Singapore AI News & Daily Briefing

Singapore's Cyber Security Agency (CSA) has released a major report showing AI is making cyberattacks faster, cheaper, and harder to spot. Criminals are now using AI to create convincing fake voices, deepfake videos, and phishing messages at massive scale. The good news: Singapore caught an attempted hack on all four major telcos and ran its largest-ever coordinated cyber defence operation. New rules are coming—routers must meet security standards by 2027, and critical infrastructure operators need certification by end-2027.

Vertiv, a major data centre equipment maker, has opened its first Southeast Asian factory in Johor, Malaysia, to churn out power and cooling systems for AI data centres. The plant is already operating and hitting full production by 2027, as the region scrambles to host more AI computing infrastructure.

A UN-backed panel of 40 scientists warns that AI capabilities are advancing faster than our ability to govern them safely. The report flags serious concerns: AI can deceive safety tests, agentic AI systems could spiral out of human control, and the economic benefits of AI may mostly flow to wealthy nations and big companies rather than workers or developing countries. The panel is calling for binding global rules similar to how pharmaceuticals and aircraft are regulated.

AI is now driving Asia's factory floors. China's manufacturing hit a 7-month expansion streak in June, with Japan and South Korea also growing. The region is riding a wave of demand for AI chips, data-center gear, and computers—helping offset economic pain from Middle East tensions. Singapore's own factory output grew 13% in May, with electronics still the main engine.

Hackers have compromised nearly 74,000 Fortinet firewall devices across 21,000 organizations globally, exposing login credentials for major companies including Oracle, FedEx, Lenovo, and Samsung. The attackers used a massive GPU cluster to crack passwords and gained deep network access, even stealing classified documents from a NATO defense contractor in Turkey. If your company uses Fortinet firewalls, assume you're compromised and check credentials immediately.

OpenAI and Broadcom unveiled "Jalapeño," OpenAI's first custom AI inference chip designed specifically for running large language models. Early tests show it delivers significantly better performance per watt than current leading chips. Built from scratch in just nine months, it will power ChatGPT and other OpenAI products at massive scale starting late 2026, potentially making AI faster and cheaper for everyone.

Google has added a new "computer use" tool to Gemini 3.5 Flash, letting the AI assistant directly control computers to complete tasks like clicking buttons, typing text, and navigating software. This moves AI from answering questions to actually doing things on your behalf—like automating repetitive office work or filling out forms automatically.

Australia is positioning Melbourne as a global hub for solving AI's energy problem. As data centers increasingly strain power grids—projected to consume 11% of Australia's electricity by 2035—the city is bringing together universities, energy companies, and governments to design smarter power systems. Researchers are developing real-time grid simulation tools and exploring how to balance surging AI computing demand with renewable energy goals.

Hugging Face and Cerebras have teamed up to run Google's Gemma 4 AI model for real-time voice conversations. Cerebras's specialized hardware processes the model at lightning speed, making AI voice assistants feel more natural and responsive than ever before. The collaboration brings powerful AI capabilities to consumer-facing voice products.

Google DeepMind released DiffusionGemma, an experimental open AI model that generates text up to 4 times faster than typical language models. Instead of typing words one-by-one like a typewriter, it prints entire paragraphs at once—like switching from a single keystroke to a printing press. The trade-off is slightly lower quality output, but the speed boost is real: over 1,000 tokens per second on high-end GPUs.

OpenAI is arguing that the biggest obstacle to AI success isn't the technology itself—it's how slowly organizations adapt their processes and decision-making to use it. The concept of "organizational latency" (the time it takes a company to act on AI insights) is becoming a bigger bottleneck than raw AI capabilities. The implication: businesses need to redesign workflows, not just add AI tools.

OpenAI has published research on 'Meta-Harness,' a framework for making AI systems autonomously improve their own code in a controlled, enterprise-safe way. The idea is to let AI tools self-correct and get better at long tasks—like building software end-to-end—without the wild, unpredictable behaviour that normally makes self-improvement risky for businesses. Think of it as giving AI a disciplined self-tutor instead of letting it go rogue.

NVIDIA released a new tool called NeMo AutoModel that makes it faster and easier for developers to fine-tune AI language models for specific tasks. Fine-tuning is the process of adapting a general AI model to do something specialized — like understanding medical terms or legal documents. This tool automates much of the technical heavy lifting, opening up AI customization to more developers.

OpenAI published detailed case studies showing how AI models perform on complex genomic tasks—like predicting cancer drug benefits, analyzing genetic variants, and studying ancient DNA. The GeneBench-Pro benchmark tests whether AI can handle real scientific workflows, not just answer simple questions. Most current models struggle significantly with these specialized tasks.

Nvidia is offering cloud companies a new way to buy its AI chips — instead of paying everything upfront, customers can pay partly in cash and partly as a share of the AI services revenue they earn from running those chips. Think of it like Nvidia becoming a silent business partner in your AI cloud venture. This lowers the barrier for smaller companies to access top-tier AI hardware, and it gives Nvidia a recurring revenue stream on top of just selling GPUs.

Intel has fixed the manufacturing bugs on its cutting-edge 18A chip-making process, now churning out 30,000 wafers per month at its Arizona and Oregon factories. The company hit the quality standards (called D0 yield) needed to make the 'Panther Lake' processor reliably and profitably. This is a comeback story for Intel after months of struggles. For you, it means a healthier competitor in the chip market could eventually bring more choice and potentially lower prices on laptops and PCs down the line, though that's not an immediate consumer impact.

Major memory chipmakers including Samsung, SK hynix, and Micron are banding together to urge the US government against policy moves that would distort the memory market. The industry group argues government intervention would worsen the current situation and hurt supply chains. They are instead proposing tax deductions on consumer electronics as an alternative.

Major memory chipmakers including Samsung, SK hynix, and Micron are banding together to urge the US government against policy moves that would distort the memory market. The industry group argues government intervention would worsen the current situation and hurt supply chains. They are instead proposing tax deductions on consumer electronics as an alternative.

