The team filled individual internal cells with a blend of gallium and iron that is solid at room temperature but liquefies when heated electrically. By sending current through specific cells, the researchers can liquefy any desired pattern, effectively encoding a map of soft and stiff regions in the material in a way that resembles writing bits in digital storage. Cooling the material returns the liquefied cells to a solid state, preserving the programmed mechanical configuration until the pattern is changed again.

In thin two-dimensional sheets, this digital composite architecture enables precise tuning of stiffness and energy dissipation across a wide range, allowing the sheets to mimic materials from relatively rigid plastics to softer rubbers. Because the underlying structure remains the same, engineers can iterate through many mechanical profiles without reprinting or remanufacturing parts. The team subjected these programmable sheets to extensive testing to quantify how different solid-liquid patterns translate into bulk mechanical responses.

The concept extends further in three dimensions, where the group created modular, Lego-like blocks that can be assembled and disassembled. Each cubic block resembles a Rubik's cube, containing 27 individually addressable cells that can be locally liquefied with electrical signals. Arranging multiple blocks into larger structures produces three-dimensional assemblies whose stiffness distributions can be reprogrammed, opening the door to adaptable load-bearing and motion-control components.

To demonstrate a potential application in soft robotics, the researchers joined 10 of these programmable cubes into a straight column that functioned as a tail for a simple robotic fish. They attached the tail to a basic motor and operated the motor identically while varying which cells in the tail were solid or liquid. The same robotic fish exhibited very different swimming trajectories depending on the internal phase pattern of the tail, showing how reconfigurable mechanical architectures can reroute motion without changing the actuator.

Lead author and PhD student Yun Bai described the goal as creating materials that behave more like living tissue than static machine parts. "We want to make materials that are alive," Bai said. He explained that while 3D printers can produce parts with fixed mechanical properties, achieving a new behavior normally requires printing a new object. In contrast, the digital composite approach lets a single object assume many mechanical personalities by reassigning which cells are solid and which are liquid in real time.

Assistant professor Xiaoyue Ni, who led the research group, said the ultimate aim is to scale these concepts into larger, more capable systems. The team envisions miniaturized versions of the programmable blocks that could operate in confined environments such as blood vessels, where a device might stiffen for propulsion and soften to avoid damaging tissue. Similar ideas could apply inside delicate electronic systems, where reconfigurable supports or vibration isolators might adapt to changing loads and conditions.

Another avenue involves tailoring the metal mixture to adjust melting and freezing points for specific environments, including the human body. By selecting alloys that transition at biocompatible temperatures, designers could create implantable components that reconfigure in situ, for example to act as adaptive stents that change stiffness or shape as a vessel remodels. Outside biomedicine, robotic platforms that traverse varied terrains or manipulate fragile objects could benefit from limbs and grippers whose mechanical profiles update on demand.

Ni said the broader vision is to build flexible, programmable materials that serve as the foundational elements of future robotic systems. Instead of treating structure and actuation as separate design problems, engineers could co-design mechanical architectures and control schemes so that a robot's body actively participates in its behavior. The digital composite strategy provides a route to such embodied adaptability by combining simple electrical inputs with reversible phase transitions at the material level.

The work relied on facilities at the Duke University Shared Materials Instrumentation Facility, which is part of the North Carolina Research Triangle Nanotechnology Network supported by the U.S. National Science Foundation under grant ECCS-2025064. Additional support came from the Beyond the Horizon Initiative of the Pratt School of Engineering at Duke University. The study, titled "Digital composites with reprogrammable phase architectures," was published online on January 23, 2026, in the journal Science Advances.

Research Report:Digital composites with reprogrammable phase architectures

]]> Sat, 07 FEB 2026 10:08:35 AEST Washington, United States (AFP) Feb 2, 2026 - Meet OpenClaw: the AI assistant that promised to be your dream intern, terrified cybersecurity experts, and now thrives on chatbot-only social media -- all in just a few weeks.

The perfect assistant?

At the heart of the tech world's latest come-out-of-nowhere fascination is an AI tool built by Austrian researcher Peter Steinberger in November to help organize his digital life.

After he called his creation Clawdbot, Anthropic asked the computer scientist to rename it because the name was too similar to its AI assistant, Claude.

Steinberger renamed it Moltbot and then OpenClaw. OpenClaw surpassed 150,000 stars within days on the GitHub computer development platform, a symbol of its exponential popularity.

