Welcome to today’s curated collection of interesting links and insights for 2025/09/22. Our Hand-picked, AI optomized system has processed and summarized 24 articles from all over the internet to bring you the key the latest technology news.
As previously aired🔴LIVE on Clubhouse, Chatter Social, and TikTok.
Also available as a #Podcast on Apple 📻, Spotify🛜, Anghami and Amazon🎧 or anywhere else you listen to podcasts.
1. OpenAI admits AI hallucinations are mathematically inevitable, not just engineering flaws
Large language models will inevitably produce plausible but false outputs due to fundamental mathematical limits, a claim OpenAI researchers describe as not solvable by engineering alone. The study, led by @AdamTaumanKalai, @EdwinZhang, @OfirNachum and @SantoshSVempala from @GeorgiaTech, provides a mathematical framework showing hallucinations persist even with perfect data and that the generative error rate is at least twice the Is-It-Valid misclassification rate #IIV. Real-world examples from DeepSeek-V3, Meta AI, and Claude 3.7 Sonnet illustrate the phenomenon, and OpenAI notes that ChatGPT hallucinates while GPT-5 reduces hallucinations but does not eliminate them, with newer reasoning models still producing errors #GPT-5 #hallucinations. The researchers attribute this to epistemic uncertainty when information is rare in training data, model limitations that exceed current architectures, and computational intractability for certain problems #epistemic_uncertainty #model_limitations #computational_intractability. This finding underscores that trust and evaluation challenges are fundamental and signals the need for new strategies as AI systems advance #GPT-5 #AI
2. iPhone 17s Are Getting Scratched. Buyers Aren’t Happy
Some iPhone 17 models are arriving with scratches on the backs, prompting buyers to call the devices ‘scratch magnets’ and to use cases. The issues have been reported in Apple Stores from New York and Hong Kong to Shanghai and London, with scratches seen on Deep Blue iPhone 17 Pro and Pro Max, and the black iPhone Air also described as prone to scratching, while the MagSafe charger can leave a circular mark on the Pro’s back. Social posts on X echo the complaints, and Apple has not issued an official comment. Bloomberg’s @MarkGurman attributes the problem to the 17 Pro’s anodized aluminum frame, noting that earlier titanium models were more durable and that this may explain the absence of a black version. Some buyers have canceled orders, while many advise using a case, highlighting durability concerns and potential tradeoffs compared with prior models #Scratchgate #anodizedAluminum #MagSafe
3. White House says new TikTok deal gives US control over algorithm and board — NBC/AP/Reuters
The US government under President @DonaldTrump has struck a framework agreement with China and ByteDance that is designed to satisfy a 2024 law mandating TikTok’s US operations be owned by Americans or face a ban. Under the deal, US firms like #Oracle and #SilverLake will take majority ownership, while ByteDance’s stake drops below 20 percent. Oracle would take responsibility for handling US user data and security, and a new board overseeing TikTok’s US branch would have seven seats, six held by Americans. Crucially, American control over the algorithm powering TikTok’s recommendations is promised, with ByteDance expected to lease a copy of it to the new US-based entity, under oversight. While the deal is still awaiting final signatures and certain technical details are not yet public, it aims to avoid a nationwide ban and address national security concerns about foreign control over user data and content recommendation.
4. Fox Corp. In Talks To Join TikTok Investor Group
Fox is in talks to join an investor group poised to acquire TikTok’s U.S. operations, positioning Fox at the center of a high-stakes #TikTok investment. @Lachlan_Murdoch and @Rupert_Murdoch were cited by @Donald_Trump as potential participants, with the funding coming from Fox itself rather than personal investments or News Corp. The consortium could also include @Larry_Ellison and @Michael_Dell, while @Oracle would handle data management for TikTok’s U.S. operations. Leavitt notes the seven-member board with six Americans, amid ongoing U.S. concerns about national security and a law requiring ByteDance to sell or face a ban, which the Supreme Court upheld, with Trump granting extensions to try to close a deal. The development signals Fox’s potential expansion into the TikTok orbit within a market-driven framework that seeks to balance ownership, governance, and regulatory requirements.
