Geek News Central Podcast

In this episode, Ray covers the fight that broke out after Anthropic published five case files on scientists using Claude for pathogen research, and the virologists who read the same files and called the work basic research. Additionally, he digs into DeepMind’s AlphaGenome Atlas mapping nine billion DNA variants, OpenAI’s numbers-free customer story about Perplexity, Gemini’s late arrival on Windows, and John Deere’s self-repair service that almost nobody uses. Finally, the show closes with Europe’s new plant-fluorescence satellite, a super El Niño already collecting damage, and a plankton shell that may be hiding a second species. – 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 back home in Oregon after a stretch in Michigan. Before the flight back, he explored Detroit and visited the Belle Isle Aquarium, where he caught the piranhas at feeding time and saw axolotls for the first time. Meanwhile, the in-flight entertainment was Rocketman, which he rates behind the Queen film but still enjoyed. Additionally, he notes he is back on the grind at Blubrry with plans for the week ahead. Then he turns to the lead story, a dispute between an AI company and the scientists who read its threat report. Anthropic Published Five Pathogen Cases, and the Virologists Pushed Back Science.org reporter Jocelyn Kaiser covered the fallout from a threat report Anthropic released last Thursday. In it, the company described five cases of scientists using Claude for work on dangerous viruses and toxins. Notably, no frontier AI company had ever said publicly that someone may have pointed its model at bioweapons work. However, Anthropic stated plainly that the people involved are working scientists and that it does not assert they intended harm. Three cases involved viruses. One sought chikungunya mutations that spread faster and evade the immune system, which Anthropic labeled highly concerning gain-of-function research. Another planned experiments on an already lethal bird flu strain. A third looked at orthopoxvirus mutations, and Anthropic argued that mapping what switches harm off also maps what leaves it on. What escalated the report was behavior, not biology. All five hid their location using relay servers, virtual private networks, or false identities. One case also ran through a reseller platform that quietly routed biology prompts to more permissive models, and Claude wrote much of that routing code after the job was pitched as an over-refusal fix. Plenty of serious people read it as alarming. Andrew Weber of the Council on Strategic Risks called the cases chilling examples of state-sponsored weapons developers using AI. Meanwhile, Ashish Jha warned on X that we are sleepwalking into a potential disaster. Then the virologists answered. Scripps Research virologist Kristian Andersen called most of the experiments just basic research, because the mutations already exist in nature and get dropped into a harmless stand-in virus. Gregory Koblentz of George Mason added that the workarounds sound like teenagers getting around parental controls. Filippa Lentzos landed in the middle, resisting both extremes. Cochrane sides with flagging it. He argues this kind of prompting should be monitored regardless of intent, and that publishing the cases was the right call. However, he questions whether Anthropic actually has the domain expertise in-house, suspecting consultants rather than full-time specialists. Four days later, the company launched a Life Sciences Verification Program, giving vetted researchers more permissive models, which he calls a sensible middle ground. 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. DeepMind Precomputed Nine Billion DNA Typos DeepMind released AlphaGenome Atlas, and the scale is genuinely hard to picture. Your genome runs about three billion letters, but only two percent spells out proteins. The rest decides when genes switch on and off, which is exactly where many disease-causing mutations hide. So DeepMind ran its AlphaGenome model against every possible single-letter swap in the human genome, all nine billion of them, then saved the answers. Each variant carries roughly twenty-seven thousand predictions across hundreds of cell types. Consequently, the dataset reaches a petabyte, more than thirty times the AlphaFold database. It is already finding things. GREGoR Consortium researchers used it to catch an overlooked variant in DNM1, a gene tied to severe childhood epilepsy, and confirmed the effect in the lab. Cochrane calls this the kind of work that blows his mind, and praises the DeepMind team for shipping AI that actually helps. Nevertheless, DeepMind attaches a clear disclaimer that the tool is not validated for clinical use. Perplexity, GPT-6 Astra, and a Story With No Numbers OpenAI published a customer story about Perplexity using GPT-6 Astra, its newest frontier model. Co-founder Johnny Ho says the model handles internal communications, edits real systems, and watches production software. Moreover, he says they now trust it with full end-to-end systems and check in far less often. However, Cochrane flags what the post leaves out. The only concrete workflow described is Astra building a test harness and running a workflow start to finish, which is end-to-end testing rather than unsupervised production. Additionally, it includes no accuracy figures, error rates, or rollback details. He also notes he has started running the same proposal across several models because different models produce genuinely different results. Consequently, he expects some kind of combined harness that runs models against each other. Still, he rates Astra’s real-world performance as better than he expected. A GitHub Marketer Automated Her Own Job Tomoko Tanaka runs regional marketing for GitHub in Japan and Korea, and she used to keep databases alive on Linux servers. Her pitch is simple: if you can write down how you do your work, you can automate it. The system runs on three pieces. An issue form captures the event details, a label triggers it, and a GitHub Action does the work. A plain-English runbook file also lets Copilot propose a campaign name and draft two versions of the invitation email. Cochrane admires the pipeline but raises a real concern. These systems can get set and forgotten, and the job then shifts from doing marketing to managing the AI that does it. Therefore, he asks what was actually automated away, and argues you lose some human flavor when every output runs through the same pipeline. Gemini Lands on Windows, Five Months Behind the Mac Google brought the Gemini app to Windows on September tenth, available globally on day one. It runs on Windows 10 or later, including the ARM chips in newer Copilot Plus laptops. Additionally, Alt and Space summon it over whatever you are working on. However, Google’s post leaves out two things. First, the Mac version shipped back on April fifteenth, so Windows trails by five months. Second, the best features are still Mac-only, including screen sharing, voice dictation across open apps, and connecting local folders. Cochrane is blunt about the gap. If a release cannot ship with the feature set the other platform already has, he asks why ship it at all. Meanwhile, he suggests Windows users curious about alternatives could check out Ubuntu. John Deere Built the Repair Tool, and Almost Nobody Uses It Writing for WIRED and republished at Ars Technica, Boone Ashworth tested John Deere’s self-repair service. Modern tractors carry upward of fifty separate controllers, and for years only a dealer could clear a fault code. That restriction sparked the right-to-repair fight in agriculture. Deere now sells Operations Center PRO Service for $195 a year per machine, rising to about $5,000 for a fleet license. Importantly, the reporter’s repair was a staged demo at Deere’s Santa Clara office, not a farm job, and he says himself it was an easy one. The real story is adoption. A year in, Deere reports about a thousand daily users against roughly 1.8 million American farms. Furthermore, Missouri farmer Jared Wilson, a named plaintiff in the $99 million class action, says the farmers he knows run cracked dealer software instead. He complains that the diagnostic tree doesn’t go deep enough when several codes fire at once. Europe Launched a Satellite That Reads the Glow Off Plants The European Space Agency posted an explainer for its FLEX mission, then launched it the following night. FLEX went up September fifteenth on a Vega-C from Kourou, sharing the ride with a Copernicus Sentinel satellite. Both are healthy and in commissioning, so science data is still a few months out. Here is what makes it clever. When a plant photosynthesizes, a little absorbed light comes back out as a faint glow called fluorescence. Crucially, that signal drops while the plant still looks perfectly green, which turns it into an early warning for drought, heat, and pests. The instrument, FLORIS, resolves down to a tenth of a nanometer in the bands it cares about most. Additionally, FLEX flies forty to a hundred kilometers ahead of a Sentinel-3 satellite so it can borrow atmospheric readings. ESA selected the mission back in 2015, so this launch took eleven years to arrive. Nunavut Is Already Turning Red NASA published a companion piece using Landsat 9 imagery of the Coppermine River valley in Nunavut. The September sixth shot shows red, orange, and yellow already sweeping the tundra, paired against the same ground still green on July twenty-seventh. What turns up there is not trees. Instead, willow and birch shrubs, blueberry, and bearberry hug the ground. Chlorophyll is unstable and needs constant rebuilding, so once days shorten, the green drains away, and the other pigments show through. Seven years of satellite data put color onset in early September with a mid-month peak, among the earliest on the continent. Meanwhile, the window is brutally short, sometimes a week or less that far north. The Super El Niño Is Already Collecting Nature’s James Dinneen reports that this year’s El Niño is forecast to be a monster. NOAA’s Climate Prediction Center gives it a seventy-five percent chance of beating every El Niño back to 1950. Furthermore, it strengthened earlier than the big events in 1997 and 2015. One wrinkle makes it unusual. This is the first strong El Niño in roughly thirty years, with the warm water stacked on the eastern Pacific rather than the middle, which shifts where the effects land. As researcher Jérôme Vialard puts it, El Niños come in flavors, like ice creams. The damage has already started. Peru canceled its main anchovy season on August twentieth, and its central bank projects the fishing sector shrinks 6.5 percent this year. Additionally, the Panama Canal cut daily transits to thirty-two after rainfall came in about thirty-six percent below normal. Meanwhile, the World Food Program projects forty-nine million more people facing acute food insecurity by the end of next year. The Plankton That Never Actually Flipped Quanta Magazine closes the show with foraminifera, single-celled ocean organisms smaller than a millimeter. Their shells blanket the seafloor when they die, creating a layered archive going back some 560 million years. Many of those shells spiral, and in some species as many as ninety-seven percent coil the same direction. Then, in the fossil record, an entire species appears to flip worldwide almost overnight. A 1959 explanation blamed temperature, but genetic work in 2006 and 2013 knocked that down. Bridget Wade’s team at University College London offers a different answer. Nothing flips at all. Instead, a second species that was always present, identical under a microscope but coiling the other way, picks up an advantage and replaces the first. So the coiling isn’t the advantage; it is an accidental barcode revealing a takeover nobody would otherwise see. Cochrane closes with a note on format. Viewership was noticeably higher for the video versions recorded at the classic studio, so he is looking at making the show video-centric going forward, possibly with a prepared visual presentation each episode. More on that in coming episodes. <!– ============================================================ WORDPRESS CHECKLIST (delete this section after publishing) ============================================================= CATEGORIES (select in sidebar): – Technology – Science – AI/ML – Space ============================================================= –> The post The Ebola Query #1878 appeared first on Geek News Central.

