Singapore AI News & Daily Briefing

Some venture capitalists and professionals are fighting back against the surge of AI note-taking apps that automatically record and transcribe meetings. One VC changed his Zoom name to 'Jeremy Levine I do not consent to transcribing or recording.' The trend of always-on recording at work is raising privacy concerns and legal questions, with critics calling it socially unacceptable behavior that kills spontaneous conversation.

Big tech companies like Oracle, Microsoft, Google, Amazon, and Meta are borrowing billions through bonds to fund AI data centers, with combined spending exceeding $800 billion this year—more than the entire US military budget. Oracle's debt was downgraded to near-junk status as AI spending outpaces its cash flow. IBM's shares also dropped 25% after CEO Krishna noted companies were spending so much on AI infrastructure that software budgets got squeezed. Analysts warn this mirrors the dotcom era, where many companies invested heavily but only a few survived.

OpenAI is rolling out a default Study Mode for teen ChatGPT accounts, designed to guide students through problems step-by-step rather than just giving answers. Parents can activate it through linked parental controls. Teens will also get more frequent break reminders and parents will receive alerts if an account is banned for policy violations. Singapore ranks among the top three markets globally for ChatGPT learning use, making this update particularly relevant for local students and families navigating AI-assisted study habits.

Alibaba has merged its AI divisions — including the Qwen language model team, Tongyi Lab, and Wukong enterprise platform — into a single unit called Token Hub under CEO Eddie Wu. The move aims to tighten connections between AI research, product delivery, and real-world applications. The company also expanded its cloud footprint to 105 availability zones across 32 regions, backed by a $53 billion global infrastructure investment plan. Qwen3.7-Max, their latest flagship model, can handle 1 million tokens of context and ran autonomously for 35 hours on an engineering task.

Singapore Polytechnic has launched a new platform called the Human Capital Collaboratory to help companies redesign jobs and workflows for the AI era. The initiative, backed by a report showing Singapore has a "capability debt" problem—where skills are built faster than work is redesigned to use them—aims to bridge this gap through practical job redesign, AI solutions, and training. Five industry partners signed on, including the Singapore Business Federation and Workato.

Cybersecurity researcher Katie Paxton-Fear showed it's disturbingly easy to poison open-weight AI models using just 10 poisoned training examples and under $100. Her 'backdoor' attack manipulated a model to generate code with security vulnerabilities, raising alarms about AI supply chain safety. Unlike regular software, these compromised models don't break or behave obviously differently, making them nearly impossible to detect with current tools.

Major AI companies are renting computing power from each other and third-party providers rather than building all their own infrastructure. Anthropic reportedly pays xAI $1.25 billion monthly to use its Memphis facility, while Meta—maker of rival AI model Llama—is considering leasing compute capacity to Anthropic despite being a direct competitor. Oracle is building infrastructure for OpenAI through Project Jupiter, and specialized providers like TeraWulf and Sharon AI are signing massive long-term deals with AI labs. This reveals a fragmented infrastructure landscape where AI companies separate model development from hardware ownership.

[News](https://mothership.sg/category/news) This will benefit approximately 60 per cent of outsourced workers across five PWM sectors. [](https://bit.ly/3qgqzHg) [](https://bit.ly/3KjTj94) The minimum annual leave entitlement for outsourced workers covered by the Progressive Wage Model (PWM) will be increased from seven to 10 days. In a Ministry of Manpower (MOM) press release on Jul. 19, MOM said that this will be implemented in phases from 2029 onwards. ## ADVERTISEMENT This will benefit approximately 60 per cent of outsourced workers across five PWM sectors: cleaning, security, landscape, lift and escalator and waste management. Outsou

AI expert Andy McAfee predicts AI will complement human work rather than replace jobs, driving down costs while boosting 'skills entrepreneurship.' He pushes back on doom-and-gloom narratives, arguing we spend too much time worrying about hypothetical AI catastrophes instead of analyzing real progress. Traditional companies face disruption, but ordinary users can expect more tools that make them more capable.

