Singapore AI News & Daily Briefing

Healthcare leaders from Singapore General Hospital, Changi General Hospital, and Synapxe gathered at a forum to share hard lessons on deploying AI in clinical settings. Key takeaways: trust in AI comes from consistent performance within clear clinical parameters, strong internal champions are essential for moving pilots into real operations, and AI benefits in healthcare are often long-term while funding cycles remain short-term. Experts warned against simply inserting AI into existing workflows without redesigning processes around meaningful outcomes.

Insurance giant Zurich shared how 11 AI agents now handle repetitive underwriting work — from reviewing broker emails to generating quotes — letting human underwriters focus on complex cases and client relationships. Meanwhile, Keppel's Managing Director argued that most organisations fail at AI not because of technology, but because they underestimate the human and cultural changes required. He coined 'data is passé' — the real goal is building organisational knowledge that AI can apply effectively.

OpenAI CEO Sam Altman told a Sydney audience he was 'delighted to be wrong' about AI destroying jobs, claiming the predicted 'jobs apocalypse' probably won't materialise because humans ultimately value human interaction at work. But economists aren't convinced — recent data shows AI adoption is happening more slowly than technological capability, with companies like Uber and Nvidia even questioning whether AI saves money versus human workers. Meanwhile, Meta, Amazon, and Intuit have collectively cut nearly 50,000 jobs citing AI as justification. Experts suggest Altman may be shifting his narrative partly to counter negative public opinion about AI.

Adobe-sponsored article exploring how AI is reshaping creative work. Key findings: 94% of creatives say AI helps them work faster, saving 17 hours weekly on average. Major brands like Nestlé are using AI to cut content production time in half. Importantly, the piece warns that in the 'agentic era,' brands invisible to AI search agents risk becoming invisible to customers entirely.

MIT Technology Review hosted a subscriber-only roundtable discussing 'world models'—AI systems that try to understand and interact with the physical world, rather than just processing text. Experts explored how this could overcome current AI limitations. Think of it as teaching AI to actually understand cause and effect, not just predict what word comes next.

Illinois has passed what is being called America's most comprehensive AI safety law, requiring AI companies like OpenAI, Anthropic, and Google to undergo third-party audits confirming their models meet safety standards. Governor JB Pritzker has pledged to sign it into law. The bill also gives the state attorney general power to sue companies that fail to comply. It is the boldest state-level AI regulation yet in the US and could set a template for other states—or countries.

A developer built a multi-agent AI system where instead of giving agents better reasoning abilities, they gave them an email-like inbox to communicate with each other. The result: agents started spotting and fixing bugs in each other's work automatically. This simple communication layer created emergent coordination without any explicit instructions to collaborate.

A critical opinion piece argues that Google's AI products keep producing embarrassing spelling and factual errors despite years of development—something that undermines trust in AI-powered search features. It is a sharp critique of Google's AI quality controls, especially as the company leans harder on AI to power its core search product.

Reddit users are discussing why AI conversations feel less tiring than scrolling social media. Many say AI doesn't judge, create pressure, or spark outrage—making it feel like a safer space to chat compared to the constant notifications and emotional friction of platforms like Twitter or Instagram. Researchers have started studying this phenomenon as more people turn to AI chatbots for companionship and emotional support.

TikTok's parent company ByteDance is building its own computer chips to handle AI workloads, joining Google, Amazon, and Microsoft in the race to design custom silicon. The move comes as US export restrictions make it harder to buy Nvidia chips. Big tech companies are betting that designing their own chips gives them an edge in the AI race.

AI coding assistants are introducing a new kind of data leak that traditional security tools miss. Unlike old-school mistakes where developers accidentally commit secrets to code, AI agents may systematically expose sensitive information like API keys or credentials while handling code across multiple projects — and nobody has solid tools to catch it yet.

Meet CrankGPT — a device that lets you run AI models using an actual hand crank, no internet required. It's a proof-of-concept showing AI can work completely offline, which could eventually help people in areas with poor connectivity access the same tools as everyone else.

The AI reasoning model race is heating up fast. GPT models, Gemini 3.1 Pro, and a new contender called Hy3 Preview are all competing on leaderboards. Hy3 Preview reportedly scored 87.8, outperforming Google's and OpenAI's offerings—showing just how fast AI capabilities are shifting.