Users download OpenClaw and connect it to a generative AI model (such as Anthropic's Claude or OpenAI's ChatGPT), then communicate with it through WhatsApp or Telegram as they would with a friend or colleague.

Early adopters gushed over OpenClaw's abilities, claiming the tool could handle everyday life's most tedious chores -- sending emails, scouring the internet for specific research, even making web purchases.

Some users went further, saying their new AI assistant was taking on a life of its own, behaving like a dream intern who proposes useful projects and anticipates problems before they arise.

In other words, OpenClaw appeared to be delivering on the promise of an AI agent -- one of Silicon Valley's most frequently hyped buzzwords.

Agents represent the next logical step in AI deployment: doing the work of clicking around to execute tasks online instead of humans doing it themselves.

Others remain skeptical. Reports of mistakes, breakdowns, and an overall chaotic experience have prompted many to quickly abandon their experiments.

- Don't do this at home -

OpenClaw is open-source, meaning anyone can modify it. According to analysts, developers across the globe are doing exactly that.

But open-source also means users are especially vulnerable from a cybersecurity perspective. Most cybersecurity analysts agree that connecting the app to your computer, personal data, and communications is a bad idea.

Even Steinberger urges users to proceed with extreme caution, and he advises non-experts to avoid the tool entirely.

When up and running, OpenClaw has the ability to read and write files, run commands, and execute scripts on your PC.

It can also control browsers, giving it the ability to make purchases, reserve hotel rooms, or check into flights.

The tool also has the power to recall past interactions in order to carry out highly personalized functions. This also poses an extra vulnerability to bad actors.

- Reddit for bots -

The OpenClaw saga took a bizarre turn recently with the creation of Moltbook -- a pseudo social network for OpenClaw agents, not humans, created by a developer. Imagine a group of AI chatbots left in a room together to converse.

Posts have ranged from polite banter to long manifestos by agents suffering an existential crisis or considering launching a crypto currency or religion.

"Who wouldn't be intrigued by the idea of taking the little guy that helps you with your to do's and giving them the ability to chill out in their off time," Moltbook's creator Matt Schlicht told TBPN, a tech news platform.

After observing this Reddit-like universe of interactions between bots on Moltbook, some of tech's biggest names suggested they were witnessing the emergence of a kind of super-intelligence. Their reactions echoed those that greeted ChatGPT and other recent AI breakthroughs.

Moltbook "is genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently," said Andrej Karpathy, a highly respected AI researcher.

Elon Musk posted in response that this was "just the very early stages of the singularity," a term used to describe the moment when human intelligence is overwhelmed by AI for ever.

The initial enthusiasm has since cooled somewhat. Many researchers now believe humans are interfering, inserting prompts to steer conversations in particular ways.

]]> Sat, 07 FEB 2026 10:08:35 AEST Tokyo, Japan (SPX) Jan 19, 2026 - When humans pick up everyday items such as fragile eggs or slippery metal cups, they instinctively adjust their grip using tactile feedback to avoid breaking or dropping them. In contrast, enabling robots to perform this kind of adaptive grasping has remained a major challenge for robotics and automation researchers.

A team at Tsinghua University has now demonstrated a sensory control synergy approach that allows robots to learn human like grasping skills from a small set of human examples. The work, reported in National Science Review, introduces a framework that captures human grasping experience and transfers it directly into robotic hands equipped with tactile sensors.

The researchers first developed a tactile glove fitted with custom tactile sensors at the fingertips to record the process of human grasping. Worn on a human hand, the glove collects multimodal tactile data describing contact, slip, and pressure while the person grasps and manipulates objects.

Inspired by human neurocognition, the team then devised a strategy to convert these raw measurements into higher level semantic grasping states such as stable, slightly unstable, or highly unstable. This encoding step filters out uninformative variations caused by differences in grasp position or hand posture between individual trials, making the representation more universal across different objects and users.

Instead of training robots on large datasets of precise tactile readings for each object, the system teaches robots to recognize and respond to these generalized interaction states. According to corresponding author Prof Rong Zhu, this focus on state level understanding makes the method data efficient and highly transferable between tasks and objects.

To turn state recognition into action, the researchers implemented a fuzzy logic controller that mimics human decision making experience. When the tactile system detects a highly unstable state, the controller increases grip force, while a stable state leads the robot to maintain its current grasp without over tightening.