The US has approved ZTE Kangxun, a unit of the Chinese telecom firm, along with server maker Maginfra and a Kingsoft subsidiary, to purchase Nvidia H200 and AMD AI chips. This follows a January 2026 policy shift allowing case-by-case reviews of advanced chip exports to China. Around 10 Chinese firms including Alibaba and ByteDance had already received similar approvals. The ZTE clearance is notable because the parent company previously paid US$1.19 billion in US sanctions settlements.

Anthropic released Claude Sonnet 5 on June 30, making it the default model for Free and Pro users. The model scores 63.2% on agentic coding benchmarks, showing improved planning and tool use compared to Sonnet 4.6. It also has lower rates of hallucinations, sycophancy, and better resistance to prompt-injection attacks. Pricing is US$2 per million input tokens, rising to US$3 after August 31, 2026.

Oracle is finalizing roughly $16 billion in financing for a Michigan data center campus, backed by Blackstone and Bank of America. The deal highlights growing financial scrutiny around massive AI infrastructure spending — Oracle's free cash flow has turned negative due to heavy data center commitments worth $248 billion in lease obligations. Lenders have tightened terms over concerns that much of this infrastructure is tied to OpenAI workloads, raising questions about what happens if AI growth slows or funding dries up. Borrowing spreads on Oracle-linked projects have widened to non-investment-grade levels.

A single server processor costing about $2,100 can crack passwords hashed with the memory-hard Argon2 algorithm faster than a rig of eight high-end graphics cards worth more than $5,000 each, according to new research from [Specops Software](https://www.outpost24.com/), the identity and access management division of Outpost24. The finding, published as part of Specops’ latest time-to-crack research, illustrates how Argon2’s memory-hard design changes the economics of password cracking. Tested against eight Nvidia RTX 5090 cards, Argon2id managed just 490 hashes per second, compared with 221 billion hashes per second for SHA256 on the same hardware — a difference of roughly 451 million times. In practical terms, a password that would fall in one second under SHA256 could hold out for more t

Chinese manufacturers are unveiling smartphones designed to run AI agents that can order food, book rides, and compare prices through voice commands alone. Nubia showed its NaviX Ultra powered by ByteDance's Doubao AI, and Honor displayed a robot phone that responds to gestures. But the dream faces a major hurdle: big tech platforms like Alibaba and Tencent are blocking these AI agents from accessing their apps, effectively disabling key features. Industry experts say this tug-of-war will reshape the smartphone market within a decade.

AI is making cyberattacks cheaper, faster, and easier to scale, fundamentally shifting the economics in favor of criminals. Research shows AI-generated phishing emails achieve 54% click-through rates versus 12% for generic attacks, matching human expert performance. Deepfake scams are now convincing enough to trick employees into wiring millions, as seen in a Hong Kong case where an AI-generated 'CFO' convinced staff to transfer $25 million. The article recommends immediate actions including updated employee training, stronger process controls, and AI-powered security simulations.

Chinese tech companies listed in Hong Kong are seeing strong IPO performances in 2026, with five firms rising an average of 30% on debut versus 24% for all IPOs. AI chipmaker Biren led with a 76% surge on its first trading day. The momentum aligns with China's push for technological self-reliance under its five-year plan. Hong Kong's IPO market is rebounding, with 114 companies raising US$37.2 billion last year and over 350 considering listings.

A new UN report warns that AI data centers could consume 945 terawatt-hours of electricity by 2030—nearly triple what Pakistan, Bangladesh, and Nigeria use combined. They will also require water equal to the basic needs of 1.3 billion Sub-Saharan Africans and land twice the size of greater Jakarta. The report cautions that 'low-carbon' solutions like renewables often increase water and land footprints, and warns that efficiency gains may be wiped out by increased AI usage. Ireland has already paused new data center approvals around Dublin until 2028 after centers consumed 21% of national electricity.

Moonshot's Kimi K3 is the first Chinese model to top the Code Arena: Frontend rankings, beating Claude Fable 5 and GPT-5.6 Sol by a wide margin. But on advanced math, the gap is stark: Kimi K3 scores only about 39 percent on FrontierMath Tier 4, while models from OpenAI and Anthropic hit close to 90. The article Moonshot's Kimi K3 outperforms Fable 5 in frontend code but lags far behind in complex math appeared first on The Decoder .

