Agregátor RSS

Čínský startup Kimi představil open-source model umělé inteligence Kimi K2.5. Nová verze pracuje s textem i obrázky a poskytuje 'paradigma samosměřovaného roje agentů' pro rychlejší vykonávání úkolů. Kimi zdůrazňuje vylepšenou schopnost modelu vytvářet zdrojové kódy přímo z přirozeného jazyka. Natrénovaný model je dostupný na Hugging Face, trénovací skripty však ne. Model má 1 T (bilion) parametrů, 32 B (miliard) aktivních.

Cybersecurity teams increasingly want to move beyond looking at threats and vulnerabilities in isolation. It’s not only about what could go wrong (vulnerabilities) or who might attack (threats), but where they intersect in your actual environment to create real, exploitable exposure. Which exposures truly matter? Can attackers exploit them? Are our defenses effective? Continuous Threat Exposure

Cybersecurity teams increasingly want to move beyond looking at threats and vulnerabilities in isolation. It’s not only about what could go wrong (vulnerabilities) or who might attack (threats), but where they intersect in your actual environment to create real, exploitable exposure. Which exposures truly matter? Can attackers exploit them? Are our defenses effective? Continuous Threat Exposure [email protected]

Hezký konvertibilní notebook od HP je levnější než kdy dřív. • OmniBook X Flip 14 má dobrou výbavu a záruku s opravou na místě. • Nabízí dotykový OLED se stylusem, Wi-Fi 7 a hodně konektorů.

Alibaba Cloud’s latest AI model, Qwen3-Max-Thinking, is staking a claim as one of the world’s most advanced reasoning engines after posting benchmark results that delivered competitive results against leading models from Google and OpenAI.

In a blog post, Alibaba said the model was trained using expanded capacity and large-scale computing resources, including reinforcement learning, which led to improvements in factual accuracy, reasoning, instruction following, alignment with human preferences, and agent-style capabilities.

“On 19 established benchmarks, it demonstrates performance comparable to leading models such as GPT-5.2-Thinking, Claude-Opus-4.5, and Gemini 3 Pro,” the company said.

Alibaba said it has added two key upgrades to Qwen3-Max-Thinking: adaptive tool use that lets the model retrieve information or run code as needed, and test-time scaling techniques that it says deliver stronger reasoning performance than Google’s Gemini 3 Pro on selected benchmarks.

Analysts offer a cautious approach to the announcement. Benchmark results evaluate the performance under specific conditions, “but enterprise IT leaders may be deploying foundation models across various use cases under different IT environments,” said Lian Jye Su, chief analyst at Omdia.

“As such, while Qwen models have shown themselves to be legitimate alternatives to Western mainstream models, their performance still needs to be evaluated in domain-specific tasks, along with their adaptability and customization,” Su said. “It is also critical to assess scalability and efficiency when these models run on Alibaba Cloud infrastructure, which operates differently from Google Cloud Platform and Azure.”

More options for vendor diversification

The launch of Qwen3-Max-Thinking is likely to add momentum to AI model diversification strategies within enterprises.

“Now that Qwen models have demonstrated themselves as legit alternatives to Western models, CIOs should consider them when evaluating pricing models, licensing terms, and the total cost of ownership of their AI projects,” Su said.  “Running on Alibaba Cloud, the cost of ownership is likely more efficient, especially in the Asia Pacific, which is great news for global companies looking to make inroads into the Chinese market or China-friendly markets.”

Competitive reasoning scores from Qwen models expand the pool of viable suppliers, making diversification more attractive, according to Charlie Dai, principal analyst at Forrester.

“For CIOs managing digital sovereignty and cost efficiency, strong alternatives change the strategic equation, and rising model parity increases the viability of mixed portfolios that balance sovereignty, compliance, and innovation speed,” Dai said.

Others said benchmark momentum is also influencing how CIOs think about multi-model strategies.

“These benchmarks are a good yardstick not just to monitor performance, but also to assess which companies are serious and consistent in investing in foundation model capabilities and adoption,” said Neil Shah, VP for research at Counterpoint Research. “This is shaping how CIOs look at diversifying to multi-model strategies to avoid putting all their eggs in one basket, while weighing performance, cost efficiency, and geopolitical headwinds.”

That said, CIOs will need to consider the availability of these models outside of APAC alongside other factors such as export controls and compliance with local regulations.

“The bigger question is how CIOs adopt US versus non-US models based on AI use cases,” Shah said. “Where reliability and compliance are critical, enterprises, especially in Western markets, will favor proprietary US models, while highly capable Chinese models may be used for non-critical workloads.”