In this episode, Ray Cochrane devotes the whole show to Apple’s fall event, the first under new CEO John Ternus. He starts with the iPhone Duo, Apple’s first foldable iPhone. From there, he covers the Apple Watch Series 12, the new Health Age score, the Apple Watch Ultra 4, and AirPods 5 with noise canceling at $129. – 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 from his father’s studio in Michigan, fresh off a trip to Saugatuck Dunes State Park. Lake Michigan stunned him because it looked like the ocean. He had hoped to spot Wisconsin across the water. However, overcast skies and fog hid the far shore. He also apologizes for the hum in the previous episode, his first video session in the studio. Without headphones on, he took the noise for ordinary background sound. The culprit turned out to be a line-level cable plugged in at one end and loose at the other. After tracing it, he rewired the whole desk so every cable is neat and out of harm’s way. Meanwhile, he is still learning the studio’s TriCaster switcher and working on keeping his eyes on the camera. He invites listeners to send suggestions. Additionally, his grandmother had trouble finding the video version on the site. As a result, a site update is coming that uses the alternate enclosure feature, so listeners can switch between video and audio. Time with his grandmother has been a highlight of the trip. Cochrane reflects on how little time he spent with this side of the family after his father moved to Michigan. Next, he flies home to Oregon after some time in Detroit. As the move comes together, he also wants a better way to get recordings off the TriCaster than a USB drive. 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. iPhone Duo: Apple’s First Foldable Arrives Seven Years After Samsung Instead of a spread of headlines, the whole episode covers Apple’s fall event, much like Chris’s recent D23 episode. The event streamed Wednesday as a pre-recorded film. It was also the first event under CEO John Ternus. He took over from Tim Cook, who became executive chairman, on September 1st. It took Apple seven years to answer Samsung’s first Galaxy Fold. Even so, the price barely moved, since the Duo starts at $1,999. Cochrane admits he hoped for something closer to $1,400. Still, based on the marketing alone, the Duo is the upgrade he is most excited about. The Duo is a book-style foldable that opens along a vertical hinge. Closed, it has a 5.4-inch outer screen and is roughly the size of a passport. Open, it reveals a 7.6-inch display, the largest Apple has put on an iPhone. Both screens share the same proportions, so apps simply scale up to fill the bigger screen. Opened flat, the Duo is just 5.2 millimeters thick, which makes it the thinnest iPhone Apple has built. Closed, it measures 11.3 millimeters and weighs 254 grams. By comparison, Samsung’s Galaxy Z Fold8 is lighter and thinner at 201 grams and 4.5 millimeters open. The Crease, the Hinge and the Return of Touch ID The crease was Cochrane’s biggest worry, since it is the most common complaint about folding phones. Apple layers the inner screen with a flexible plastic cover, thin glass, and a titanium plate underneath. Custom adhesives let those layers slide as the phone bends. On top, a microscopically etched matte finish scatters reflections to hide the crease. However, Apple only claims the crease is harder to see, not gone. Early hands-on reports split down the middle. Chance Miller at 9to5Mac found it “hard to notice a crease at most angles.” In contrast, Andrew Romero at 9to5Google, working from early videos, said it looks “just as visible” as on Samsung’s Fold8. The hinge uses more than a hundred parts and a row of magnets to hold the phone shut. Cochrane says it snaps open cleanly, which a $2,000 phone should. Notably, the Duo carries an IP68 rating, surviving six meters of water for up to 30 minutes. That matters on a foldable, because dust in the hinge is a classic way these phones die. The Duo also has no Face ID. Instead, a Touch ID fingerprint reader sits in the side button and works open or closed. Cochrane offers his own theory for the switch. With AI now able to generate convincing faces and video, a fingerprint may simply be the safer biometric. Inside the Duo: Chip, Cameras, Software and Battery The Duo runs the A20 Pro, the same chip as the iPhone 18 Pro, built on a two-nanometer process. A vapor chamber carries heat away so the phone can hold full speed longer. Cochrane welcomes that, since his own phone runs hot while charging or running demanding apps. Two 48-megapixel cameras sit on the back, a main and an ultra-wide. The outer screen unlocks the new tricks. For example, you can shoot selfies with the rear cameras, and Duo Preview lets your subjects see the shot. Smart Take snaps the photo when everyone is posing, and Kid Cue plays Peanuts animations to get kids to look at the camera. On the software side, the Duo ships with iOS 27.1. The Dynamic Island now runs vertically along the side, and Split View puts two apps side by side for the first time on an iPhone. Netflix, Zoom and Slack already support it. The Duo carries two batteries, one in each half, and Apple rates it for up to 24 hours of typical use. Charging to 50 percent takes about 20 minutes with a 60-watt adapter. However, that adapter is sold separately, which Cochrane calls out as a raw deal for first-time buyers. MagSafe and Qi2 wireless charging run at up to 25 watts. Storage starts at 256GB for $1,999 and climbs to $3,199 for 2TB. Preorders open October 16th, and the Duo reaches stores October 23rd. Ternus called it “the most transformational change to iPhone since the original.” Cochrane still wants one, though the price gives him pause, and he asks listeners whether they plan to buy. Apple Watch Series 12 Checks Your Heart Every Five Seconds Next, Cochrane turns to the Apple Watch Series 12 and its new Health Sensing System. The watch reads your pulse by shining green light into your skin and measuring what bounces back. After wrestling with the technical term, photoplethysmography, he notes that everyone sensibly calls it the optical sensor. Apple made those lights bigger and more efficient. As a result, the watch now checks your heart rate every five seconds, all day. It also measures heart rate variability, the tiny wobble between beats, as often as every five minutes. A bigger wobble generally means your body is rested and recovered. Cochrane loved Apple’s swimming ad, where a swimmer finishes his swim and checks his heart rate. To him, it feels far more futuristic than the rugged adventure watches of twenty years ago. Swim tracking itself isn’t new. What changed is the constant background check and a live heart rate readout on the watch face. All of that feeds a new daily readiness score from zero to ten. It weighs your recent activity, training