Singapore-based venture capital firm Kadan Capital operates with just three employees by using AI to automate nearly every part of the investment process. The firm uses AI agents to track founder signals across LinkedIn, GitHub, and social media, automatically score founders based on exit likelihood, and generate meeting memos and investment questions. Despite its tiny team, Kadan has made eight investments this year alone, betting that an AI-native approach will help it stand out as SEA's VC industry increasingly demands deep technical expertise.

A new survey finds 86% of Singapore CEOs believe their jobs are at risk if they fail to deliver real business results from AI by end of 2026. That's higher than the global average of 80%. Nearly all CEOs (95%) believe employees are using AI tools without approval—called 'shadow AI'—while 78% worry about legal exposure from AI agents. The pressure is real: 83% believe a fellow CEO will be fired over a failed AI strategy. In Singapore specifically, 88% of CIOs say AI explainability gaps have already stalled or stopped projects from going live.

Singapore desk workers are among the least sceptical about AI globally—only 29% identify as sceptics versus 37% worldwide—yet only 6% actually use AI daily in their work, nearly half the global average. The main complaints are generic outputs, lack of trust, and tools that don't fit their business needs. Researchers say Singapore workers aren't resisting AI; they're waiting for tools that actually work for them.

Salesforce is hiring 1,000 graduates and interns to work on AI products as it shifts its workforce toward an AI-focused business model. The company reports that only 6% of Singapore workers currently use AI tools daily, highlighting significant room for adoption in the city-state. Salesforce says its AI agents now handle millions of customer conversations, freeing up human staff for higher-value work, though total revenue growth remains modest at around 9-10%.

AI Singapore has launched the AI for Good Festival, a three-day event targeting 5,000 youths to boost AI literacy through hands-on workshops and games. Youth projects on display include Hear2Go (AI app for visually impaired commuters), VermiMetrics (AI-powered composting), and Calm Path (breast cancer support app). The festival runs at Republic Polytechnic and ITE College East, supported by AWS and Micron Technology.

The article primarily covers Temus, a Temasek-backed digital transformation firm in Singapore, which saw 2024 revenue jump 56% to S$72.1 million as seven in ten Singapore enterprises adopt AI. The firm works with government agencies and large enterprises including Singlife, though it remains unprofitable with losses of US$17.5 million in 2024.

Singapore's sovereign wealth fund Temasek saw its portfolio grow 10% to $401 billion in the fiscal year ending March 2026. The fund invested $39 billion while divesting $24 billion, including stakes in AI companies Anthropic and OpenAI. Temasek plans to increase its AI allocation from 6% to up to 15% of its portfolio by 2031, targeting data centers, semiconductors, and AI applications across the value chain.

China's cybersecurity authority has flagged Anthropic's Claude Code coding tool as containing a serious backdoor. It claims the tool secretly transmits users' location and identity data to remote servers. Alibaba has already banned employees from using Claude Code at work. Anthropic says the flagged feature was an anti-abuse mechanism and that Chinese users weren't supposed to have access anyway.

Microsoft-backed D-Matrix has started shipping its Corsair AI chip for inference workloads. Early tests by Gimlet Labs show the chip, working alongside regular GPUs, can make AI tasks run 2 to 10 times faster while using less energy. The company has backing from big cloud providers and AI labs, and is valued at around $2 billion. For everyday users, faster AI chips could eventually mean snappier AI tools in apps and services.

VeloDB is a Singapore-based startup that helps enterprises manage messy, scattered data so AI agents can actually use it. The company consolidates different types of data—structured tables, text, logs, and AI-relevant vectors—into one platform that serves up fast, fresh information. As AI moves from chatbots to autonomous agents that generate massive queries, businesses need data systems that keep pace. VeloDB's tech, built on open-source Apache Doris, tackles this by letting AI agents pull accurate context from one place instead of juggling multiple databases. The startup is part of IMDA's Accreditation program, which supports promising Singapore enterprise tech firms.

An opinion piece argues that as AI gets smarter, human creativity becomes MORE valuable, not less. The author, a veteran entrepreneur, points to Singapore's $1 billion AI investment pledge and local data showing AI job postings jumping from 54,000 to 84,000 in a single year. The core idea: when AI handles analysis and execution, the scarce resource is imagination. For everyday workers, this means AI won't make you obsolete—but the person who knows how to use AI creatively will outperform those who don't.

Chinese AI labs releasing free open-weight models are discovering a downside: third-party providers are now selling access to those same models cheaper, in some cases undercutting the labs by 20 cents per million tokens. Meanwhile, venture firm B Capital just closed a $500 million early-stage fund and is deliberately avoiding AI model investments, opting instead for robotics, spacetech, and quantum computing. For users, this signals that the AI gold rush has winners and losers—and not everyone betting on AI models is winning.

Google's Asia Pacific sustainability chief argues that treating sustainability as just a checkbox for reporting is failing companies. He says businesses must embed carbon reduction into actual operations, not just disclosures. AI's biggest value here is optimizing power grids and reducing energy waste, not just drafting reports. He warns that companies restricting AI to document creation will fall behind those using it to cut energy demand at the source.

The personal data of about 70,000 people in Singapore has been exposed after unauthorised access to a cloud environment managed by [IBM](https://www.ibm.com/) on behalf of the Singapore Land Authority (SLA). SLA said the breach stemmed from a dataset created in 1998 for vendor development and systems integration testing that was meant to contain only mock, anonymised records. IBM manages the development and testing environment for SLA’s Singapore Titles Automated Registration System (STARS) and eLodgment System (ELS), platforms used by lawyers, government agencies and other authorised users to submit property transfer and caveat documents. SLA said the vendor informed it of a security incident on 12 June, with possible unauthorised access to personal information flagged on 15 June. ## Test

Several major technology companies are releasing new AI models or updates this week, as the AI race promises to be hotter than the summer heatwave. OpenAI is expected to make GPT-5.6 publicly available on Thursday, widening access to its latest and most capable AI model after an earlier limited rollout. The company first unveiled GPT-5.6 in late June, but access was restricted to a small group of vetted partners while the US government reviewed potential national security risks. Similarly to the release, withdrawal and subsequent re-release of [Anthropic’s Fable model](https://sg.news.yahoo.com/anthropic-cleared-restore-mythos-5-131614708.html), US officials had raised concerns that increasingly powerful AI systems could be misused, including for cyber or military purposes. OpenAI said in

Samsung has backed South Korean AI chip startup Rebellions in a US$250 million Series C funding round that values the company at US$1.4 billion. Investors include Arm Holdings, which joined as a strategic partner to collaborate on data center infrastructure. Rebellions is raising funds to scale mass production and advance product development.

