Nadella tells Microsoft staff AI must stay under human control

source:

[1] Microsoft CEO Satya Nadella posts an internal warning to employees about AI safety (https://www.businessinsider.com/microsoft-sat ...)[2] AI development: Microsoft CEO Satya Nadella backs need for evaluators for AI systems amid 'slowdown' debate - The Economic Times (https://m.economictimes.com/tech/artificial-i ...)[3] Satya Nadella backs AI slowdown for safety and warns against monopoly (https://www.msn.com/en-in/technology/artifici ...)[4] AI leaders want to hit the brakes after years of reckless speed (https://arstechnica.com/ai/2026/09/ai-leaders ...)

Related News

Sep 15 2026, 00:16

76% of retirees regret delaying savings as 401(k) gaps widen

Retirees are telling the workforce something uncomfortable: the money they set aside was not enough, and the numbers suggest most current savers are on the same path. A TIAA Institute survey found 76 percent of retirees regret not starting to save earlier, and 71 percent wish they had put away more. "The retirees in this study are sending a clear and urgent message to everyone still in the workforce: what happens today will define the retirement you experience tomorrow," Surya Kolluri, head of the TIAA Institute, said in a press release. The regret is not confined to the size of the nest egg. Nearly half of retirees, 49 percent, said they underestimated healthcare and long-term care costs, and a separate 49 percent regretted not planning for late-life disruptions such as health problems, career shifts, job loss, and caregiving. For today's workers, the top stated concern is having enough to cover unexpected expenses and emergencies. ## The gap between the target and the median balance Fidelity's widely cited guideline calls for saving the equivalent of your annual salary by age 30, three times salary by 40, six times by 50, and 10 times by 67. Measured against actual balances, the shortfall is large at every age band. Full-time workers ages 35 to 44 earn a median of about $74,700 a year, according to Bureau of Labor Statistics data. Fidelity's target would put them near $224,100 by age 40. The median 401(k) balance for that group is $46,919, roughly a fifth of the goal, according to Vanguard data. At middle age the arithmetic does not improve. Workers ages 45 to 54 earn a median of nearly $73,900, implying a target near $443,400 at age 50. Their median 401(k) balance is about $79,000, just over a year's salary. For workers ages 55 to 64, median earnings of about $71,100 sit against a median balance of roughly $107,000 — about one and a half times salary, versus a target of eight times as much. The mechanism behind the gap is compounding, and it is unforgiving of delay. A dollar contributed at 25 has four decades to grow; the same dollar at 45 has two. That is why the earliest contributions carry the most weight, and why the retirees in the survey describe the delay itself, not any single bad decision, as their central regret. Saving consistently is also hard when competing goals claim the same dollars. Half of American adults report leaving the job market for more than a year because of career changes, layoffs, health issues, or — most commonly — caring for children, according to the TIAA Institute survey. Women disproportionately take career breaks for caregiving, which means they are more likely to pause retirement contributions and typically reach retirement with smaller balances than men. ## Social Security is no longer the backstop it was The second thread running through the survey is a generational split over how much to rely on Social Security. Almost all baby boomers surveyed, 94 percent, said the program would be a viable source of retirement income. Only about half of Generation Z, 51 percent, said the same. That skepticism has a documented basis. Unless Congress acts, the trust fund paying Social Security retirement benefits is projected to run short in late 2032, leaving payroll taxes to cover 78 percent of scheduled benefits — a 22 percent cut, according to the 2026 Social Security Trustees Report. An Investopedia analysis estimates a single retiree would need roughly $130,000 more in savings to absorb what could be trimmed from benefits. The last time the program's finances drew comparable attention, in the 1983 amendments, Congress legislated a combination of tax increases and a gradual rise in the full retirement age rather than allow an automatic reduction. Whether a similar fix arrives before the projected 2032 date is the single largest variable in any long-term household plan. For working households, the practical read is that the savings target is doing more of the work than it used to. Someone who reaches 67 with 10 times salary is far less exposed to a benefit adjustment than someone relying on Social Security for the majority of retirement income. The levers available are unglamorous and well documented: capture any employer match in full, raise the deferral rate with each pay increase, and treat a career break as a trigger to resume contributions rather than a permanent exit. The figures cited here come from the TIAA Institute survey, Fidelity's published guidelines, Bureau of Labor Statistics earnings data, Vanguard 401(k) data, and the 2026 Social Security Trustees Report. Savings targets and program projections change; readers should verify current figures against the latest official releases before acting. This article is for informational purposes only and does not constitute investment advice.