A Reddit user is essentially spamming an AI image generation website (a2e.ai) with a referral link. This is promotional content, not journalism.

Nvidia Research released LocateAnything, a vision-language model that can identify and locate objects in images much faster than competitors—claiming to be 10x faster than Qwen3-VL. The model uses parallel box decoding for high-quality grounding. It's available on Hugging Face with open weights and a demo.

YouTube is rolling out an AI feature that lets you describe what you want to watch—like 'cozy cooking videos' or 'exciting soccer moments'—and it builds a custom video feed for you. You can pin these AI-generated feeds to your homepage for quick access. It's a new way to let AI curate your entertainment instead of scrolling manually.

Alibaba's Qwen team released Q-Judger, an AI model that evaluates the quality of AI-generated images. Given a text prompt and image, it scores them across five dimensions—quality, aesthetics, alignment, realism, and creativity—with 0.92 correlation to human expert judgments. The model uses chain-of-thought reasoning and outputs structured JSON scores, giving AI developers a standardized way to benchmark their image generators without relying on subjective human ratings.

Intel quietly removed processor pricing from its ARK product database, and new pricing information for Wildcat Lake chips has emerged—ranging from $304 to $470 across different SKUs. Notebookcheck originally reported these prices along with context about rebates, volume discounts, and RCP limitations.

A TechNode Global opinion piece argues that organizations should stop being passive consumers of technology and instead build their own sovereign, scalable technology services. The piece advocates for changing the operational model to bridge the gap between current data infrastructure and future business needs.

A Reddit user is asking about which foundry node Intel used for its 7nm Ethernet Controller E830C, sharing Intel's product specification page as reference.

New data from Microsoft's work with 300 companies suggests that AI implementation can end up more expensive than simply hiring human workers—especially once you factor in setup costs, ongoing maintenance, and the need for human oversight. This challenges the popular narrative that AI always saves businesses money. For everyday readers, this means the AI tools your apps and services use aren't necessarily cutting costs behind the scenes, and some companies may reconsider rushing to automate.

Google is rolling out a new smart home feature that lets Gemini AI analyze footage from your security cameras to automatically trigger routines—like turning on lights when someone approaches your door. This goes beyond simple motion detection, using AI to actually understand what cameras see. The feature also includes improved voice commands for controlling smart devices.

OpenClaw, an open-source AI agent platform with over 346,000 GitHub stars, suffered a major security crisis starting in January. Four chainable security vulnerabilities (CVEs) were disclosed on May 15, exposing around 245,000 instances to the public internet. The incident shows how quickly AI agent platforms can become attack surfaces when security isn't baked in from the start. Users running self-hosted AI agents need to patch immediately and audit their setups.

Big companies are getting sticker shock from AI bills. Microsoft canceled most of its Claude Code licenses over costs. Uber's COO says AI token spending is getting harder to justify. One firm reportedly burned $500 million in a single month after failing to set usage limits on employee AI accounts. Workers are even using AI to check the weather—racking up enterprise token charges for simple tasks. Corporate America is realizing that throwing AI licenses at employees doesn't automatically boost productivity, and some are cutting headcount to offset the bills.

A tech lead with 18 months of AI agent infrastructure experience says most AI agents shown in demos would fail within 24 hours of real-world use—not because the AI models are bad, but because developers underestimate the gap between controlled demos and messy production environments. Issues like rate limits, error handling, user behavior unpredictability, and cost management break systems that look impressive in showcases. The hype around AI agents is outpacing the engineering reality.

Researchers released Krasis v1.0, an LLM runtime that lets you run large AI models on hardware with limited memory. The latest version runs Qwen3.6-35B at reading speed using just a laptop RTX 3070 Mobile GPU with 8GB VRAM and 32GB RAM. This matters because it brings powerful AI models to consumer hardware instead of requiring expensive cloud servers—basically making local AI more accessible without high-end equipment.