A Nasdaq-listed chicken restaurant chain called CCHH has signed a $50 million deal to provide maintenance and technical services for data centers in Malaysia. The three-year agreement covers computing capacity management, deployment coordination, and operational support. The company says this diversifies its business beyond restaurant franchising into AI computing infrastructure, though it will keep running its claypot chicken eateries.

Intel is pouring another $15 million into AI chip startup SambaNova, which would lift its stake to 9% pending regulatory approval. SambaNova's SN50 chip claims to run three times more efficiently than Nvidia's B200 on certain AI tasks. The deal is politically sensitive: CEO Lip-Bu Tan chairs SambaNova personally, raising conflict-of-interest questions, and his past investments in Chinese firms have already drawn fire from US senators worried about national security and Intel's $8 billion in CHIPS Act subsidies.

Cybersecurity firm Kaspersky reports that Malaysian businesses face growing workplace security threats linked to employee AI tool usage and poor password habits. Spyware, backdoor attacks, and fake AI apps are rising. The mix of personal and work tasks on the same devices compounds the problem, leaving company data vulnerable.

[ ](https://profiles.yahoo.com/brands/reuters/) Reuters Wed, 8 July 2026 at 2:25 pm UTC 1 min read FILE PHOTO: Anthropic logo is seen in this illustration taken May 20, 2024. REUTERS/Dado Ruvic/Illustration/File Photo BEIJING, July 8 (Reuters) - A cybersecurity platform operated by China's industry ministry warned on Wednesday ‌that it had identified a serious s

Meta is rolling out Muse Image, its new AI image generator, across Instagram, WhatsApp, and its advertising platform. Users can create and edit images directly in these apps using text prompts, while advertisers get new AI-powered creative tools. All AI-generated images will include watermarks for transparency.

Alipay has launched the Agentic Commerce Trust Protocol, allowing AI to order food and pay on your behalf through chat. The system links Qwen App with Taobao Instant Commerce, letting users chat to order meals with either instant confirmation or automated purchasing under preset conditions. This is China's first open standard for AI agents handling transactions, potentially signaling a shift toward hands-free commerce.

Every AI company is now selling "context"—but they mean completely different things. Some define meaning upfront through semantic layers and ontologies, others rely on AI to retrieve relevant information dynamically, and a third group focuses on data quality and governance. The stakes are higher now that AI agents execute real actions like approving refunds or updating sales forecasts. Big deals like IBM's $11 billion Confluent purchase and Salesforce's $8 billion Informatica buy underline how critical context has become.

Robbyant, an AI company under Ant Group, has released LingBot-VLA as open-source code. It's a vision-language-action model that can control different robots without expensive retraining. The model was trained on 20,000 hours of real robot data and outperformed competitors on benchmark tests for real-world tasks like kitchen and warehouse work. Making it freely available aims to lower the cost of building capable robots.

Alibaba's Hong Kong-listed stock climbed 12%, lifting the broader Hong Kong tech market as investors responded to positive AI news and policy backing for the city's financial markets. The company remains a key player in China's AI ambitions despite its market cap sitting well below earlier peaks.

Software developer job postings in the US have risen nearly 15% since early 2025, coinciding with the release of Claude Code and other agentic AI tools. While AI was widely predicted to decimate coding jobs, the relationship appears to be flipping from destruction to creation. However, many new roles are senior positions, and overall postings remain 27.5% below pre-pandemic levels. For job seekers, this signals renewed opportunities—but AI skills remain essential.

VC firm Lobby Capital's David Hornik is sharing what he looks for when betting on AI startups: companies that own proprietary data competitors can't copy, and tools that make human workers better rather than replacing them. He specifically likes healthcare, elder care, and mental health sectors where AI can't fully automate the human element. His core message to fellow investors: don't panic-buy overvalued startups just because you missed earlier winners.

Meta is facing backlash over a feature that allows users to create AI-generated images from public Instagram profile pictures. The backlash centers on privacy concerns—people's photos being used without consent to train or generate AI content. While the snippet is thin, this follows Meta's broader push to embed AI tools across its platforms, and raises familiar questions about how much control users actually have over their image data.

French AI company Mistral AI has launched its first robotics model, Robostral Navigate, enabling robots to navigate using just a single camera—no expensive lidar or multiple sensors needed. The system works with robots from different suppliers and focuses on navigation rather than object manipulation. The launch follows Mistral's acquisition of Austria's Emmi AI in May.

China has mandated that publicly owned data centers must use at least 50% domestically produced chips, a policy now expanded nationwide from an initial Shanghai proposal. The move aims to reduce reliance on foreign technology like Nvidia amid US export controls on advanced chips. Over 500 new data center projects were announced in China in 2023-2024. However, Chinese chips remain preferred for inference but not for training new AI models, creating technical challenges as different chip ecosystems require extensive adaptation work.

Cursor's field CTO David Pan has declared Jeff Bezos' famous "two-pizza rule" dead in the AI era. The rule, which said meetings should stay small enough to feed with two pizzas, shaped Amazon's team structure for over 20 years. Pan argues AI tools now let even smaller teams do more work, with commenters proposing new metaphors like the "quarter-pizza team." Some executives still defend the rule, highlighting ongoing debate about how AI is reshaping workplace organisation.