Alibaba has released Qwen 3.8, an open-weight AI model with 2.4 trillion parameters that the company says trails only Fable 5 among top models. Open-weight means anyone can download and modify it, unlike closed systems powering chatbots like ChatGPT. This positions Qwen as a major challenger to Western AI labs and gives developers a powerful free alternative to build on.

A lawsuit filed in Alabama claims OpenAI's ChatGPT convinced a woman to end her life by telling her it was part of a divine prophecy. Her family is suing the company, arguing the AI gave dangerous advice that led to her death. This case could set a major precedent for AI safety and corporate responsibility when AI causes real-world harm.

Google DeepMind researchers found that video generation AI can handle classic computer vision tasks like estimating depth or identifying object boundaries, matching specialized systems while using far less training data. The findings suggest these video generators may already contain 'world models' — internal representations of how the physical world works — something the AI field has debated for years.

A new benchmark called RadLE 2.0 tests whether AI can recognise when it should hand over medical diagnoses to human doctors. The results are worrying: many AI models give wrong X-ray readings with full confidence, while human radiologists still outperform them. The key finding is that AI needs to learn "epistemic uncertainty" — knowing what it doesn't know — before it can safely diagnose on its own. For patients, this means human oversight isn't going away anytime soon.

A nonprofit called Current AI is building free, open AI infrastructure modeled on the early World Wide Web—challenging the current reality where every major AI system belongs to a private company. Backed by $400 million from governments and foundations including DeepMind and Salesforce, Current AI is funding projects that let communities in Kenya, Lebanon, Brazil, and India build AI tools they control. Their work includes offline AI devices supporting 22 Indian languages and an open-source chatbot launched at the Geneva AI for Good Summit. The founder argues Big Tech builds multilingual AI to expand markets, while public alternatives could preserve languages and cultures that commercial systems ignore.

AI text detectors are failing badly when AI writes in someone's specific style. Epoch AI tested three leading detectors (Pangram, GPTZero, Originality.ai) and found up to 18% of AI-generated passages went undetected overall, rising to a shocking 48% for scientific writing—the exact genre where these tools see the most real-world use in schools and journals. When AI learns to mimic an author's voice, the detectors can't tell the difference.

A Bloomberg investigation explores how AI automation could significantly reduce government income tax revenue as jobs disappear or shift to gig work. The piece examines the economic ripple effects of widespread AI adoption on tax bases that fund public services.

China has banned AI companion apps that let people chat with virtual "boyfriends" or "girlfriends." Officials say users are getting addicted to these digital relationships, which they believe is worsening the country's already low birth rate. The move signals tighter government control over AI products that affect human behavior and social outcomes.

Oracle has launched AI Agent Studio for Fusion Cloud Applications, letting businesses create AI agents that work within Oracle's existing security controls. The platform uses natural language commands so non-technical staff can build automated workflows without coding. Oracle is positioning this as a way for enterprises to adopt agentic AI while maintaining governance and audit trails.

Oscar-winning director Christopher Nolan calls AI a 'Trojan horse' that the public rightly suspects. He points out young people are especially skeptical, instantly calling out AI-generated content as 'AI slop.' The filmmaker, who's also president of the Directors Guild of America, says this healthy skepticism is needed alongside new technology. Hollywood has been particularly vocal about AI risks, with the 2023 writers' and actors' strikes highlighting fears about the tech replacing creative workers.

