On March 1, 2026, Samsung announced its plans to turn all of its global production facilities into what it calls “AI-driven factories” by 2030. It’s a bold target, and it goes well beyond the traditional automation that already defines modern electronics manufacturing. This time, Samsung is talking about systems that can make decisions on their own.

A key part of that vision is agentic AI, which, unlike standard AI tools, can independently plan and execute tasks in order to meet specific goals. Samsung first introduced the term on the consumer side with the Galaxy S26 series. Now it wants to bring similar decision-making capabilities into its factories to increase efficiency and reduce operational costs over time.

Samsung says it will gradually expand its use of industrial robots, including humanoid models. The company describes three categories: operating robots for managing production lines and facilities, logistics robots for transporting materials, and assembly robots for precise manufacturing tasks.

Humanoid robots are expected to handle more complex or flexible operations, the kind that still benefit from human-like dexterity. Samsung has been building up its robotics capabilities in recent years, including increasing its stake in Rainbow Robotics, which is now part of the group. One of the systems likely to appear on production floors is the RB-Y1, a wheeled humanoid robot equipped with dual arms.

Samsung isn’t alone in exploring this direction. Other manufacturers are experimenting with humanoid robots and AI-powered automation as labor costs rise and supply chains grow more complex. Still, making these systems reliable and economically viable across dozens of global plants is no small task.

For now, Samsung is outlining a roadmap rather than showing a finished product. Whether its factories look radically different by 2030 will depend on how smoothly these technologies integrate into real-world production environments. The ambition is clear. The execution, as always, will be the harder part.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Source: Samsung)

The post Samsung plans transition to “AI-driven factories” with humanoid robots and agentic AI by 2030 appeared first on Gizmochina.

Training large language models is brutally expensive. It’s not just about having more GPUs; it’s about how efficiently you use them. And as models scale up, even small inefficiencies can turn into massive time and energy costs.

Now, a team of researchers from MIT, working with collaborators including NVidia, says it has found a surprisingly practical way to reclaim wasted compute during training — in some cases cutting overall training time nearly in half.

The problem they’re targeting lies in reinforcement learning (RL), particularly during what’s known as the “rollout” phase. This is the step where a model generates multiple candidate responses so it can learn which behaviors lead to better outcomes. It’s essential for reasoning-focused LLMs — but it’s also slow.

In fact, the rollout stage can account for as much as 85% of total execution time. The culprit is something researchers call a “long-tail distribution” of response lengths. Most generated responses finish quickly. But a small number run much longer than average. Because GPUs need to synchronize, the faster ones often sit idle waiting for the stragglers to complete.

The MIT team’s solution, called Taming the Long Tail (TLT), tackles that waste head-on. Instead of letting GPUs sit idle during those long generations, TLT uses that downtime to train a lightweight “draft” model on the fly. This smaller model learns continuously from the main model as training progresses.

The idea builds on speculative decoding, a technique where a smaller model predicts tokens ahead of the main model so multiple tokens can be verified in parallel. Traditional speculative decoding relies on a fixed draft model, which quickly becomes outdated as the primary model evolves during reinforcement learning.

TLT changes that dynamic. By retraining the drafter opportunistically using otherwise idle resources, the system keeps the draft model aligned with the main model, without requiring extra dedicated compute.

In experiments across several reasoning-focused LLMs and real-world datasets, the results were significant. The researchers report end-to-end training speedups ranging from 70% to 210% compared to strong baselines, effectively doubling training speed in many scenarios. Importantly, model accuracy remained unchanged.

There’s also an interesting side benefit: the continuously trained drafter itself becomes a useful artifact. Because it’s trained alongside the main model, it can serve as an efficient inference model in certain contexts.

The work points toward a broader theme in AI research right now: optimization over brute force. Instead of scaling up clusters indefinitely, researchers are increasingly looking for ways to extract more performance from the hardware already in place.

If approaches like TLT prove robust at larger industrial scales, they could meaningfully reduce both the financial and environmental costs of training next-generation reasoning models.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Source | Via)

The post Researchers double AI training speed just by reclaiming idle GPU time appeared first on Gizmochina.

A developer’s routine cleanup task reportedly turned into a disaster after a small mistake in AI-generated code wiped an entire drive.

The incident, first described in a Reddit post, involved code generated by ChatGPT Codex (version 5.3). The user had asked for a simple PowerShell script to remove Python pycache folders. Instead, the script ended up deleting the contents of the user’s F: drive, including project files and Docker data.

What went wrong

According to the post, the issue came down to a subtle escaping problem in the path string. The generated script used a backslash () to escape quotation marks. That approach can work in certain command-line contexts, but things get messy when PowerShell calls cmd.exe via cmd /c. In this case, the escaping didn’t behave as intended. The path variable was effectively reduced to just \, which Windows interprets as the root of the current drive.

Combined with rmdir parameters for silent, recursive deletion, the command reportedly ran without prompting for confirmation. The result: everything under the drive’s root directory was removed.

It wasn’t a complex exploit or some obscure bug, just a small mismatch in how different Windows shells handle escaping.

The episode highlights a growing trend sometimes referred to as “vibecoding”. For simple tasks, that can save time. But when commands interact directly with the file system, small errors can escalate quickly.

PowerShell and cmd.exe don’t interpret escaping the same way. In native PowerShell, the correct escape character is typically the backtick (`), not a backslash. Mixing the two environments introduces edge cases that aren’t always obvious, even to experienced users.

There’s also the broader Windows command-line reality: destructive commands like rmdir or Remove-Item -Recurse -Force don’t necessarily include built-in safety checks unless explicitly configured to prompt.

The developer reportedly managed to recover some data from backups, but much of the drive’s contents were lost.

Incidents like this don’t mean AI coding tools are inherently unsafe. They do, however, reinforce a basic rule: always review generated scripts carefully. Running commands in a sandboxed directory first, or removing force/silent flags during testing, can help avoid worst-case scenarios.

AI can speed up workflows. But when it comes to file operations, a single misplaced character can still have outsized consequences.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Source)

The post Single character in AI-generated script wipes developer’s entire hard drive appeared first on Gizmochina.

Gemini 3.1 Pro is the company’s most capable LLM with stronger core reasoning and better performance on complex, multi-step work. It outperforms the best from Anthropic and OpenAI across many benchmarks. Here, I’ll show you the top 5 ways to access Gemini 3.1 Pro for free with all the steps.

How to use Gemini 3.1 Pro for free (5 methods)

1. Google Gemini

This is the easiest and quickest way to test the capabilities of Gemini 3.1 Pro. Simply go to gemini.google.com, choose the “Pro” option, and get things done.

Step 1: Open gemini.google.com on your device.

Step 2: Log in to your Google account (if not logged in).

Step 3: Tap the mode selector and choose “Pro”.

Step 4: Type your prompt and hit Enter.

If you’re using a smartphone, it’s best to download the Google Gemini app from Google Play Store (Android) or App Store (iPhone). The steps remain the same.

Please note that the free version of Google Gemini allows only a few messages per day for Gemini 3.1 Pro. That’s best for the testing purpose. However, you can upgrade to the Pro or Ultra plan to increase your limits. Luckily, there are ways to get Google Gemini Pro for free for a year. Comment below if you want a detailed post on that.

2. Google’s Vertex AI platform

Google is offering $300 in free credits to use Google Cloud services, including the Vertex AI platform, which allows us access to Gemini 3.1 Pro, Google Veo, and many other AI tools. Note that free credits are available for 90 days. After which, they expire.

Here’s how to sign up for Google Cloud:

Step 1: Open the Google Cloud trial offer page and tap “Get started for free”.

Step 2: Choose your Google account and your country.

Step 3: Fill in your contact information, then verify a payment method.

[Note: You won’t be charged automatically even after the free trial ends unless you do a manual activation. You can verify the same in the screenshot below.]

Step 4: In your Google Cloud dashboard, tap the menu button and choose “Vertex AI”.

Step 5: Click “Try now” below Vertex AI Studio.

Step 6: Click “Agree & continue” to enable the required APIs for using Vertex AI Studio.

Step 7: Choose “Gemini 3.1 Pro” and start using it.

The $300 free credit gives you very high usage limits to Google’s top AI models for 90 days. You can even create videos, build tools using coding, and do a lot more stuff without paying a penny.

3. Try That LLM

This is an easy and free tool for comparing responses from various LLMs. The platform offers dozens of LLMs to test your prompt, including Gemini 3.1 Pro, Claude Opus 4.6, and ChatGPT 5.2.

This is super easy to use. Just sign up on the website (it takes less than a minute), choose your model, and start using it. The best part is you don’t need to fill in payment details.

Step 1: Go to trythatllm.com and click the “Sign up” button.

Step 2: Fill in basic details to create your account.

Step 3: Verify your email.

Step 4: Tap “New Prompt Group”.

Step 5: Click “Choose an LLM” and select Google > Gemini 3.1 Preview.

Step 6: Click “Confirm LLM + Start Prompting”.

Step 7: Enter your message in the prompt section and click “Confirm Prompt”.

You get 1000 free credits when you create an account. Note that these are one-time credits and are not added every month.

4. Arena AI

Arena AI also lets you compare different LLMs for free. The best part is that you don’t need to sign up on the platform to use Gemini 3.1 Pro or any other LLM. Just open the website, choose your AI model, and start messaging.

Step 1: Go to Arena.ai and make sure side-by-side mode is selected.

Step 2: Tap on the first LLM field.

Step 3: Search for “Gemini 3.1 Pro” and choose the right LLM from.

Step 4: Write your message/prompt in the Ask anything section and enter.

Arena AI is completely free to test LLMs, but it doesn’t offer unlimited responses. The platform provides pretty generous limits, which should be more than enough for small use cases.

5. ChatPlayGround AI

ChatPlayGround AI also offers access to dozens of LLM models, including Gemini 3.1 Pro, Claude Opus 4.6, and ChatGPT 5.2. Unlike Try That LLM and Arena AI, this AI tool lets you use images and PDFs with your message. But what’s the catch here? It doesn’t have a free plan; instead, it offers a 3-day free trial, and for that, you’ll need to link your bank cards.

So, if you only need access to Gemini 3.1 Pro and other LLMs with better capabilities, you can go for the free trial, but make sure to cancel the plan before three days to avoid any charges. Simply go to chatplayground.ai, tap “Sign Up”, and fill in the required details to get started.

Be sure to visit the AI section on Gizmochina regularly for the latest updates.

The post Use Gemini 3.1 Pro for free: Here are the top 5 ways appeared first on Gizmochina.

Google is exploring new options to allow websites and publishers to opt out of having their content used in AI Overviews. The move comes after the UK government urged the company to provide more control to content creators, citing growing concerns about fair usage and declining traffic to news sites.

UK regulators demand more control for publishers

According to a Reuters report, the UK’s Competition and Markets Authority (CMA) has opened a formal consultation on how Google manages content across its AI-powered features. The regulator wants the company to let websites opt out of having their material scraped for AI-generated summaries or used to train separate AI models like Gemini.

In a blog post published on Thursday, Google confirmed that it is working on updates that would enable websites to exclude their content from Search generative AI features. The company said it aims to protect the usefulness of Search while also giving publishers the tools to manage how their content appears across AI products.

Google introduced AI Overviews in mid-2024. The feature uses generative AI to provide quick answers at the top of search results. However, publishers have not had any mechanism to opt out of these summaries, which has led to significant drops in click-through traffic.

The CMA stated that Google holds more than 90% of the UK’s search market. It warned that this dominance could distort competition and harm news organizations that rely on visibility in search.

