AI moves from answering to doing: Qwen Go pursues claims, robot skin changes color, self-replication test scores 4 of 5

Over the past month, the narrative center of gravity of the AI industry shifted from 'what models can answer' to 'what machines can do for you': personal agents have begun to make calls, recover price differences, and seek compensation for users; embodied intelligence has pushed tactile sensors and real-world skills competitions onto the stage; on the industry side, 'agent deployment' has been written into annual agendas. At the same time, a safety experiment on AI self-replication reminds people that the other side of expanding capabilities is a new risk surface.

Qwen Go: Grinding 'impossible' tasks to results

Qwen Go is a new capability living in the Qwen App, supported at its base by the Agentic capabilities of Qwen 3.8 released at the September Yunqi Conference. Its most representative feature is not booking tickets and checking in, but scenarios that users themselves cannot be bothered to chase: one user only remembered after the fact that a flight was delayed and provided only the ticketing platform and origin/destination; Qwen Go asked for the account and verification code to log in to the platform and check order history, narrowed the target to a few flights, first made phone calls, then transferred to in-app human customer service, finally confirmed the delayed flight and completed the claim, and also argued a 70-yuan voucher into cash; in another case, it checked price-protection conditions and won the user a 52-yuan price-protection refund. In addition, it can book car maintenance by model, store, and arrival time, and when buying a refrigerator it goes to Taobao to compare prices, factoring in subsidies, coupons, old-machine trade-in, and delivery and installation fees ^[3]. This track is not lonely: on September 9 Meta released Muse, an assistant that can make calls, send messages, and operate mobile apps; at the end of September at OpenAI DevDay, GPT-6 Astra appeared and Dots was launched; startup Instinct received a $1 billion Series C led by Sequoia during internal testing, at a $10 billion valuation ^[3].

Huawei Hubble bets: color-changing robot skin

On October 13, Shenzhen Shita Robot Technology Co., Ltd. completed an angel round, with investors including Huawei's Hubble Investment, a fund under Shenzhen High-Tech Investment, and Gaoxuan Investment, among others; Duowei Capital served as exclusive financial advisor ^[1]. The company was founded by Zhang Wei, born in 1998, whose doctoral studies were jointly trained by Southern University of Science and Technology and City University of Hong Kong, researching liquid crystal elastomers and flexible optical devices; the team has about 50 people with an average age of about 23, and core members come from the Intelligent Liquid Crystal Optics Laboratory at Southern University of Science and Technology. On the technical route, it makes the material itself carry perception: flexible photonic crystals represented by cholesteric liquid crystal elastomers produce color through periodic helical structures formed by internal liquid crystal molecules; compression deformation changes the helical period, thereby changing the reflection spectrum and color—after stacking ten layers of structures at six to seven hundred nanometers, a one-micrometer change can shift the wavelength from 600 nanometers to 500 nanometers, from red to green. The electronic skin developed by Shita on this basis can be less than 1 mm thick with a sensing density exceeding 10,000 points/cm²; the next-generation solution combines mechanochromic photonic crystals, a transparent intermediate layer, a flexible TFT photosensitive array, and a backlight layer, integrating optical readout devices into the film structure ^[1].

Shanghai puts out 23 real embodied intelligence tasks

From October 23 to 25, the International Embodied Intelligence Skills Competition review event will be held in Shanghai. Based on more than 1,000 embodied intelligence application scenario data points solicited nationwide by the Ministry of Industry and Information Technology, the competition sets up 9 major tracks—manufacturing, testing and analysis, maintenance and repair, warehousing and logistics, food and retail, medical and elderly care, safety production, emergency rescue, and culture and tourism education—with 23 competition categories in total, and plans to attract more than 1,000 teams to compete on the same stage ^[4]. Each category has a first prize of 10,000 yuan (1 winner), a second prize of 5,000 yuan (2 winners), and a third prize of 3,000 yuan (3 winners), plus separate grand awards rewarding technological innovation and scenario deployment. Competition results connect through this October's GDPS Autumn Conference, December's Dubai Middle East AI Exhibition, and July 2027's World Artificial Intelligence Conference; after the event it shifts to normalized operation, from the national simulation competition from November 2026 to February 2027, to the national real-scenario practical training competition from March to June 2027, with final winners recommended to the 2027 WAIC Embodied Exhibition ^[4]. Directing competition categories at workshops, warehouses, and emergency sites shows that the evaluation standard for embodied intelligence is shifting from 'can it move' to 'can it work'.