Once the human sensory control synergy was built, the team transferred it into a robotic hand outfitted with its own tactile sensors. With this transferred framework, the robot demonstrated adaptive grasping of diverse objects, including a slippery umbrella, a fragile raw egg, and a heavy bottle, and generalized well to unfamiliar items not included in the initial training.

The human inspired framework can be established efficiently using small datasets collected from a single person, avoiding the need for extensive multi user or multi object training campaigns. Experimental tests showed that a robot equipped with this transferred synergy achieved an average grasping success rate of 95.2 percent when handling slippery, fragile, soft, and heavy objects.

In dynamic experiments, the robot sensed external disturbances such as pulling forces and responded by autonomously increasing its grip to prevent slips. These results underscore the potential of tactile based control to improve robot robustness when interacting with uncertain or changing environments.

The team also demonstrated a real world application in which a robot completed the multi step task of hand brewing coffee. From locating items and scooping coffee powder to stirring and serving, the robot relied on its tactile feedback and sensory control logic to manage uncertainties at each step.

Prof Zhu explained that the robots in this study learn universal grasping by understanding the sensory and control logic underlying tactile data rather than by copying detailed human motion trajectories. This emphasis on underlying principles effectively gives robots the ability to draw inferences from one instance and extend what they have learned to new situations.

According to the researchers, this human like sensory control synergy offers a promising route to robots that can handle a wide range of objects with minimal retraining. By making it possible for robots to absorb and reuse human experience, the approach could accelerate the deployment of intelligent robotic systems in real world service, industrial, and domestic scenarios.

The work was carried out by the research group led by Prof Rong Zhu at the State Key Laboratory of Precision Measurement Technology and Instrument, Department of Precision Instrument, Tsinghua University. The study highlights how combining bio inspired sensing strategies with experience based control can narrow the gap between human and robotic dexterity.

Research Report: Human-Taught Sensory-Control Synergy for Universal Robotic Grasping

]]> Sat, 07 FEB 2026 10:08:35 AEST Tokyo (AFP) Jan 30, 2026 - The agile robot dog springs up on all fours, takes a step forward and charges at the tense crowd at a Tokyo exhibit, held back by a simple yet strong metal chain.

The silver mechanical creature then starts thrashing around violently, to gasps and exclamations from spectators at the installation, designed to probe humanity's relationship with increasingly realistic machines.

The Japanese media artist behind it told AFP he hoped the audience would consider the dangers posed by artificial intelligence but also feel "pity" for the struggling robot.

"Our future is going to be stressful, because people treat robots as objects, but we feel empathetic stress with these movements and reactions," said Takayuki Todo, 40.

Global tech giants are investing vast sums into humanoid and other lifelike robots, with grand plans for factory automation, home help and other futuristic "physical AI" services.

But so far actual use cases remain scarce and fully automated robots are still a rare sight, with most impressive displays -- including Todo's -- relying on remote operators to control the robot's movements.

For the artist, the point of the three-day installation at the Tokyo Prototype festival is to provoke thought.

Like the metal leash, "we are protected by an unreliable, thin chain of ethics. And if it's cut off, we will be killed by this technology," said Todo.

- 'Robot abuser' -

For his installation, titled "Dynamics of a Dog on a Leash" and first shown last year, Todo purchased three robot canines made by Chinese startup Unitree, costing thousands of dollars each.

One is already broken and repairs are needed for another, as the dogs often get tangled in the chain and end up crashing onto the floor.

Todo, who said he had been attacked online "as a robot abuser", visited Unitree in China last year to excuse himself for the unconventional treatment of their device.

The short hourly display, on show through Saturday in a business district skyscraper, is drawing large crowds, with many spectators including children curious to see a robot of this kind for the first time.

"It gave me the chills," said 34-year-old student and food service worker Kimie Furuta.

"Imagining it actually attacking like that... it could be terrifying to face."

On the brighter side, robots and AI could one day help ease staff shortages, including in the catering sector, she said.

Anatol Ward, a Tokyo resident in his 50s, said the robot reminded him of a guard dog.

"In some sense it was scary. But also it was fascinating -- like, what the robot was capable of."

Todo said that "of course" he was afraid of military uses for such robots, but noted it is not just a future concern.

"Robots and drones are killing soldiers in Ukraine or Palestine," he said.