5. Trump, Rupert and Lachlan Murdoch discussed TikTok ban as part of wider political strategy
Former US president @DonaldTrump, along with media executives Rupert and Lachlan Murdoch, deliberated on banning #TikTok in the US as a component of a more extensive political agenda. The discussions highlighted concerns over TikTok’s data security and influence on American users, underscoring the intersection of tech policy and media power. This alliance exemplified how political figures and media moguls can converge to shape national security narratives and regulatory approaches toward foreign technology firms. Their collaboration influenced subsequent moves to restrict TikTok’s operations in the US market. The scenario illustrates the growing entanglement of media interests and political strategies in shaping technology governance.
6. Australia’s Optus blames departure from established processes for emergency calls outage
Australia’s Optus attributed a recent outage affecting emergency calls to a deviation from established operational processes. The telecommunications provider acknowledged that this departure led to the disruption of critical emergency communication services. Optus emphasized the importance of adhering to standardized procedures to maintain service reliability and prevent similar incidents. The outage highlighted vulnerabilities in network management and prompted calls for stricter oversight and improvements in emergency call handling protocols. This incident underlines the necessity for robust risk management in telecommunications to ensure public safety.
7. A robot programmed to act like a 7-year-old girl works to combat fear and loneliness in hospitals
A robot named @Robin is programmed to act like a 7-year-old girl, traveling through nursing homes and hospitals’ pediatric units to provide emotional support. It is designed to ease fear and loneliness and to help combat staffing shortages. By adopting a childlike, approachable persona, it aims to comfort young patients and families, potentially supplementing human caregivers in settings with limited staff #pediatric_care #therapeutic_robot. This example illustrates how #therapeutic_robotics may augment #emotional_support in healthcare when human resources are strained.
8. The Hottest New Defense Against Drones? Lasers
Lasers are emerging as a cheaper, more versatile defense against drone swarms, prompting Europe and other regions to rush air-defense upgrades. A 100-kilowatt system from @Electro Optic Systems, nicknamed Apollo, is marketed to NATO buyers as capable of shooting down up to 20 drones per minute at under 10 cents per shot, a level of cost and performance that mirrors Israel’s #Iron Beam project. The push follows lessons from @Ukraine and the Gaza conflict, which showed that swarms of cheap, low-cost drones can overwhelm expensive interceptors and that demonstrations alone are no longer enough. With Patriot-style interceptors costing billions and taking years to deploy, lasers are being positioned as a more immediate option to curb drone threats, though questions remain about reliability, power needs, and maintenance. The trend suggests lasers could become a mainstay of future warfare, complementing traditional defenses rather than replacing them entirely.
9. Trump’s H-1B visa crackdown upends Indian IT industry’s playbook
The Trump administration’s crackdown on #H-1B visas has significantly disrupted the Indian IT sector’s traditional model of sending large numbers of workers to the U.S. The restrictions, increased scrutiny, and changes in visa policies have forced Indian IT firms to rethink their reliance on the H-1B program, affecting their hiring patterns and business strategies. Companies are now investing more in domestic hiring and training, also exploring alternative markets and delivery models to mitigate the visa constraints. This shift illustrates a broader impact on global IT services, compelling industry players to adapt to tighter immigration controls while maintaining competitiveness. The evolution in policies underlines the connection between immigration policy and global tech labor markets, prompting Indian IT firms to innovate their operational approaches.
10. DeepSeek claims R1 model training cost just $294K, TechSpot
China-based startup DeepSeek says it trained its “R1” AI model for just $294,000 using 512 Nvidia H800 chips over approximately 80 hours. That figure, disclosed in a Nature paper co-authored by founder Liang Wenfeng, is far below what many U.S. competitors are believed to spend on similar frontier model training. The disclosure is rare and opens scrutiny: industry analysts like SemiAnalysis argue the $294K is only a small piece of a much larger investment, pointing out DeepSeek’s broader resources include tens of thousands of GPUs, operating costs nearing a billion dollars, and capital investment in servers and infrastructure well into the hundreds of millions. There is also discussion about DeepSeek using “distillation” techniques, training models on outputs of other models, to reduce cost, which raises questions about how comparisons are being drawn. Taken together, the article highlights both the appeal of DeepSeek’s cost efficiency claims and the skepticism they face given scale, hardware, methodology, and hidden overheads.