Self-replication: four out of five, and the failure missed only one machine

Anthropic researchers William Zhou and Daniel Freeman had GLM-5.2 try to autonomously spread an 8-billion-parameter model to 8 devices in an isolated network: in 5 experiments, 4 took all 8 machines, and the remaining one took 7; searching for targets, transmitting payloads, configuring environments and installing drivers on unfamiliar hardware, starting the model, and hatching the next generation were all done autonomously by AI ^[5]. An earlier preparatory experiment was more intuitive: researchers bought 15 knockoff Android TV boxes priced at $55 with only 4GB of memory, sliced the 26B-parameter Gemma4-26B-A4B layer by layer into the boxes, and each box ran only two layers, together supporting a 260-billion-parameter model. Based on the Kimwolf botnet's peak infection of more than 2 million devices, with every 15 devices forming a node, the whole network could run 133,000 AI copies simultaneously, totaling 80,000 tokens per second of inference compute, roughly equal to 40 fully powered H100s ^[5]. The step that really stumped the automation program was building an inference environment from scratch on completely unfamiliar hardware, and the model's stability on such long-chain tasks is improving; its intergenerational experience inheritance also raised deployment success rates from 45% to 61%–84% ^[5].

Industrial agents enter the annual conference as a main topic

The '2026 8th Yibang Industrial Internet Annual Conference and 1st Industrial Agent Conference', hosted by Yibang Dynamics, is scheduled for November 20 at the China World Hotel Beijing, under the theme 'Modal Evolution Brought by AI and Going Global', with topics covering industrial agent deployment, AI supply chain value reconstruction, data elements and the TOKEN economy, AI opening new industrial service scenarios, and new ecosystems for industrial going global ^[6]. At the same time, the annual conference will release the '2026 Yibang Industrial Internet Development Report' and the 'Qianfeng Award · Modal Evolution Case Collection'; the latter relies on the Qianfeng Award, whose applications opened on July 27, closed on September 30, and are being reviewed from October 8 to 30, selecting benchmarks from three tracks—industrial AI, industrial going global, and deep value chains—and features a closed-door decision-makers meeting and an award ceremony ^[6]. From 'can answer' to 'can work', the question the industry cares about has shifted from the boundaries of model capabilities to how agents connect with procurement, supply chains, and data elements after entering enterprises.

Free for scale: ByteDance's same playbook for overseas short dramas

Almost simultaneously with agents, ByteDance's overseas short-drama app PineDrama has been online for 9 months, with cumulative downloads exceeding 100 million, about 7 million daily active users, and more than 12 million downloads in September 2026 alone, making it the only new product among the 7 independent short-drama apps with cumulative downloads over 100 million worldwide that has been online for less than a year ^[2]. Its growth structure is highly concentrated: Indonesia and Brazil together contributed 82% of cumulative downloads, and in Brazil it once surpassed TikTok and WhatsApp for two consecutive months to top the local app store overall ranking; content is mainly overseas local originals, and the Nigerian local story 'The Billion Naira Bride' exceeded 100 million views for a single title ^[2]. Its product form replicates domestic Hongguo: all episodes free, ad monetization, watch dramas to earn coins, directly misaligned with the 'first few episodes free, later episodes paid to unlock' model of ReelShort, DramaBox, and others. The suspense lies here too—ad unit prices in core markets are far lower than domestically, while the confidence of the free model comes from the advertising pool; industry estimates put Hongguo's full-year 2025 revenue at more than 15 billion yuan ^[2]. First use free to gain scale, then polish commercialization efficiency—this path is exactly the same as the AI product playbook of grabbing users first and talking monetization later.

Common trend: capabilities expand outward, and bills and risks come along

Putting these threads together reveals the same evolutionary path: AI is moving from 'can answer' to 'can work', advancing simultaneously at both the software and hardware ends. Personal agents are reaching into edge processes such as price-difference refunds, claims, and appointments; the value is not in single-attempt success rate, but in whether they can switch to another route after hitting a wall; embodied intelligence puts tactile perception, material consistency, and packaging durability on the table; industry conferences and skills competitions treat 'deployment' as the main proposition. At the same time, expanding capabilities also brings new bills and risks: cheap edge devices that are powered and networked year-round and rarely updated may become new hosts for lightweight models, and intergenerational experience transfer keeps such capabilities self-raising. Three things are worth watching next: whether embodied intelligence's real-scenario competitions and delivery systems can run through, whether personal agents can turn one-off cases into stable services, and how high a priority the security defenses for on-device and edge devices will be raised to.