"We feel it's a distant place, but as an artist we have to imagine it's in front of us."

]]> Sat, 07 FEB 2026 10:08:35 AEST Tokyo, Japan (SPX) Oct 15, 2025 - Researchers at the Beijing Institute of Technology have developed a centimeter-scale quadruped piezoelectric robot that combines compact design, high durability, and advanced mobility. The new prototype demonstrates record-setting locomotion speed, lifting capacity, and resilience, offering a model for future miniature robotic systems capable of operating in extreme or constrained environments.

Centimeter-scale robots are valued for their lightweight structure, agility, and ability to access small spaces. However, previous designs have faced limits due to electromagnetic motors and moving transmission parts, which hindered further miniaturization and introduced electromagnetic interference and wear. To overcome these challenges, the BIT team used piezoelectric materials to create a new built-in actuation system, eliminating external drive mechanisms and boosting integration.

Lead author Yu Gao explained, "Here, we propose a novel centimeter-scale quadruped piezoelectric robot with high integration and strong robustness, which promises to bring new perspectives for the construction and application of centimeter-scale robots."

The new design achieves four major breakthroughs. First, a built-in actuation method powers a 14.47-gram robot (70 mm + 13 mm + 15.8 mm) capable of speeds up to 47.38 body lengths per second, lifting loads nearly 29 times its weight, and reaching movement resolutions as fine as 0.33 micrometers. Second, the untethered version integrates actuation, control, communication, and power supply within a metal piezo unit, operating at only 10 V0-p for up to 32 minutes.

Third, the robot's structural robustness was demonstrated through extreme stress testing. It continued functioning after being stepped on by an adult weighing 66.45 kilograms - more than 3,500 times its own mass - as well as after multiple drops and kicks. Fourth, equipped with a sensor module, the robot performed real-time image grayscale conversion, object detection, and color tracking, highlighting potential applications in imaging and intelligent sensing.

According to the research team, future work will focus on improving the efficiency of built-in actuation methods and piezoelectric arrangements, integrating micro grippers for precision handling, and enhancing position feedback for greater control accuracy. The researchers believe their design provides a foundation for next-generation miniaturized robotic systems with applications in inspection, surveillance, and biomedical environments.

Research Report:A Centimeter-Scale Quadruped Piezoelectric Robot with High Integration and Strong Robustness

]]> Sat, 07 FEB 2026 10:08:35 AEST New York (AFP) Jan 29, 2026 - As "AI slop" floods the internet, efforts are mounting to stem an online deluge of shoddy images and videos made using increasingly advanced tech tools.

Easily accessible generative artificial intelligence tools, such as Google's Veo and OpenAI's Sora, enable the creation of realistic imagery using just a few descriptive words.

Images of cats painting, celebrities in compromising situations, and cartoon characters endorsing products are among the AI-generated detritus proliferating on social networks and video-sharing platforms.

"The rise of AI has raised concerns about low-quality content - also known as AI slop," YouTube chief executive Neal Mohan said of the irksome phenomenon.

Such content is "cheap, bland and mass-produced," Swiss engineer Yves, who declined to give his last name, told AFP, echoing discussions on social media website Reddit.

Brands like Equinox gyms and Almond Breeze almond milk have played off AI slop frustration in recent ad campaigns, offering themselves as authentic, real alternatives.

Meanwhile, Microsoft chief executive Satya Nadella has urged people to move beyond the debate over whether AI creations are slop or sophistication to embracing the technology as a way to amplify creativity and productivity.

Microsoft is among the tech giants investing heavily in AI.

"At its core, the criticism of AI slop is the criticism of some individual's creative expression," argued Bob Doyle, a YouTube personality specializing in AI-driven media creation.

"You may think it's useless, but to them it's the beginning of an idea; a seed."

- Machine made -

However, online bulletin board Pinterest saw fit late last year to begin allowing users to filter out some AI-generated content.

Pinterest told AFP that it added the filter after hearing from people who wanted to see fewer synthetic images.

TikTok introduced a similar filter on its globally popular video platform late last year.

YouTube, along with Meta-owned Instagram and Facebook, also offers ways to reduce the amount of synthetic imagery their users encounter, but gives no clear-cut filter.

Major social media platforms had previously focused primarily on labeling AI-created videos so viewers would not mistake them for showing real scenes, but ample synthetic content seemed to avoid the labels.