More governance and compliance challenges

Geopolitical tensions are adding another layer of complexity for enterprises evaluating models such as Qwen3-Max-Thinking. According to Dai, this requires closer scrutiny of operational details, particularly around system logs, model update mechanisms, and how data moves across borders.

He added that enterprise evaluations should go beyond performance testing to include red-team exercises, strict isolation of sensitive data, and alignment with internal risk and compliance frameworks.

“Enterprises evaluating Alibaba Cloud-hosted models need to scrutinize how AI safety controls, data isolation, and auditability are implemented in practice, not just on paper,” Su said. “While most cloud providers now offer in-region or on-premise deployments to address sovereignty rules, CIOs still need to assess whether those controls meet internal risk thresholds, particularly when sensitive IP or regulated data is involved.”

Victim and Big Brother Watch will argue the Met's policies are incompatible with human rights law

The High Court will hear from privacy campaigners this week who want to reshape the way the Metropolitan Police is allowed to use live facial recognition (LFR) tech.…

Zemi zasáhla nejsilnější radiační bouře za poslední dvě desetiletí • Tento jev vyvolal polární záře pozorovatelné dokonce i z území České republiky • Astronauti na oběžné dráze museli vyhledat úkryt v silněji stíněných částech stanice

Microsoft on Monday issued out-of-band security patches for a high-severity Microsoft Office zero-day vulnerability exploited in attacks. The vulnerability, tracked as CVE-2026-21509, carries a CVSS score of 7.8 out of 10.0. It has been described as a security feature bypass in Microsoft Office. "Reliance on untrusted inputs in a security decision in Microsoft Office allows an unauthorized

Microsoft on Monday issued out-of-band security patches for a high-severity Microsoft Office zero-day vulnerability exploited in attacks. The vulnerability, tracked as CVE-2026-21509, carries a CVSS score of 7.8 out of 10.0. It has been described as a security feature bypass in Microsoft Office. "Reliance on untrusted inputs in a security decision in Microsoft Office allows an unauthorized Ravie Lakshmananhttp://www.blogger.com/profile/[email protected]

A critical security flaw has been disclosed in Grist‑Core, an open-source, self-hosted version of the Grist relational spreadsheet-database, that could result in remote code execution. The vulnerability, tracked as CVE-2026-24002 (CVSS score: 9.1), has been codenamed Cellbreak by Cyera Research Labs. "One malicious formula can turn a spreadsheet into a Remote Code Execution (RCE) beachhead,"

A critical security flaw has been disclosed in Grist‑Core, an open-source, self-hosted version of the Grist relational spreadsheet-database, that could result in remote code execution. The vulnerability, tracked as CVE-2026-24002 (CVSS score: 9.1), has been codenamed Cellbreak by Cyera Research Labs. "One malicious formula can turn a spreadsheet into a Remote Code Execution (RCE) beachhead," Ravie Lakshmananhttp://www.blogger.com/profile/[email protected]

Another actively abused Office bug, another emergency patch – Office 2016 and 2019 users are left with registry tweaks instead of fixes.

Updated  Microsoft has issued an emergency Office patch after confirming a zero-day flaw is already being used in real world attacks.…

Abyste nadále od Air Bank dostali slevu 300 Kč na vyúčtování služeb od 02, budete muset s bankou více „bankovat“. Pro některé zákazníky se ale nic nezmění.

For companies looking to upgrade PCs this year, gear up: laptops will be more costly and configurations slower, analysts said.

PC makers “are already signaling price increases across the board and likely memory spec downgrades, especially in entry-level devices,” said Rishi Padhi, principal analyst at Gartner.

Low-cost laptops offering decent performance won’t come anytime soon, thanks to shortages in chips and memory. “With the RAM required for Windows 11 and Copilot+ now being much costlier, the price increase and scarcity in supply makes it difficult to deliver sub-$600 laptops without compromising user experience,” Padhi said.

The price increase for components is accelerating when compared to the last quarter of 2025. Tariffs are also a factor.

“Many PC components come from overseas suppliers, and the PC vendors can’t just eat all the tariffs that are imposed on them,” said Jack Gold, principal analyst at J. Gold Associates.

The 2026 PC market “is shaping up to be extremely volatile,” said IDC research vice president Jean Philippe Bouchard in a press release. Like Gartner, IDC attributes this year’s price shock to shortages of components that include CPUs and memory components.

This supply constraint is structural and persistent, not cyclical, and may last well into 2027, Gartner’s Padhi said.

“This dynamic is delaying true AI PC experiences into the premium segment, limiting mass market adoption, and forcing vendors to either raise prices or reduce specifications to protect margins,” Padhi said.