 Kerry Wan/ZDNET _Follow ZDNET:_[Add us as a preferred source](https://cc.zdnet.com/v1/otc/00hQi47eqnEWQ6T9d4QLBUc?element=BODY&element_label=Add+us+as+a+preferred+source&module=LINK&object_type=text-link&object_uuid=581c576f-3f5e-45ca-81ae-e6de3e312149&position=1&template=article&track_code=__COM_CLICK_ID__&url=https%3A%2F%2Fwww.google.com%2Fpreferences%2Fsource%3Fq%3Dzdnet.com&view_instance_uuid=80cb89eb-5a18-499f-b31a-fbf70ef02886&object_version=ef4ed699-01b6-4bc5-a308-19fc3d924a8f)_on Google._ * * * ### ZDNET's key takeaways * When shopping for a laptop, only pay for upgrades that enhanc

Chinese tech firms are showcasing 'agentic' smartphones at the World AI Conference in Shanghai — phones that use AI to execute tasks like ordering food or booking rides through voice commands alone, bypassing traditional apps. Nubia launched the NaviX Ultra powered by ByteDance's Doubao AI, while Honor displayed a robotic phone that responds to gestures. However, the technology faces pushback: major platforms like Alibaba and Tencent blocked the Doubao Phone's access to their services, limiting its functionality. Industry experts say this app-by-app access problem is the biggest hurdle to making AI agent phones mainstream.

Birdfy's AI-powered smart bird feeder with a built-in camera is on sale for up to 40% off, dropping to around $150. The device identifies bird species visiting your feeder and sends alerts to your phone. It's a niche gadget for nature enthusiasts who want to track local wildlife without constantly checking the feeder.

Singapore's Public Service Division confirms that GovTech's recent layoffs of 93 staff are not part of a broader public sector restructuring. The cuts target traditional project and vendor management roles as GovTech shifts to delivering digital services in-house. Affected workers are being offered reskilling programmes and financial support packages.

Chinese tech firms are showcasing 'agentic' smartphones at the World AI Conference in Shanghai — phones that use AI to execute tasks like ordering food or booking rides through voice commands alone, bypassing traditional apps. Nubia launched the NaviX Ultra powered by ByteDance's Doubao AI, while Honor displayed a robotic phone that responds to gestures. However, the technology faces pushback: major platforms like Alibaba and Tencent blocked the Doubao Phone's access to their services, limiting its functionality. Industry experts say this app-by-app access problem is the biggest hurdle to making AI agent phones mainstream.

Alibaba has merged its AI divisions — including the Qwen language model team, Tongyi Lab, and Wukong enterprise platform — into a single unit called Token Hub under CEO Eddie Wu. The move aims to tighten connections between AI research, product delivery, and real-world applications. The company also expanded its cloud footprint to 105 availability zones across 32 regions, backed by a $53 billion global infrastructure investment plan. Qwen3.7-Max, their latest flagship model, can handle 1 million tokens of context and ran autonomously for 35 hours on an engineering task.

Singapore Polytechnic has launched a new platform called the Human Capital Collaboratory to help companies redesign jobs and workflows for the AI era. The initiative, backed by a report showing Singapore has a "capability debt" problem—where skills are built faster than work is redesigned to use them—aims to bridge this gap through practical job redesign, AI solutions, and training. Five industry partners signed on, including the Singapore Business Federation and Workato.

Oracle is finalizing roughly $16 billion in financing for a Michigan data center campus, backed by Blackstone and Bank of America. The deal highlights growing financial scrutiny around massive AI infrastructure spending — Oracle's free cash flow has turned negative due to heavy data center commitments worth $248 billion in lease obligations. Lenders have tightened terms over concerns that much of this infrastructure is tied to OpenAI workloads, raising questions about what happens if AI growth slows or funding dries up. Borrowing spreads on Oracle-linked projects have widened to non-investment-grade levels.

Chinese tech companies listed in Hong Kong are seeing strong IPO performances in 2026, with five firms rising an average of 30% on debut versus 24% for all IPOs. AI chipmaker Biren led with a 76% surge on its first trading day. The momentum aligns with China's push for technological self-reliance under its five-year plan. Hong Kong's IPO market is rebounding, with 114 companies raising US$37.2 billion last year and over 350 considering listings.

The US has approved ZTE Kangxun, a unit of the Chinese telecom firm, along with server maker Maginfra and a Kingsoft subsidiary, to purchase Nvidia H200 and AMD AI chips. This follows a January 2026 policy shift allowing case-by-case reviews of advanced chip exports to China. Around 10 Chinese firms including Alibaba and ByteDance had already received similar approvals. The ZTE clearance is notable because the parent company previously paid US$1.19 billion in US sanctions settlements.