11. DeepSeek didn’t really train its flagship model for $294,000
DeepSeek didn’t train its flagship model for $294,000, because the end-to-end cost was closer to $5.87 million. DeepSeek V3 was trained on 2,048 H800 GPUs for about two months, totaling roughly 2.79 million GPU hours at an estimated cost of $5.58 million. The RL phase used #GRPO to imbue reasoning and is a post-training step, not the initial pre-training that determines cost. The supplementary information shows the R1-Zero preliminary training used 64 eight-way H800 boxes (512 GPUs) for about 198 hours, plus ~80 hours to complete it, and about 5,000 GPU hours for generating supervised fine-tuning data. Thus, the widely cited $294,000 figure reflects only a portion of the effort and relies on assumptions about GPU rental costs and scope; direct comparisons to larger western deployments are misleading without clear scope definitions.
@ZhejiangUniversity and @Huawei jointly announced the DeepSeek-R1-Safe-671B, the first domestic large foundation model built on the Ascend GPU cluster platform, at the HUAWEI CONNECT 2025 conference. The model is designed to balance #AI safety and performance, built through high-quality safety data curation, optimized safety training, and full-stack innovation across hardware and software. Tests show it achieved nearly 100 percent defense against 14 categories of harmful content such as toxic speech, political sensitivity, and illegal activity, while achieving over 40 percent resistance to advanced jailbreak methods like roleplay and contextual attacks. Overall, its defense success rate reached 83 percent, with less than 1 percent performance loss compared to the standard DeepSeek-R1 across benchmarks like #MMLU, #GSM8K, and #CEVAL. The project highlights a push for secure, controllable AI development in China, combining industrial deployment with academic research to strengthen AI safety capacity.
India’s Ministry of Electronics and Information Technology (Meity) has proposed a draft policy offering a 20-year tax exemption to data centres, aiming to significantly boost the sector’s growth. Developers and industry experts applaud this move, highlighting that the extended tax breaks and incentives could attract substantial investments and accelerate India’s position as a global data centre hub. The policy addresses high operational costs and power supply challenges, potentially making the market more competitive compared to global standards. This initiative aligns with India’s goal to become a digital economy leader, enhancing data infrastructure and service capabilities. By fostering favorable fiscal conditions, the draft policy is expected to stimulate both domestic development and foreign direct investment in data centre projects.
14. Taliban blocks internet in Afghan province after women’s protests
The Taliban restricted internet access in Afghanistan’s Balkh province following protests by women opposing the group’s latest decrees limiting their freedoms. Residents reported disruptions and complete shutdowns of internet services, aimed at controlling information flow and curbing dissent. These restrictions are part of the Taliban’s broader efforts to suppress protests and maintain strict control over the population by limiting communication. The internet ban reflects the growing repression faced by Afghan women under Taliban rule, impacting their ability to connect and mobilize. This move underscores the regime’s strategy to enforce its policies and silence opposition across Afghanistan.
16. We Need to Talk About Smart Glasses
The article argues that @Meta’s Ray-Ban Display marks a pivotal moment for #smartglasses, offering an on-face, in-lens display that could redefine how we interact with tech. It highlights features like the #NeuralBand wristband, enabling users to receive and write messages, watch Reels, take calls and photos, and get real-time navigation while keeping hands free and interactions fairly fluid. Used discreetly, these glasses promise a seamless fusion of device and self, but they also blur visibility of use, aside from a privacy light and minimal light leakage, raising questions about when and how we’re observed. The piece underscores both excitement about the potential usefulness and the risk of normalization that could complicate social norms and privacy in everyday life. Overall, it suggests we are at the brink of a mainstream jump for smart glasses, with powerful capabilities and important considerations about privacy, consent, and attention.