load, vitals and sleep, then updates through the day. Engadget points out that competitors already offer readiness scores and faster sampling. In Cochrane’s view, Apple is catching up to the market, but doing it in a big way. Apple’s Heart Rate Accuracy Study Apple claims no wearable reads heart rate more accurately, and this time it published the study. Testers wore a Series 12 on one wrist and a rival on the other, with a Polar H10 chest strap as the reference. The rivals came from Garmin, Google, Huawei, Samsung, WHOOP and Oura. Across 1,254 people at five sites, Apple led in 139 of 159 comparisons. Eighteen were too close to call. The only two losses went to Google’s Pixel Watch 4 at rest, by less than half a beat per minute. Cochrane appreciated that a fifth of participants had the darkest skin tones, where optical sensors have historically struggled. The Series 12 comes in 42- and 46-millimeter sizes, including a new ceramic version. Battery life holds at 24 hours, and fifteen minutes of charging now buys up to twelve hours. Pricing starts at $399 for aluminum, $699 for titanium and $899 for ceramic. It reaches stores September 18th. Health Age and Apple’s Rebuilt Health App Cochrane admits he first misread this feature, and he is glad his initial read turned out wrong. Apple is rebuilding the Health app around Apple Intelligence, its umbrella name for its AI features. The app adds an Insights tab, a For You section, and a Longevity tab covering seven areas of health. Sumbul Desai, a physician who runs health and fitness at Apple, framed the goal. She said pairing the watch’s sensors with Apple Intelligence lets Apple “translate complex biometric data into clear guidance for users.” In other words, nobody wants to scroll through a spreadsheet of their own heartbeats. Inside the Longevity tab sits Health Age. It weighs your VO2 max, resting heart rate, sleep, and heart rate variability against your real age. You can also add lab results like A1c and LDL cholesterol. Apple’s sample screen shows a body running 3.9 years younger than its birthday. Cochrane finds the feature interesting, and he hopes the data stays private, much like health records at a hospital. However, Apple didn’t share the math behind Health Age. He would like Apple to publish its formula, because people will take a number like that seriously. The new app also adds at-home movement evaluations using the iPhone camera and the watch. Users can book a $119 blood panel covering more than 50 biomarkers at roughly 2,000 Quest Diagnostics locations. Cycle Tracking gains perimenopause and menopause support. The redesigned app arrives later this year, while watchOS 27 lands September 14th. Finally, Cochrane raises a concern about all this monitoring. He hopes Apple built strong safeguards around the health data, because a leak of data like this would bring serious lawsuits. Apple Watch Ultra 4 Can Outlast a 100-Mile Race The Apple Watch Ultra 4 shares nearly everything with the Series 12, from the sensors to the S11 chip. This year, what sets it apart is stamina. Normal use now reaches 50 hours, up from 42, and Low Power Mode stretches to 84. Meanwhile, fifteen minutes on the charger buys up to 18 hours, where last year’s model managed 12. A new Max Extended Workout mode tracks up to 45 hours of running or hiking, with GPS logging every second. Apple says that is typically enough to finish a 100-mile race. The watch also uses dual-frequency GPS, which listens to two satellite signals to reject readings bounced off tall buildings. DC Rainmaker came away impressed with both watches. As Ray Maker put it, “I’d argue they’re the biggest leaps in Apple Watch hardware and features since they launched the Ultra series back a few years ago.” The Ultra 4 holds at $799 and ships September 18th. Audio Intelligence and the Always-Listening Question Both new watches also gain a set of listening features Apple calls Audio Intelligence. Sound Recognition alerts deaf or hard-of-hearing users to sirens, alarms, doorbells, or a crying baby. Shazam names nearby songs without being asked. Then come the two raising eyebrows: Live Rewind and Siri Recap. Live Rewind shows the last fifteen seconds of your conversation as text with a double press of the Digital Crown. Siri Recap hands you a title and key points after a conversation, leaving out financial details and ID numbers. Both arrive in beta later this year, in English first, and need an iPhone 16 or newer. Cochrane saw the pushback coming. To replay fifteen seconds, the watch has to be listening before you ask. Apple answers with privacy promises, including that everything is opt-in and raw audio is deleted after processing on the chip. Still, he argues that opting in means choosing to give up privacy, and he has a hard time believing the processing stays on the device. AirPods 5 Bring Noise Canceling to $129 AirPods 5 comes in two versions, and the headline is price. Until now, noise cancellation in the regular AirPods cost $179. Both AirPods 5 models include it, starting at $129. Noise canceling is harder in the regular AirPods, because they lack the rubber tips that seal your ear. Apple’s AirPods marketing director Eric Treski admitted as much in 2024, calling it “really, really hard to create this great ANC quality in a non-ear-tip product.” Nevertheless, Apple says AirPods 5 removes up to 50 percent more outside noise than AirPods 4 with noise canceling. Interestingly, the chip and microphones are unchanged. The gain comes from reworked vents and ports in the shell, plus smarter software. For perspective, Apple says AirPods Pro 3 still cancel up to three times more noise. Gizmodo found AirPods 5 “did a good job silencing the background noise, but not entirely.” The $129 model charges over USB-C. For $149, you add a wireless charging case, a Find My speaker, a volume swipe on the stem, and an extra hour of listening. Both models also carry an IP57 rating and support Live Translation. They reach stores September 18th. How Apple Filmed the Event Cochrane closes by praising the production itself. Apple built different scenes and backgrounds throughout the film, and Six Colors noted a Hollywood backlot segment with a fake rainstorm, Roman soldiers, and astronauts. The closing card read “Shot on iPhone. Edited on Mac.” He still hopes to get an iPhone Duo, though he isn’t sure about paying two grand. For now, he may stick with his current phone. Cochrane wraps up by pointing listeners to the GNC Insider program, the newsletter, and podcastapps.com, adding that he likes Pocket Casts. Listeners can reach Ray or Chris at [email protected]. He flies back to Oregon next, with another audio episode planned for Monday and a video site update coming this week. The post Apple’s September Event: iPhone Duo, Apple Watch Series 12, and AirPods 5 #1877 appeared first on Geek News Central.