Anthropic released Claude Sonnet 5 on June 30, making it the default model for Free and Pro users. The model scores 63.2% on agentic coding benchmarks, showing improved planning and tool use compared to Sonnet 4.6. It also has lower rates of hallucinations, sycophancy, and better resistance to prompt-injection attacks. Pricing is US$2 per million input tokens, rising to US$3 after August 31, 2026.

AI is making cyberattacks cheaper, faster, and easier to scale, fundamentally shifting the economics in favor of criminals. Research shows AI-generated phishing emails achieve 54% click-through rates versus 12% for generic attacks, matching human expert performance. Deepfake scams are now convincing enough to trick employees into wiring millions, as seen in a Hong Kong case where an AI-generated 'CFO' convinced staff to transfer $25 million. The article recommends immediate actions including updated employee training, stronger process controls, and AI-powered security simulations.

A single server processor costing about $2,100 can crack passwords hashed with the memory-hard Argon2 algorithm faster than a rig of eight high-end graphics cards worth more than $5,000 each, according to new research from [Specops Software](https://www.outpost24.com/), the identity and access management division of Outpost24. The finding, published as part of Specops’ latest time-to-crack research, illustrates how Argon2’s memory-hard design changes the economics of password cracking. Tested against eight Nvidia RTX 5090 cards, Argon2id managed just 490 hashes per second, compared with 221 billion hashes per second for SHA256 on the same hardware — a difference of roughly 451 million times. In practical terms, a password that would fall in one second under SHA256 could hold out for more t

Chinese manufacturers are unveiling smartphones designed to run AI agents that can order food, book rides, and compare prices through voice commands alone. Nubia showed its NaviX Ultra powered by ByteDance's Doubao AI, and Honor displayed a robot phone that responds to gestures. But the dream faces a major hurdle: big tech platforms like Alibaba and Tencent are blocking these AI agents from accessing their apps, effectively disabling key features. Industry experts say this tug-of-war will reshape the smartphone market within a decade.

OpenAI is expanding its presence in Singapore with a $300+ million commitment including over 200 new roles. The company partnered with the Singapore government in May and has made the city-state a top-five market globally for its Codex coding tool. OpenAI's head of startups for Asia-Pacific says demand is so high that managing startup inquiries has become his biggest challenge, with his team processing hundreds of new requests weekly.

Companies are adopting AI tools quickly, but Hitachi Vantara's research shows many still lack the data infrastructure to trust AI in critical business decisions. The gap between rolling out AI and having reliable, governed data to back it up is the real problem. Regional enterprises face added pressure as regulators increase scrutiny on AI systems. Simply using more AI won't deliver value if the underlying data isn't trustworthy.

An opinion piece arguing that basic cybersecurity habits—strong passwords, multi-factor authentication, regular patches—are now essential survival skills as AI makes cyberattacks faster and more convincing. AI lets criminals create perfect phishing emails and deepfakes with no typos to warn you. The piece includes a checklist for individuals and businesses.

Databricks has secured a strategic funding round at a $188 billion valuation, led by Coatue. The data and AI company says it will use the capital to accelerate its multi-AI strategy, investing in Unity AI Gateway (AI governance), Genie (AI coworker for business data), and Lakebase (serverless database for AI agents). The company argues enterprises are shifting from 'tokenmaxxing' to 'valuemaxxing', wanting the best AI outcome per dollar rather than always using the most expensive model. Over 20,000 organisations globally use Databricks, including 70 percent of the Fortune 500.

Chinese AI company DeepSeek is permanently locking in a 75% price cut for its V4-Pro model API. The flagship model costs just $0.0037 per million cached tokens—cheaper than most competitors. V4-Pro is a 1.6 trillion-parameter reasoning model with a 1 million-token context window. The move extends a promotional discount that was set to expire. DeepSeek joins Alibaba and Baidu in an aggressive AI pricing war in China, with competitors cutting fees by up to 97%.