The higher prices put companies in a bind on upgrading to PCs with AI capabilities, said Gold. AI is still mostly done through browsers in the cloud, and not on PCs, so companies may not be tempted to upgrade to AI PCs.

“There really isn’t any ‘killer app’ outside of what I can do through a web browser. Copilot [Microsoft’s AI assistant] is fine, but do I want to go out and spend $1,000 on a new machine to use that feature?” Gold said.

The price increase is also partially due to a heavier concentration on supplies of higher-end units, not the lower-priced systems that consumers often buy. That’s like SUVs dominating car supplies, Gold said.

“Even Chromebooks are experiencing issues as the vendors are moving to supply higher-price PCs rather than building lower-end Chromebooks,” he said.

PC shipments blew past expectations in 2025, growing by 9.1% to 270 million units compared to 2024, according to Gartner. The expiration of Windows 10, users’ and companies’ desire to futureproof their systems, and a rush to upgrade hardware before tariffs took effect drove shipments.

AI PCs benefitted from the Windows 11 upgrade push in 2025. Microsoft coaxed companies to upgrade to systems with the Copilot+ moniker, indicating PCs capable of running AI tools locally.

Microsoft will embed more AI features into its Windows 11 PCs over 2026 and 2027, but Copilot+ will likely become a broader user experience, as opposed to a hardware selling point, Gartner’s Padhi said.

AI demand in servers and general large language model (LLM) requirements are also driving PC prices up, Padhi said.

Memory makers are concentrating on the production of memory for data centers, which are using the far more lucrative high-bandwidth memory (HBM) for AI applications. The scarcity leads to higher memory prices and limited PC features as computer makers pass the cost on to consumers, Padhi said.

Intel and AMD have already said their PC processors are in short supply and they can’t meet the full demand. The companies are working to increase supplies.

According to Gartner, Lenovo was the top PC vendor in 2025, with shipments totaling 73.6 million units and a 27.2% market share. Lenovo’s 2025 shipments went up by 17.6% compared to 2024.

HP was the second-largest PC maker, with shipments totaling 57.4 million units in 2025, growing by 8.3%. The company had a 21.3% market share.

Third-place Dell shipped 41.3 million units, growing by 4.9%, and had a 15.3% market share. Apple was in fourth place with 25 million units, growing by 10.3%, with a 9.2% market share. Asus and Acer were in fifth and sixth place, respectively.

Related reading:

• Intel’s AI pivot could make lower-end PCs scarce in 2026
• Memory shortage threat helps trigger PC buying spree
• Buyer’s guide: How to choose the right business laptops
• Microsoft’s Copilot+ PC hype needs to end, analysts say
• What is an AI PC? The power of artificial intelligence on your desktop

Podle uživatelských hlášení bylo poškozeno bootování do Windows 11 24H2/25H2. • Problémy se objevily po instalaci povinné lednové servisní aktualizace. • Microsoft sbírá hlášení a situaci prověřuje.

Cybersecurity researchers have discovered a JScript-based command-and-control (C2) framework called PeckBirdy that has been put to use by China-aligned APT actors since 2023 to target multiple environments. The flexible framework has been put to use against Chinese gambling industries and malicious activities targeting Asian government entities and private organizations, according to Trend Micro

Cybersecurity researchers have discovered a JScript-based command-and-control (C2) framework called PeckBirdy that has been put to use by China-aligned APT actors since 2023 to target multiple environments. The flexible framework has been put to use against Chinese gambling industries and malicious activities targeting Asian government entities and private organizations, according to Trend MicroRavie Lakshmananhttp://www.blogger.com/profile/[email protected]

Indicie, ze kterých různé weby dělaly závěry o brzkém vydání Arc B770 se dost možná týkaly jiného produktu. Rovněž postaveného na GPU Battlemage G31, ale profesionálního: Arc Pro B70…

Aplikace řidiče včas upozorní na nebezpečné zpomalovací prahy a ostré zatáčky • Waze nyní přehledně zobrazí správný jízdní pruh pro průjezd kruhovým objezdem • Systém nově varuje před snížením rychlosti a upřednostňuje vaše oblíbené trasy

V Raspberry Pi OS lze nově snadno povolit USB Gadget Mode a díky balíčku rpi-usb-gadget (CDC-ECM/RNDIS) mít možnost se k Raspberry Pi připojovat přes USB kabel bez nutnosti konfigurování Wi-Fi nebo Ethernetu. K podporovaným Raspberry Pi připojeným do USB portu podporujícího OTG.