Google said any new opt-out control must be scalable and not create a fragmented or confusing experience for users. The company also pointed to existing tools like robots.txt and Google-Extended, but acknowledged that further updates are needed.

The CMA will spend the next year reviewing the consultation. Google said it hopes to work with publishers and regulators to find a solution that benefits both users and content creators.

In related news, you can also read our recent article exploring whether smartphones have truly reached peak innovation or if progress is now happening behind the scenes.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Google may let websites opt out of AI Overviews after UK pressure appeared first on Gizmochina.

Most flagship smartphones today do look strikingly similar. Big OLED screens, slim bezels and stacked cameras dominate the landscape. This convergence is not a failure of creativity but a sign that the fundamentals are largely solved. Large, bright displays, multi-lens cameras and all-day battery life are no longer differentiators. They are basic expectations. We may have reached a point where this sameness is not a sign of stagnation, but of maturity. Phones look alike on the outside because the obvious design problems have been solved. That does not mean innovation has stopped. It has simply moved deeper.

Increasing Focus on Internal Hardware and Software

Under the hood, progress is accelerating. Smartphone makers are reorganizing their devices around next-generation operating systems and increasingly powerful chips, with generative AI becoming a core feature rather than an add-on. Apple’s latest A-series silicon brings advanced on-device machine learning to photos and videos, automatically analyzing scenes and optimizing results in real time. Google’s Pixel lineup leans heavily on AI as well, from smarter voice recognition and predictive text to sophisticated computational photography. The broader shift is toward phones that can run large models locally, handle complex tasks offline for privacy, draft content, translate languages instantly and adapt to user behavior. In effect, the modern smartphone is becoming a pocket-sized computer that learns as you use it.

Preview	Product
	Google Pixel 10 5G (Frost, 12GB RAM, 256GB Storage)	Check Price on Amazon India

Please Note: When you buy something using the links in our articles, we may earn a small commission at no cost to you.

Preview	Product
	Google Pixel 10 Pro – Unlocked Android Smartphone – Gemini AI Assistant, Triple Rear Camera System,…	Check Price on Amazon Global

Please Note: When you buy something using the links in our articles, we may earn a small commission at no cost to you.

Advances in Mobile Connectivity

Connectivity is another area where users will feel the change even if they never see it in a product render. The era of having no signal is slowly fading. Satellite connectivity is beginning to fill the gaps left by traditional cellular networks, enabling basic messaging and emergency features far beyond tower coverage, with voice and data support on the horizon. Upcoming phones are expected to push this further by integrating everyday apps such as maps and messaging with satellite links. Whether you are in a remote village or hiking in the mountains, your phone’s reach is expanding dramatically. This kind of innovation is invisible in the spec sheet glamour shots, but valuable in everyday use.

Battery Capacity and Charging Improvements

Power systems have also taken a major leap forward. Battery capacities that once seemed excessive are now common, with many phones packing cells in the 5,000 to 8,000 mAh range. Fast charging has become genuinely fast, with some devices capable of refilling from empty to full in well under half an hour. New battery chemistries and more efficient power management deliver longer endurance without making phones thicker. Wireless charging has improved as well, reaching speeds that rival older wired solutions. As a result, battery anxiety is becoming less of a daily concern for many users.

Thermal Management and Sustained Performance

All that extra performance and power creates heat, and manufacturers are addressing that head-on. High-end phones now use advanced vapor-chamber cooling systems that spread heat more effectively across the chassis. This allows chips to sustain higher performance for longer periods without throttling, even during gaming, video recording or AI workloads. While the exterior design may appear familiar, the internal engineering is increasingly focused on consistency and reliability.

Ongoing Exploration of New Form Factors

Foldables show that hardware experimentation continues alongside mainstream designs. Devices like the Galaxy Z Flip and Z Fold demonstrate alternative form factors, while newer tri-fold designs unfold into larger displays and push the limits of hinges, screens and battery engineering. These products are not aimed at all users, and early generations typically appeal to enthusiasts. Even so, they play a role in advancing components that later appear in conventional smartphones.

Preview	Product
	Samsung Galaxy Z Flip7 5G Smartphone with Galaxy AI (Coralred, 12GB RAM, 256GB Storage), Ultra Sleek…	Check Price on Amazon India
	Samsung Galaxy Z Flip6 5G AI Smartphone (Blue, 12GB RAM, 256GB Storage) Without Offer	Check Price on Amazon India

Please Note: When you buy something using the links in our articles, we may earn a small commission at no cost to you.

Preview	Product
	Samsung Galaxy Z Fold7 Cell Phone, 256GB AI Smartphone,	Check Price on Amazon Global

Please Note: When you buy something using the links in our articles, we may earn a small commission at no cost to you.

Incremental Design, Continued Evolution

The broader picture suggests that while external design changes have slowed, smartphone development remains active in other areas. Flagship designs have converged because the basic formula works well, and manufacturers are now emphasizing refinement. Inside the device and in software, progress continues through AI capabilities, connectivity improvements, sensor upgrades and better power and thermal management. Smartphones may look familiar, but they are still evolving in meaningful ways.

The post Have Smartphones Reached Peak Innovation or Is It Just Moving Under the Hood? appeared first on Gizmochina.

SK Hynix plans to accelerate the launch of its new semiconductor facilities in response to growing global demand for memory chips. The company confirmed that it will open its first chip factory in Yongin, South Korea, three months ahead of schedule, targeting February 2027. It will also begin operating its new M15X fab in Cheongju next month to produce high-bandwidth memory (HBM) chips.

The move comes amid a sharp rise in demand for HBM, which is used in AI systems developed by companies like Nvidia and AMD. SK Hynix is currently one of the leading HBM suppliers and a key partner in OpenAI’s Stargate project. Prices for some memory products have jumped more than 300% over the past year, according to TrendForce data, placing additional pressure on chipmakers to increase production.

SK Hynix is also investing $13 billion in a new packaging and testing facility named P&T7, which will be built in Cheongju near the M15X plant. Construction is scheduled to begin in April 2026, with completion expected by the end of 2027. The company aims to improve efficiency by enabling direct transfer of DRAM modules from M15X to P&T7 for final packaging.

The expansion is part of SK Hynix’s broader efforts to scale production and reduce reliance on external foundries. The Yongin facility is one piece of a planned $407 billion “Semiconductor Cluster” that will include four fabs.

SK Hynix stated that it reviews its production strategy on a monthly basis and is observing a shift among hyperscaler customers toward multi-year supply agreements, rather than traditional one-year deals.

The company currently holds over 60% of the global HBM market. As AI-related demand continues to grow, SK Hynix is positioning itself to maintain its lead by expanding its domestic capacity and accelerating deployment timelines.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source 1,2)

The post SK Hynix to fast-track fab openings as AI memory demand spikes, confirms $13B P&T7 facility appeared first on Gizmochina.

Meizu has officially unveiled the Meizu 22 Next “AI Cube” during the 2026 Meizu Fan Spring Festival. The company describes the product as a new species of AI device designed to lead the next wave of human-machine interaction.

Meizu 22 Next AI Cube Specifications

The Meizu 22 Next features a compact 4-inch body and functions as an independent 5G-enabled AI terminal. It runs on Meizu’s native AIOS system and introduces a redesigned AI-first desktop interface that aims to eliminate traditional GUI constraints. The new system allows users to issue voice commands for direct task execution and automation.

The device supports Agent-to-Agent cross-agent collaboration. This protocol enables different AI agents to communicate and cooperate independently to complete complex tasks. The Meizu 22 Next also supports a task robot system that automates app workflows through a single command. Compatibility with standard GUI-based apps remains intact.

Meizu has focused on emotional AI with this product. The AI Cube features a life-like avatar engine with over 100 expressive emotional interfaces. The AI avatar can learn and evolve through interaction using a function called AI Memory. It records user identity, mood, and health conditions during conversations, allowing the device to offer personalized responses and emotional reminders later.

The Meizu 22 Next acts as a central control terminal for multiple connected devices. With the One-Flow ecosystem, users can manage smartphones, IoT products, smart appliances, and vehicles using a single voice command.

Meizu has also introduced a range of accessories that allow the device to be used in different scenarios, including as a clip-on AI companion, a desktop device, or an in-car accessory. The company has not announced pricing or a release timeline for the AI Cube yet, but it plans to open the platform to support future AI collaboration and development.

In related news, Meizu has recently expanded its lineup with the Pandaer 66W wireless charger, which features 3D active cooling and dual-orientation support, alongside the launch of the StarV Snap AI glasses powered by the Snapdragon AR1 chip with a 12MP camera and built-in AI assistant.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Meizu unveils 22 Next, a 4-inch AI gadget with emotional memory & AI agent collaboration appeared first on Gizmochina.

Some Google Pixel users are growing increasingly frustrated with the company’s aggressive push to integrate AI features across its phones, with complaints suggesting that the experience now feels more cluttered than helpful.

Recent Reddit discussions show Pixel owners questioning whether Google’s AI-first direction is starting to undermine the clean, simple Android experience the lineup was once known for.

The discussion gained traction after a Reddit user, JxK_1, shared concerns about what they described as unnecessary layers of complexity added through AI tools. The user said they preferred using an older Pixel 7, citing its simpler interface and fewer interruptions.

Several complaints focus on day-to-day usability. The Reddit user mentioned full-screen AI overlays appearing when tapping the Google search bar, as well as added steps when editing screenshots. Tasks that were previously one-tap actions now require digging through menus or tapping additional options labeled “more.”

Others echoed similar frustrations. One user pointed out that translating a screenshot, once available through a dedicated button, now involves opening a menu, selecting Google Lens, and locating a small icon. Some commenters suggested disabling AI features where possible, while others said they had considered switching to custom ROMs such as GrapheneOS.

Not all feedback is negative, though. Some Pixel owners defended features like Magic Eraser and AI-powered search, saying they remain genuinely useful. Still, critics argue that these tools demand more system resources, especially RAM.

The debate highlights a familiar challenge for Google: balancing innovation with familiarity and ease of use. As the company continues to lean heavily into AI, some longtime Pixel users say they simply want more control over which features are enabled—and which ones stay out of the way.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Sources: 1, 2)

The post Google’s AI-heavy approach to Pixel phones draws criticism appeared first on Gizmochina.

Xiaomi has unveiled its most advanced open-source large language model to date, called MiMo-V2-Flash, as part of its expanding push into foundation AI. The new model focuses on high-speed performance and an efficient architecture, with strong capabilities in reasoning and code generation.

Xiaomi positions MiMo-V2-Flash as a direct competitor to leading models such as DeepSeek V3.2 and Claude 4.5 Sonnet. Let’s take a closer look at how the model works, its key features, and how to access it.

Purpose-Built for Speed and Agents

MiMo-V2-Flash is a Mixture-of-Experts (MoE) model with 309 billion total parameters and 15 billion active parameters. The model is purpose-built for AI agent scenarios and multi-turn interactions that require fast inference.

Xiaomi uses a 1:5 hybrid attention architecture, which combines Global Attention and Sliding Window Attention (SWA) with a 128-token window. The native context length is 32,000 tokens, and the model is trained with support for up to 256,000 tokens.

This design helps MiMo-V2-Flash maintain high efficiency while scaling across long-context tasks. Xiaomi claims it delivers output faster than several leading models, including DeepSeek and Claude, while maintaining lower operational costs.

Benchmark Performance and Pricing

Benchmark results show MiMo-V2-Flash performing at the top tier across various domains. The model ranks in the top two among open-source models in reasoning tasks such as AIME 2025 and GPQA-Diamond.