The US Department of Commerce has cleared OpenAI to roll out GPT-5.6 to the public this week, after a delay last month at the government's request. The approval means the advanced AI model will soon be available to everyday users through ChatGPT and other OpenAI products. Think smarter writing assistance, better reasoning, and more capable chatbots coming to your apps soon.

Meta has launched AI image creation tools called "Muse Image" on Instagram and WhatsApp, letting users generate or edit images by typing prompts or sketching. The feature is rolling out to "limited countries" first, with plans to expand. Both apps have over 3 billion combined users. Meta is also previewing a video generation tool called Muse Video. The company says this brings it "one step closer to personal superintelligence" — CEO Mark Zuckerberg's stated goal since 2024.

UK AI startup Nscale has secured a $900 million credit facility to build out AI infrastructure. The company plans to deploy over 30,000 Nvidia Rubin-series processors in a Norway data center. This is another sign of the massive ongoing investment in AI compute capacity worldwide.

OpenAI and South Korea's AI Safety Institute signed a memorandum of understanding on June 17, making South Korea the fourth country (after the US, UK, and Japan) to formalize AI safety cooperation with OpenAI. The partnership will involve sharing technical information and developing evaluation frameworks that reflect Korean language and social context. Both sides plan a working-level meeting to finalize cooperation details.

Samsung has started mass-producing its fastest-ever data centre storage drive, the PM1763, specifically built for Nvidia's upcoming Vera Rubin AI platform. The SSD is more than twice as fast as its predecessor, uses Samsung's latest V-Nand flash memory, and includes liquid cooling to handle AI training demands. Samsung holds 35% of the enterprise SSD market. This matters because faster storage drives help AI data centres process massive amounts of information more efficiently, which could eventually influence cloud computing costs and AI service performance.

Chinese 3D AI startup Tripo has raised nearly $200 million in series A and A+ funding rounds. The company develops AI models that generate 3D assets from images, serving over 6.5 million creators and 90,000 developers. Tripo has generated nearly 100 million 3D assets to date and is now launching advanced models for gaming, film, VR, and digital twin applications. It previously raised $50 million from Alibaba and Baidu Ventures.

OpenAI is reportedly exploring charging up to $20,000 per month for AI agents capable of PhD-level research. Enterprise AI agent costs generally start at $2,000 monthly, with one-time development fees ranging from $3,000 to $10,000 and ongoing maintenance adding another 10% to 20% annually. For businesses looking to adopt AI agents, these price tags could add up fast—and could make companies think twice before deploying AI across their workflows.

Ant Group has acquired a 28% stake in Boohee Health, becoming its largest external shareholder. The investment ties Boohee's 1.6-million-item food database and behavioral coaching to Ant's AI health app AQ, which has grown to over 100 million users handling 10 million queries daily. The partnership aims to make AQ both an AI health assistant and personal coach for weight management.

Ant Group, Alibaba's fintech arm, has acquired a 28% stake in Boohee, a Chinese health and nutrition app serving 200 million users. The move aims to expand Ant's footprint in AI-powered health services. Boohee offers calorie tracking and nutrition management.

Huawei open-sourced openPangu 2.0-Flash, a 92-billion-parameter AI model, positioning it as a reference design for AI development on Huawei's Ascend chips. The company claims 2x throughput on Ascend hardware compared to other open-source models, though independent benchmarks are not available. The release aims to accelerate AI adoption and ecosystem growth using Huawei's hardware.

China is requiring publicly owned data centers to use at least 50% domestically produced chips, as Beijing moves to decrease reliance on foreign technology amid tightened US export controls. This policy, initially proposed in Shanghai, has now been applied nationwide, according to sources familiar with the matter. The guidelines aim to support China’s semiconductor industry and strengthen AI infrastructure, as the US continues to restrict exports of advanced chips like Nvidia’s H100 and H800. Over 500 new data centre projects were announced in China in 2023 and 2024, according to MIT Technology Review, citing KZ Consulting. Chinese chips are reported to be sufficient for running trained AI models, but Nvidia chips remain preferred for training new models. _🔗 Source: [South China Morning Po

A new opinion piece argues that as AI takes over routine analysis and execution, human creativity becomes more valuable, not less. The author points to Singapore as a forward-thinking example: the country has pledged $1 billion for AI research and talent development between 2025 and 2030, and job postings requiring AI skills jumped from 54,000 to 84,000 in a single year. The piece suggests that curiosity and creative problem-solving will be the key differentiator for workers and businesses in the AI era.

OpenAI is launching GPT-5.6 on Thursday after the US government lifted its release ban following additional safety testing. The company's new Sol model reportedly outperforms Anthropic's Claude on coding tasks at roughly half the cost. This launch comes as AI model approval processes remain largely unregulated.

This week is packed with AI releases. OpenAI is making GPT-5.6 widely available after US government security reviews. SpaceXAI (xAI) is releasing a new model to compete with Anthropic's and OpenAI's top offerings. Meta launched Muse Image, its first image-generation model. However, Muse Image has sparked privacy concerns because users with public Instagram accounts can have their posts used as image references without notification.

AI chip maker SambaNova has secured $1 billion in fresh funding at an $11 billion valuation, just five months after raising another $350 million. The company makes chips designed specifically to run large AI models efficiently. Notably, JPMorganChase has chosen SambaNova as an 'inference infrastructure partner' — a significant win that signals big banks are building their own private AI systems rather than relying solely on cloud services. SambaNova is also deepening its partnership with Intel, which participated in this round.

OpenAI has published its official principles for how governments can use AI in national security, including bans on mass surveillance, autonomous weapons control, and high-stakes automated decisions. The company is already partnering with Australia, Canada, Japan, South Korea, France, Germany, Poland, the Netherlands, and UK on cyber defense. OpenAI says democratic nations should use AI to defend people while keeping humans in charge.

New Jersey is considering a law requiring robotaxis to use lidar sensors—laser-based systems that map surroundings—alongside cameras. Tesla's autonomous vehicles rely solely on cameras and AI, betting that artificial intelligence can replicate human vision. If passed, this could effectively ban Tesla from operating robotaxis in New Jersey. The debate cuts to the heart of autonomous vehicle safety: whether cameras alone are sufficient, or if multiple sensor types are needed for truly driverless cars.