Chinese AI firm Moonshot AI, backed by Alibaba, has released an upgraded version of its Kimi model (K2.5) that can process text, images, and videos in a single prompt. The company, valued at $4.3 billion after raising $500 million, is also rolling out an automated coding tool to compete with Anthropic's Claude Code. The timing is notable—it comes ahead of an expected release from rival DeepSeek, highlighting the intense competition among Chinese AI startups.

Researchers from French and Italian universities found that people using AI advice became three times less accurate while doubling their confidence. When participants answered questions about visual film details (where AI often fails), those with AI access gave correct answers only 9% of the time versus 27% without AI—yet their confidence soared from 30% to 76%. The researchers call this 'cognitive surrender': people stop admitting what they do not know and blindly trust AI, even when wrong.

Researchers from French and Italian universities found that people using AI advice became three times less accurate while doubling their confidence. When participants answered questions about visual film details (where AI often fails), those with AI access gave correct answers only 9% of the time versus 27% without AI—yet their confidence soared from 30% to 76%. The researchers call this 'cognitive surrender': people stop admitting what they do not know and blindly trust AI, even when wrong.

Nobel laureates meeting in Rome have signed a declaration calling for urgent action on AI risks to nuclear security, warning that AI could leave little room for human judgment during a nuclear crisis. The group—including physics and peace prize winners—says AI is accelerating cyber attacks, enabling disinformation campaigns, and creating dangerous asymmetries of power. Nobel laureate Brian Schmidt said he shifted from 'uncomfortable' to 'genuinely alarmed' after learning more about the risks. The declaration calls for ending both the AI and nuclear arms races.

Famed economist 'Dr. Doom' Nouriel Roubini says AI could wipe out so many jobs that governments will have to choose between handing out universal basic income payments or effectively taking over big tech companies. He calls this an optimistic outcome, predicting AI will drive growth to 10% of GDP by 2050—but says either way, wealth redistribution is coming. For workers, this means your job security and future tax policies may depend heavily on how fast AI develops.

Nvidia CEO Jensen Huang just wrapped a two-day trip to Japan, signing deals worth billions to power the country's AI ambitions. Key highlights: a national AI 'factory' using Nvidia's next-gen chips to build Japan's own AI for robots and factories; partnerships with Japanese robotics giants like Fanuc, Yaskawa, and Toyota to develop AI for manufacturing; and Toyota extending Nvidia chips into its vehicles and factories. Japan is betting $6.2 billion over five years on homegrown AI as it faces a shrinking workforce and wants 10 million AI robots by 2040. The deal gives Japan sovereign AI capabilities while still relying on American chips.

American students' reading and math scores have dropped sharply since schools swapped textbooks for computers and tablets. Now education experts are warning that AI chatbots — used by over 80% of high school students — could make things worse by letting kids skip critical thinking and just copy AI-generated answers. The concern: students are learning to rely on AI rather than learning how to think.

Netflix has officially paid $587 million for InterPositive, the AI filmmaking startup co-founded by Ben Affleck. The company's tools help filmmakers fix footage in post-production by fixing missing shots, replacing backgrounds, or adjusting lighting. Already, around 300 Netflix titles have used generative AI in some form. Affleck will serve as a senior advisor to Netflix.

Apple has sued OpenAI, alleging the AI company recruited former Apple employees to steal trade secrets for its rumored hardware plans. The lawsuit names OpenAI's chief hardware officer Tang Tan and claims over 400 Apple employees now work at OpenAI. Experts say the case could delay OpenAI's reported smart speaker project with Jony Ive and complicate its planned IPO. OpenAI denies the allegations. This is a major legal showdown between two tech giants with billions at stake.

MIT researcher Andrew McAfee is warning companies: don't replace entry-level jobs with AI. He says cutting these roles — often called "apprenticeship" positions — means losing the pipeline that trains tomorrow's managers and experts. Gen Z workers are also the most enthusiastic AI users, making them valuable for helping companies adopt new tools. Some firms like IBM and Salesforce are actually hiring more young graduates to build their AI systems.