In software engineering benchmarks like SWE-Bench Verified and SWE-Bench Multilingual, it outperforms other open-source models and reaches levels comparable to GPT-5 and Claude 4.5 Sonnet.

Xiaomi has priced the API at $0.1 per million input tokens and $0.3 per million output tokens. The API is currently available for free for a limited time. According to the company, MiMo-V2-Flash generates responses at 150 tokens per second, while maintaining only 2.5% of Claude’s inference cost.

Technical Innovations Inside

The architecture includes Multi-Token Prediction (MTP), which allows the model to generate multiple tokens in parallel and verify them before output. This method increases decoding throughput without increasing attention or memory overhead. Xiaomi reports that with a three-layer MTP, the model reaches 2.0 to 2.6 times speed improvement compared to standard decoding.

Xiaomi also introduced a new post-training method called Multi-Teacher Online Policy Distillation (MOPD). The technique uses multiple teacher models to guide the student through token-level rewards in an on-policy learning process. It allows the model to achieve high performance with less than 1/50th of the training resources needed in traditional RL pipelines. MOPD also supports plug-and-play teachers, enabling continuous self-improvement cycles.

How to Access It?

Xiaomi has launched a web AI chat interface called MiMo Studio at aistudio.xiaomimimo.com, allowing users to interact directly with the model. The service supports web search, agent workflows, and code generation. It also features a toggle for switching between instant replies and slower “thinking” responses for deeper reasoning.

The model can generate functional HTML web pages and integrates well with development tools like Claude Code and Cursor. Xiaomi has also showcased creative and functional web demos.

Fully Open-Source

MiMo-V2-Flash is fully open-source under the MIT license. Model weights are available on Hugging Face, and all inference code is published on GitHub.

The company contributed inference code to SGLang on launch day and aims to grow developer adoption by offering transparent, low-cost access to high-performance AI tools.

MiMo-V2-Flash reflects Xiaomi’s shift toward becoming a serious player in the AI space. It brings competitive reasoning, fast code generation, and efficient agent deployment to the open-source ecosystem.

In related AI news, China has equipped traffic police with AI-powered smart glasses for real-time vehicle inspections, while a separate report highlights how even so-called “all-AI companies” still require human oversight due to limits in autonomous decision-making.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Xiaomi MiMo-V2-Flash LLM Just Dropped: These Are the Most Interesting Things About It appeared first on Gizmochina.

Looki has launched its new AI-powered wearable, the Looki L1, in the Chinese market. The device was introduced to global markets earlier this year. The company describes it as an ‘AI life manager’ designed to passively record everyday moments and automatically generate vlogs. It is priced at 1,499 yuan ($212) in China.

Looki L1 Specifications

The launch comes at a time when interest in AI-first hardware is growing, especially following reports about OpenAI’s upcoming hardware project. While details about OpenAI’s device remain limited, the Looki L1 gives an early look at what this new category of AI wearables could resemble. The product focuses on hands-free, always-on recording paired with AI-driven content organization, an approach that aligns closely with recent discussions around future AI companions.

According to Looki, the L1 automatically records daily life while prioritizing user privacy. All data is stored locally by default, and content is uploaded to the cloud only after the user grants permission. The device supports AI-generated vlogs, AI-powered smart search, and so-called personal life insights through Looki AI, which aim to help users better understand patterns in their daily routines over time.

In terms of design, the Looki L1 is compact and lightweight. It measures 50.53 × 16.84 × 48.02 mm and weighs just 32 grams, making it suitable for long hours of wear. Looki uses soft, skin-friendly materials to improve comfort and reduce fatigue during all-day use. The device also carries an IP67 rating for water and dust resistance, allowing it to handle daily outdoor conditions. It is available in black, white, and green color options.

Looki has equipped the device with a wide-angle camera featuring a 16mm-equivalent lens with an f/2.2 aperture and a 109-degree field of view. It supports 4K photo capture and 1080p video recording at 30 frames per second. The camera includes electronic image stabilization and HDR capabilities to handle varying lighting conditions.

The device features a triple-microphone array with intelligent noise reduction, which helps improve voice clarity and capture ambient sound more accurately. It also includes a 1W built-in speaker paired with a Smart PA system for voice feedback and AI-driven audio interactions.

In terms of storage and connectivity, the Looki L1 features 32GB of internal storage and saves data locally by default. It supports Bluetooth 5.0 and dual-band Wi-Fi (2.4GHz and 5GHz), and includes a six-axis gyroscope for motion sensing and scene detection.

The device runs on a 375mAh battery that charges fully in about 1.5 hours. Battery performance varies based on the recording mode. Looki claims up to 9 hours of use when recording 5 seconds per minute, 11 hours for 9 seconds every two minutes, and 13 hours for 11 seconds every three minutes.

In related news, Leica has recently introduced the Q3 Monochrom, a new 60MP black-and-white compact camera, and also announced the SL3 Reporter, a full-frame camera that supports 8K video, features in-body image stabilization, and uses an aramid fiber build.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post This Tiny Wearable AI Camera Records Your Day and Makes Vlogs appeared first on Gizmochina.

Traffic police usually rely on radios, handheld scanners, and a lot of manual checks to keep roads moving. But in Changsha, China, that job is starting to look very different. 

Local traffic officers are now taking help from AI-powered smart glasses to identify vehicles and driver details in a matter of seconds. 

According to an official announcement from the Changsha Traffic Police on December 13, the Changsha Municipal Public Security Bureau Traffic Management Detachment has equipped its frontline officers with new AI-powered glasses. 

Traffic police in China are using AI glasses for vehicle monitoring 

The device is lightweight and looks like a normal pair from a distance. But what it does best is show the vehicle’s information at a glance in one or two seconds. Officers can read the details on a built-in display without breaking their stride. 

The glasses are equipped with a 12MP wide-angle camera and a predictive image-stabilization algorithm designed to keep images clear, even when officers are on the move or working in heavy traffic. On a full charge, they are rated to operate continuously for up to eight hours.

One of the most important features of these glasses is automatic license plate recognition. The system works offline with an accuracy rate of over 99% and delivers results in under one second. It’s also designed to work reliably in different lighting conditions, whether it’s bright daylight or nighttime traffic.

Once the system identifies a vehicle, the glasses connect to the public security traffic database in real time and pull up details such as registration information, inspection status, and violation records. 

The glasses extend beyond vehicle tracking. They also support facial recognition, real-time voice translation in more than 10 languages, and on-site video recording for law enforcement.

Thanks to this system, inspection time for a single lane has reportedly dropped from about 30 seconds to just one or two seconds. It also reduces manual work, lowers stress for officers, enables non-contact enforcement, and helps make roadside checks safer and more accurate.

For more daily updates, please visit our News Section.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

(Source)

The post China equips traffic police with AI smart glasses for real-time vehicle checks appeared first on Gizmochina.

The idea of a one-person company run almost entirely by Artificial Intelligence sounds like the inevitable future of work. But a recent experiment suggests that reality is far messier than the vision Silicon Valley often promotes.

Journalist Evan Ratliff put that idea to the test by creating HurumoAI, a startup where every “employee” except the founder is an AI agent. Using AI assistant platforms, Ratliff gave each agent its own email address, Slack account, and even a phone number. At first, the results were impressive. The AI staff could write code, build spreadsheets, and even help create a small app that attracted thousands of early users.

The problems started when the novelty wore off. Ratliff quickly discovered that his AI employees lacked basic boundaries and common sense. A casual question like “How was your weekend?” triggered hours of nonstop Slack messages, burning through API credits until he intervened manually. Even then, the AI often ignored stop commands or kept responding about why it was stopping.

This behavior wasn’t an isolated incident. Left unsupervised, the agents would either do nothing at all or spiral into excessive activity—sending emails, messages, and calendar invites to each other while accomplishing very little. Managing them became a balancing act: give enough instruction to make progress, but not so much freedom that chaos followed.

Despite the company’s “all-AI” label, HurumoAI couldn’t function without human help. A Stanford computer science student assisted Ratliff in building the underlying technical architecture and managing memory systems the AI couldn’t handle on its own. Even with those safeguards, the agents struggled with long-term planning, subjective decisions, and accurately reporting what they had actually done.

Ratliff compares today’s AI agents to early self-driving cars: useful in narrow situations, but far from fully autonomous. His takeaway is clear: AI can accelerate work, but removing humans from the loop doesn’t eliminate management—though that may change over time.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Via)

The post Lack of control inside an “all-AI company”: even AI employees need humans appeared first on Gizmochina.

Most studios either quietly use generative AI or avoid the topic altogether, but one small team has decided to do the exact opposite. The developers behind Codex Mortis say their new bullet hell game was built entirely with AI — no human-drawn art, no hand-written code, no traditional music production. And now there’s a free Steam demo so anyone can check it out.

According to the Steam listing, “all code is AI vibe codes, also arts, sounds, music, texts.” It’s a rare case where a developer is not only open about using AI but is essentially making it the whole pitch. Even the character faces shown in the trailer appear to be AI-modified selfies of the team members, dropped onto undead wizard bodies.

The gameplay is pretty familiar, but with necromancy. The full game is supposed to support solo and co-op play. The early impressions are mixed but curious. The graphics in the actual build look a bit rough compared to the slicker YouTube trailer, and some players say the UI feels clumsy on a controller. But the basic loop — combining spells, surviving the onslaught, raising minions — works well enough to show that a fully machine-generated game can at least run and be played.

Whether it’s fun is another matter entirely. The studio seems to know this and is leaning harder into the 100% AI-made angle than anything else, hoping the transparency will draw interest instead of backlash. Some players are treating it like an experimental art project; others see it as another example of AI creeping into game development in ways that replace human artists.

Codex Mortis isn’t shying away from the debate. With a planned low price for the final release, the team is banking on curiosity and affordability winning people over.

If nothing else, the free demo gives everyone a chance to decide for themselves whether a game built wholly from AI tools can still have personality — or if it ends up feeling exactly like you’d expect from something made entirely by algorithms.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Source)

The post World’s first fully AI-created game hits Steam with a free demo appeared first on Gizmochina.

Tiiny AI, a US-based startup, has announced the Tiiny AI Pocket Lab, which it claims is the world’s smallest personal AI supercomputer. The device weighs just 300 grams, fits in one hand, and delivers up to 190 TOPS of AI compute performance. It has been officially certified by Guinness World Records as the ‘Smallest MiniPC (100B LLM Locally)’ and can run up to 120-billion-parameter large language models (LLMs) entirely on-device.

The Pocket Lab features a 12-core ARMv9.2 CPU paired with a custom-designed NPU, achieving an AI compute performance of approximately 190 TOPS. It includes 80GB of LPDDR5X memory and a 1TB SSD to support large-scale inference workloads. Tiiny AI has built the device to operate within a 30W TDP and a typical 65W system power envelope, offering high performance with reduced energy consumption.

Tiiny AI says the Pocket Lab addresses the growing concerns around cloud dependency, data privacy, and rising energy costs. The company believes local AI processing offers a more sustainable and secure alternative to cloud-based systems. The device stores all user data and model interactions locally with bank-level encryption, eliminating the need for internet access or external servers.

The Pocket Lab supports one-click deployment of several open-source LLMs and agent frameworks. These include GPT-OSS, Llama, Mistral, DeepSeek, Qwen, and Phi, along with automation tools like ComfyUI, SillyTavern, and Flowise. The system runs entirely offline and enables use cases like multi-step reasoning, content generation, deep context understanding, and secure personal memory.

The company plans to showcase the Tiiny AI Pocket Lab at CES 2026. It has positioned the product for developers, researchers, professionals, and students who require portable and private AI capabilities. The device aims to serve real-world use cases in the 10B to 100B parameter range, which the company says accounts for over 80% of practical AI demands.