Sep 14 2026, 23:55

Samsung backs Euclyd in $231 million round to challenge Nvidia inference

Samsung co-led a 200-million-euro ($231 million) Series A for Dutch chip designer Euclyd, giving the world's largest memory maker an equity stake in a startup building inference silicon on a non-GPU architecture that will not reach customers until 2028. "AI is becoming a foundation of economic growth, scientific discovery and national competitiveness, but its potential will remain constrained unless we fundamentally change the infrastructure beneath it," Euclyd CEO Bernardo Kastrup said. Somerset Capital Partners, the Scaleup Europe Fund managed by EQT, and Innovation Industries co-led the round alongside Samsung, Kastrup told CNBC. Euclyd, founded in 2024, designs both the processor and the memory architecture for inference — the process of running trained AI models to generate answers, as distinct from the training phase that consumes most of Nvidia's GPU supply today. The company says its systems will cut the energy and cost requirements of AI data centers, though it has not published independent benchmarks and its hardware has yet to be proven in commercial deployment. The strategic logic runs through Samsung's balance sheet rather than its venture arm. "Samsung can help us in more ways than money," Kastrup said. "They are one of the biggest memory manufacturers in the world. They do a lot of engineering, they know a lot about systems, they know the supply chain, they have a huge network." That points to high-bandwidth memory (HBM, the stacked DRAM that feeds AI accelerators) and potentially Samsung Foundry capacity — the two inputs where Nvidia's own supply chain is most constrained. ## Nvidia's inference moat is the target, not its training lead Nvidia became the world's most valuable company by repurposing gaming GPUs for AI, and its H100 and H200 accelerators remain the reference point every challenger is measured against. The H100 carries 80GB of HBM3 at roughly 3.35 terabytes per second of bandwidth, according to Nvidia's published specifications — a figure that determines how fast a model can be served. Euclyd has not disclosed comparable bandwidth, transistor count, process node or thermal design power for its silicon, and did not publish the test conditions behind its efficiency claims. That gap matters because inference is where the money is shifting. Training runs are episodic and concentrated among a handful of labs; inference scales with every user query and every deployed agent, which makes cost per token the metric hyperscalers now manage most aggressively. OpenAI said in August that its first in-house chip, Jalapeño, had "industry-leading speed and efficiency." Google has run TPUs for a decade, AWS builds Trainium and Inferentia, and Meta is developing its own accelerators. Each of those programs removes volume from Nvidia's addressable market, and Euclyd is betting that enterprises wanting self-hosted inference — rather than hyperscalers — represent the unclaimed segment. Euclyd is targeting two revenue streams: selling hardware and physical rack systems to enterprises that want secure, self-hosted inference, and licensing its intellectual property to companies building their own chips. The second stream is the faster path to revenue and the one that most directly threatens Nvidia's pricing power, because it converts a hardware purchase into a design win. ## 2028 rollout leaves a long window for incumbents The timeline is the constraint. Euclyd aims to begin rolling out physical chip systems in 2028, with thousands of enterprise customers served by 2030, Kastrup said. Between now and then, Nvidia is expected to ship at least two further accelerator generations, and the in-house programs at Google, AWS, Meta and OpenAI will have moved from first silicon to volume deployment. A startup that has not taped out production silicon is competing against roadmaps that are already funded and staffed. For investors, the read-through is split. Nvidia's near-term fundamentals are barely touched by a $231 million round — the sum is smaller than a single day of the company's data center revenue run rate — but the direction of capital is the signal. Samsung's participation ties a top-three memory supplier to an architecture that, if it works, would consume HBM and foundry capacity outside the Nvidia ecosystem. Samsung Electronics shares trade in Seoul and the company reports memory results quarterly; any expansion of the Euclyd relationship into a supply agreement would show up first in foundry and memory order commentary rather than in the venture line. The competitive question is not whether Euclyd beats Nvidia on raw performance — it almost certainly will not at first — but whether it can deliver acceptable inference throughput per dollar at a fraction of the power draw. Until the company publishes silicon specifications and third-party benchmarks, that claim remains unverified against the H100's 3.35 TB/s of memory bandwidth and the efficiency figures OpenAI has claimed for Jalapeño. This article is for informational purposes only and does not constitute investment advice.