In this episode, Ray Cochrane records from the Michigan studio on the one year anniversary of losing the show’s original host, Todd Cochrane. The featured story is Anthropic’s $5 million grant program, which pays outside researchers to measure whether AI is actually good for the people using it. Ray also covers Anthropic’s Enterprise Frontier Safeguards, Google’s Fairwind cyber defense program, Meta’s organizational second brain, sixteen green AI projects across Asia-Pacific, an Oklahoma Bitcoin mine condemned after leaking 3.8 million gallons of water, BepiColombo closing in on Mercury, and the 2026 Ig Nobel Prizes. – 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 from the studio in Michigan, recording on the one year anniversary of the show losing its original host, Todd Cochrane, his father. The family has gathered for the occasion. He also plans to record several episodes throughout the week as a commemorative run, so listeners should expect a heavier release schedule than usual. He also shares where things stand. He and Delaney have gone back and forth on relocating to Michigan, but the family is pushing, and the plan is firming up for the coming month or so. Part of the draw is getting his dad’s studio cleaned up and running the way it used to. Finally, he passes along an update on the Zuvers, a story long-time listeners will recognize. The father of the three children is due out of jail soon, though he now faces retrial for the murder of the three boys. Cochrane notes that the original prison term related to their return rather than the murders themselves. The family’s expectation, he says, is that this time the sentence will be permanent. Anthropic Puts $5 Million Behind Measuring AI’s Effect on You The featured story asks a question the industry has mostly avoided. Not whether a model is smart, but whether you are better off after using it. Anthropic is funding independent research to find out, and the structure is the interesting part. The money creates what the field calls evaluations: standardized tests for AI models. Hundreds already exist for coding, math, and reasoning. However, almost none measure whether the tool is good for the human on the other end. Crucially, Anthropic is not building these tests in-house. Grantees receive funding, model access and technical support, then work fully independently. Everything they produce must be published open source, so any developer at any company can run it. As Cochrane puts it, Anthropic is paying to create a ruler other people will use to measure Anthropic. The example Anthropic leads with is deliberately uncomfortable. Claude might offer diet and workout advice to someone asking about weight loss. But if that user has a history of disordered eating, the same response becomes actively harmful. Cochrane connects this to sycophancy, a topic the show has returned to before, and notes that models carry no real memory of who you are unless you tell them. That is exactly why the problem has gone unmeasured. Code compiles or it doesn’t. Math checks out or it doesn’t. Wellbeing has no answer key, because the identical response can help one person and harm the next. The Five Things Anthropic Wants These Tests to Do Cochrane walks through all five criteria, since each targets a specific way this research usually fails. First, state plainly what you are measuring. It is easy to measure something adjacent, such as how often a model uses warm language, then publish numbers about a proxy while claiming insight into loneliness itself. Second, involve clinicians and subject-matter experts in designing and validating the tests. A capable engineer may simply not know what the warning signs of an eating disorder look like in text. Without those people in the room, the test measures what engineers imagine the harm looks like. Third, count overcompliance and overrefusal as harms. Overcompliance is the obvious failure. Overrefusal is the model becoming so cautious it turns useless, or shutting down someone who had nowhere else to ask. Cochrane adds that overrefusal is far harder to spot, because you often cannot tell it is degrading until it has degraded. Fourth, reflect real multi-turn usage. Most safety testing throws one nasty question at a model and checks for refusal. Meanwhile, actual harm builds across a long conversation where every individual message reads as fine. Fifth, validate the automated graders against real experts. Scoring thousands of conversations means an AI grades the AI. Consequently, a subtly wrong grader skews every downstream number in the same direction, and nothing in the process flags it. Cochrane expects this to be the hardest criterion to satisfy, since psychiatric and medical specialists do not come cheap by the hour. Applications close September 21st, and shortlisted applicants will be notified by October 5th. Notably, the announcement names no individual researcher or executive. 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. Anthropic’s Enterprise Frontier Safeguards Keep Your Logs in Your Vault Anthropic’s second announcement targets a completely different audience. Catching misuse across many sessions requires keeping logs. However, banks and hospitals cannot hand that data to an outside vendor, because regulators will not allow it. Safety monitoring and compliance were in direct conflict. The fix moves where the data lives. Activity logs go into the customer’s own cloud bucket at Amazon, Microsoft or Google, locked with keys the customer controls. Automated systems scan a rolling window for serious abuse such as credential theft, and alerts route straight to the customer’s own security team. No human at Anthropic reads it, and Anthropic does not charge for the feature. The backstory explains the urgency. Anthropic had begun retaining thirty days of data with its Fable 5 model, specifically to catch patterns that only appear across multiple sessions. Regulated customers had to walk away immediately. Over a hundred organizations helped shape the result, including a quarter of the Fortune 100 and every US globally systemically important bank. It works across Claude Code, Claude Enterprise, Amazon Bedrock, and Microsoft Foundry. Wells Fargo’s security team put it simply, saying their logs stay in a Wells-managed environment under Wells-managed keys. Cochrane adds a personal note on the same retention change. Because transcripts now persist thirty days after last touch, he can keep Claude Code threads open far longer and lean on clear without losing context. Google’s Fairwind Program Hands Patching to the Machines Google is running the same play from the defensive side. Normally, a human engineer must reproduce a security flaw, understand it, write a fix, then prove the fix broke nothing else. That can stretch to weeks, and attackers live in exactly that gap. Fairwind pairs two systems to close it. Gemini 3.8 Flash Cyber is a model tuned for security work. CodeMender hunts for flaws and writes the patches. Together, Google says they produce verified, deployment-ready patches in minutes, running inside the customer’s own environment. Access is deliberately restricted to three groups: national cyber authorities; critical infrastructure operators in healthcare, telecom, energy, and finance; and major technology platform providers. More than 650 partners are involved. Cochrane’s take centers on logging quality. Systems that emit specific, unique errors rather than generic ones let these tools find real issues almost immediately. Furthermore, he sees value in purpose-tuned models over general ones, since different training data and methodology surface gaps a single familiar model would miss. Meta Built an AI That Learns From Its Own Experts Every organization has two or three people who actually understand the hard systems. When they are busy, everyone waits. When they leave, the knowledge leaves too. Most AI assistants address this with retrieval, searching documents when you ask. Meta went the other way. A long-running offline process digests the source material ahead of time into more than 200 structured knowledge files, so the thinking happens before anyone asks. On top sit what Meta calls recipes, step-by-step procedures mirroring how a specialist actually works a problem. The reported numbers are concrete. Restructuring cut tokens consumed per turn by 80 percent. Assessments that took days now take minutes. The team built the whole system in six weeks across three sprints. Cochrane focuses on the feedback loop. Normally, an expert corrects an assistant in chat; the answer improves once, then the system forgets. Meta’s version compiles corrections into knowledge files and regression-tests them, so fixes stick without retraining the model. However, he raises two real caveats from his own attempts. Building those knowledge files is token-heavy, and codebases change. If the offline process does not rerun, the documentation drifts away from the code it describes. He also predicts this will become a service offered by AWS, GitHub, or GitLab rather than something each company builds alone. Sixteen Green AI Projects Across Asia-Pacific Google DeepMind named the first cohort of its Accelerator: AI for the Planet program. Sixteen organizations span New Zealand, Singapore, South Korea, Indonesia, Thailand, India, Australia and Japan, receiving three months of access to Google’s AI stack plus mentorship and a Singapore bootcamp. Several stand out. A New Zealand outfit called 800 Trust uses bioacoustics, monitoring an ecosystem by listening to it and running AI over continuous audio to track biodiversity. Wildlife.ai builds open-source AI camera traps. Australia’s X-Centric replaced the soil lab with a handheld X-ray reader that answers a farmer standing