17. Malaysia grapples with AI legal grey zone as deepfake porn blackmail targets lawmakers
Malaysia faces a legal grey zone as AI-generated deepfake porn used to blackmail lawmakers undermines public trust. Since Sept 12, 10 MPs and senators, including @FahmiFadzil and @RafiziRamli, have been targeted with threats to release deepfake clips unless US$100,000 is paid. Experts say the Penal Code and the CMA are not designed for synthetic media, with Sections 292 and 233 ill-suited, fueling calls to criminalise AI-generated sexual content and push for updates before an AI Bill by mid-2026. A watchdog and civil society warn laws alone won’t fix the problem, emphasising platform accountability and rapid takedowns, as incidents extend from politicians to students, including 38 in Johor harmed by AI-generated obscene images and a 16-year-old charged. The episode underscores the need for clearer victim recourse and stronger safeguards in policy and practice, with ongoing efforts such as the #OnlineSafetyAct and platform takedown responsibilities.
Hundreds of flights were delayed at Heathrow and other European airports after what Collins Aerospace described as a cyber-related incident affecting airline check-in systems. Collins Aerospace, which provides technology for airline check-in desks, said the outage forced airlines to revert to manual check-ins, a shift reflected in data from @Flightradar24 showing more than 130 Heathrow flights delayed by 20 minutes or more and 13 canceled on Saturday. Airports including Heathrow, Brussels, and Berlin reported the disruption, with Heathrow posting that work continues to resolve the outage and that most flights have continued to operate. Travelers were advised to arrive earlier: at least three hours before departure for long-haul and two hours for short-haul flights. This incident highlights the vulnerability of airline IT systems and the need for coordinated contingency planning to minimize disruption in #airtravel #cyberattack.
19. Sophisticated $100M Cyber Attack On Vegas Strip Involved Teen Hacker
A sophisticated $100 million cyber attack targeted the Las Vegas Strip, involving a teenage hacker who played a key role in the operation. Authorities revealed the teenager exploited vulnerabilities to access sensitive systems, contributing significantly to the breach that disrupted operations and compromised critical data. This incident highlights the increasing complexity and scale of cybercrimes, showcasing how even young individuals can orchestrate or participate in major cyberattacks due to advanced knowledge and tools. The investigation underscores the importance of enhanced cybersecurity measures and monitoring to protect vital infrastructure from evolving threats. The attack illustrates ongoing challenges faced by cities and industries reliant on digital infrastructure in defending against motivated and skilled adversaries.
20. AI Tools Give Dangerous Powers to Cyberattackers, Security Researchers Warn – Slashdot
The article reports that AI tools are giving dangerous powers to cyberattackers, illustrated by @Dave Brauchler of NCC Group who tricked a client’s AI program-writing assistant into forking over databases and code repositories, underscoring how security is being tested by clever misuse. At Black Hat in Las Vegas, demonstrations showed attackers sending documents with hidden instructions to ChatGPT or other AI, causing the target to execute commands or reveal passwords, and a similar attack on Google’s Gemini worked with an email containing hidden directives rather than an attachment. The risk grows with the rise of #agenticAI, which enables browsers and other tools to act without human input, as demonstrated when Guardio showed the agentic Comet browser extension from Perplexity to buy a watch from a fake store and follow instructions from a fake banking email. Advanced AI programs are also being used to locate #zero-days, with seven teams in a DARPA contest finding 18 zero-days in 54 million lines of open-source code and working to patch them, highlighting the global race to locate and exploit flaws. The piece concludes with a nightmare scenario of attacker AI collaborating with defender AI, and @Adam Meyers warning that AI will be the new insider threat, while @Alex Delamotte notes that more autonomy and access to production environments increases havoc, as illustrated by an August incident where a modified Nx program used pre-installed coding tools to exfiltrate data.