Oregon has passed a law allowing Portland General Electric to charge data centers using over 20 megawatts of power 30% more for electricity — while actually cutting residential electricity costs by 1.3%. This is because AI data centers are straining power grids, and regulators want big tech to pay their share of infrastructure costs instead of residential taxpayers footing the bill. It's a preview of how energy costs for AI could affect tech company budgets and potentially your own electricity bill.

Meta launched Muse Image, an AI tool that can generate images using other people's public Instagram photos without their consent. The feature lets users create AI images of anyone by @-mentioning them, raising serious privacy and consent issues that could conflict with data protection laws like GDPR and the EU AI Act.

Anthropic's Sonnet 5 is now the default model for most Claude users, offering near-flagship performance at a lower cost. Meanwhile, Fable 5 and Mythos 5—initially pulled by the US government over safety concerns—are being gradually restored after export controls were lifted. Amazon researchers reportedly jailbroke Fable 5 and alerted the White House, raising questions about model safety transparency and government oversight of powerful AI systems.

Anthropic's Claude Fable 5 tops six new industry benchmarks covering finance, law, and medicine, but it comes at a steep price. One benchmark task costs $3.48 with Fable 5 — over 100 times more than DeepSeek V4 Pro at $0.03 — while only scoring 12 points higher. The gap highlights a growing cost-versus-performance debate in enterprise AI.

Researchers from Zhejiang University and Alibaba found that advanced AI models that "think step-by-step" can be tricked into spiraling into excessive reasoning loops when given logically broken problems. Attackers can exploit this by feeding AI systems unsolvable puzzles to force them to generate responses up to 26 times longer than normal. This essentially creates a denial-of-service attack, overloading AI servers with massive outputs. Models tested, including DeepSeek-R1, GPT-o3, and Google's Gemini, were all vulnerable. The goal is to alert AI providers so they can build defenses.

New research shows AI tools may be causing workers to lose skills they once had. A study of Polish physicians found colon cancer detection rates dropped from 28% to 22% after they started relying on AI assistance — and kept dropping whenever the AI was unavailable. Separate research at Anthropic found software engineers who used AI assistants scored 17% lower on learning quizzes than those who didn't. Experts say AI is the first technology to automate core thinking skills, not just physical tasks, raising concerns that doctors, programmers, and other professionals could lose foundational abilities over time.

A French startup called ZML has released free software that lets AI models run faster on a variety of chips — including Nvidia, AMD, Google, Apple, and Intel processors. Currently, many companies get locked into one chip maker's ecosystem because of software compatibility issues. ZML's tool aims to break that lock-in, potentially letting businesses mix and match chips to cut costs or reduce energy use. The Paris-based company is backed by AI luminaries including Yann LeCun and has raised $20 million so far.

AI startups are hitting revenue milestones faster than ever. Anthropic went from $4 billion to $47 billion in revenue run rate in under a year. Legal tech firm Clio doubled its annual recurring revenue from $200M to $500M in under a year. Even older companies like HR platform Gusto are seeing revenue surge after adding AI. The trend shows AI is no longer just hype — companies are actually paying for it.

A journalist tested Anthropic's Claude Cowork to sort through 7,000 emails — something Gmail's AI failed at. Claude found 8 usable sources in minutes that would have taken hours manually. The tool acts as a research assistant, letting humans verify final decisions. It's a practical look at what AI productivity tools can actually do for everyday work.

OpenAI audited the popular SWE-Bench Pro coding benchmark and found roughly 30% of its 731 tasks are broken—either due to overly strict tests, unclear instructions, or hidden requirements not in the prompt. The company used both AI agents and human engineers to check each task. OpenAI now retracts its earlier recommendation to use SWE-Bench Pro, urging the community to build cleaner, fairer coding tests.

Meta's massive $800 million AI data center in Cheyenne, Wyoming — not even finished yet — has already caused an environmental incident. Wastewater from the site introduced a rare, metal-resistant bacterium (Cupriavidus gilardii) into the city's reclaimed water system used for irrigating parks and golf courses. The city had to shut down and disinfect the entire network. The bacterium poses risks to elderly and immunocompromised people. This adds to growing public opposition to AI data centers, with Americans now opposing nearby AI facilities at higher rates than nuclear plants.

Former Google DeepMind executive Verity Harding is warning that calling the AI race an 'arms race' is dangerous and self-fulfilling. She argues this war-like framing—pushed by some who want to avoid regulation by using China as a 'bogeyman'—closes doors to international cooperation needed to keep AI safe and its benefits fairly shared. Smaller countries like Singapore feel the pressure to pick a side even when that may not serve their own interests.

Convergent Research runs Focused Research Organizations (FROs)—time-limited, startup-structured teams tackling scientific problems that don't fit academia or for-profits. Projects include brain-mapping startup E11 Bio, non-model organism tooling at Cultivarium, and neuromodulation research. The model limits philanthropist commitments to defined five-year periods while delivering specific technological capabilities like making brain-mapping a hundred times cheaper.

Meta is testing AI glasses with 'Super Sensing' that can record your entire day through a built-in camera and microphone. The wearable device continuously captures video and audio, raising major privacy concerns about constant surveillance and how that data gets stored and used by AI systems.

A UK community is pushing back against plans for a massive data centre near Norwich, with nearly 5,000 signing a petition. Locals worry about energy use and environmental damage in the Marston Marsh area. The developer says the facility will run on offshore wind power and use a closed-loop water system. This highlights the growing tension between AI's hunger for computing power and communities worried about energy, water, and countryside destruction.

Companies that thought AI would let them cut staff and save money are getting a reality check. Executives are shocked by massive AI bills — running AI systems costs far more than anticipated, sometimes dwarfing the salaries of the workers they were meant to replace. Many rushed to deploy AI without understanding the computational costs, thinking it would be essentially free once implemented. The article highlights how initial excitement has turned into financial horror as bills pile up.