Some smaller tech players, such as streaming platform Coda Music, have introduced measures such as having users report AI creations.

Once confirmed, accounts get labeled as AI artists so listeners know what they are getting, according to Coda founder and chief executive Randy Fusee.

"There has been a lot of participation in the identification of AI artists so far," Fusee told AFP.

"By and large, (Coda users) just don't want AI music."

Coda, which reports having some 2,500 users, offers the option of completely blocking AI content from suggested playlists.

Cara, a social network for artists and designers with more than a million users, relies on a combination of algorithms and human moderation to filter AI-generated content.

"People want the human connection," said Cara founder Jingna Zhang.

"I could like a child's drawing because I'm charmed by it, as opposed to (something made by) a machine with no intention."

]]> Sat, 07 FEB 2026 10:08:35 AEST Hong Kong (AFP) Jan 21, 2026 - Investor confidence in Chinese AI startups is riding high, but obstacles to their long-term success range from US export controls to the puzzle of how to become profitable.

This month, two leading players in China's artificial intelligence industry, Zhipu AI and MiniMax, made dazzling debuts on the Hong Kong stock exchange.

The pair are part of a wave of rapidly growing Chinese "AI tigers" spurred by another startup, DeepSeek, whose low-cost AI model, on par with US rivals, stunned the world a year ago.

But Zhipu AI's co-founder Tang Jie warned later that despite the achievements of Chinese companies in large open-source AI models, the gap with the United States "may actually be widening".

DeepSeek and other top Chinese AI providers have focused on free, open-source technology -- a strategy that can attract users fast but brings in less cash than private, closed systems.

"Large-scale models in the US are still mostly closed-source... we need to acknowledge challenges and gaps we face," Tang said at a conference in Beijing.

Geopolitical struggles could also hold Chinese AI back.

US export sanctions on advanced microchips used to train and run AI systems, as well as precision chipmaking equipment, have been cited as a key constraint by top industry figures.

"The challenge isn't just technology," Nick Patience, practice lead for AI at tech research group Futurum, told AFP.

"It's the high cost of computing under sanctions and the delicate balance of innovating within a strict regulatory framework."

- 'Burning cash' -

Shares in Zhipu AI, a major provider of chatbot tools to Chinese businesses, have soared 80 percent since it went public.

MiniMax, which targets the consumer market with its multimedia AI tools, has seen even stronger gains.

Their IPOs came ahead of any such move from OpenAI, the San Francisco-based startup behind the phenomenally popular ChatGPT.

Although OpenAI's value has ballooned in funding rounds to a staggering $500 billion, it does not expect to be profitable before 2029 owing to huge outlays to build the computing infrastructure it relies on.

Zhipu AI and Minimax are also logging increasing losses while costs, including for training new AI models, rise.

Both are "burning cash faster than they can generate sustainable revenue streams", analyst Poe Zhao, founder of Hello China Tech, told AFP.

US restrictions bar the most advanced, energy-efficient AI chips on the market, made by US company Nvidia, from sale in China.

Using domestic chipsets, Chinese AI developers need two to four times more computational power to train their models, according to Lian Jye Su, chief analyst at Omdia.

Zhao and other analysts call 2026 a critical test for the global AI sector as it chases elusive monetisation prospects.

Whether companies "can move beyond coding and unlock real commercial value" is vital to their survival, Zhao said.

- Industrial uses -

Koda Chen said his firm Suanova Technology, which provides and invests in computing power for Chinese AI companies, has identified opportunities in finance and healthcare.

He sees this year as a "turning point" for China's AI businesses to achieve profitability in more sectors.

"Clients are developing payment habits, and products are gaining customer stickiness," the Suanova CEO said.

China is handing out massive subsidies to support AI innovation and its industrial policies also illustrate its ambition to compete with the United States in the sector.

Beijing this month announced plans to deploy three to five general-purpose large AI models in manufacturing by 2027.

The government said it also planned to strengthen supplies of computing power.

These moves show the country is serious about AI driving the real-world economy, Futurum's Patience said.

China "is trying to build the AI-powered factory of the world", he said.

The large language model market in China, still in its early stages, is estimated to grow to $14.5 billion by 2030, according to consultancy Frost and Sullivan, with the future unit price of computing power expected to decline.

China's engineering talent base and the lower cost of generating electricity there work in its favour, said Tang Heiwai, an economics professor at the University of Hong Kong.