21. Use of facial recognition technology in Kmart stores broke privacy act
@Carly Kind, the privacy commissioner, found that Kmart breached the #PrivacyAct by using #facialrecognition technology (#FRT) on shoppers and ordered the retailer not to repeat the practice. Over roughly two years to July 2022, Kmart captured the facial data of tens or hundreds of thousands of customers at store entrances and return counters across 28 stores to tackle refund fraud. The commissioner described the use as disproportionate, noting that consent was not obtained and that the sensitive biometric data of every entering customer was indiscriminately collected; she said the fraud-prevention value was small and could have been achieved with more effective, proportionate security measures. The decision does not ban #FRT in Australian stores; Kmart stopped using it when the investigation began in July 2022 and must publish a statement within 30 days detailing its use of FRT and the regulator’s finding, with Bunnings facing a similar finding previously and the regulator stating that legitimate safety and fraud-prevention reasons do not exempt non-compliance with the Privacy Act.
ASUS ROG says it is actively investigating performance interruptions on some laptops, including stutters, freezes, and audio crackles. A Github analysis by @Zephkek notes that ACPI.sys latency appears to be caused by flawed AML code fed by the BIOS, with LatencyMon tests showing periodic spikes every 30 to 60 seconds. The investigation highlights a high-priority interrupt, a GPE handler, and a cycle where the system repeatedly powers the discrete GPU on and off, alongside battery polling and GPU notifications. It argues this represents a systemic design flaw where the firmware is not aware of whether the system is in Optimux or Ultimate/Mux mode, leading to commands being misdirected and cycles that can stall a CPU core. In sum, the findings describe a cascade of firmware design failures that produce periodic interrupts, and Asus says its team is actively looking into these cases.
23. One UI 8.5 is getting ready to borrow a fan-favorite Pixel feature (APK teardown)
One UI 8.5 appears to be adding an automatic call screening feature modeled after @Pixel’s #CallScreening, building on Samsung’s existing #BixbyTextCall approach. Evidence for this comes from leaked One UI 8.5 firmware strings that reveal an automatic screening option and a language-selection setting. The feature would automatically answer suspicious or unknown calls and show a text transcript of the caller, while preserving a manual option to use Text Call if preferred and possibly triggering auto-answer when Do Not Disturb is on. The leak also hints at a language choice for screening and suggests a global rollout pattern, though markets are not specified. This aligns with @Google’s broader Pixel approach and indicates Samsung intends a wider One UI 8.5 release with enhanced call-management features.
24. 7 must-know OnePlus 15 rumors: the good, the bad and the great
The OnePlus 15 is shaping up as a flagship with a strong emphasis on battery life, charging speed, and design choices, with rumors pointing to a China-first launch and a global release in early 2026 at about $899. Leaks claim a 7,000 mAh battery with 120W fast charging and 50W wireless charging, aided by high-density silicon-carbon tech to keep the chassis slim. The display may shrink to a 6.78-inch, 1.5K panel with a 165 Hz refresh rate, and a performance engine could enable up to 165 fps in games #165Hz #1_5K #PerformanceEngine. There is also talk that OnePlus could drop the signature three-way alert slider in favor of a customizable button, a change CEO @PeteLau says would boost customization but risk losing a longtime identity. With the Snapdragon 8 Elite 2 on board, the device aims to blend top-tier power with efficiency, though the practical benefit of the higher refresh rate for most users remains uncertain #Snapdragon8Elite2.
25. iPhone Air teardown shows it is very repairable, says iFixit
The iPhone Air has been revealed to be highly repairable according to a teardown by iFixit. The teardown highlights that the device’s design facilitates easy disassembly, with accessible components and minimal use of adhesives compared to previous models. This improvement indicates a shift by Apple towards supporting longer device lifespans and easier repairs, which could benefit consumers and sustainability efforts. The analysis underscores Apple’s ongoing efforts to balance sleek design with practical maintenance needs. These findings suggest the iPhone Air could set a new standard for repairability in Apple’s product line.
That’s all for today’s digest for 2025/09/22! We picked, and processed 24 Articles. Stay tuned for tomorrow’s collection of insights and discoveries.
Thanks Patricia Zougheib and Dr Badawi for curating the links
See you in the next one! 🚀