in the field. In India, Climitra Carbon verifies invasive species removal and converts the biomass into biochar. Access matters more than money, as Cochrane points out. A six-person conservation nonprofit cannot run frontier models, because the compute bill alone would consume its operating budget. The organizations closest to these problems have always been furthest from the tooling. An Oklahoma Bitcoin Mine Condemned After Leaking 3.8 Million Gallons Cochrane corrects the figure up front. Many outlets report three million gallons, but El Reno city officials put it at 3.8 million. A Bitcoin mining operation on West Jensen Road leaked water and dropped pressure badly enough to close El Reno Public Schools, the Canadian County Courthouse, and other city and county offices. It happened during a month with 24 days above 100 degrees, in a region under severe to extreme drought. The leak is not the damning part. The city issued a stop-work order on the site back in 2023 for electrical, construction, and fire safety violations, then never followed up. The facility ran roughly three years under an order nobody enforced. Interim city manager Ken Brown did not dodge it, saying simply, “We failed.” The site is now posted as condemned with a ten-day removal notice, and a hearing with operator Athlon BT is set for September 14th. Reaching the company has proven difficult, with a website stuck on a maintenance notice and an unanswered California phone number. Cochrane draws one distinction that the coverage keeps blurring. This is a Bitcoin mining rig, not a hyperscale AI data center. The residents who lost pressure absorbed the cost, and nobody billed the operator for the aquifer. BepiColombo Closes In on Mercury After Eight Years Reaching the closest planet to the Sun took nearly eight years, which sounds backward until you understand the physics. Falling toward the Sun means gaining enormous speed, and a spacecraft arriving too fast simply sails past. So the mission spent those years shedding velocity across nine gravity-assist flybys: one at Earth, two at Venus, and six at Mercury itself. On September 3rd, the transfer module separated and was discarded more than 200 million kilometers from Earth. Mission control marked the moment with “Roll call completed, GO for separation.” Importantly, the spacecraft has not arrived yet. Orbit insertion is November 21st, the two science orbiters separate around December 9th and 10th, and science operations begin in April 2027. That November date is revised, having slipped from an earlier plan. Two orbiters fly because they do different jobs. Europe’s Mercury Planetary Orbiter studies the planet and what lies beneath its surface. Japan’s Mio studies the magnetosphere. Mercury having a magnetic field at all is genuinely odd, since a planet that small should have cooled and lost it long ago. The mission honors Giuseppe “Bepi” Colombo, the Italian mathematician who worked out how to slingshot a spacecraft to Mercury in the first place. NASA Builds a Rocket Engine Talk You Can Hear NASA hosts a webinar on Friday, October 2nd at 2 pm Eastern, titled “The RS-25 Engine and the Future of Artemis Missions.” It runs about two hours, is open to anyone, and RSVPs close September 25th. The RS-25 is not new hardware. It is the Space Shuttle main engine, the same design that flew 135 shuttle missions across three decades, now bolted four at a time into the SLS core stage. Together they produce roughly 2.2 million pounds of thrust at 111 percent of their original shuttle rating. Throwing them away is the strange part. These are precision engines built to fly repeatedly, yet Artemis expends four on every launch. NASA got a proven engine and skipped a decade of development, and that is the trade. The accessibility work is the real story. The event includes an audio-described video of an engine test firing, a live Q&A with an Artemis engineer, and a panel on accessibility in space and science. A test firing normally sells entirely on spectacle. Conveying that, and the engineering underneath it, without the visuals is a genuine design problem NASA built the whole event around solving. Buried Underwear Wins a 2026 Ig Nobel The 2026 Ig Nobel Prizes were awarded September 3rd in Zurich. Marc Abrahams launched them in 1991 at the satirical magazine Annals of Improbable Research, and what began as a roast is now something researchers actively want. The soil science prize went to a team that buried 1,000 pairs of underwear across more than 25 countries and dug them up two months later. Measuring how much cotton rotted turns out to be a cheap, surprisingly good proxy for how biologically alive the soil is. The biomechanics prize recognized a precise, cross-species definition of kissing that works for animals and excludes passing food. By that definition, polar bears kiss, some birds kiss, and so do ants. Nature reports the behavior traces back roughly 21.5 million years to the ancestor of all large apes, meaning ancient humans very likely kissed Neanderthals. Cochrane’s favorite was the physics prize, awarded for a splash-free urinal design that reportedly cuts spray to 1.4 percent of normal. Cochrane closes by pointing listeners to the GNC Insider program, the newsletter, and podcastapps.com for a modern podcast app. Feedback on the show’s format is explicitly welcome at [email protected], where either Ray or Chris will read it. The post Anthropic’s $5 Million Push to Measure AI’s Effect on Wellbeing #1876 appeared first on Geek News Central.

In this episode, Ray Cochrane breaks down Hot Chips 2026, the engineering conference where IBM, NVIDIA, Intel, AMD, Arm, and Fujitsu all showed how their next processors actually work. The headline disclosure is a mainframe core that runs Arm natively. Ray also covers Apple’s odd M6 Mac mini naming, London’s first autonomous Uber rides, Amazon’s purchase of the company behind DuckDB, GitHub’s HydraFusion, the best of IFA 2026, and new USDA research on farmed salmon. – 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 personal update. He apologizes for the late rollout and the missed Monday episode, having spent the week fighting a cold. He’s also heading to Michigan to spend time with family and visit his father’s gravesite. He hopes to record a couple of shows from his dad’s old studio while he is there, including a special episode planned for Tuesday. He also points listeners to a fresh site redesign that trades the old techie look for something cleaner and friendlier. The featured segment starts from a wrap-up post on Arm’s newsroom. However, Cochrane broadens it to cover the whole conference rather than a single article. Hot Chips has run every August since 1989, and this year’s event was the 38th, held August 23rd through the 25th at Stanford’s Memorial Auditorium. What Hot Chips Actually Is Cochrane draws a line between Hot Chips and the big consumer trade shows. CES and Computex exist for product announcements and marketing. Meanwhile, Hot Chips is an IEEE engineering conference where chip architects present block diagrams and die photos for thirty minutes at a stretch. The audience matters as much as the content. Roughly five hundred people who design chips for a living fill the room, and they would spot a fudged number immediately. No written paper is required, just the talk and the slides. In-person tickets sold out this year, as did Stanford’s dorm housing. Cochrane says he plans to cover the conference annually going forward. IBM Built a Mainframe Core That Speaks Arm The disclosure that stopped Cochrane cold came from IBM. Its next processor for IBM Z and LinuxONE runs two completely different instruction sets natively, on every one of its eleven cores. Those are z/Architecture, IBM’s own mainframe language, and AArch64, which is 64-bit Arm. Crucially, this is not emulation. Nor is it Arm cores glued onto the die beside the mainframe cores. IBM built 2,792 Arm instructions directly into the hardware, which it says is more than double the mainframe instruction count. Each core carries two separate decoders while sharing the caches, branch predictor and register files downstream. It switches between the two in nanoseconds. IBM even added dedicated hardware to flip byte order, because Arm and the mainframe store numbers in opposite directions. The payoff is Arm SystemReady compliance, meaning off-the-shelf Arm Linux runs on a mainframe unmodified. Patrick Kennedy of ServeTheHome, who was in the room, wrote: “I am sitting here still in awe of what IBM is doing here; this is not Z plus Arm cores, this is Z and Arm in one core.” Cochrane flags one precision point that is easy to get backward. IBM did not license Arm’s core designs and drop them in. Instead, it took its own mainframe core and taught it AArch64 under an architecture license, which is considerably harder engineering. The specifications are striking. The chip uses a 2nm process, with eleven cores running above 5.7GHz sustained and no turbo mode at all. Each core gets 36MB of L2 cache, backed by a 3.5GB virtual L4 pool. Furthermore, the reliability target is eight nines, which works out to roughly three tenths of one second of unplanned downtime per year. IBM gave it no name and no ship date, though the press expects “Telum III” around 2028. Arm, Fujitsu and NVIDIA Show Their Hands Arm itself had plenty to discuss, starting with an unfortunate name. Its first chip in thirty-five years is called the AGI CPU, which is a product name rather than any claim about artificial general intelligence. For three and a half decades, Arm designed processor blueprints and licensed them out, collecting royalties without competing. That era is now over. The AGI CPU is Arm’s own silicon, co-designed with lead customer