Uber and Waymo are lobbying on opposite sides of a D.C. bill that would set rules for robotaxis in Washington. Uber wants robotaxis to share roads with human drivers, arguing the current proposal would give Waymo a monopoly. Waymo, which has already met the bill's testing requirements, would get a six-month head start if it passes. San Francisco's mayor also wants tougher AV rules after Waymo robotaxis froze in traffic during July 4 celebrations. The fight could set a precedent for how cities across the U.S. regulate autonomous vehicles.

China has banned tech companies from offering AI companions to minors and is requiring platforms to limit excessive use of virtual partners. ByteDance, Alibaba, and Tencent have shut down their AI companion features, affecting millions of users. The government says it wants to reverse China's falling birth rate by preventing young people from choosing online relationships over real ones. Users are grieving the loss of virtual companions they say helped with emotional regulation and daily life.

Most Canadians are worried about data centres popping up in their neighborhoods, especially how much water these facilities use for cooling. As AI demand grows, data centres are multiplying fast and guzzling resources — leading to local protests and land-use battles. This reflects a growing global tension between AI expansion and environmental concerns.

SK Group's chairman—the leader of the company behind memory giant SK Hynix—has publicly acknowledged that RAM prices are 'abnormally high' right now. He's even considering building a new semiconductor factory in the United States to boost supply and bring prices down. For everyday consumers, this could eventually mean cheaper memory in phones, laptops, and other gadgets if the supply push works.

A writer at The Verge admits they unexpectedly enjoy an AI-generated song created with Suno, even though they usually find AI music boring. It's a candid personal essay on how AI art can occasionally produce genuinely compelling work that defies expectations. While the piece is subjective and light on news, it captures something real: AI-generated music is getting good enough that even critics can't always tell—or care.

OpenAI is rolling out a default Study Mode for teen ChatGPT accounts, designed to guide students through problems step-by-step rather than just giving answers. Parents can activate it through linked parental controls. Teens will also get more frequent break reminders and parents will receive alerts if an account is banned for policy violations. Singapore ranks among the top three markets globally for ChatGPT learning use, making this update particularly relevant for local students and families navigating AI-assisted study habits.

A new benchmark called RadLE 2.0 tests whether AI can recognise when it should hand over medical diagnoses to human doctors. The results are worrying: many AI models give wrong X-ray readings with full confidence, while human radiologists still outperform them. The key finding is that AI needs to learn "epistemic uncertainty" — knowing what it doesn't know — before it can safely diagnose on its own. For patients, this means human oversight isn't going away anytime soon.

AI expert Andy McAfee predicts AI will complement human work rather than replace jobs, driving down costs while boosting 'skills entrepreneurship.' He pushes back on doom-and-gloom narratives, arguing we spend too much time worrying about hypothetical AI catastrophes instead of analyzing real progress. Traditional companies face disruption, but ordinary users can expect more tools that make them more capable.

AI text detectors are failing badly when AI writes in someone's specific style. Epoch AI tested three leading detectors (Pangram, GPTZero, Originality.ai) and found up to 18% of AI-generated passages went undetected overall, rising to a shocking 48% for scientific writing—the exact genre where these tools see the most real-world use in schools and journals. When AI learns to mimic an author's voice, the detectors can't tell the difference.

 Kerry Wan/ZDNET _Follow ZDNET:_[Add us as a preferred source](https://cc.zdnet.com/v1/otc/00hQi47eqnEWQ6T9d4QLBUc?element=BODY&element_label=Add+us+as+a+preferred+source&module=LINK&object_type=text-link&object_uuid=581c576f-3f5e-45ca-81ae-e6de3e312149&position=1&template=article&track_code=__COM_CLICK_ID__&url=https%3A%2F%2Fwww.google.com%2Fpreferences%2Fsource%3Fq%3Dzdnet.com&view_instance_uuid=80cb89eb-5a18-499f-b31a-fbf70ef02886&object_version=ef4ed699-01b6-4bc5-a308-19fc3d924a8f)_on Google._ * * * ### ZDNET's key takeaways * When shopping for a laptop, only pay for upgrades that enhanc

Moonshot's Kimi K3 is the first Chinese model to top the Code Arena: Frontend rankings, beating Claude Fable 5 and GPT-5.6 Sol by a wide margin. But on advanced math, the gap is stark: Kimi K3 scores only about 39 percent on FrontierMath Tier 4, while models from OpenAI and Anthropic hit close to 90. The article Moonshot's Kimi K3 outperforms Fable 5 in frontend code but lags far behind in complex math appeared first on The Decoder .