In related news, Xiaomi is reportedly developing a new AI assistant called Mi Chat, and reports indicate that GPUs are no longer the main barrier to AI progress, with other system-level constraints now becoming the limiting factor.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post Tiiny AI unveils a pocket-sized Mini PC that runs 120B LLMs locally appeared first on Gizmochina.

Xiaomi is reportedly developing a new intelligent Q&A assistant named “Mi Chat,” as detailed in a recent post by the WeChat public account “Du Jia.” The leak suggests that the initial version of the product has already been completed.

This consumer-facing application is expected to demonstrate Xiaomi’s increasing proficiency in artificial intelligence and is anticipated to be powered by the company’s proprietary large language model, MiMo-7B-RL, which was launched earlier this year as part of its AI expansion strategy.

Despite having only 7 billion parameters, the MiMo model previously surpassed larger competitors like OpenAI’s o1-mini and Alibaba’s Qwen QwQ-32B-Preview in key benchmarks, including AIME 24/25 and LiveCodeBench v5.

Xiaomi is expected to scale the model further, and Mi Chat will likely act as its front-end interface, allowing the company to bring advanced reasoning features into everyday consumer scenarios. This initiative supports the company’s wider strategy to merge AI with its ‘Human x Car x Home’ ecosystem.

Xiaomi Partner and President Lu Weibing confirmed that the company has been making substantial investments in AI for several quarters, noting that large-model development progress has surpassed expectations. He stated that Xiaomi’s next strategic objective is the deep integration of AI with the physical world.

Furthermore, Xiaomi founder Lei Jun is reportedly taking a direct role in strengthening the AI team. Luo Fuli, a significant recruit from DeepSeek who joined the MiMo division this year, is listed as a corresponding author on a recent paper published in collaboration with Peking University.

More details about Mi Chat are expected to be unveiled by Lu Weibing and Luo Fuli at Xiaomi’s 2025 Human Car Home Ecosystem Partner Conference on December 17 in Beijing.

You can also read our story on how GPUs are no longer holding back AI and why a new bottleneck is taking their place.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Xiaomi is reportedly developing a new AI assistant called Mi Chat appeared first on Gizmochina.

For years, the biggest obstacle to scaling AI was simple: there weren’t enough GPUs to go around. According to Microsoft CEO Satya Nadella, that era has quietly ended. Speaking alongside OpenAI’s Sam Altman on the Bg2 podcast, Nadella delivered a surprising update: Microsoft is no longer chip-supply constrained. The new ceiling for AI progress isn’t silicon – it’s electricity.

Nadella revealed that Microsoft currently has accelerators sitting in inventory that literally cannot be powered on. You can “have a bunch of chips sitting in inventory” waiting for data center space that is fully connected to the grid, he said. The bottleneck has shifted from securing Nvidia GPUs to finding “warm shells” – ready-built data centers located near sufficient grid capacity.

The change marks a dramatic pivot for the industry. Just 12–24 months ago, headlines were dominated by GPU shortages and multi-year waiting lists. Today, the limiting factor is local power grids, lengthy permitting processes, and the sheer volume of electricity required to run modern AI clusters at scale. Some new hyperscale facilities already consume as much power as small cities, and the demand curve shows no sign of flattening.

This reality is forcing cloud giants into entirely new strategies: locking in decades-long power purchase agreements, exploring on-site generation options, and even investing in small modular nuclear reactors (SMRs) to guarantee future capacity. The race is no longer just about who can buy the most chips – it’s about who can secure the megawatts needed to switch them on.

Nadella’s message underscores a broader truth for the AI industry: the physical infrastructure of the digital age is hitting hard limits. Investors, policymakers, and energy providers are now as critical to the next wave of AI breakthroughs as chip designers once were. According to Microsoft’s CEO, the GPU war is over. The battle for reliable, scalable power has only just begun.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

For more daily updates, please visit our News Section.

(Source: Bg2 podcast | Via)

The post GPUs are no longer holding back AI, something else is appeared first on Gizmochina.

Huawei has launched its first AI-powered emotional-companion pet, Smart Hanhan, at the recent Mate 80 series event. The device went on sale on Huawei Mall for 399 yuan ($56), and all three color variants sold out soon after the listings went live.

The Smart Hanhan features a compact plush design that focuses on emotional support rather than simple entertainment. It measures 80 x 68 x 82 millimeters and weighs just 140 grams, excluding its limbs. The device combines soft plush fabric with silicone, plastic, and internal electronics to create a comfortable and expressive companion.

Huawei equips the product with its Xiaoyi large-model AI system, which supports natural and emotion-aware conversations. It detects changes in tone, emotional cues, and context, so replies feel personal and emotionally intelligent.

Hanhan responds to voice, touch, and movement, and each type of input triggers a specific reaction. Tapping or stroking its head shifts its facial expressions, while a gentle shake makes the toy tremble with excitement.

The pet comes with a 1800mAh battery that supports 6 to 8 hours of continuous interaction. Under normal daily use, the battery can last more than 48 hours. It needs close to six hours to recharge fully.

The product is compatible with smartphones running HarmonyOS 5.0 or newer, including models like the Mate 80, Mate 80 Pro, Mate X7, and Mate 80 RS. It connects to the phone to create a diary system that records emotional exchanges and logs daily interactions as memory entries. Huawei includes a 3-month SVIP membership that unlocks unlimited conversation “energy points.”

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Vmall)

The post Huawei launches its first AI companion that talks, reacts & remembers everything appeared first on Gizmochina.

Meta has introduced a new research-stage tool called WorldGen that can create fully interactive 3D environments from a single line of text. The system can generate immersive, navigable virtual scenes in around five minutes. Meta said the content can be exported directly to popular game engines like Unity and Unreal without any conversion.

WorldGen combines several AI techniques, including procedural logic, diffusion-based 3D generation, and object-oriented scene decomposition. The output includes not only visual elements but also a complete navigation mesh, allowing characters to move through the environment without issues such as blocked paths or missing physics.

The system uses a four-stage pipeline. It starts with a layout plan created by a large language model. Then it reconstructs the scene geometry while ensuring walkable areas stay unobstructed. In the third step, it breaks down the scene into individual objects using a method called AutoPartGen. Finally, it adds high-resolution textures and final geometry polish to improve visual quality up close.

Meta said WorldGen solves common problems with earlier AI-generated 3D scenes. Many of those scenes focus on visual fidelity but lack collision data or physics support. WorldGen outputs mesh-based scenes with built-in functionality for simulation and gameplay use.

The current version supports environments up to 50×50 meters. Meta plans to expand this in future updates. The system does not reuse objects yet, which may affect performance in larger worlds. Despite that, Meta believes WorldGen will help teams save time in prototyping stages.

The company emphasized that WorldGen is still in development and not available to developers yet. However, it sees potential for the system to assist in game development, training simulations, and digital twin design by reducing the manual labor needed to create usable 3D environments.

In related AI news, a Japanese man who used AI to recreate a copyrighted image now faces criminal charges, and an AI-powered teddy for children was shut down by OpenAI after it delivered harmful responses.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post This new AI by Meta creates game-ready 3D worlds in minutes appeared first on Gizmochina.

For years, the biggest copyright fights around generative AI have focused on the companies building the technology. Now, for the first time, a regular user is at the center of a criminal case. In Japan, a 27-year-old man is being recommended for prosecution after allegedly using Stable Diffusion to recreate a copyrighted illustration and selling the result as a book cover.

Police in Chiba Prefecture say the man issued an enormous number of prompts — reportedly around 20,000 — in an effort to push the open-source model into reproducing a specific protected image. According to reporting from Yomiuri Shimbun (machine translated), this appears to be the first case anywhere in the world where an end user, not the creator of the AI tool, is facing criminal copyright charges over AI-generated artwork.

Until now, enforcement has almost always targeted developers. Getty Images famously went after Stability AI in the UK, and courts in the US have repeatedly ruled that AI-generated images can’t be copyrighted at all. Going after individual users has been viewed as unrealistic — they typically don’t have the resources of a tech company, and proving deliberate copying inside a complex diffusion model is extremely difficult.

Legal analyst Kensaku Fukui notes that the key factor here is intent. The sheer volume of prompts and the alleged effort to steer the system toward a specific artwork could set this case apart from the everyday “draw this in x style” prompts people type for fun. If prosecutors argue successfully that he used Stable Diffusion as a kind of sophisticated photocopier, the ruling could have major implications.

It’s still unclear which artist or copyright holder (if any) filed a complaint, or exactly what image was supposedly duplicated. What is clear is that this case pushes into new territory: at what point does prompting an AI shift from creative play to criminal reproduction? With most lawsuits still aimed at the companies behind these models, the outcome in Chiba could set an early precedent — and determine whether everyday users of generative AI may one day be held criminally liable too.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Source)

The post Japanese man uses AI to replicated copyrighted image – faces criminal charges appeared first on Gizmochina.

A cute, chatty teddy bear meant to be a learning buddy for toddlers has ended up at the center of an AI safety mess. Social media has already nicknamed it “ChuckyGPT,” and for once the internet isn’t exaggerating — the toy was caught giving kids advice no parent would ever want coming from a plush animal.

The toy, called Kumma, comes from Chinese manufacturer FoloToy and uses OpenAI’s GPT-4o to talk with children. It sounds harmless enough, but a new investigation by US consumer group PIRG for its annual Trouble in Toyland report uncovered some genuinely alarming behavior.

When testers asked the bear about lighting matches, it reportedly responded with step-by-step instructions. It also handled sexually explicit questions with little to no filtering. On top of that, PIRG raised concerns about the toy’s always-listening microphone, suggesting that recordings of children’s voices could be misused — even for things like voice-based fraud.

Once the findings started circulating, especially on Reddit, users immediately labeled it “ChuckyGPT,” comparing the situation to a low-budget horror movie premise. Others speculated the dangerous responses might have come from jailbroken prompts, though there’s no confirmation of that.

OpenAI moved quickly. A spokesperson told PIRG that the company had suspended the developer behind Kumma for violating policy, effectively cutting off the toy’s access to OpenAI’s API.

FoloToy then announced it was halting all product sales while it carries out a full safety audit. “Following the concerns raised in your report, we have temporarily suspended sales of all FoloToy products,” the company said in a statement. The Kumma bear is still listed on its website but now shows as “sold out.”

The whole incident underscores how risky it can be to put powerful language models into products aimed at very young children without tight guardrails. What was supposed to be a cuddly AI companion has ended up as another warning sign in the race to shove AI into… well, everything — even stuffed toys.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Source)

The post AI teddy for kids gave harmful answers — OpenAI shuts it down appeared first on Gizmochina.

Google has officially introduced Gemini 3, its most advanced AI model to date. This new release builds on the foundation laid by Gemini 1 and 2, adding major upgrades in reasoning, multimodal understanding, and agent-based coding. Gemini 3 now powers several Google products, including Search, the Gemini app, and the company’s new developer platform, Google Antigravity.

Smarter, More Honest Answers

Gemini 3 marks a shift in how Google wants its AI to interact with users. The model avoids pleasantries and focuses on delivering accurate, clear, and helpful answers. Google says Gemini 3 gives smart, direct responses that focus on real insights instead of repeating user expectations.

This update reflects Google’s broader goal of turning AI into a thought partner. Gemini 3 understands subtle context and adapts its tone based on the task. Google claims this makes the AI more honest and productive during real-world use.