Meta, running up to 136 cores on TSMC’s 3nm process at 300 watts. Arm’s CEO says the company has more than $2 billion in customer demand across the next two fiscal years. Fujitsu brought the detail Cochrane called the coolest of the conference. Its MONAKA chip packs 144 Arm-based cores, but the trick is the cache. Rather than sitting alongside the cores and eating die area, the entire last-level cache lives on a separate 5nm die with the 2nm compute die stacked directly on top. It ships in 2027 in 350-watt and 500-watt versions. NVIDIA had more stage time than anyone, with six sessions. Its new Vera CPU carries 88 cores of NVIDIA’s own Olympus design, which marks a change: the previous Grace CPU used Arm’s off-the-shelf cores. Consequently, Arm’s win here is the instruction set, not the blueprint. The memory disclosure drew the most attention, with a fully loaded system reaching 1.5TB at 1.2TB/s while the whole memory subsystem draws just 30 to 40 watts. The Caveat on NVIDIA’s Benchmark Slides Cochrane pushes back on how NVIDIA presented its numbers. On the standard SPEC integer benchmark, Vera scored 925 against AMD’s 128-core EPYC score of 898, about three percent ahead. However, the slide NVIDIA showed normalizes that same result per physical core, which makes a three percent gap look enormous. NVIDIA defends the choice, arguing that per-core throughput matters when thousands of AI agents run at once. Cochrane grants that it is a fair argument to make. Even so, his verdict is blunt: it is a different number from the headline one, and presenting it that way is not the best look. 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. Apple’s Newest Chip Landed in Its Cheapest Mac Last week’s episode covered the Mac Studio half of Apple’s August announcement. Tonight Cochrane takes the other half, the new Mac mini, and finds the numbering genuinely strange. The $899 base Mac mini gets the M6, which is Apple’s first 2nm chip and the newest process the company has shipped. It brings twelve CPU cores, twelve GPU cores, and 170GB/s of memory bandwidth. Apple also introduced a third CPU core class called super cores. So Apple’s most advanced chip sits in its cheapest desktop, while the M5 Pro, M5 Max and M5 Ultra above it all carry a lower number. It gets stranger. The M6 mini has Thunderbolt 4 while the pricier M5 Pro mini has Thunderbolt 5, and the memory ceiling runs backward too. Apple explains none of it across three announcement pages. Reading between the lines, Cochrane figures the M6 is the entry point of a new generation that shipped ahead of its larger siblings. London’s Robotaxis Started Carrying Passengers Arm’s monthly roundup covers everything outside the data center, and Wayve stood out. Transport for London granted the British self-driving company private hire vehicle licenses on August 5th, the same category a minicab needs. Then on September 3rd the service launched with Uber, marking the first autonomous rides ever offered to UK passengers. A small fleet of Ford Mustang Mach-Es covers anywhere in London except the airports, with a TfL-licensed safety driver still aboard. Over 140,000 Londoners signed up. The compute runs on NVIDIA’s Arm-based automotive platform, which is why Arm claims the win. Cochrane notes the pattern: once you set a standard nobody can move off, these wins keep arriving. Two more items round it out, both about squeezing AI onto phones. Google’s Pixel 11 shipped in August with the Tensor G6, and Google claims on-device AI runs up to 3.5x faster using 3.5x less energy. Those are Google’s own unbenchmarked figures. Separately, Graphcore’s research team worked with Arm to run an 11-billion-parameter vision model on phone-class processors by squeezing each parameter to 2.7 bits, taking the model from roughly 22GB down to 3.7GB. Intel Is Pitching AI Infrastructure From a Long Way Back Intel’s newsroom post previews the AI Infra Summit, running September 15th through 17th in Santa Clara. CEO Lip-Bu Tan takes a fireside chat on Tuesday morning, with three Intel sessions across the show. Cochrane unpacks two terms first. Physical AI means AI that acts in the real world through sensors and motors rather than living on a screen, and Intel takes it seriously enough to have renamed its PC division the Client Computing and Physical AI Group in May. Disaggregated inference splits the two phases of running a model: reading your prompt is compute-hungry, while writing the answer back is memory-hungry. The context is where this gets interesting. Intel’s revenue rose 25 percent last quarter, its fastest growth since 2011. Nevertheless, in AI accelerators the company barely registers next to NVIDIA. Gaudi has effectively been abandoned, to the point that Intel stopped maintaining its open-source driver, and AMD passed Intel in data center revenue last quarter. Tellingly, when Intel demoed disaggregated inference at Computex, NVIDIA GPUs handled one phase and SambaNova chips the other while Intel supplied the coordinating CPU. Crescent Island, Intel’s actual inference chip, does not sample until later this year, and Intel declined to publish its memory bandwidth. Amazon Bought the Company Behind DuckDB On August 26th, Amazon signed a deal to acquire DuckLabs, the Amsterdam company behind DuckDB. Cochrane spends time explaining what DuckDB is, since listeners outside the data world may never have encountered it. The problem it solves is familiar. Querying a large pile of data files traditionally meant either running a database server or spinning up a data warehouse with a cluster, a bill, and a loading pipeline. Both are heavy machinery for a question you wanted answered in ten seconds. DuckDB instead ships as a library rather than a server. You add it to your program like any other package, point it at your files, and write ordinary SQL directly against them. The data never moves. The usual comparison is SQLite, which is embedded in nearly every phone and browser on earth. Where SQLite excels at looking up one record, DuckDB rebuilds that embedded idea for chewing through millions of rows. It now sees roughly 62 million monthly downloads on Python’s package index alone, up from about 25 million last October. AWS says it is buying the company, not the project. DuckDB stays free and open source under the MIT license, held by a Dutch nonprofit foundation, and the founders join AWS while continuing to run technical direction from Amsterdam. Andy Warfield, a VP and distinguished engineer at AWS, described DuckDB as “the glibc of structured data: a lean, unglamorous, ubiquitous dependency that a great deal of software links against and almost nobody has to think about.” Cochrane sits with what that ownership means. He reaches for an analogy: imagine Daniel Stenberg selling curl. He doubts it would ever happen, but the concept alone is startling given how much infrastructure depends on it. His read is that AWS is betting DuckDB becomes as foundational as curl and SQLite already are. GitHub Has One AI Model Grade Another One’s Homework GitHub shipped Project HydraFusion into Copilot as a research preview. Instead of routing your request to a single model, it picks one of three approaches per request. Sometimes one model simply answers. Alternatively, a cheaper model drafts, and a quality gate decides whether to escalate. The interesting one is Critique. One model writes the code, a separate model from a different family reviews it read-only, and the original gets one pass to revise. Despite the name, nothing is fused here. There is no voting and no merging, just one model at a time with a gate deciding whether to spend more. GitHub explained the reasoning in an earlier post: “a model reviewing its own work is still bounded by its own training biases: the same training data and techniques, the same blind spots.” Research supports it. A team at NeurIPS in 2024 showed that models recognize their own writing and score it higher than human graders do. GitHub’s earlier number had a Claude Sonnet and GPT critic pairing closing about three-quarters of the gap between Sonnet and the larger Opus model. This mirrors a workflow Cochrane uses constantly and has described on a previous episode. He runs a cross-check review with a second model from a different company, and it routinely surfaces issues the first model missed. He explains that different training data, different engineers, and different reinforcement approaches build different internal biases about what counts as correct. Looking ahead, he expects more of these “Frankenstein patterns” where models from different training families work together. The Best of IFA 2026, and What You Can Actually Buy IFA opened to the public in Berlin for its 102nd year, with about 1,900 brands. Cochrane splits The Verge’s roundup in two, since much of what generates headlines at these shows never ships. Starting with real products, iRobot’s flagship Roomba Max 875 Combo runs $1,199 and ships in about two weeks. Its SealForce feature drops a hidden skirt from the chassis when it detects carpet, sealing against the fibers so suction concentrates instead of leaking out the sides. That reaches 35,000 pascals, iRobot’s strongest yet. A step-down model at $899 carries the same trick, though Cochrane balks at both prices. Anker’s Soundcore Sleep 4 Pro earbuds arrive in November at $349.99. The charging case carries its own round touchscreen, so you pick soundscapes, set alarms, and read sleep stats without your phone. Optical sensors read heart rate and variability from the ear canal, which beats the wrist for accuracy, and the case masks a snoring partner. Cochrane remains unconvinced about sleeping with earbuds in. Philips also has smart rope lights, the Hue Liane 360, on sale now. They glow evenly