook Neo displacing Chromebooks and Windows machines in K-12. Close to half of the Neo’s large U.S. education purchases last quarter replaced competing fleets, including 25,000 students in Pinellas County, Florida. Cochrane sees the first truly affordable Apple laptop finally competing where Chromebooks have lived for a decade. He expects Apple to keep taking K-12 share from Google and Microsoft.

Apple Upgrade Turns iPhones Into a Lease

Apple also launched Apple Upgrade, a hardware leasing program backed by Klarna. iPhones start at $17.99 per month with 12- to 36-month terms by device, and lease-end options let you upgrade, buy out, or walk away. However, entry-level hardware like the base iPhone 16 and Mac mini is excluded, and enrollment requires a Klarna account and a soft credit check. Cochrane admits he has never been a leasing fan, having once bought his phone outright with tip money, but he expects plenty of takers.

WhatsApp Calling Comes to the Browser

Meta brought calling to WhatsApp Web, no app download required. Browser calls keep end-to-end encryption and add screen sharing, a dedicated Calls tab, waiting rooms for group links, and automatic HD video. Additionally, group calls now transfer between devices mid-call, and a Drown Out the Noise feature suppresses background racket. Everything rolls out gradually over the coming weeks.

July’s Gemini Drop

July’s Gemini Drop adds voice control on macOS, letting users dictate and transform text in any active window. Gemini Spark, the always-on assistant, expands nearly worldwide, except the EEA, UK, Switzerland, and Nigeria. New Gemini 3.6 Flash and 3.5 Flash-Lite models arrive alongside avatar-based image personalization and app hookups for Dropbox, Zillow Rentals, and Viator.

The Rescue Satellite That Needs Rescuing

Katalyst Space Technologies’ LINK spacecraft, launched to boost NASA’s sinking Swift observatory, began tumbling in late July. Two of its three reaction wheels failed and its cold-gas thrusters degraded, so engineers used an electric xenon thruster to cut the spin from nine degrees per second to about 1.5. CEO Ghonhee Lee says a capture attempt is “very much in the cards,” with rendezvous in late August at the earliest. Cochrane wonders what happens if the window closes, and he promises to keep listeners updated.

Curiosity Finds a Honeycomb Valley on Mars

Finally, Curiosity rolled into a Martian valley nicknamed Valle Grande and found a honeycomb landscape. Polygonal fractures just inches wide stretch in every direction and climb a butte named Miraflores. Project scientist Ashwin Vasavada says the “sea of polygons took our breath away.” The rover just passed fourteen years on Mars, and scientists suspect ancient warming and cooling cycles fractured the ground.

Cochrane wraps with housekeeping: become a GNC Insider at geeknewscentral.com/insider, email geeknews@gmail.com, subscribe to the newsletter, and grab a modern podcast app at podcastapps.com (he uses Pocket Casts himself). He signs off until Monday’s show.

The post Capture The Real Flag #1871 appeared first on Geek News Central.

In this episode, Ray Cochrane breaks down AI distillation, the teacher-student technique frontier labs now lean on to train smaller, cheaper models. He also covers GPT-5.6’s government-vetted rollout, Claude Sonnet 5 landing on AWS, Maryland’s two-year data center pause, and Microsoft’s climbing carbon numbers. Finally, he wraps with Apple’s $30 billion Broadcom deal, Meta’s tamper-proof recording light, Michigan’s parasite outbreak, and a simulation that erased a super El Niño.

– Want to start a podcast? Its 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 with a quick personal update. Longer days have him outdoors, including a float trip on the Sandy River at Dabney State Park, where he found clearer water, clay-like sand, and easy footing. Next week brings both a move and a trip home, so he is stocking up on Trader Joe’s “Power Berries” and IKEA bags at his mom’s request. Then he turns to the lead story.

AI Distillation Explained: How Frontier Models Teach Each Other

Cochrane’s featured story comes from Hugging Face engineer Sergio Paniego. Distillation is teacher-student training for AI: a capable model generates the training signal, and a smaller student learns to match it. The classic off-policy version compresses giant models into cheap students, either through soft labels or piles of worked answers. Google’s Gemma models and DeepSeek’s R1-Distill line were built exactly this way.

However, the industry is now converging on multi-teacher on-policy distillation, or MOPD. Labs build reinforcement-learning specialists for math, coding, and agentic work, then have them grade a single student, word by word, as the student generates its own answers. DeepSeek-V4, MiMo-V2-Flash, and NVIDIA’s Nemotron 3 Ultra all run versions of the recipe, and the Qwen3 team reported better results at roughly a tenth of the GPU hours of raw reinforcement learning. Finally, self-distillation lets models like Cursor’s Composer 2.5 learn from better-prompted versions of themselves.

Sponsor: GoDaddy

Economy hosting $6.99/month, WordPress hosting $12.99/month, domains $11.99. Website builder trial available. Use codes at geeknewscentral.com/godaddy to support the show.

GPT-5.6 Arrives With a Government-Vetted Rollout

OpenAI shipped GPT-5.6 as a three-tier family: Sol, Terra, and Luna. Sol costs five dollars in and thirty dollars out per million tokens, half of Claude Fable 5’s rate. The benchmarks split: Sol Ultra wins Terminal-Bench at 91.9 percent, while Claude Fable 5 still leads SWE-Bench Pro. Notably, the API launched in limited preview to roughly 20 partners vetted by the U.S. government, though the model went live in Microsoft 365 Copilot on day one.