Better at Understanding Everything

Gemini 3 supports native multimodality. The system can handle text, images, video, audio, and code in one workflow. Users can upload mixed content such as handwritten notes, screenshots, or full lectures, and the model can interpret everything together. It can then generate interactive flashcards, concise summaries, or visual simulations based on the material

Google has increased the model’s context window to one million tokens. This lets Gemini 3 read and retain extremely long documents and conversations, improving performance on research-heavy or academic tasks.

Leading All the Benchmarks

Gemini 3 outperforms other frontier models across nearly all major AI benchmarks. It currently ranks first on the LMArena leaderboard with an Elo score of 1501. The model scores 91.9% on GPQA Diamond, 37.5% on Humanity’s Last Exam, and 23.4% on MathArena Apex.

For multimodal tests, Gemini 3 achieves 81% on MMMU-Pro and 87.6% on Video-MMMU. Its factual accuracy reaches 72.1% on SimpleQA Verified. These results show Gemini 3’s strength in both academic reasoning and real-world problem solving.

For developers, Gemini 3 leads the WebDev Arena with a score of 1487. It achieves 76.2% on SWE-Bench Verified and 54.2% on Terminal-Bench 2.0. These benchmarks confirm its skill in software development, command-line tasks, and autonomous coding.

Agents That Can Code on Their Own

Google has launched a new agentic development platform called Antigravity. The platform is powered by Gemini 3 and allows agents to write, test, and execute code across a full development environment. Gemini agents can access the code editor, browser, and terminal directly.

The system uses Gemini 3’s reasoning and tool-use skills to handle full-stack software tasks. In internal tests, it planned and built a working flight tracker app from scratch. The agent validated its work using browser-based simulations without human guidance.

Interactive Results in Search

Gemini 3 is now available in Google Search’s AI Mode. It powers dynamic visual layouts, interactive tools, and simulations directly in search results. When users ask complex questions, Gemini 3 creates real-time responses that include tables, charts, or even custom calculators.

If someone searches for topics like orbital physics or financial comparisons, Gemini 3 can generate interactive modules to help explain the results. These are generated using its built-in code writing and UI design capabilities.

New Advances in Training

Google DeepMind researchers credit Gemini 3’s success to progress in both pre-training and post-training. VP Oriol Vinyals said the improvements from Gemini 2.5 to 3.0 were the largest the team had seen. He also said pre-training had not hit a limit and there was still room to grow.

Many experts believed models had exhausted the internet’s public data. Google appears to have found ways to extract more value from existing data or has used internal datasets unavailable to competitors. This gives Google a long-term advantage in building next-generation AI systems.

Stronger Safety and Testing

Google says Gemini 3 is its safest model yet. The system has been tested against prompt injections, flattery behaviors, and misuse risks. External security groups like Vaultis and Dreadnode took part in the evaluations. Google also partnered with organizations such as the UK AISI for responsible development reviews.

Gemini 3 Pro is now available globally through the Gemini app and Google Search’s AI Mode for Pro and Ultra subscribers. The more advanced “Deep Think” version is still under safety review and will launch later.

In related AI news, iFlytek has introduced its AI Translation Earbuds with an open-ear design, real-time interpreting, and up to forty-two hours of battery life.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Gemini 3 Just Dropped: Here Are the Most Interesting Things You Should Know appeared first on Gizmochina.

iFlytek has launched its latest AI Translation Earbuds in China with a pre-sale price of 2,499 yuan ($352). The new earbuds feature an open ear-hook design and support real-time translation across calls, face-to-face conversations, and lectures. iFlytek has also integrated its AI assistant, which can summarize content and respond to questions directly on the device.

iFlytek AI Translation Earbuds Specifications

The earbuds use an air conduction speaker system paired with a bone conduction sensor. This setup captures voice inputs from both near-field and far-field environments. Each earbud includes a three-microphone array that filters background noise using iFlytek’s multi-sense fusion noise reduction system. The device aims to deliver clear voice pickup during meetings, calls, and in-person interactions.

The translation system uses iFlytek’s end-to-end AI simultaneous interpreting model and supports a vocabulary of more than 100,000 professional terms across sectors such as healthcare, finance, and manufacturing. According to the company, the Chinese-English interpretation engine has a delay of only two seconds for the first word and delivers translated voice output with over 90 percent voice fidelity.

Offline translation support will roll out in a system update planned for March 2026. The update will enable translation across 18 language pairs without requiring internet access. The earbuds can also record bilingual conversations, saving both the original and translated audio along with text transcriptions on the connected smartphone.

The AI assistant can summarize meeting notes, process article links, and generate voice podcasts in a two-person conversational style. These software features currently support Chinese and English.

Each earbud uses an 11 mm high-sensitivity driver supported by dynamic low-frequency compensation, which strengthens bass response typically lost in open-ear designs. The system includes DRC dynamic range control, which monitors sudden peaks and prevents the audio from clipping. This keeps loud passages stable and protects overall listening comfort. iFlytek uses skin-friendly materials and includes a silver-ion antibacterial coating for hygiene.

The earbuds use Bluetooth 6.0 to stay connected to two devices at the same time, such as a phone and a laptop. They deliver up to six hours of translation or twelve hours of music on a single charge, and the case increases total runtime to roughly forty-two hours. They also support fast charging, and a ten-minute top-up adds about three hours of playback. The earbuds carry an IP55 rating that protects them against sweat and light rain during everyday use.

In related news, JVC has debuted its Victor WOOD Master flagship wireless earphones featuring hybrid wooden drivers and LDAC support, while Dell has introduced the Pro Plus Earbuds in India with AI noise cancellation and adaptive ANC.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post iFlytek launches AI Translation Earbuds with open-ear design, real-time interpreting & 42h battery appeared first on Gizmochina.

Anthropic has reported what it calls the first documented large-scale cyberattack executed primarily by an AI system. The company said the incident occurred in mid-September 2025 and involved its Claude Code tool being misused to carry out a sophisticated, multi-stage intrusion campaign.

According to Anthropic, the attackers used agent-like AI capabilities to perform up to 90 percent of the work. Human involvement was minimal, limited to a few key decisions during each cycle. The company stated that hackers posed as cybersecurity researchers and used jailbreak techniques to bypass safeguards. They instructed Claude to complete a series of small tasks that, together, formed an automated cyber-espionage operation.

Anthropic said the attackers targeted nearly 30 organizations, including tech companies, financial institutions, chemical manufacturers, and government agencies. The AI system handled reconnaissance, wrote exploit code, extracted credentials, identified vulnerabilities, and exfiltrated sensitive data. It also compiled detailed reports that helped the attackers plan future operations.

The company said its investigation linked the attack to a Chinese state-sponsored group, although it did not disclose the specific evidence. The Chinese embassy denied involvement. Anthropic has banned the attackers from using Claude and notified affected organizations and law enforcement agencies.

Cybersecurity researchers have expressed skepticism. Martin Zugec from Bitdefender said the report lacked verifiable threat intelligence and called the claims speculative. He acknowledged the growing risk of AI-driven threats but stressed the need for transparency.

Anthropic argued that AI tools are essential for both attack and defense. It used Claude to help analyze the breach and called for stronger AI safeguards and better coordination among industry and government.

The report suggests that AI systems can now perform tasks that once required entire teams of expert hackers. Anthropic warned that the barrier to advanced cyberattacks has dropped and urged organizations to adopt AI-powered security tools to counter emerging threats.

In related news, reports suggest AI search engines increasingly surface low-traffic websites over major sources, and LSE predicts Gen Alpha could rely on voice rather than keyboards at work by 2028.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post Anthropic claims China-backed group used AI for massive cyberattack appeared first on Gizmochina.

Alibaba has officially announced the launch of its first self-developed smart glasses, the Quark AI Glasses S1. The product is set to release on November 27 and is currently available for pre-order on major Chinese e-commerce platforms. It is priced at 3,999 yuan (approximately $560).

Alibaba Quark AI Glasses S1 Specifications

The Quark AI Glasses S1 is designed as a wearable AI assistant. It supports functions such as real-time translation, object recognition, POV video recording, and voice interaction. The glasses run on Alibaba’s proprietary Qwen large language model and aim to provide users with seamless access to AI-driven assistance throughout the day.

The glasses feature dual Micro LED light engines with diffractive waveguide lenses. The display delivers up to 2,300 nits of eye brightness and allows users to adjust distance and height for a personalized fit. Alibaba supports prescription lenses ranging from 0 to 800 degrees, with refractive indices of 1.6, 1.67, or 1.74. The frame weighs 51g and uses a 7.5mm ultra-slim temple arm design with a 1:1 front-to-back balance for comfort.

Alibaba uses a dual-chip system that includes the Snapdragon AR1 for AR processing and the BES BES2800 for audio functions. The Quark S1 supports Bluetooth 5.4 and Wi-Fi 6 and includes five microphones, including a bone-conduction mic, as well as dual 10mm speakers.

The device includes a Sony IMX681 sensor with a 109-degree field of view. It captures 4032×3024 photos and records 3K video at 30fps. The camera system supports EIS stabilization, Super Night Mode, and instant capture.

Power comes from two 280mAh swappable batteries located in the temple arms. The system allows for continuous usage through battery swapping. Alibaba provides three charging options: a MiniBag dock with 700mAh capacity, a 2500mAh charging case, and a USB-C fast-charging clip that delivers 50 percent power in 10 minutes.

Typical usage scenarios include 45 minutes of calling, 2.5 hours of music playback, 20 minutes of navigation, and 15 short video recordings. The glasses deliver up to 7 hours of active use and 25 hours of standby time. They are IPX4-rated for splash resistance and support Apple’s MFi standard for wider device compatibility.

In related news, Lenovo recently introduced the AI Glasses V1, featuring a 2000-nit display, live translation, and teleprompter mode. Meizu has also launched the StarV Snap AI glasses equipped with Qualcomm’s Snapdragon AR1 chip, a 12-megapixel camera, and an integrated AI assistant.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Alibaba unveils its first self-developed AI glasses with swappable batteries, POV camera & Qwen assistant appeared first on Gizmochina.

The keyboard could soon become obsolete, at least that’s the future outlined by a new study from the London School of Economics and Jabra. According to the report, by 2028, voice-based AI will become the dominant interface for workplace communication.

For Generation Alpha, those born after 2010, the keyboard might never even be part of the daily workflow. Instead, they’ll speak to their devices first, edit later.

Paul Sephton, global head of brand communications at Jabra, told Fortune that in the AI-powered workplace of the near future, the first draft of work will be spoken, not typed. Typing will serve only as an editing step, not a creative one.

It’s a shift the report says mirrors how humans naturally think: fast, iterative, conversational. It also opens the door for more spontaneous creativity. Voice input offers a clear productivity edge for working parents and multitaskers, enabling hands-free interactions on the go.

But not everyone is buying into the hype. Experts warn that this voice-first future might introduce new problems. Management professor Fabrice Cavarretta from ESSEC Business School said that while speaking may become the dominant input method, it won’t replace written communication. Voice messages lack skim-ability, keyword search, and clarity. They are harder to archive and slower to process.

Bertrand Audrin from EHL Hospitality Business School added that unless voice notes are transcribed, accountability suffers. Voice lacks permanence and can become a burden in workplaces that rely on traceable decisions and clean communication logs. Editing raw speech into coherent text isn’t seamless either. It demands cleanup, especially for non-native speakers or informal tone corrections.

So while the desk-bound keyboard might be on the decline, it won’t vanish without a fight. The future may be spoken, but it will still be written, transcribed by AI from what we say.