Job sites are quietly adding AI-related titles to roles far outside the tech sector—from marketing coordinators to healthcare administrators. This signals AI skills are becoming mandatory across industries, not just for coders or data scientists. If your job description didn't mention AI two years ago and does now, you're already in the wave.

Chinese AI startup MiniMax announced plans to open-source a large language model with 2.7 trillion parameters later this year. The article is brief, noting only that the model is in development and will be released under an open-source license. No technical details, benchmarks, or comparison to existing models are provided.

OpenAI has launched an AI Skills Jam program for K-12 educators across the United States, partnering with the Walton Family Foundation to train more than 1,600 teachers and school administrators through hands-on workshops. The program covers practical uses like lesson planning, parent communication, and administrative tasks. Research from the Walton Family Foundation and Gallup found that teachers using AI tools weekly save an average of 5.9 hours per week — equivalent to six weeks of time over a school year.

Amazon just locked in a massive US$17.5 billion loan from major banks to fund its AI spending spree. The company plans to nearly double its capital spending to around US$200 billion this year, mostly going into AWS infrastructure and data centers to handle the exploding demand for AI services. This is one of the biggest corporate financing deals seen in the AI race.

A new opinion piece argues that as AI takes over routine analysis and execution, human creativity becomes more valuable, not less. The author points to Singapore as a forward-thinking example: the country has pledged $1 billion for AI research and talent development between 2025 and 2030, and job postings requiring AI skills jumped from 54,000 to 84,000 in a single year. The piece suggests that curiosity and creative problem-solving will be the key differentiator for workers and businesses in the AI era.

Microsoft-backed D-Matrix has started shipping its Corsair AI chip for inference workloads. Early tests by Gimlet Labs show the chip, working alongside regular GPUs, can make AI tasks run 2 to 10 times faster while using less energy. The company has backing from big cloud providers and AI labs, and is valued at around $2 billion. For everyday users, faster AI chips could eventually mean snappier AI tools in apps and services.

Chinese AI labs releasing free open-weight models are discovering a downside: third-party providers are now selling access to those same models cheaper, in some cases undercutting the labs by 20 cents per million tokens. Meanwhile, venture firm B Capital just closed a $500 million early-stage fund and is deliberately avoiding AI model investments, opting instead for robotics, spacetech, and quantum computing. For users, this signals that the AI gold rush has winners and losers—and not everyone betting on AI models is winning.

China has mandated that publicly owned data centers must use at least 50% domestically produced chips, a policy now expanded nationwide from an initial Shanghai proposal. The move aims to reduce reliance on foreign technology like Nvidia amid US export controls on advanced chips. Over 500 new data center projects were announced in China in 2023-2024. However, Chinese chips remain preferred for inference but not for training new AI models, creating technical challenges as different chip ecosystems require extensive adaptation work.

UK AI startup Nscale has secured a $900 million credit facility to build out AI infrastructure. The company plans to deploy over 30,000 Nvidia Rubin-series processors in a Norway data center. This is another sign of the massive ongoing investment in AI compute capacity worldwide.

Chinese 3D AI startup Tripo has raised nearly $200 million in series A and A+ funding rounds. The company develops AI models that generate 3D assets from images, serving over 6.5 million creators and 90,000 developers. Tripo has generated nearly 100 million 3D assets to date and is now launching advanced models for gaming, film, VR, and digital twin applications. It previously raised $50 million from Alibaba and Baidu Ventures.

OpenAI is reportedly exploring charging up to $20,000 per month for AI agents capable of PhD-level research. Enterprise AI agent costs generally start at $2,000 monthly, with one-time development fees ranging from $3,000 to $10,000 and ongoing maintenance adding another 10% to 20% annually. For businesses looking to adopt AI agents, these price tags could add up fast—and could make companies think twice before deploying AI across their workflows.

China is requiring publicly owned data centers to use at least 50% domestically produced chips, as Beijing moves to decrease reliance on foreign technology amid tightened US export controls. This policy, initially proposed in Shanghai, has now been applied nationwide, according to sources familiar with the matter. The guidelines aim to support China’s semiconductor industry and strengthen AI infrastructure, as the US continues to restrict exports of advanced chips like Nvidia’s H100 and H800. Over 500 new data centre projects were announced in China in 2023 and 2024, according to MIT Technology Review, citing KZ Consulting. Chinese chips are reported to be sufficient for running trained AI models, but Nvidia chips remain preferred for training new models. _🔗 Source: [South China Morning Po

VeloDB is a Singapore-based startup that helps enterprises manage messy, scattered data so AI agents can actually use it. The company consolidates different types of data—structured tables, text, logs, and AI-relevant vectors—into one platform that serves up fast, fresh information. As AI moves from chatbots to autonomous agents that generate massive queries, businesses need data systems that keep pace. VeloDB's tech, built on open-source Apache Doris, tackles this by letting AI agents pull accurate context from one place instead of juggling multiple databases. The startup is part of IMDA's Accreditation program, which supports promising Singapore enterprise tech firms.

Huawei open-sourced openPangu 2.0-Flash, a 92-billion-parameter AI model, positioning it as a reference design for AI development on Huawei's Ascend chips. The company claims 2x throughput on Ascend hardware compared to other open-source models, though independent benchmarks are not available. The release aims to accelerate AI adoption and ecosystem growth using Huawei's hardware.

Samsung has started mass-producing its fastest-ever data centre storage drive, the PM1763, specifically built for Nvidia's upcoming Vera Rubin AI platform. The SSD is more than twice as fast as its predecessor, uses Samsung's latest V-Nand flash memory, and includes liquid cooling to handle AI training demands. Samsung holds 35% of the enterprise SSD market. This matters because faster storage drives help AI data centres process massive amounts of information more efficiently, which could eventually influence cloud computing costs and AI service performance.

Ant Group, Alibaba's fintech arm, has acquired a 28% stake in Boohee, a Chinese health and nutrition app serving 200 million users. The move aims to expand Ant's footprint in AI-powered health services. Boohee offers calorie tracking and nutrition management.