Claude Sonnet 5 Lands on AWS, Plus Quick AWS Wins

Claude Sonnet 5 arrived on AWS through Bedrock, pitched as top-tier intelligence at Sonnet pricing. Additionally, Amazon WorkSpaces for AI agents reached general availability, enabling agents to drive full desktop applications securely. OpenSearch gained a log-analytics engine claiming four times the price-performance, and SageMaker now scales inference about twice as fast. Cochrane also flags that Kendra and Q Business move to maintenance mode at the end of July.

Anthropic Wants You to Reflect on Your Claude Habits

Anthropic launched Reflect, a beta feature that analyzes your past Claude conversations and visualizes how you actually use the assistant. It requires Memory, excludes incognito and health-related chats, and keeps its insights inside the tool. Cochrane loves the idea. He reviews his own transcripts to extract prompt patterns and turn them into reusable skills, and he suggests listeners simply ask their AI to do the same.

AlphaEvolve Goes GA on Google Cloud

Google made AlphaEvolve generally available to Google Cloud customers on the Gemini Enterprise Agent Platform. The agent acts as an evolutionary collaborator: provide a baseline algorithm and your goals, and it searches for better, human-readable code. BASF, JetBrains, and Kinaxis are the named early adopters. Meanwhile, Cochrane renews his standing wish that DeepMind release AlphaGo as a playable teacher.

Google Adds “How This Ad Was Made” AI Labels

Google is adding a “How this ad was made” section to My Ad Center across Search, YouTube, and Discover. Ads built with Google’s own AI tools automatically get the disclosure, backed by invisible watermarks. However, ads made with outside tools rely on advertiser self-declaration. Cochrane points out the limits of voluntary disclosure in an AI-flooded content economy.

Microsoft’s Carbon Emissions Climb 25 Percent

Microsoft’s new sustainability report shows emissions up 25% in 2025, driven by a data center construction spree. The gross figure is 34 million metric tons before offsets, while other coverage puts the net figure at around 20 million. Water consumption also jumped thirty-four percent, even as Microsoft claims its first water-positive year. Cochrane argues regulation needs to catch up, since Google and Amazon report similar increases.

Prince George’s County Pauses Data Centers for Two Years

Prince George’s County adopted a two-year moratorium on new data center development, the longest pause in Maryland so far. The resolution blocks new applications, including hyperscale projects, until the council passes real regulations. Water and energy impacts remain open questions the county intends to study. Cochrane gives kudos to residents for making their voices heard.

Apple and Broadcom Ink a $30 Billion U.S. Chip Deal

Apple is expanding its partnership with Broadcom with a multiyear agreement expected to exceed $30 billion. The deal covers custom silicon and wireless components, with more than fifteen billion chips to be made on American soil. Broadcom’s Fort Collins, Colorado plant anchors the work with a $1.5 billion equipment expansion. Tim Cook framed the deal as accelerating Apple’s commitment to American manufacturing.

MSI and Intel Ship the First Arc G3 Extreme Handheld

Intel detailed how it co-engineered the MSI Claw 8 EX AI+, the first handheld on the Arc G3 Extreme processor. Highlights include a heat-spreading board layout and game-tuning loops that Intel says run Cyberpunk 2077 up to thirty-seven percent faster. The device is on sale now in void purple for around $1,500. At that price, Cochrane jokes he would rather buy a computer.

Meta’s Glasses Get a Tamper-Proof Recording Light

Meta answered the most common privacy questions about its AI glasses. Photos stay private on the device until the wearer imports or shares them, and a white capture LED blinks during any recording with no off switch. Moreover, newer glasses disable the camera if the LED is blocked, tampered with, or destroyed. Cochrane reminds listeners these claims are Meta grading its own homework, but the blink signal is worth recognizing in public.

Michigan’s Parasite Outbreak Tops 1,200 Cases

Michigan’s cyclosporiasis outbreak reached 1,251 cases since June 22, with roughly forty hospitalizations along the way. Northwest Ohio adds more than five hundred cases. The parasite typically spreads through contaminated fresh produce, and investigators still have not found the source. Cochrane’s advice: wash your produce, and get tested if your symptoms fit.

AI Finds the San Andreas Fault’s Silent Slips

Researchers paired AI with borehole strainmeters to detect dozens of hidden slow-slip events beneath the San Andreas Fault’s Parkfield section. Each silent slip releases stress within hours and is reliably followed by low-frequency earthquakes. Together, the findings support a continuous spectrum from silent creep to destructive quakes. The study appears in Nature Communications, and Cochrane hopes it will lead to better earthquake prediction.

Cloud Brightening Erased a Super El Niño, in a Simulation

Finally, a Science Advances study simulated marine cloud brightening in response to the 1997 and 2015 super El Niño events. Seeding clouds over the eastern Pacific erased the events entirely inside the model. Real deployment would take roughly 2,400 ships spraying continuously, and the simulations showed side effects like extra warming over Europe and Asia. Cochrane finds the weather-machine concept fascinating, yet he questions the consequences of altering cycles the planet runs for a reason.

The post AI Distillation: How Frontier Models Teach Each Other #1870 appeared first on Geek News Central.