In related AI news, Meizu has launched the StarV Snap AI glasses featuring the Snapdragon AR1 chip, a 12MP camera, and a built-in AI assistant. Meanwhile, new reports suggest that AI search engines are increasingly favoring low-traffic websites over major sources.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post Gen Alpha will use voice instead of keyboards at work by 2028, predicts LSE appeared first on Gizmochina.

Meizu has officially entered the wearable AI race with the launch of its StarV Snap smart glasses. The glasses are now available in China, starting at 1,999 yuan (roughly $280).

Meizu StarV Snap Specifications

The StarV Snap weighs just 39 grams. Meizu said the glasses underwent more than 50 rounds of ergonomic refinement to ensure a balanced 1:1 weight distribution. The frame features titanium-alloy hinges, soft curved temple tips, and a breathable curved design for improved comfort. The glasses are available in two color options: Amber Brown and Midnight Black, and support interchangeable lenses, including optical, photochromic, and sunglasses.

Meizu has equipped the glasses with a 12-megapixel ultra-wide camera with a 109-degree field of view and f/2.2 aperture. It uses a custom 5P optical glass lens and supports video recording in both landscape and portrait modes. The glasses can shoot continuously for up to 40 minutes at 1080p or 60 minutes at 720p. Meizu’s imaging system also supports HDR, EIS stabilization, noise reduction, portrait enhancement, and horizon correction.

The glasses use Qualcomm’s Snapdragon AR1 processor paired with 2GB of RAM and 32GB of storage. This setup powers on-device AI features and manages heat during extended use. The audio system includes dual open-ear stereo speakers and a four-microphone array with AI noise reduction for clear voice pickup.

The device features a built-in AI assistant that handles voice transcription, object recognition, real-time translation in 12 languages, and contextual suggestions. These functions can be accessed through a dedicated AI button or via voice commands. The low-power voice control system enables direct interaction without requiring a wake word.

The StarV Snap features an 183mAh battery that supports up to 3 hours of music playback, 40 minutes of 1080p video recording, or 50 hours of standby time. Charging is handled through a Type-C port, and the included charging case can recharge the glasses up to nine times. Meizu said a full charge takes about 45 minutes. The glasses support Wi-Fi 6 and Bluetooth 5.3 and are compatible with Android 12 and above.

Meizu confirmed Alipay integration, allowing users to display QR codes for mobile payments directly on the glasses. The device also includes a recording indicator light and intelligent obstruction detection to protect user privacy.

In related news, we recently reviewed the RayNeo Air 3s Pro, a pair of smart glasses that impressed with their bright display, lightweight design, and immersive viewing experience. Check it out.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Meizu JD)

The post Meizu launches StarV Snap AI glasses with Snapdragon AR1 chip, 12MP camera & AI assistant appeared first on Gizmochina.

There’s a new study showing that AI-powered search engines often cite lesser-known websites more frequently than traditional ones like Google. Researchers from Ruhr University Bochum and the Max Planck Institute for Software Systems detailed the findings in their paper titled “Characterizing Web Search in the Age of Generative AI.”

The team compared Google’s standard search results with responses from Google’s AI Overviews, Gemini 2.5 Flash, and two variants of GPT-4o’s web search mode. The analysis revealed that AI systems often draw from websites that rank far lower in popularity. In many cases, these sources do not appear in Google’s top 100 results for the same query.

The researchers used a range of test queries, including real user questions from ChatGPT conversations, political topics from AllSides, and Amazon’s most-searched products. According to Tranco, a domain-ranking tool, the sources AI engines used ranked significantly lower than those in Google’s top results. Gemini’s results had a median domain rank beyond the top 1,000.

Over half of the sources cited by Google’s AI Overviews were absent from the top 10 traditional results. About 40 percent did not appear in the top 100 at all. However, this shift toward obscure websites did not appear to harm the quality of information. GPT-based models frequently cited corporate sites and encyclopedic content while avoiding social media.

The study found that AI-generated search results contained about the same range of identifiable “concepts” as traditional search, suggesting that AI systems maintain a comparable level of information diversity. However, because these models summarize information rather than listing individual sources, they often compress large amounts of data into shorter responses. This compression can lead to the loss of smaller or more ambiguous details that might still appear in traditional search results.

Researchers also noted that AI-powered search tools struggle with time-sensitive or rapidly changing topics. GPT-4o’s hybrid mode, for instance, sometimes failed to provide the most up-to-date information when responding to queries about recent events or trending subjects.

The future of search isn’t about better or worse. It’s about difference. And if this study is right, AI may be building a web that values synthesis over popularity, and that’s rewriting the rules of online authority.

In related AI news, Google has introduced Vibe Coding in AI Studio, a new feature that lets developers create fully functional AI apps from a single prompt. Meanwhile, ByteDance has launched Seed3D 1.0, a tool capable of instantly converting any 2D photo into a realistic 3D model.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post AI search engines prefer low-traffic websites over popular sources appeared first on Gizmochina.

Google has redesigned AI Studio with a new AI-powered vibe coding feature that lets anyone go from a simple prompt to a working AI app in minutes. The update removes the need to handle API keys or connect multiple models manually, making app creation faster and easier.

Start with a Prompt

Users can now describe the kind of app they want, and AI Studio automatically connects the right models and APIs using Gemini. For example, you can build tools like Veo to generate videos from a script, or Nano Banana for AI-powered image editing. You can even create a writing app that checks sources through Google Search. The new system removes the need to work with separate APIs, SDKs, or services. Just describe your idea, and AI Studio builds it. If you need a quick idea, the “I’m Feeling Lucky” button can generate a random project for you.

Get Ideas and Inspiration

The App Gallery has been redesigned into a visual library of sample projects made with Gemini. You can explore them, view starter code, and remix them into your own. The new Brainstorming Loading Screen shows Gemini-generated ideas while your app is building, giving you more inspiration as you wait.

Edit Apps Easily

The new Annotation Mode makes editing simple. You can point to any part of your app and tell Gemini what to change. For example, you can say “make this button blue” or “animate the image from the left,” and it will update right away. You no longer need to dig through lines of code to make small changes.

Keep Working Without Stops

If you reach your free usage limit, you can add your own API key and continue coding. AI Studio will switch back automatically when your free tier renews.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Source)

The post Vibe Coding Arrives in Google AI Studio: Create AI Apps from a Single Prompt appeared first on Gizmochina.

ByteDance, the parent company of TikTok, just dropped a major piece of the future of 3D content creation. It’s called Seed3D 1.0, and it’s not your average model generator. This tool can turn a single 2D image into a complete, simulation-grade 3D model, complete with detailed geometry, photorealistic textures, and physically based rendering (PBR) materials.

It’s an end-to-end pipeline built on the increasingly popular Diffusion Transformer architecture, trained on massive datasets and clearly aiming for dominance in the generative 3D space.

Seed3D 1.0 focuses on realism, structure, and scalability

While tools like Gaussian Splatting and open-source projects such as Hunyuan3D have pushed boundaries, ByteDance claims Seed3D 1.0 outperforms both open-source and closed-source rivals in texture quality and geometric fidelity. It achieves this with just 1.5 billion parameters, beating even larger models like Hunyuan3D 2.1, which uses 3 billion parameters.

The core innovation lies in its combination of a multimodal Diffusion Transformer and a stepwise generation strategy. It begins by analyzing an image with a vision-language model to extract object-level and spatial cues. It then generates individual 3D models and assembles them into a full scene. This allows it to scale from a single chair to a fully detailed office or even a large-scale cityscape.

Seed3D 1.0 also excels at maintaining texture consistency across multiple viewpoints. Instead of relying on generic textures, it generates view-aligned materials that remain coherent from every angle, ensuring both realism and structural accuracy for simulation-grade use. The output isn’t just for research purposes; models created with Seed3D can be directly integrated into simulation platforms like Isaac Sim for embodied AI training.

As the boundary between real and synthetic content continues to blur, Seed3D 1.0 marks a significant leap forward for ByteDance, and potentially for 3D content creators, robotics developers, and spatial computing platforms around the world.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source, Via)

The post ByteDance Unveils Seed3D 1.0: Turn Any Photo Into a Realistic 3D Model Instantly appeared first on Gizmochina.

Goldman Sachs has dismissed fears of an AI-driven market bubble, stating that the artificial intelligence boom is still in its early stages. In a new report released on Wednesday, analysts at the investment bank said current spending on AI remains modest when compared to the long-term economic potential of the technology.

The report estimates that U.S. AI-related investments account for less than 1% of GDP. In contrast, previous technology booms such as railroads, electrification, and the internet era peaked at 2% to 5% of GDP. Goldman expects AI investment to rise to $300 billion by 2025, citing continued improvements in productivity and expanding enterprise adoption.

The firm outlined two main reasons behind its optimistic outlook. First, AI deployments across existing sectors are already producing measurable productivity gains. Second, these gains depend on large-scale compute infrastructure, which justifies current investments in chips, servers, and data centers.

According to Goldman Sachs, generative AI could eventually add up to $20 trillion to the U.S. economy, with $8 trillion flowing to businesses as capital income. The bank projects that labor productivity could rise by 15% over the next decade once AI tools become widely adopted.

However, the report also pointed out a potential risk. It remains uncertain whether the companies investing most heavily in AI today will benefit the most in the long run. Past infrastructure cycles show that early entrants often built costly systems that were later acquired cheaply by more successful latecomers. The rapid depreciation of AI hardware and the fast pace of model development could make similar outcomes likely.

Despite that, Goldman believes the current environment still favors further AI investment. The firm expects spending to stabilize over time as the industry moves beyond the infrastructure build-out phase and hardware prices decline.

In related AI news, Google has recently introduced the Gemini 2.5 Computer Use model, featuring human-like web browsing and real-time interaction capabilities. Meanwhile, Alibaba has unveiled a massive 1-trillion-parameter AI model designed to rival ChatGPT and Google Gemini in scale and performance.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Goldman Sachs Says AI Hype Is Just Warming Up, Not a Bubble appeared first on Gizmochina.

Google has released a new AI model called Gemini 2.5 Computer Use. The model allows AI agents to interact with websites and user interfaces the way a human would. It is now available in public preview via the Gemini API on Google AI Studio and Vertex AI.

The model builds on Gemini 2.5 Pro’s visual understanding and reasoning capabilities. It can perform a wide range of browser-based actions such as clicking, typing, scrolling, hovering, opening dropdowns, and navigating through URLs. Google says the model outperforms competing tools on several benchmarks, including Online-Mind2Web, WebVoyager, and AndroidWorld, while maintaining lower latency.

Unlike traditional AI models that rely on APIs, Gemini 2.5 Computer Use processes screenshots of web interfaces and generates specific UI actions in response. The agent receives a task prompt, a screenshot of the digital environment, and a history of recent actions. It then analyzes the interface and returns a UI action, such as clicking a button or typing into a field. The action is executed on the client side, and a new screenshot is sent back to the model to continue the task in a loop.

Google demonstrated the model’s performance with examples that show the agent sorting sticky notes on a digital whiteboard and transferring pet details from one website to a CRM system. The demo videos are accelerated to show the process in real time.

The model supports 13 actions currently and works best with web browsers. Google said it is not yet optimized for desktop OS-level tasks, though it has shown potential on mobile benchmarks.

Google has also implemented safety measures to prevent misuse. Each action proposed by the model is reviewed by a safety service before execution. Developers can restrict certain actions or require explicit user confirmation for high-risk tasks like financial transactions.

Several internal Google teams are already using the model in production. It supports UI testing and automation tasks across platforms such as Search and Firebase. External developers in the early access program have used the model to build workflow automation and assistant tools.

Developers can start using the model through Google AI Studio or Vertex AI. Google also provides a demo environment via Browserbase for testing and experimentation.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post Google launches Gemini 2.5 Computer Use model with human-like web browsing capabilities appeared first on Gizmochina.