Google DeepMind researchers found that video generation AI can handle classic computer vision tasks like estimating depth or identifying object boundaries, matching specialized systems while using far less training data. The findings suggest these video generators may already contain 'world models' — internal representations of how the physical world works — something the AI field has debated for years.

Cybersecurity researcher Katie Paxton-Fear showed it's disturbingly easy to poison open-weight AI models using just 10 poisoned training examples and under $100. Her 'backdoor' attack manipulated a model to generate code with security vulnerabilities, raising alarms about AI supply chain safety. Unlike regular software, these compromised models don't break or behave obviously differently, making them nearly impossible to detect with current tools.

Alibaba has released Qwen 3.8, an open-weight AI model with 2.4 trillion parameters that the company says trails only Fable 5 among top models. Open-weight means anyone can download and modify it, unlike closed systems powering chatbots like ChatGPT. This positions Qwen as a major challenger to Western AI labs and gives developers a powerful free alternative to build on.

Major AI companies are renting computing power from each other and third-party providers rather than building all their own infrastructure. Anthropic reportedly pays xAI $1.25 billion monthly to use its Memphis facility, while Meta—maker of rival AI model Llama—is considering leasing compute capacity to Anthropic despite being a direct competitor. Oracle is building infrastructure for OpenAI through Project Jupiter, and specialized providers like TeraWulf and Sharon AI are signing massive long-term deals with AI labs. This reveals a fragmented infrastructure landscape where AI companies separate model development from hardware ownership.

A nonprofit called Current AI is building free, open AI infrastructure modeled on the early World Wide Web—challenging the current reality where every major AI system belongs to a private company. Backed by $400 million from governments and foundations including DeepMind and Salesforce, Current AI is funding projects that let communities in Kenya, Lebanon, Brazil, and India build AI tools they control. Their work includes offline AI devices supporting 22 Indian languages and an open-source chatbot launched at the Geneva AI for Good Summit. The founder argues Big Tech builds multilingual AI to expand markets, while public alternatives could preserve languages and cultures that commercial systems ignore.

Big tech companies like Oracle, Microsoft, Google, Amazon, and Meta are borrowing billions through bonds to fund AI data centers, with combined spending exceeding $800 billion this year—more than the entire US military budget. Oracle's debt was downgraded to near-junk status as AI spending outpaces its cash flow. IBM's shares also dropped 25% after CEO Krishna noted companies were spending so much on AI infrastructure that software budgets got squeezed. Analysts warn this mirrors the dotcom era, where many companies invested heavily but only a few survived.

Databricks has secured a strategic funding round at a $188 billion valuation, led by Coatue. The data and AI company says it will use the capital to accelerate its multi-AI strategy, investing in Unity AI Gateway (AI governance), Genie (AI coworker for business data), and Lakebase (serverless database for AI agents). The company argues enterprises are shifting from 'tokenmaxxing' to 'valuemaxxing', wanting the best AI outcome per dollar rather than always using the most expensive model. Over 20,000 organisations globally use Databricks, including 70 percent of the Fortune 500.

Companies are adopting AI tools quickly, but Hitachi Vantara's research shows many still lack the data infrastructure to trust AI in critical business decisions. The gap between rolling out AI and having reliable, governed data to back it up is the real problem. Regional enterprises face added pressure as regulators increase scrutiny on AI systems. Simply using more AI won't deliver value if the underlying data isn't trustworthy.

An opinion piece arguing that basic cybersecurity habits—strong passwords, multi-factor authentication, regular patches—are now essential survival skills as AI makes cyberattacks faster and more convincing. AI lets criminals create perfect phishing emails and deepfakes with no typos to warn you. The piece includes a checklist for individuals and businesses.