Stay ahead with the latest tech breakthroughs, industry trends, and cutting-edge news on our Innovation Feed channel.
https://samirb.com.np/
🔍 Samsung to Upgrade Bixby with Perplexity for Galaxy S26

Samsung is preparing a major update for Bixby ahead of next year's Galaxy S26 launch. While Galaxy AI is powerful, the Bixby assistant has lagged, currently leaving complex questions to Google Gemini while it handles only basic tasks.

Reports suggest Samsung will partner with Perplexity to solve this. Under the new plan, Bixby would still manage system settings. At the same time, Perplexity would handle complex, web‑based queries with citations, similar to Apple's integration with ChatGPT.

Read More
Meta acquires AI startup Manus to boost commercial revenue

Meta Platforms agreed to acquire Singapore-based AI startup Manus, which sells autonomous AI agents to small and medium-sized businesses, as part of CEO Mark Zuckerberg's push to build commercial products around the company's massive AI investments.

Manus reached $125 million in annual revenue run rate just eight months after launching in March 2025, making it the fastest startup to hit $100 million in annual recurring revenue, with the platform processing over 147 trillion tokens since launch.

The acquisition follows Meta's December purchase of AI wearables maker Limitless and comes as the company commits at least $70 billion in 2025 capital expenditures for AI infrastructure, with projections exceeding $100 billion next year.
AI boom drives RAM shortage, threatening PC sales and console launches

IDC warned in December that PC prices could rise 4% to 8% in 2026 as a global memory shortage driven by AI data center demand threatens to contract the PC market by up to 8.9%, with Dell and Lenovo signaling price increases of up to 15%.

The crisis stems from memory makers prioritizing high-bandwidth memory production for AI infrastructure over consumer chips, with OpenAI securing agreements in October 2025 for Samsung and SK Hynix to supply up to 900,000 DRAM wafers monthly—representing over 40% of projected global capacity.

Gaming console makers are debating whether to delay the PlayStation 6 and next Xbox from their planned 2027-2028 launch window due to tight memory supplies, with industry analysts indicating relief may not arrive until late 2026 or 2027 at the earliest.
ByteDance plans $5.6B Huawei chip buy after Nvidia cutoff

ByteDance is planning to purchase more than RMB 40 billion ($5.6 billion) worth of Huawei's Ascend AI chips in 2026, a sharp increase from nearly zero domestic chip procurement this year, with the first batch valued at tens of billions of yuan expected to enter delivery soon.

The massive order is driven by the computing power gap that emerged after Nvidia's H20 chip supply was halted in April 2025 due to U.S. export restrictions, and by explosive growth in token usage across ByteDance's Volcano Engine cloud platform and Doubao AI application, which processed over 30 trillion tokens daily by September.

The deal signals China's accelerating shift toward semiconductor self-sufficiency, with ByteDance deepening cooperation with Huawei to secure domestic computing power for its rapidly expanding AI services that command a 49.2% market share for large model services on China's public cloud.

Read detailed
Visa and Mastercard build payment systems for AI agents

Visa and Mastercard have launched competing payment frameworks to enable AI shopping agents to autonomously search, compare, and complete purchases on behalf of consumers, with Visa completing hundreds of pilot transactions and predicting agentic commerce will become mainstream in 2026.

Both payment giants unveiled their protocols in October 2025, with Visa's Trusted Agent Protocol using cryptographic signatures developed with Cloudflare and Mastercard's Agent Pay framework completing U.S. rollout in November, ahead of planned global expansion in early 2026.

The infrastructure push follows OpenAI's September launch of Instant Checkout in ChatGPT, which allows purchases from Etsy and Shopify merchants without leaving the chat interface, though questions remain about liability when AI agents make purchasing errors.
CD Projekt sells GOG storefront to co-founder

CD Projekt sold its DRM-free PC storefront GOG to co-founder Michał Kiciński for approximately $25.2 million, ending the platform's 17-year tenure within the CD Projekt Group.

The divestment allows CD Projekt to focus on developing The Witcher 4, Cyberpunk 2, and a new intellectual property, with GOG continuing to operate independently under Kiciński's ownership while maintaining its DRM-free philosophy.

The companies signed a distribution agreement ensuring future CD Projekt Red titles will launch on GOG, with Kiciński pledging to preserve classic games and support titles with a retro spirit.

Read detailed
AI chatbots discriminate against non-standard dialect speakers, studies find

AI chatbots including ChatGPT and Meta's Llama systematically discriminates against non-standard dialect speakers, showing 19% more stereotyping and 25% more demeaning content compared to standard English, according to multiple 2024-2025 studies from UC Berkeley, Cornell, and German universities.