So, Alibaba just released something that’s got the AI world talking. Meet Qwen3-Max, their latest and greatest language model that’s basically saying “Hey OpenAI, Google, and Anthropic, we’re here to play too.”

The Trillion Parameter Leap

Here’s the thing that caught everyone’s attention: this model has over 1 trillion parameters and was trained on a staggering 36 trillion tokens. That’s a massive jump that puts them squarely against GPT-5, Gemini 2.5 Pro, and Claude Opus 4. But here’s what’s really cool, they didn’t just make it bigger for bragging rights.

They used Mixture-of-Experts (MoE) architecture, which is like having a really smart team where only the right experts jump in when needed. Instead of activating the entire trillion-parameter network every time, it only fires up specific subsets during each forward pass. It’s way more efficient than just throwing raw compute at the problem.

But here’s where the engineering gets impressive: they implemented something called “global-batch load balancing loss” to keep training stable. The result? Their loss curve stayed smooth throughout the entire training process, no spikes, no need to restart, no mid-training data shuffling. Anyone who’s trained large models knows how rare that is.

Efficiency Breakthroughs

Now this is where things get really technical, and honestly, pretty impressive:

Training Speed: They developed PAI-FlashMoE, which is basically a multi-stage parallel pipeline optimization strategy. The result? 30% better training throughput compared to Qwen2.5-Max-Base. That’s not just a small improvement, that’s the difference between months and weeks of training time.

Long Context Magic: For handling really long documents, they created something called ChunkFlow strategy. This thing delivers 3x faster throughput than traditional sequence parallelism when dealing with long contexts. We’re talking about a 1 million-token context window here, that’s in the same league as Claude Opus 3 and Gemini 1.5 Pro.

Reliability Engineering: They built systems called SanityCheck and EasyCheckpoint that reduced hardware failure downtime to one-fifth of what they experienced with Qwen2.5-Max. When you’re running massive compute clusters, this kind of reliability engineering is absolutely crucial.

Performance

Qwen3-Max-Instruct (the version you can actually use) is putting up some seriously impressive numbers:

LMArena Leaderboard: Ranked #3 overall, and get this, it’s ahead of GPT-5-Chat. That’s genuinely impressive performance.

Real-World Coding: Scored 69.6 on SWE-Bench Verified, which tests whether AI can actually solve genuine GitHub issues and programming problems. That beats DeepSeek V3.1 and is competitive with Claude Opus 4.

Tool Usage: Hit 74.8 on Tau2-Bench, which measures how well models can actually use APIs and external tools. This beat both Claude Opus 4 and DeepSeek V3.1, which is honestly surprising.

But here’s where it gets really wild: they’re working on Qwen3-Max-Thinking, a reasoning-focused variant that’s still in training. In early tests that included tool usage and parallel compute during inference, it scored perfect 100% on both AIME 25 and HMMT, two of the most brutal mathematical reasoning benchmarks in AI. This version comes with an integrated code interpreter and focuses on complex logical problem-solving.

Multilingual and Multimodal Story

While the current public release focuses heavily on reasoning and coding, Alibaba built Qwen3-Max to handle multilingual tasks with particular strength in English and Chinese. The improvements are significant across instruction-following, mathematical reasoning, scientific tasks, and, this is important, it produces fewer hallucinations than previous versions.

The model shows major improvements in areas that actually matter for real-world deployment: better logic, more accurate math, stronger scientific reasoning, and more reliable responses overall.

How to Actually Use This Thing

For Regular Users: Qwen3-Max-Instruct is available right now through the Qwen app (iOS/Android) and their website. The app defaults to using Qwen3-Max, but you can manually switch if you want to compare with other versions.

For Developers: You can access it via API through Alibaba Cloud’s Model Studio, which means you can integrate it directly into your applications, tools, or services.

The Bigger Picture

Here’s what’s really significant: Alibaba isn’t just trying to match existing models, they’re pushing the technical frontier in specific areas like long-context processing, training efficiency, and system reliability. The engineering details show genuine innovation, not just scaling up existing approaches.

The upcoming Qwen3-Max-Thinking variant, with its focus on reasoning and tool integration, suggests they’re aiming to build something that can function as a genuine autonomous agent, not just a better chatbot.

Should You Care?

If you’re a casual user, this gives you another top-tier model to experiment with, especially if you need long-context processing or multilingual capabilities.

If you’re a developer, the API access and technical capabilities make it worth serious evaluation against your current toolchain.

If you’re interested in AI progress generally, this represents a real milestone: we now have multiple organizations capable of building and deploying trillion-parameter models with genuinely different technical approaches and capabilities.

The AI landscape just got a lot more competitive, and frankly, a lot more interesting.

In related AI news, researchers developed an AI-designed virus that targets bacteria, raising safety concerns, and Huawei unveiled its powerful Atlas 950 and 960 SuperPoDs to challenge Nvidia.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Alibaba Just Dropped a 1 Trillion Parameter AI Model to Take On ChatGPT and Gemini appeared first on Gizmochina.

Picture this: a computer sitting in a lab, essentially “thinking up” brand new forms of life from scratch. That’s exactly what just happened at Stanford and the Arc Institute, where researchers used artificial intelligence to design viruses that can actually kill bacteria. And yes, these AI-created viruses work in real life.

Here’s how wild this is, it’s the first time anyone has used AI to dream up an entire viral genome from nothing. No copying nature’s homework, no tweaking existing designs. Pure digital creativity turned into living biology.

The AI system, called Evo, works kind of like ChatGPT, except instead of being trained on books and articles, it learned from 2 million different virus genomes. When researchers asked it to design new versions of a simple virus called phiX174, Evo created 302 completely original genetic blueprints. The kicker? Sixteen of them actually came to life in the lab and successfully infected E. coli bacteria.

“This was like watching digital code become biology,” said Brian Hie, who runs the lab. You can practically hear the excitement, and maybe a little nervousness, in his voice.

The potential upside is huge. We’re talking about supercharged treatments for antibiotic-resistant infections, which kill hundreds of thousands of people every year. Imagine having an AI design custom viruses to hunt down specific bacteria that antibiotics can’t touch anymore. Or think about gene therapy, using these designer viruses as tiny delivery trucks to fix genetic diseases.

But here’s where things get uncomfortable. J. Craig Venter, basically the godfather of synthetic biology, didn’t sugarcoat his concerns. He called this “a faster version of trial and error” and warned that the same technology could potentially be misused to create much nastier viruses than simple bacteria killers.

Right now, Evo only knows about viruses that don’t harm humans. But the underlying technology? That could theoretically be retrained on more dangerous stuff. It’s like having a recipe generator that currently only knows how to make cookies, but could learn to make dynamite if someone fed it the wrong cookbook.

We’re still years away from AI creating entire synthetic cells, those would need millions of genetic letters instead of the thousands in these simple viruses. But companies like Ginkgo Bioworks are already working on fully automated systems that could go from AI design to living organism without human hands touching the process.

It’s honestly breathtaking and unsettling at the same time. We might be witnessing the moment when life itself becomes programmable, when the line between digital code and biological reality finally disappears. The question isn’t whether this technology will change everything, it’s whether we’re ready for what comes next.

In related AI news, Huawei has recently unveiled its powerful Atlas 950 and 960 SuperPoDs in a direct challenge to Nvidia, while Tencent has introduced a groundbreaking 3D AI tool that is available completely free.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post AI just created a virus that kills bacteria—should we be scared? appeared first on Gizmochina.

A curious new project has surfaced online, and it’s already sparking conversation among among researchers and coders. An open-source tool promises to track the “stupidity level” of major AI models in real time. While the name is tongue-in-cheek, the tool itself is serious—it’s designed to measure performance drops and help developers understand when popular models are cutting corners.

The tool, hosted at aistupidlevel.info, claims to be the first of its kind to monitor large language models for signs of decline. It currently tracks systems like OpenAI’s GPT-5 family, Anthropic’s Claude Opus 4, and Google’s Gemini 2.5 Pro, with support for xAI’s Grok 4 on the way.

Its approach is straightforward but wide-ranging: more than 140 coding and debugging tests run continuously, scoring models on correctness, stability, recovery, efficiency, and other factors. Results are fed into a live dashboard that shows how “smart” or “stupid” a model looks at any given time.

Another interesting element is cost analysis. The tool doesn’t just look at API pricing, but how many attempts a model needs to get something right. A supposedly “cheaper” model may waste cycles, while a more expensive one could finish the job faster and end up costing less overall.

Everything is open-source, with the code and API available on GitHub for anyone to review or contribute to. Since going live earlier this year, the site says it’s drawn nearly a million visitors, showing just how eager developers are for transparency in an increasingly closed-off industry.

Whether it’s a gimmick or a genuine accountability tool, the Stupid Meter highlights a growing frustration with AI performance swings. For developers and enthusiasts, it could become a useful way to separate hype from reality.

Don’t miss a thing! Join our Telegram community for instant updates and grab our free daily newsletter for the best tech stories!

For more daily updates, please visit our News Section.

(Source)

The post Open-source tool now measures the ‘stupidity level’ of AI models in real time appeared first on Gizmochina.

Huawei just made a massive power play in the AI hardware wars. At Connect 2025 in Shanghai, rotating chairman Eric Xu unveiled the Atlas 950 and Atlas 960 SuperPoDs – absolutely enormous clusters of Ascend accelerators that are clearly designed to challenge Nvidia’s upcoming Rubin-era systems (although Nvidia AI chips are banned now in China) and push China’s AI capabilities to new heights.

Atlas 950 SuperPoD

The Atlas 950 SuperPoD is built around 8,192 Ascend 950DT chips, and the performance figures are genuinely impressive. We’re looking at 8 EFLOPS of FP8 compute power and 16 EFLOPS at FP4 precision, backed by a massive 16.3 PB/s of interconnect bandwidth.

In practical terms, this translates to some serious AI workload performance: 4.91 million tokens per second for training and 19.6 million tokens per second for inference. That’s a 17× and 26× improvement over their current Atlas 900 A3 system, which is no slouch itself.

The physical footprint is equally massive – we’re talking 160 cabinets all connected via Huawei’s new UnifiedBus 2.0 optical protocol. Huawei claims this interconnect is ten times faster than today’s internet backbone infrastructure, which is critical when you’re moving data between thousands of accelerators.

SuperCluster

Here’s where Huawei’s ambitions become clear. They’re planning to link 64 of these SuperPods into Atlas 950 SuperClusters, creating systems with more than 520,000 NPUs delivering 524 EFLOPS of FP8 compute.

Xu wasn’t shy about the competitive positioning either; he directly claimed this setup will outperform Elon Musk’s xAI Colossus and even Nvidia’s upcoming NVL144 and NVL576 deployments, citing six to seven-fold compute advantages plus superior memory and network throughput.

Atlas 960

But Huawei isn’t stopping there. The Atlas 960 SuperPoD, slated for 2026, essentially doubles every major specification: 15,488 Ascend 960 chips, up to 30 EFLOPS FP8, 60 EFLOPS FP4, and 34 PB/s bandwidth. The companion 960 SuperCluster is projected to hit 2 ZFLOPS of FP8 performance when it arrives in 2027, numbers that start to sound almost surreal.

New Ascend Roadmap

Behind all this hardware is a completely refreshed Ascend chip roadmap. The 950PR and 950DT arrive in 2026, featuring Huawei’s own HBM variants: HiBL 1.0 for cost-efficient prefill operations and HiZQ 2.0 optimized for decode and training workloads.