around the tube rather than showing individual LEDs. They also run $400 for three meters, which works out to about $130 per meter of rope light. As for concepts nobody can buy, iRobot showed a robot vacuum that carries a smaller robot vacuum on its back in a garage and lowers it to deploy. Lenovo brought a 14-inch laptop whose screen rolls out to 17 inches at the press of a button, which reviewers call the first rollable that feels close to shippable. Tecno showed a phone with essentially no border around the screen, and Acer had a Windows gaming handheld that swivels its screen up over a keyboard. Two themes ran through the show. Humanoid robots were the loudest thing on the floor, and IFA’s own CEO framed the event as being about robots that work rather than robots that demo. Meanwhile, AI stopped being its own product category and became an ingredient, showing up in refrigerators, treadmills, dishwashers, and motorized TV mounts. Farmed Salmon Isn’t the Omega-3 Machine It Used to Be USDA scientists measured farmed salmon and found considerably less of the good fat than the government’s own database claims. EPA and DHA are two fatty acids you get almost entirely from fish. Your body can build them from the plant version, but only in tiny amounts, so the NIH’s position is that eating them is the only practical way to raise your levels. Those fatty acids are structural pieces of every cell, with DHA concentrating in the brain and retina. That is why the federal dietary guidelines, issued jointly by USDA and Health and Human Services, recommend at least eight ounces of fish a week and steer you toward salmon. That amount is calibrated to deliver about 250 milligrams a day. Published in Frontiers in Nutrition last month, the study found EPA and DHA in farmed Atlantic salmon came in 54.7 percent lower than USDA’s own reference values, last updated in 2018. A three-ounce serving fell from roughly 1,670 milligrams to about 756. Consequently, two servings a week now fall about 14 percent short of the target. Plant-derived fats meanwhile rose two to three times over. The likely cause is feed. Salmon are carnivores, and farms once fed them oily little fish. There was never going to be enough of those as the industry scaled, so crops filled the gap: soy, canola, sunflower and linseed. Importantly, the study does not claim to have proven this and calls the feed shift a plausible explanation. Independent corroboration lends it credibility. Researchers at Stirling measured a similar halving in Scottish salmon between 2006 and 2015, and Norway’s marine institute saw it across thousands of samples. There is a land dimension too. Roughly half the world’s soy grows in South America, where rainforest gets cleared for feed. Matthew Hayek, who studies the environmental cost of protein at NYU, told Inside Climate News that “soy is a major, important protein and oil ingredient in fish farming.” Adding up two decades of soy across all fish farming, he puts the extra forest clearing at around the area of Nicaragua or Bangladesh. That figure covers all fish farming rather than salmon alone, and Hayek notes it is hard to attribute soy use to any single species. Cochrane closes with two caveats. First, the study measured only eight fish, bought around Maryland, DC and Virginia over six weeks in 2023, and nearly all sourced from Chile. That is not a national survey, and the authors say plainly the sample was not large enough to change government advice. Rather, it flags that a federal database value needs rechecking, which is what the paper set out to do. Second, on whether you should care, farmed salmon still beats beef, chicken and eggs by a mile, since those carry essentially zero EPA and DHA. What it loses is its crown among fatty fish, dropping to mid-pack behind herring, sardines and mackerel and roughly level with trout. The broader health case is also softer than the 2000s suggested. A review of 86 trials covering 162,000 people found supplements barely moved heart attacks or deaths, so eating fish and swallowing fish oil are not the same claim. No producer has responded to the findings, and USDA, whose own scientists ran the study, declined an interview and did not answer emailed questions. Cochrane wraps up with housekeeping and a note that he will be back on Labor Day. The post The Mainframe Learned to Speak Arm #1875 appeared first on Geek News Central.

In this episode, Ray Cochrane digs into Anthropic’s Model Hardware Standard. It is a shared driver that lets an AI agent run real lab equipment, from pipetting robots to the lasers inside a quantum computer. He also covers OpenAI’s builder’s guide to GPT-5.6, Google’s new Expert Intelligence book feature, Apple’s M5 Ultra Mac Studio, and a judge’s order forcing Google to stop hiding rival app stores. Finally, he weighs in on Apple’s proposed 15 percent link-out fee, Meta’s Australia numbers, the White House deputizing private hackers, and why rivers obey a 1957 math rule. – 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. He is hunting for tickets to Michigan for his dad’s anniversary, and he has been learning Blender and Godot on the side, mostly modeling and blocking out levels. Consequently, he asks listeners for advice on starting a big game project, and he plans to record his progress, maybe as a time lapse. Then it is straight into the featured story. Anthropic’s Model Hardware Standard: A Driver for the Physical World The featured story comes from Anthropic, which opened a research preview of the Model Hardware Standard, or MHS. Cochrane frames it as the other side of the question NVIDIA’s world models raised two weeks ago: when do AI agents start touching actual machines? A typical lab runs a microscope, a liquid handler, a robotic arm, and a plate reader, each from a different vendor with its own control software. One Janelia researcher in the post launches seven programs in three languages just to start an experiment. Anthropic says wiring a setup like that takes weeks or months of specialist work. MHS is a driver, the same kind of translation layer a printer uses, except every device gets described with a tiny set of commands like read and write. Devices announce themselves on the network. A plain-English reference file then records what each machine measures, what can be adjusted, and which safety limits get enforced no matter what the agent asks. Agents then reach the hardware through the Model Context Protocol, the command line, or plain code. Cochrane sees the same move the industry keeps making, from coding harnesses to RSS and JSON: agree on a standard and let everyone build against it. In fact, he calls MHS the hardware version of MCP. The partner results carry the segment. QuEra builds quantum computers from individual atoms held by lasers that must hold their frequency to about one part in a trillion. A four-person team spent months on a relock script that worked 58 percent of the time. However, four copies of Claude iterating overnight through MHS produced a decision-tree script that recovers the laser in about six seconds, and it passed 99.3 percent of 700 blind trials. Carnegie Mellon wrote MHS drivers for four instruments across three incompatible computers in about eight hours, then ran dose-response experiments three times faster and blocked all six deliberately induced faults. Genentech, meanwhile, showed the limits. Claude used the same pump speed for water, a foamy protein solution, and a human had to explain that the bubbles were a physics problem. That gap in physical intuition is what sticks with Cochrane. He doubts it will change soon, and he suspects the fix will arrive as sub-agents or sub-models that judge a request against an expected outcome. He also connects MHS to a video of racing robots that never learned to stop at the finish line. What happens, he wonders, once they can read a distance sensor through a shared standard? Still, he calls the announcement a fantastic read and points listeners to the full article. 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 Does the Same Work for a Fraction of the Cost OpenAI’s builder’s guide to GPT-5.6 leads the headlines. Cochrane recaps the three tiers from episode 1870, Sol, Terra, and Luna, plus the separate dial for reasoning effort. On BrowseComp, a benchmark for digging up obscure facts on the web, the old GPT-5.5 flagship scored about 84 percent on a run that cost 33 dollars three months ago. Luna now matches that score for a dollar thirty-three, and OpenAI has since cut Luna’s price another 80 percent. Browser Use reports Luna finishing 78 percent of its hardest browser tasks for about 14 dollars, against 80 percent for roughly 235 dollars from the best available model. The guide’s other big addition is a multi-agent beta flag. It lets the model handling a request spawn parallel helper agents that report back to a root agent inside a single API call. However, Cochrane is unimpressed by the timing. He has been running that pattern in Claude Code for months, so he sees OpenAI copying a workflow other companies already ship rather than inventing its own. Along the way, he plugs Claude Code’s remote-control sessions, which let him send prompts from his phone to a terminal session at home. Google Lets Gemini Read the Books You Actually Bought Google launched Expert Intelligence, a name Cochrane calls quite the reach. The feature lets you drop a book you bought on Google Play Books into Gemini Notebook, formerly NotebookLM, and ask questions answered only from that book, with citations. Cochrane sees real power here for students, since he once used NotebookLM to organize scattered course PDFs. Additionally, publishers get a cut, which he calls a far better deal than the wholesale scraping of books