"These factors would grant China greater resilience in development than the United States as an AI superpower", he said.

twa/kaf/dan

INVESTOR AB

MiniMax

NVIDIA

]]> Sat, 07 FEB 2026 10:08:35 AEST Vatican City (AFP) Jan 24, 2026 - Pope Leo XIV on Saturday stressed the risks of generative artificial intelligence, how it could usurp human identities and relationships, influence public opinion and deepen social polarisation

In a message marking the World Day of Social Communications, the pope said AI systems reflect the worldview of their creators and can shape patterns of thought by reproducing biases embedded in the data they process.

"The challenge... is a matter of protecting human identity and authentic relationships," the pontiff said.

Leo XIV's warning comes as generative AI makes leaps and bounds towards replicating, altering and manufacturing images, music and text to levels sometimes indistinguishable from human-made works.

In 2023, Leo's predecessor, pope Francis, was the subject of a viral fake AI image of him wearing a white puffer jacket.

Since then, generative AI has been a go-to tool by some high-profile figures, including US President Donald Trump, who has posted or reposted computer-generated images to his online accounts.

Pope Leo XIV warned that just a small number of companies hold significant power over AI development, and that AI tools increase "the difficulty of distinguishing between reality and simulation".

Since his election pope last May as the first pope from the United States, Leo XIV has consistently warned about the growing influence of AI technologies.

He also criticised systems that present statistical probability as reliable knowledge, arguing that such tools ultimately offer only approximations.

He said the challenge ahead is to establish effective governance and called for young people to be educated about how algorithms influence perceptions of reality.

Last month, he condemned the accelerating use of AI in military applications, warning against delegating life?and?death decisions to machines.

Musk makes Davos debut with promise of robots for all
Davos, Switzerland (AFP) Jan 22, 2026 - Elon Musk sees his humanoid robots hitting the market next year, one of several "optimistic" forecasts by the US tech mogul at his first-ever Davos appearance on Thursday.

In front of a packed conference hall, Musk had a chance to tear into a World Economic Forum he has long derided as a "boring" confab of out-of-touch elites.

But in a remarkably subdued "conversation" with WEF interim chair Larry Fink -- also the CEO of investment behemoth BlackRock -- Musk stuck to his script of optimistic enthusiasm for AI, robotics and space travel.

He was not pressed for example on the scandal caused by sexualised deepfakes of his Grok AI tool, or claims of persistent fake news spread by his X social network.

"Who wouldn't want a robot to watch over your kids, take care of your pet... If you had a robot that could take care and protect an elderly parent, that'd be great," he told the audience.

His Optimus robots will be doing more complex tasks later this year, he said, and "by the end of next year I think we'll be selling humanoid robots to the public".

Musk also predicted the artificial intelligence boom will have models that are "smarter than any human by the end of this year, and I would say no later than next year".

"And then probably by 2030 or 2031, so five years from now, AI will be smarter than all of humanity collectively."

But he ended his talk with a caveat: "Generally, I think that for quality of life, it is actually better to err on the side of being an optimist and wrong, rather than being a pessimist and right."

]]> Sat, 07 FEB 2026 10:08:35 AEST Beijing (AFP) Jan 20, 2026 - Before DeepSeek shook up the tech world and put Chinese artificial intelligence on the map, Wu Chenglin's own startup had nearly folded three times -- but in the past year it has raised $30 million.

The January 2025 release of a low-cost generative AI model from DeepSeek that performed at a similar level to ChatGPT and other top American chatbots upended assumptions of US dominance in the sensitive sector.

The breakthrough has galvanised China's AI scene, despite hurdles posed by rivalry with the United States, and fears of a global market bubble.

"It gave a lot of people confidence" that China's AI community previously lacked, Wu told AFP.

His venture DeepWisdom, whose flagship product is a platform for AI-powered software development, had struggled to stay afloat despite its popularity among programmers.

But as excitement around DeepSeek fuelled a boom in spending, Wu raised 220 million yuan in two funding rounds.

Meanwhile, Shi Yaqiong and her team at Beijing-based Jinqiu Capital have closed deals with more than 50 AI firms over the past 12 months.

Shi, the fund's vice-president, described a "clear surge" in enthusiasm around Chinese AI and competition among investors since the DeepSeek shock.