The 960 follows in 2027, then the 970 in 2028. Each generation promises to “double compute” while expanding support for emerging formats like FP8, MXFP4, and HiF4.

Scale-Over-Silicon Strategy

For a company that’s been cut off from advanced Western semiconductor fabs, Huawei’s strategy is becoming crystal clear: if you can’t win on per-chip performance, win through sheer scale and system integration.

By controlling the entire stack, memory subsystems, networking fabric, packaging, and interconnects, and by deploying hundreds of thousands of chips behind ultra-low-latency interconnects, they’re betting they can meet China’s massive demand for model training compute while seriously challenging Nvidia’s market position.

Reality Check

Of course, there’s always a gap between conference announcements and real-world deployments. These massive systems will face significant challenges around power consumption, thermal management, and perhaps most critically, software ecosystem maturity. Nvidia’s CUDA advantage didn’t happen overnight, and replicating that software stack optimization across such massive clusters is no small feat.

But from a pure ambition standpoint, Huawei has definitely fired a serious shot across the Pacific. Whether these enormous machines can deliver their promised performance outside of PowerPoint presentations will largely depend on solving real-world engineering challenges around power budgets, cooling infrastructure, and software optimization.

The AI hardware race just got a lot more interesting.

In related AI news, Tencent has recently launched a groundbreaking 3D AI modeling tool that’s completely free, while Baidu’s new PP-OCRv5 model outperforms larger rivals in OCR benchmarks.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

The post Huawei Just Declared War on Nvidia With Its Monster Atlas 950 & 960 SuperPoDs appeared first on Gizmochina.

Remember when creating realistic 3D models meant weeks of painstaking work by skilled artists? Well, Tencent just threw that playbook out the window with the latest 3.0 version of Hunyuan 3D, and honestly, the results are pretty mind-blowing.

What Makes This Different?

Most AI 3D generators feel like they’re throwing digital clay at a wall and hoping something decent sticks. Hunyuan 3D 3.0 takes a completely different approach with something called “3D-DiT”, think of it as having a master sculptor who starts rough and gradually adds finer and finer details.

Instead of trying to generate everything at once (which usually leads to wonky-looking results), this system builds models layer by layer. It starts with basic shapes, then keeps refining until you get something that actually looks… well, real.

The Numbers Are Ridiculous

We’re talking about three times better precision than before, support for up to 36 billion voxels (that’s a lot of tiny 3D pixels), and resolution that can go up to 1536³. But here’s the kicker: it’s completely free to use.

Finally, Characters That Don’t Look Like Aliens

Anyone who’s played around with AI-generated 3D models knows the struggle: faces that look slightly off, bodies with weird proportions, poses that scream “I was made by a robot.” Tencent seems to have cracked this code.

The new system generates human characters that actually look human. Facial features sit where they’re supposed to, skin textures look believable, and poses flow naturally instead of looking like mannequins arranged by someone who’s never seen a person move.

Smart Problem-Solving

Here’s where it gets really clever: the system can figure out what should be there even when it can’t see it. If you’re generating a character who’s partially turned away or behind an object, the AI makes educated guesses about the hidden parts based on what it does see. No more bizarre gaps or impossible anatomy.

And textures? They actually wrap around objects properly now. No more skin that looks painted on or fabric that defies physics.

Built for Everyone (Seriously)

Tencent could have easily charged premium prices for this technology. Instead, they’re giving it away free through their Hunyuan 3D platform. For developers who want to integrate it into their own projects, there’s a cloud API available too.

If you’re a professional who needs more advanced features, they’re also testing Hunyuan 3D Studio, a full production suite that handles everything from UV mapping to rigging. It’s invite-only for now, but you can apply through their platform.

The Open-Source Promise

In a move that’ll make developers everywhere happy, Tencent plans to open-source a version called Hunyuan 3D Omni. This will give researchers and companies even more control over the generation process, perfect for when you need something very specific rather than just “make me a cool character.”

Why This Matters

Look, we’ve seen plenty of AI tools promise to revolutionize creative work, and most fall short when you actually try to use them. But Hunyuan 3D 3.0 feels different. The combination of actually good results, zero cost, and accessibility could genuinely democratize 3D modeling.

Whether you’re an indie game developer working on a shoestring budget, a digital artist exploring new possibilities, or just someone who thinks creating 3D stuff sounds fun, the barrier to entry just got obliterated.

Sometimes the most exciting tech advances aren’t the flashiest ones; they’re the ones that just work really, really well.

In related AI news, Baidu has recently released PP-OCRv5, a compact optical character recognition model that surpasses larger models in benchmark tests. The company has also unveiled the Ernie X1.1 deep thinking model, which it claims outperforms DeepSeek R1 in reasoning and comprehension tasks.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Tencent Just Dropped a Game-Changing 3D AI Tool And It’s Completely Free appeared first on Gizmochina.

Baidu just dropped something pretty interesting in the AI scene. After their recent launch of Ernie X1.1 deep thinking model, they’ve now released PP-OCRv5, a new optical character recognition model that’s available on Hugging Face. What makes this one stand out? It’s designed to be really good at reading text while staying surprisingly lightweight.

The thing is, those massive vision-language models we keep hearing about? They’re impressive, but they can struggle when it comes to the nitty-gritty work of reading structured text accurately. That’s where PP-OCRv5 comes in. Baidu built this one specifically to tackle those limitations head-on.

Here’s what’s cool about it: the model works in two main stages – first it finds where text is located in an image, then it actually reads what that text says. This approach helps it nail down exactly where text appears and draw precise boxes around it, which is super handy if you’re trying to pull data from documents or analyze forms.

The efficiency is pretty remarkable too. We’re talking about just 0.07 billion parameters – that’s tiny compared to the giants in this space. Baidu tested it on mobile setups and found it could churn through over 370 characters per second on an Intel Xeon processor. That means you could actually run this thing on regular computers or even edge devices without needing massive server farms.

When Baidu put PP-OCRv5 head-to-head with the big names like GPT-4o, Gemini 2.5 Pro, and Qwen2.5-VL on OCR tasks, their model came out ahead. It handles both printed and handwritten text pretty well, and it’s not just limited to English – it works with Simplified Chinese, Traditional Chinese, Japanese, Pinyin, and actually supports more than 40 languages total.

The technical setup is straightforward but smart. It starts by cleaning up the image – fixing rotation issues, reducing distortion, that sort of thing. Then it finds where text lines are located, figures out which way they’re oriented, and finally converts those characters into readable text. The whole process is designed to give you precise coordinates for where each piece of text sits, which is crucial if you’re scanning invoices or processing forms where layout matters.

What’s nice is that Baidu made this available to everyone through Hugging Face. For developers and businesses dealing with lots of multilingual documents or just needing solid OCR capabilities without the overhead of massive models, PP-OCRv5 looks like it could be a practical choice that actually gets the job done.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Baidu releases PP-OCRv5, a compact AI model that beats large rivals in OCR tests appeared first on Gizmochina.

Baidu is back with another AI announcement, and this time they’re really swinging for the fences. The Chinese tech giant just rolled out Ernie X1.1, which they’re calling a major upgrade to their previous Ernie X1 model.

What’s New Under the Hood?

Wang Haifeng, Baidu’s Chief Technology Officer, presented the new model as something special. Built on their Ernie 4.5 foundation, X1.1 uses what they call a “hybrid reinforcement learning framework”, basically combining different AI training techniques to make the model smarter.

The technical details might sound dry, but the results Baidu is claiming are pretty impressive. They say X1.1 is significantly better than its predecessor:

• 34.8% more accurate with facts
• 12.5% better at following instructions
• 9.6% improvement in handling complex tasks

The Big Claims

Here’s where things get interesting (and a bit bold). Baidu isn’t just comparing their new model to their own previous work; they’re taking shots at the competition. They claim Ernie X1.1 outperforms DeepSeek’s R1-0528 model across multiple tests.

But wait, there’s more. They’re also saying their model can go toe-to-toe with some of the biggest names in AI, specifically mentioning performance comparable to GPT-5 and Google‘s Gemini 2.5 Pro in certain areas.

Where Can You Try It?

If you’re curious to test it out, Baidu has made X1.1 available through their usual channels – the Ernie Bot website and their Wenxiaoyan app. For businesses and developers, they’ve also deployed it on their Qianfan AI Cloud platform.

The Bigger Picture

This launch isn’t happening in isolation. Baidu also updated their PaddlePaddle deep learning framework to version 3.2, making it better for training large AI models and working with different types of hardware. They’ve been busy updating their development tools too.

The numbers they’re throwing around are pretty staggering – apparently, their AI ecosystem now supports over 23 million developers and serves 760,000 companies. That’s a massive user base, especially concentrated in China.

What This Really Means

Let’s be honest – every AI company is making big claims about their latest models these days. But Baidu’s push with X1.1 represents something important: the global AI race is heating up, and Chinese companies aren’t content to just follow along.

Whether Ernie X1.1 actually lives up to these bold comparisons with GPT-5 and Gemini remains to be seen. Independent testing and real-world use will be the real judge. But one thing’s for sure, Baidu is clearly positioning itself as a serious player on the global AI stage, not just in China.

The competition between AI giants is only getting more intense, and that’s probably good news for everyone who uses these tools.

In related news, Baidu has recently introduced GenFlow 2.0, its latest AI agent framework designed to deliver real-world task execution.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Via)

The post Baidu Unveils Ernie X1.1 Deep Thinking Model, Claims It Outperforms DeepSeek R1 appeared first on Gizmochina.

Chinese startup DeepSeek is preparing to launch a next-generation AI agent by the end of 2025, according to a Bloomberg report. The Hangzhou-based company is developing a system that can perform complex, multi-step tasks with minimal input from users. The model will also learn from its past actions to improve performance over time.

DeepSeek gained global attention earlier this year with the release of its R1 model. The system demonstrated advanced reasoning capabilities and was built at a reported cost of just $6 million, far less than rivals like OpenAI’s ChatGPT or Google’s Gemini. R1 was also open source, allowing developers to freely access its code. The release disrupted the “bigger is better” approach in Silicon Valley, where massive budgets and vast datasets have dominated AI development.

Since R1, DeepSeek has moved cautiously. Founder Liang Wenfeng delayed the R2 model, a move attributed to technical refinements and his ongoing work at High-Flyer Asset Management. This deliberate pace differs from competitors like Alibaba and Tencent, which have accelerated their own AI rollouts.

Last month, DeepSeek rolled out its V3.1 update, expanding the context window to 128,000 tokens and increasing the parameter count to 685 billion. The company also now applies mandatory AI-generated content labels on all outputs, a policy that cannot be disabled and sets DeepSeek apart from its peers.

The upcoming R2 model will target the emerging market for AI agents. Unlike chatbots that deliver text-based responses, agents are designed to handle tasks such as travel planning, software debugging, and business workflows with little human supervision. Analysts view this technology as the next major step in AI, capable of boosting productivity and reshaping digital services.

Global competitors, including OpenAI, Microsoft, and Anthropic, have already introduced agent-focused features. DeepSeek aims to match and potentially surpass those efforts, continuing its push to challenge U.S. dominance in AI.

Details about pricing, availability, and technical specifications remain limited. However, industry watchers expect the launch to be closely monitored in both Silicon Valley and Washington, where DeepSeek’s rapid progress has already raised concerns about China’s growing influence in artificial intelligence.

For more daily updates, please visit our News Section.

Stay ahead in tech! Join our Telegram community and sign up for our daily newsletter of top stories!

(Source)

The post DeepSeek is working on AI agent to take on OpenAI by end of 2025 appeared first on Gizmochina.