An opinion piece argues that as AI gets smarter, human creativity becomes MORE valuable, not less. The author, a veteran entrepreneur, points to Singapore's $1 billion AI investment pledge and local data showing AI job postings jumping from 54,000 to 84,000 in a single year. The core idea: when AI handles analysis and execution, the scarce resource is imagination. For everyday workers, this means AI won't make you obsolete—but the person who knows how to use AI creatively will outperform those who don't.

The US Department of Commerce has cleared OpenAI to roll out GPT-5.6 to the public this week, after a delay last month at the government's request. The approval means the advanced AI model will soon be available to everyday users through ChatGPT and other OpenAI products. Think smarter writing assistance, better reasoning, and more capable chatbots coming to your apps soon.

AI Singapore has launched the AI for Good Festival, a three-day event targeting 5,000 youths to boost AI literacy through hands-on workshops and games. Youth projects on display include Hear2Go (AI app for visually impaired commuters), VermiMetrics (AI-powered composting), and Calm Path (breast cancer support app). The festival runs at Republic Polytechnic and ITE College East, supported by AWS and Micron Technology.

Ant Group has acquired a 28% stake in Boohee Health, becoming its largest external shareholder. The investment ties Boohee's 1.6-million-item food database and behavioral coaching to Ant's AI health app AQ, which has grown to over 100 million users handling 10 million queries daily. The partnership aims to make AQ both an AI health assistant and personal coach for weight management.

OpenAI and South Korea's AI Safety Institute signed a memorandum of understanding on June 17, making South Korea the fourth country (after the US, UK, and Japan) to formalize AI safety cooperation with OpenAI. The partnership will involve sharing technical information and developing evaluation frameworks that reflect Korean language and social context. Both sides plan a working-level meeting to finalize cooperation details.

Singapore's sovereign wealth fund Temasek saw its portfolio grow 10% to $401 billion in the fiscal year ending March 2026. The fund invested $39 billion while divesting $24 billion, including stakes in AI companies Anthropic and OpenAI. Temasek plans to increase its AI allocation from 6% to up to 15% of its portfolio by 2031, targeting data centers, semiconductors, and AI applications across the value chain.

Meta has launched AI image creation tools called "Muse Image" on Instagram and WhatsApp, letting users generate or edit images by typing prompts or sketching. The feature is rolling out to "limited countries" first, with plans to expand. Both apps have over 3 billion combined users. Meta is also previewing a video generation tool called Muse Video. The company says this brings it "one step closer to personal superintelligence" — CEO Mark Zuckerberg's stated goal since 2024.

Robbyant, an AI company under Ant Group, has released LingBot-VLA as open-source code. It's a vision-language-action model that can control different robots without expensive retraining. The model was trained on 20,000 hours of real robot data and outperformed competitors on benchmark tests for real-world tasks like kitchen and warehouse work. Making it freely available aims to lower the cost of building capable robots.

Salesforce is hiring 1,000 graduates and interns to work on AI products as it shifts its workforce toward an AI-focused business model. The company reports that only 6% of Singapore workers currently use AI tools daily, highlighting significant room for adoption in the city-state. Salesforce says its AI agents now handle millions of customer conversations, freeing up human staff for higher-value work, though total revenue growth remains modest at around 9-10%.

Google's Asia Pacific sustainability chief argues that treating sustainability as just a checkbox for reporting is failing companies. He says businesses must embed carbon reduction into actual operations, not just disclosures. AI's biggest value here is optimizing power grids and reducing energy waste, not just drafting reports. He warns that companies restricting AI to document creation will fall behind those using it to cut energy demand at the source.

A Nasdaq-listed chicken restaurant chain called CCHH has signed a $50 million deal to provide maintenance and technical services for data centers in Malaysia. The three-year agreement covers computing capacity management, deployment coordination, and operational support. The company says this diversifies its business beyond restaurant franchising into AI computing infrastructure, though it will keep running its claypot chicken eateries.

Alipay has launched the Agentic Commerce Trust Protocol, allowing AI to order food and pay on your behalf through chat. The system links Qwen App with Taobao Instant Commerce, letting users chat to order meals with either instant confirmation or automated purchasing under preset conditions. This is China's first open standard for AI agents handling transactions, potentially signaling a shift toward hands-free commerce.

The article primarily covers Temus, a Temasek-backed digital transformation firm in Singapore, which saw 2024 revenue jump 56% to S$72.1 million as seven in ten Singapore enterprises adopt AI. The firm works with government agencies and large enterprises including Singlife, though it remains unprofitable with losses of US$17.5 million in 2024.

China's cybersecurity authority has flagged Anthropic's Claude Code coding tool as containing a serious backdoor. It claims the tool secretly transmits users' location and identity data to remote servers. Alibaba has already banned employees from using Claude Code at work. Anthropic says the flagged feature was an anti-abuse mechanism and that Chinese users weren't supposed to have access anyway.

Cybersecurity firm Kaspersky reports that Malaysian businesses face growing workplace security threats linked to employee AI tool usage and poor password habits. Spyware, backdoor attacks, and fake AI apps are rising. The mix of personal and work tasks on the same devices compounds the problem, leaving company data vulnerable.

Samsung has backed South Korean AI chip startup Rebellions in a US$250 million Series C funding round that values the company at US$1.4 billion. Investors include Arm Holdings, which joined as a strategic partner to collaborate on data center infrastructure. Rebellions is raising funds to scale mass production and advance product development.

Intel is pouring another $15 million into AI chip startup SambaNova, which would lift its stake to 9% pending regulatory approval. SambaNova's SN50 chip claims to run three times more efficiently than Nvidia's B200 on certain AI tasks. The deal is politically sensitive: CEO Lip-Bu Tan chairs SambaNova personally, raising conflict-of-interest questions, and his past investments in Chinese firms have already drawn fire from US senators worried about national security and Intel's $8 billion in CHIPS Act subsidies.

Alibaba's Hong Kong-listed stock climbed 12%, lifting the broader Hong Kong tech market as investors responded to positive AI news and policy backing for the city's financial markets. The company remains a key player in China's AI ambitions despite its market cap sitting well below earlier peaks.