Real-world harms include Amazon's Rufus assistants giving incorrect responses to African American English speakers, ChatGPT changing an Indian job applicant's Dalit-associated surname to a higher-caste name, and a UK council AI failing to understand a Derbyshire dialect.

Researchers warn that as AI becomes embedded in education, hiring, and government services, these biases may amplify existing discrimination, though developers say specialized models for dialect speakers and system corrections could address the problem.

Read detailed
TSMC to raise chip prices annually through 2029

TSMC has notified customers of consecutive annual price hikes for advanced chips starting January 1, 2026, with increases of 3-10% expected through 2029 as production capacity for sub-3nm nodes remains severely constrained by AI demand.

The world's largest contract chipmaker's shares hit a record high of NT$1,530 on December 29, 2025, as investors responded to reports that tight supply and strong bookings from Nvidia, Apple, AMD, and Broadcom are driving pricing power.

TSMC's chairman stated in November that the company's advanced-node capacity falls roughly three times short of customer demand, with analysts expecting the price increases to bolster the chipmaker's traditionally weak first quarter when it holds its investor conference on January 15, 2026.

Read detailed
China deploys humanoid robots at Vietnam border crossing
Published


China has deployed UBTECH's Walker S2 humanoid robots at a Vietnam border crossing in Fangchenggang to guide travelers, manage crowds, patrol corridors, and inspect cargo under a $37 million contract that began deliveries in December 2025.

The Walker S2 stands 1.76 meters tall with 52 degrees of freedom and can autonomously swap its own batteries in about three minutes, enabling 24-hour continuous operation at the high-traffic border checkpoint.

The deployment represents a strategic test of humanoid robots in regulated public infrastructure, with Chinese officials positioning humanoid robotics as a key industry and planning to scale production to 10,000 units by 2027.
AI startups raise record $150B in 2025 amid bubble fears

AI startups raised a record $150 billion in 2025, surpassing the 2021 peak of $92 billion, with major deals including OpenAI's $41 billion funding led by SoftBank and Anthropic's $13 billion September round, according to PitchBook data reported by the Financial Times.

Venture capitalists are urging founders to build "fortress balance sheets" ahead of potential market turbulence in 2026, as leading AI firms like OpenAI grow at unprecedented rates while smaller startups face dwindling funding options amid investor caution.

The funding surge has depleted cash reserves at major venture capital firms including Thrive Capital, Andreessen Horowitz, and Tiger Global, which have begun raising new funds, while some analysts warn that concentrated infrastructure spending by companies like Microsoft and Meta poses valuation risks.

Read detailed
China launches world's first 30MW pure hydrogen turbine

China launched Jupiter I, the world's first 30-megawatt-class pure hydrogen gas turbine, which began stable power generation in Inner Mongolia on Sunday, operating entirely on hydrogen and marking a breakthrough in renewable energy storage technology.

The turbine addresses renewable energy waste by converting excess wind and solar electricity into hydrogen during off-peak hours and back to power during peak demand, with the project integrating a 500MW wind farm and photovoltaic energy systems.

Jupiter I can reduce carbon emissions by over 200,000 tonnes annually compared to thermal power units and generates enough electricity per hour to meet the daily needs of 5,500 households, according to officials at Mingyang Hydrogen Gas Turbine Technology.

Read detailed
MongoDB flaw under active exploit, 87,000 servers exposed

MongoDB disclosed CVE-2025-14847, dubbed MongoBleed, a high-severity vulnerability allowing unauthenticated attackers to extract sensitive data from server memory, with active exploitation confirmed shortly after a public exploit became available on December 26, 2025.

The flaw affects a broad range of MongoDB versions from legacy 3.6 through 8.2.2 due to improper handling of zlib-based network message decompression, with approximately 87,000 instances exposed worldwide and 42% of cloud environments hosting at least one vulnerable instance, according to Censys and Wiz.

MongoDB has released patches for all affected versions and recommends immediate upgrades or disabling zlib compression as a temporary workaround, while unverified reports link the vulnerability to a breach of Ubisoft's Rainbow Six Siege servers that forced the game offline.

Read detailed
SK Telecom unveils Korea's first 519B-parameter AI model

SK Telecom unveiled A.X K1, Korea's first hyperscale AI model with 519 billion parameters, on December 27 as part of the Korean government's Sovereign AI Foundation Model project aimed at positioning the nation among the top three global AI powers alongside the U.S. and China.

The model functions as a "Teacher Model" that transfers knowledge to smaller AI models rather than merely consuming information, and will be integrated into SK Telecom's A-Dot service with over 10 million subscribers and Liner's platform serving more than 11 million global users to advance an "AI for Everyone" framework.

The eight-member consortium—including SK hynix, Krafton, Rebellions, and Seoul National University—built the full-stack AI ecosystem using proprietary Korean technologies and plans to release A.X K1 as open-source to boost the nation's AI competitiveness.