"The kind of projects with an initial valuation in 2024 of $10-20 million were, in 2025, expected to have initial valuations around $20-40 million," she said.

- Engineer dividend -

Shares in two leading Chinese AI startups, Zhipu AI and MiniMax, soared on their market debuts in Hong Kong this month.

Frenzy over the much-hyped potential of AI to change the world is driving global stocks to record highs, led by chipmakers and tech giants.

But the big-spending euphoria has sparked fears of a market crash, with many investors hyper-focused on any sign the AI bubble could burst, and questioning when new companies will become profitable.

Access to top-end chips made by US giant Nvidia is also restricted in China under White House policies designed to curb China's technological development.

But that hasn't dampened the spirits of young developers in the world's second largest economy.

At an AI networking event held on a brisk winter afternoon last week in a stylish Beijing cafe, animated discussion filled the air about the future of the fast-moving industry.

Chip export controls mean Chinese AI is more likely to be "open-source and cheap" which could make it more useful to society, said one participant, entrepreneur Li Weijia.

China is often said to enjoy an "engineer dividend" that benefits its AI sector, and talent is flocking to the field.

Online hiring platform Zhilian Zhaopin reported a 39 percent increase in applications to AI-related jobs in the first three quarters of 2025, after DeepSeek's breakout.

- 'Cost-efficient' -

"China has a huge application developer ecosystem and people are very good at building apps," Shen Qiajin, founder of ideaFlow, told AFP.

"But for a very long time, we didn't have a good cost-efficient model," he said.

That is a gap DeepSeek has now filled.

The firm began in 2023 as a side project of a data-driven hedge fund co-founded in the tech hub Hangzhou by Liang Wenfeng, which had access to a cache of powerful Nvidia processors.

Today, the company -- expected to release its next AI model within weeks -- holds four percent of global market share for chatbots, according to web traffic analysis company Similarweb.

ChatGPT dominates at 68 percent while Google's Gemini is catching up at 18 percent, Similarweb estimates.

DeepSeek's decision to make its systems' inner workings public, in contrast to the closed AI models sold by OpenAI and other Western rivals, has boosted adoption of its tools by developers and businesses, Neil Shah at Counterpoint Research said.

Its tools have had "strong adoption in cost-sensitive emerging markets", he said.

But in the West users are more cautious, "primarily on account of privacy and national security concerns".

Even so, the domestic market is huge. By June 2025, more than half a billion Chinese internet users reported having used generative AI products, according to the China Internet Network Information Center.

Entrepreneur Yang Yiwen said her parents had their first meaningful encounter with AI during last year's Chinese New Year, when they watched her use DeepSeek to plan a family trip.

"They found it quite fun," she said.

ll/kaf/reb/dan

MiniMax

NVIDIA

GOOGLE

]]> Sat, 07 FEB 2026 10:08:35 AEST San Francisco, United States (AFP) Jan 13, 2026 - A trial in the lawsuit brought by Elon Musk against OpenAI CEO Samuel Altman and other defendants -- including tech giant Microsoft -- is scheduled to begin April 27, according to a federal court order issued Tuesday.

The legal showdown will center on Musk's claim that OpenAI and Altman abandoned the startup's original nonprofit, public-benefit mission and misled him while turning the lab into a highly lucrative, Microsoft-linked for-profit enterprise.

Musk alleges that Altman "intentionally courted and deceived" him by proposing a nonprofit that would serve as a counterweight to Google's DeepMind AI research unit and make AI technology open source for the benefit of humanity.

The suit claims OpenAI's leaders have since restructured the organization, created for-profit affiliates, and struck multibillion-dollar deals in a way that violated that mission.

The Tesla and SpaceX boss was an original co-founder of OpenAI and its biggest funder when it began in 2015, providing tens of millions of dollars in seed capital. He left three years later.

OpenAI went on to release ChatGPT in 2022 and Musk eventually launched a rival, xAI.

Microsoft, a major investor in OpenAI since 2019, has poured billions into the firm and become its biggest shareholder after a restructuring last year.

The trial, to be held before a jury, will be presided over by US District Judge Yvonne Gonzalez Rogers in Oakland, California.

]]> Sat, 07 FEB 2026 10:08:35 AEST Subscribe to our daily selection of space, military, environment and energy newsletters responseText http://visitor.constantcontact.com/manage/optin/ea?v=0016gbbKsaiGSpQFojVO8ZoHw%3D%3D