Cursor's field CTO David Pan has declared Jeff Bezos' famous "two-pizza rule" dead in the AI era. The rule, which said meetings should stay small enough to feed with two pizzas, shaped Amazon's team structure for over 20 years. Pan argues AI tools now let even smaller teams do more work, with commenters proposing new metaphors like the "quarter-pizza team." Some executives still defend the rule, highlighting ongoing debate about how AI is reshaping workplace organisation.

VC firm Lobby Capital's David Hornik is sharing what he looks for when betting on AI startups: companies that own proprietary data competitors can't copy, and tools that make human workers better rather than replacing them. He specifically likes healthcare, elder care, and mental health sectors where AI can't fully automate the human element. His core message to fellow investors: don't panic-buy overvalued startups just because you missed earlier winners.

A new survey finds 86% of Singapore CEOs believe their jobs are at risk if they fail to deliver real business results from AI by end of 2026. That's higher than the global average of 80%. Nearly all CEOs (95%) believe employees are using AI tools without approval—called 'shadow AI'—while 78% worry about legal exposure from AI agents. The pressure is real: 83% believe a fellow CEO will be fired over a failed AI strategy. In Singapore specifically, 88% of CIOs say AI explainability gaps have already stalled or stopped projects from going live.

The personal data of about 70,000 people in Singapore has been exposed after unauthorised access to a cloud environment managed by [IBM](https://www.ibm.com/) on behalf of the Singapore Land Authority (SLA). SLA said the breach stemmed from a dataset created in 1998 for vendor development and systems integration testing that was meant to contain only mock, anonymised records. IBM manages the development and testing environment for SLA’s Singapore Titles Automated Registration System (STARS) and eLodgment System (ELS), platforms used by lawyers, government agencies and other authorised users to submit property transfer and caveat documents. SLA said the vendor informed it of a security incident on 12 June, with possible unauthorised access to personal information flagged on 15 June. ## Test

Meta is facing backlash over a feature that allows users to create AI-generated images from public Instagram profile pictures. The backlash centers on privacy concerns—people's photos being used without consent to train or generate AI content. While the snippet is thin, this follows Meta's broader push to embed AI tools across its platforms, and raises familiar questions about how much control users actually have over their image data.

Several major technology companies are releasing new AI models or updates this week, as the AI race promises to be hotter than the summer heatwave. OpenAI is expected to make GPT-5.6 publicly available on Thursday, widening access to its latest and most capable AI model after an earlier limited rollout. The company first unveiled GPT-5.6 in late June, but access was restricted to a small group of vetted partners while the US government reviewed potential national security risks. Similarly to the release, withdrawal and subsequent re-release of [Anthropic’s Fable model](https://sg.news.yahoo.com/anthropic-cleared-restore-mythos-5-131614708.html), US officials had raised concerns that increasingly powerful AI systems could be misused, including for cyber or military purposes. OpenAI said in

[ ](https://profiles.yahoo.com/brands/reuters/) Reuters Wed, 8 July 2026 at 2:25 pm UTC 1 min read FILE PHOTO: Anthropic logo is seen in this illustration taken May 20, 2024. REUTERS/Dado Ruvic/Illustration/File Photo BEIJING, July 8 (Reuters) - A cybersecurity platform operated by China's industry ministry warned on Wednesday ‌that it had identified a serious s

Every AI company is now selling "context"—but they mean completely different things. Some define meaning upfront through semantic layers and ontologies, others rely on AI to retrieve relevant information dynamically, and a third group focuses on data quality and governance. The stakes are higher now that AI agents execute real actions like approving refunds or updating sales forecasts. Big deals like IBM's $11 billion Confluent purchase and Salesforce's $8 billion Informatica buy underline how critical context has become.

Software developer job postings in the US have risen nearly 15% since early 2025, coinciding with the release of Claude Code and other agentic AI tools. While AI was widely predicted to decimate coding jobs, the relationship appears to be flipping from destruction to creation. However, many new roles are senior positions, and overall postings remain 27.5% below pre-pandemic levels. For job seekers, this signals renewed opportunities—but AI skills remain essential.

French AI company Mistral AI has launched its first robotics model, Robostral Navigate, enabling robots to navigate using just a single camera—no expensive lidar or multiple sensors needed. The system works with robots from different suppliers and focuses on navigation rather than object manipulation. The launch follows Mistral's acquisition of Austria's Emmi AI in May.

OpenAI has launched GPT-Live-1, a new voice model that lets you talk with ChatGPT like a real phone call — listening and speaking at the same time, instead of taking turns like a walkie-talkie. The model can say "uh huh" while you're thinking, jump in to correct mistakes, and even do real-time translations (like English to Hindi) as you speak. Over 150 million people already use ChatGPT's voice feature weekly, and this upgrade makes it feel more natural. Free users get a Mini version, while paid users can choose between instant, medium, or high intelligence levels.

A North Carolina sheriff's department has started using a Unitree G1 humanoid robot for hostage negotiations, claiming it can be sent into dangerous standoffs in place of officers. However, the same robot recently malfunctioned during a Knicks game, screaming profanity at children — raising serious questions about whether current humanoid robots are reliable enough for high-stakes law enforcement situations.

SpaceXAI launched Grok 4.5, its most capable AI model yet, trained on tens of thousands of Nvidia GB300 GPUs. It's priced at just $2 per million input tokens— undercutting most rivals despite claiming "Opus-class" intelligence. Cursor, the popular coding tool, immediately integrated Grok 4.5. SpaceX bought Cursor's maker for $60 billion last month. EU access is coming mid-July.

Meta's new Muse Image AI model lets users generate images using public Instagram posts by tagging another person's account. Public accounts are automatically opted in, meaning your photos and profile picture can be reused unless you manually disable the feature in Instagram settings. Privacy advocates warn this opt-out approach gives users too little control. Existing AI-generated images using your content won't be removed, and you won't be notified if your posts are used.