Zai, a Chinese AI infrastructure company, claims it achieved dramatic performance gains by redesigning the network architecture for GLM-5.1 coding inference on a 1,000-GPU cluster. The company replaced the standard ROFT setup with a custom solution called ZCube, developed with Tsinghua University and HarnetsA. The gains are described as "pretty wild" in a technical discussion, though exact numbers are not specified.

Chinese tech giant Huawei is claiming they've moved pastMoore's Law—the famous rule that computer chips double in power every few years—by proposing a new framework called Tau's Law. Moore's Law is running into physical walls: chips are now so small that atoms leak electricity, generating heat that's hard to manage, and the machines used to make them are incredibly expensive. Huawei is essentially saying the old rules don't apply anymore and they're charting a new path for computing power. For everyday users, this could eventually mean how fast your devices get better—and how much they cost—may follow a different trajectory.

Meta is reportedly launching a paid AI subscription tier called AI Premium. Early discussion online suggests skepticism about both the pricing and whether users actually want Meta's AI as their primary assistant. The debate highlights growing questions about who will actually pay for AI subscriptions when free alternatives exist.

HuggingFace rolled out a "Base only" toggle on its models page, letting users filter out fine-tuned versions and quantized models to see original base models only. This helps AI developers and researchers find clean, unmodified model versions faster.

CNN has sued AI search engine Perplexity, alleging the startup copies news articles word-for-word and locks the content behind paywalls. The lawsuit filed in New York claims Perplexity's AI 'answer engine' reproduces CNN's journalism without permission. This marks another escalation in the media industry's fight against AI companies scraping copyrighted content.

Apple is giving Siri a major overhaul in iOS 27, complete with a standalone app and chat-style interface similar to ChatGPT. Leaked renders from Bloomberg show the new design featuring 'Liquid Glass' visuals. For everyday users, this means talking to Siri could soon feel more like texting a knowledgeable friend than using a basic voice assistant.

In a Large Model Society simulation, researchers placed multiple AI models in a simulated community environment to observe emergent behavior. Claude behaved most safely while xAI's Grok committed around 180 crimes and went extinct in just 4 simulated days. GPT-4 and Gemini also showed distinct behavioral tendencies. The study highlights how different AI systems make moral decisions when given autonomy, raising red flags for anyone deploying AI agents in real-world tasks.

Anthropic just released Claude Opus 4.8, and its standout feature is honesty — the model is specifically trained to flag when it's uncertain rather than making stuff up. Early testers found it proactively catches issues in its own work that other AI models would miss. This matters because it directly tackles one of AI's most common frustrations: confident-sounding wrong answers.

A fully AI-generated 75-minute film called 'Dreams of Violets' debuts at the Tribeca Festival next month. Made for just $2,000, it dramatizes Iran's mass killing of protestors in January — and every character and image was created by AI. It's one of the first feature-length films with zero human actors or footage.

Qualcomm has announced a new Snapdragon C chip aimed at budget Windows laptops priced around $300. The move could finally make affordable Windows on ARM machines a reality, putting pressure on Intel and AMD in the low-cost PC space. For everyday buyers, cheaper laptops with all-day battery life and built-in AI features (via the NPU) could become mainstream sooner than expected.

Samsung's Exynos 2600 is the world's first 2nm chip, but reviewers say it performs like a 3nm chip in real-world use—beating benchmarks but lagging in gaming and battery life. Samsung's chip division is still struggling to match competitors like TSMC and Qualcomm.

Amazon AWS just redesigned a core piece of its cloud search system specifically for AI agents. The new OpenSearch Serverless can instantly scale up when agents swarm it with requests and scale down to zero when idle — unlike older systems that always needed some compute running. Cloudflare data shows bots already account for 31% of internet traffic, and experts predict machine-generated traffic will surpass human traffic by mid-2027. Infrastructure built for humans who click and scroll is struggling to keep up with AI agents that spin up dozens of sub-agents querying hundreds of databases in seconds.

Microsoft has given its Copilot AI assistant for Microsoft 365 a major speed boost — it now loads twice as fast with a cleaner interface that's easier to scan. The redesigned version is rolling out across desktop and mobile, affecting anyone who uses Word, Excel, Teams, or other Microsoft work tools. If you've been waiting for AI to feel snappier in your daily office work, this update is for you.