Sep 14 2026, 23:36

ByteDance puts Doubao inside the phone as 147 million AI handsets loom

System-level AI assistants are moving from demonstration to shelf space in China, and ByteDance is now selling one. The company's Doubao Phone Assistant reached consumers on Sept 14, embedded in a handset maker's software rather than shipped as a standalone app, and the first device carrying it — ZTE's Nubia NaviX Ultra — goes on sale Sept 16. The timing matters because IDC expects 147 million AI phones to ship in China in 2026, more than 53 percent of a total market it forecasts will shrink 2.2 percent to 278 million units as component costs rise. "The deep integration of large models into high-frequency digital terminals will push operating systems toward AI-native design, and model vendors will keep pushing toward full-stack AI to accelerate commercialization," Zong Jianshu, an analyst at Changjiang Securities, wrote in a research note. Yan Lei's team at Ping An Securities said cooperation between model developers and hardware makers will move beyond simple software pre-installation toward operating-system-level fusion, which it expects to produce new applications and business models. The consumer edition adds capabilities the May technical preview did not ship. On-screen question answering lets the model read whatever is displayed without a screenshot or app switch; in a ByteDance demonstration, a user pointed a camera at a room and asked for a cabinet under 1.2 meters and under 1,000 yuan ($149), and the assistant matched products on an e-commerce platform. Local retrieval covers photo albums, text messages and notes, with ByteDance saying that data never leaves the device. A dedicated AI button combines press, fingerprint authentication and wake-up into one action, and recording functions tie into Feishu Minutes. Cross-app execution handles tasks such as hailing a ride by voice. The most consequential piece is the beta release of the "operate phone" feature alongside the SAEP screen-automation declaration protocol, which lets third-party apps state which AI operations they permit and which they refuse. ByteDance opened a 30-day public comment period on the rules and said it will not automate apps that explicitly reject the capability. A banking app can declare that AI may not execute transfers, while a user can still let the assistant send a WeChat message. ByteDance paired the protocol with an Agent Protection System that manages permitted actions in tiers, though it has not disclosed the technical details, and published a privacy white paper on its website. That design choice is a bet on governance as a distribution strategy. Handing operational boundaries to the apps being operated converts a contested question — whether an assistant may move money or post on a user's behalf — into a configuration problem each developer solves for itself. It also shifts the burden of enforcement away from ByteDance, which avoids becoming the arbiter of what AI can do inside other companies' products. The test arrives quickly: the comment window closes around the same time the NaviX Ultra reaches shelves, and third-party responses over the following weeks will show whether the framework holds or becomes a formality. The competitive frame is unforgiving. Apple has demonstrated comparable screen-context understanding through Siri and Apple Intelligence, and Huawei ships its own on-device assistant across its handset line, meaning ByteDance is supplying an assistant to partners that compete with both. Guolian Minsheng Securities argues the industry's competitive focus is moving from hardware specifications to "AI capability plus ecosystem interconnection," a shift that favors whoever controls the assistant layer rather than the handset. For ZTE, the NaviX Ultra is a chance to differentiate a mid-tier brand on software it does not own; for ByteDance, it is a route to a daily-use distribution channel that does not depend on app-store placement. The financial stakes sit in the mix shift rather than the unit total. A market contracting 2.2 percent to 278 million units still leaves 147 million AI phones, and the premium those devices command depends on whether buyers treat the assistant as a reason to upgrade or as a feature they ignore after a week. ByteDance has not disclosed revenue terms with handset partners, and Nubia has not published a price or pre-order volume for the NaviX Ultra, leaving the Sept 16 launch as the first hard read on whether system-level AI moves units in a shrinking market. This article is for informational purposes only and does not constitute investment advice.