Read detailed
Global RAM shortage to persist until 2027, SK Hynix warns

DRAM prices have surged nearly 300% in three months, with 16Gb DDR5 chips rising from $6.84 in September to $27.20 by December 2025 as memory manufacturers prioritize high-bandwidth chips for AI data centers over conventional RAM for consumer electronics.

Apple is concentrating its iPhone 17 memory orders with Samsung, which will supply 60% to 70% of low-power DRAM compared to a previously balanced split with SK Hynix, as 12GB modules now cost $70 versus $30 at the beginning of 2025.

Industry analysts warn the shortage could persist until 2027 or 2028 when new fabrication plants come online, with SK Hynix telling analysts that memory deficits may last until late 2027 as chipmakers struggle to balance AI infrastructure demand against traditional electronics needs.

Read detailed
Brin owns up to Glass mistakes while Google plans AI smart glasses

Google announced on December 8 that it will launch AI-powered smart glasses in 2026 through partnerships with Warby Parker, Samsung, and Gentle Monster, marking its return to the market a decade after discontinuing Google Glass.

The new glasses will feature Gemini AI assistant and come in two versions: screen-free audio glasses with speakers, microphones, and cameras, and display glasses with an in-lens screen for navigation and translation, with the audio version launching first.

Google co-founder Sergey Brin told Stanford students in late December he "jumped the gun" with the original Google Glass launch, admitting he thought he was "the next Steve Jobs" and commercialized the product too quickly before making it cost-effective and polished.
🏗 90.7 km: The Length of the Most Powerful Collider in History

In 2025, CERN confirmed the feasibility and technical viability of the Future Circular Collider (FCC), the successor to the Large Hadron Collider.

The 90.7 km ring would run beneath France and Switzerland at an average depth of 200 meters. The planned collision energy would reach up to 100 TeV, at least 6 times higher than the LHC. The main goals are ultra-precise studies of the Higgs boson and searches for physics beyond the Standard Model.

The first phase is estimated at about $17 billion. For the first time in CERN's history, the project has attracted private capital. A consortium that includes funds linked to DST Global founder Yuri Milner and former Google CEO Eric Schmidt has pledged $1 billion.

CERN's member states are expected to make a final decision around 2028. The FCC would not start operations before 2040.
OpenAI seeks safety chief for $555K amid exodus

OpenAI CEO Sam Altman announced on December 27 that the company is hiring a Head of Preparedness to address emerging AI risks including mental health impacts and cybersecurity vulnerabilities as models rapidly advance.

The role will oversee OpenAI's preparedness framework across biological, cybersecurity, and AI self-improvement threats, coming after the company revealed over one million weekly ChatGPT users send messages indicating suicidal intent.

The hiring follows significant safety team departures and lawsuits against OpenAI linking chatbots to teenage suicides, with former researchers criticizing the company for prioritizing product development over safety.

Read detailed
Wired says Alibaba's Qwen will dominate AI in 2026


Wired declared that 2026 will be dominated by Alibaba's Qwen AI model rather than OpenAI's GPT-5, following disappointing August 2025 releases of American models that suffered from basic errors and failed to meet expectations.

Chinese AI model downloads overtook U.S. models on HuggingFace in July 2025, with Qwen becoming the second-most-used open model globally and gaining adoption from Airbnb, Nvidia, and even Meta for training new models.

Qwen's rise stems from its open-weight architecture allowing easy customization, transparent research practices that earned a Best Paper Award at NeurIPS 2025, and real-world deployment in applications from smart glasses to electric vehicle dashboards, according to the article.
China releases draft rules to regulate human-like AI systems

China's cyberspace regulator released draft rules on December 27 requiring providers of human-like AI systems to monitor users for addiction and emotional dependency, with public consultation open until January 25.

The regulations mandate that users be notified they're interacting with AI at login and every two hours or when overdependence is detected, and require providers to intervene when users exhibit extreme emotions or addictive behavior.

The draft bans content endangering national security or spreading misinformation, and requires security assessments for services launching human-like AI features or reaching one million registered users.

Read detailed
AI agents fail to meet 2025 workforce predictions

OpenAI CEO Sam Altman predicted in early 2025 that AI agents would join the workforce and materially change company output, but the technology has fallen dramatically short by year's end according to industry insiders.

Andrej Karpathy, an OpenAI co-founder, said in October that AI agents are "cognitively lacking and it's just not working," estimating it will take about a decade to resolve current issues with intelligence, multimodal capabilities, and memory retention.

A Deloitte survey found only 11% of organizations currently use AI agents in production, with real-world testing revealing even basic tasks like clicking elements or selecting dropdown menus can take minutes or fail entirely.

Read more
Back to Top