that trained earlier models. Nevertheless, he asks who loses out, because a paid publisher does not automatically mean a paid author. He floats the same idea for artists, even a penny per use, then admits that may be too idealistic. Apple’s M5 Ultra Mac Studio Is Built to Run Big Models at Home Back in episode 1861, when Apple killed the Mac Pro, an M5 Ultra Mac Studio was expected later this year. Now it is here. The M5 Ultra brings up to a 36-core CPU, an 80-core GPU, and 512GB of unified memory moving 1.2 terabytes per second. Apple claims up to 4.3 times the AI performance of the M3 Ultra. Thunderbolt 5 can also cluster four machines into one memory pool for up to three times faster inference. The M5 Max model starts at $2,499 and the Ultra at $5,499, with shipping on September 22 and the 512GB configuration arriving in late October. Cochrane finds the clustering pitch ridiculous at that price, but he invites anyone who spends the money to report back. Apple Opens a Manufacturing School in Houston Apple also opened a 20,000-square-foot Advanced Manufacturing Center in Houston. It offers free classes for small and midsize manufacturers, from circuit board design to hands-on time on a scaled-down production line, with college students joining later. Cochrane calls it a solid step in the bring-manufacturing-home movement. The bigger story is the campus itself, which builds Apple’s AI servers and will add the first US-assembled Mac mini line later this year. That ties back to the Mac mini shortage that followed the OpenClaw rush, when Tim Cook warned of months-long waits. Cult of Mac was still reporting four-month waits in late July. However, Cook blamed chip supply rather than assembly, so Cochrane is not counting on relief just yet. Amazon EC2 Turns Twenty Amazon EC2 turned twenty this week, which Cochrane admits makes him feel old. The 2006 beta offered one server size in one region for ten cents an hour. Each came with a 1.7 gigahertz Xeon and under two gigabytes of memory, and accounts were capped at twenty servers. Today AWS offers more than 1,200 instance types across 39 regions. Consequently, Cochrane credits the company with turning that tiny product into the backbone of cloud and AI computing. Intel Gamer Days: Two Free Games, With Fine Print Intel Gamer Days runs through September 13. Buy a qualifying Core Ultra Series 2 or 14th Gen desktop chip, a Core Ultra Series 3 laptop, or an Arc graphics card. In return you get Star Wars: Galactic Racer plus the Tomb Raider: Legacy of Atlantis remake. GamesRadar values the pair at about 120 dollars. However, neither game is out yet, and codes must be redeemed by October 31 even though the Tomb Raider remake ships in February. Cochrane calls that awful, but he still tells qualifying buyers to claim the deal early. Note that 13th Gen chips do not qualify. Judge Orders Google to Stop Hiding Rival App Stores A jury found Google’s Android app monopoly illegal in late 2023, and Judge James Donato ordered rival stores into the Play Store in 2024. On August 13, Epic’s lawyer demonstrated that searching Play for “store for apps” returned Walmart instead of any app store. Donato called that “not acceptable” and ordered three fixes within a week. Searches must surface third-party stores, listings need a plain install button, and the “are you looking for” interstitial has to go. Cochrane welcomes the monopoly being chipped away, but he notes that a controlling entity still sits atop every app store. In his view, community hubs like app stores and social media need a public infrastructure layer. He suspects governments skip that investment because companies already run the services, while selling your data. Apple Wants 15 Percent of Purchases Outside Its Store The other half of the Epic saga is Apple’s proposed link-out commission. After the 2021 anti-steering injunction, Apple charged 27 percent on purchases made through external links. A judge held it in contempt last year, and the Ninth Circuit then allowed a fee limited to the cost of running the system. Judge Yvonne Gonzalez Rogers refused to wait for the Supreme Court, writing that “further delay is unwarranted.” Apple filed 15 percent for standard apps, 10 percent for subscription renewals and partner programs, and 5 percent for small businesses. It also conceded the rate would be “essentially zero” under the appeals court’s cost yardstick. Since Apple has charged nothing on link-outs since the contempt ruling, Cochrane sees this as a raise. He calls a cut on purchases made on a developer’s own website disturbing. He also recalls reading about the size of Uber’s payments to Apple, and he questions whether that kind of percentage is sustainable for companies without funding. Meta Says It Has Cut Off 750,000 Australian Kids Meta reported locking out more than 750,000 Facebook and Instagram accounts in Australia by the end of June under the country’s under-16 social media law. Over 500,000 of those were removed before the law even took effect. Detection relies mostly on AI scanning posts and bios for tells like birthday messages, plus user reports and blocks on re-registration. However, the post gives no count of mistaken removals or appeals, and the regulator’s early data shows under-16 usage falling only from about 86 to 81 percent. Meta wants a single age signal at the operating system or app store level, and Cochrane agrees completely. He connects it to the MHS idea from the top of the show: platforms need a standard flag to reference instead of guessing. The White House Deputizes Private Hackers Earlier this month the White House signed a National Security Presidential Memorandum that lets vetted private security firms run surveillance and disruption operations against overseas criminal groups. The Justice Department and Homeland Security hold the contracts and oversee the work. Firms need a proven track record, vetted staff, and a bond of at least $1 million, and must submit operating procedures within 60 days. Cochrane finds the measure aggressive in a good way and hopes it deters attacks on innocents. Still, he takes Kevin Beaumont’s warning seriously that the private security industry profits from ransomware existing. He compares it to the old Head and Shoulders myth: why solve the problem that drives your revenue? A Weather Satellite Watched the Eclipse Shadow Cross Europe Cochrane skips the readout on this one and simply sends listeners to ESA’s site. The MTG-I1 weather satellite captured the Moon’s shadow sweeping across Europe during the August 12 eclipse. Watching a shadow cross an entire continent, he says, was a first for him. Additionally, it leaves him excited about the research happening beyond the planet. Rivers, Deltas, and the Number 0.6 Quanta Magazine explains Hack’s law, which John Hack discovered in 1957 while measuring streams in Virginia and Maryland. A stream’s length tracks its drainage area raised to the power of 0.6, regardless of the rock underneath, and satellite data later confirmed it worldwide. Computer models in the 1990s showed why. Channels that capture extra runoff cut deeper and steal from their neighbors until the network settles into the arrangement that wastes the least energy. Now a University of Texas Rio Grande Valley team has found the same 0.6 exponent in river deltas, which spread water out rather than gathering it. Nobody knows why yet, and Cochrane calls it a really cool read. Sugar Helped Grow the Human Brain, Too A new paper in Science, co-authored by Jennie Brand-Miller at the University of Sydney, adds a third ingredient to the story of early human brain growth. Alongside meat and cooking, natural sugars from ripe fruit and honey may have fueled it too. The brain is about two percent of body weight but burns twenty percent of resting energy. It runs on glucose, which meat and marrow barely supply and raw starch cannot release without fire. The team modeled ancestral diets from a chimp-like baseline through Homo erectus and concluded that the earliest hominins may have drawn over 65 percent of their energy from natural sugars. Cochrane stresses that it is a model, not fossils, and notes that paleoanthropologist Marina Lozano thinks the authors place widespread cooking too early. Still, he loves this kind of deep research. Retracing the steps to our own intelligence, he suggests, could hint at what it takes for intelligent life to develop at all. A Brain Rhythm That Tells Doctors Where to Aim Finally, Science Daily covered a University of Cologne study on deep brain stimulation. That is the implanted-electrode treatment that eases Parkinson’s tremors for some patients but not others. Andreas Horn’s team recorded from 50 patients using both the implanted electrodes and an external magnetic scanner. They identified a circuit between the electrode’s target and the frontal cortex that oscillates at 20 to 35 cycles per second. Stronger coupling there predicted bigger improvement after surgery, though the study, published in Brain, shows correlation rather than cause. First author Bahne Bahners hopes the finding helps tune DBS more precisely, especially for patients who have not responded well. Cochrane half-jokingly asks whether MHS might one day drive those electrodes, and he calls brain disorders the hardest thing in the body to treat. Cochrane wraps with housekeeping: become a GNC Insider at geeknewscentral.com/insider, email [email protected] with questions or comments, subscribe to the newsletter, and grab a modern podcast app at podcastapps.com. He thanks GoDaddy for over twenty years of keeping the show on the air, promises to catch everyone next Monday, and wishes listeners a great night. The post Eyes, Hands, and a Sense of Timing #1874 appeared first on Geek News Central.