1T4R RRAM Chip Tapes Out, Agent Phone Goes on Sale as Embodied AI and AI Variety Shows Heat Up

In mid-October, China's AI industry saw productization milestones across several layers at once: a new type of memory chip completed Gb-level tape-out, a large-model-native agent phone officially went on sale, the data collection and competition systems for embodied intelligence continued to expand, and generative video pushed the production threshold for variety content down to a level individual creators can afford. The common thread running through these developments is that model capabilities are seeking scalable, concrete forms of deployment along four paths: chips, terminals, data, and content.

1T4R Architecture RRAM Chip Tapes Out: New Memory Moves Toward Gb-Level Productization

On October 12, 2026, MemSilicon announced the tape-out of 'MountTai 2026', its first-generation high-density, high-capacity RRAM chip based on the 1TnR technology route. The chip uses a 28nm CMOS process platform and a 1T4R cell structure, with a single-chip capacity of 1Gb, a density of 30Mb/mm², an access latency of less than 20ns, and integrated on-chip hardware error-correction circuitry. 1T4R means one select transistor corresponds to four resistive memory cells, sharing the transistor to reduce the average area overhead per memory cell. The company views this route as a milestone in moving 1TnR technology from early research toward Gb-level chip productization ^[1].

The background is that different memory technologies each have constraints. SRAM offers fast access, but its memory cells are large and density is low; embedded Flash handles program and data storage tasks, but as logic processes scale down, integration becomes more difficult; RRAM, which combines non-volatility with the potential for high-density integration, has therefore become one of the mainstream approaches to new memory. Public reports show that major global foundries including Samsung, TSMC, and GlobalFoundries have included RRAM in their advanced process node technology roadmaps; TSMC explicitly includes RRAM in its 12nm and 6nm node support scope, and Infineon has selected RRAM for 28nm-node microcontrollers. MemSilicon's technical accumulation comes from the team of Academician Liu Ming of the Chinese Academy of Sciences; as early as 2020, the research team had publicly presented 28nm 1T2R RRAM research ^[1].

Shared transistors do not mean that memory cell density and performance increase proportionally. Non-target current paths between multiple memory cells, read/write interference, and reliability under different operating conditions all need to be addressed through joint optimization of devices, processes, and circuit design. Around the core challenges of large-capacity RRAM chips based on the 1TnR structure, 'MountTai 2026' proposed and validated at the circuit and architecture levels a negative-voltage design and unified reset circuit, a multi-voltage-domain wordline driver circuit, an adaptive protection bias voltage technology, and an on-chip adaptive follower read circuit, which together with hardware error-correction circuitry form a dual read assurance mechanism ^[1].

Policy and product planning are also advancing in parallel. In September 2026, the Ministry of Industry and Information Technology and the National Development and Reform Commission jointly issued the '15th Five-Year Plan for the Development of the Electronic Information Manufacturing Industry', which already lists new memory products such as resistive memory and computing paradigms such as compute-in-memory as development directions. MemSilicon currently plans three product lines: embedded RRAM IP for display drivers, microcontrollers, SSD controllers, and security chips; high-density, high-capacity RRAM focused on increasing integration density, expanding single-chip capacity, and refining standalone product forms, exploring different structures such as 1TnR and 3D integration; and hybrid compute-in-memory combining RRAM with DRAM for embodied intelligence, AI PCs, and other edge inference scenarios, exploring ways for different memory technologies and computing units to work together ^[1].

Agent Phone Goes on Sale: System-Level Agent and Full-Stack 'Software-Hardware-Model' Pricing

On October 13, 2026, large-model-native AI terminal brand STEPX held a launch event in Shanghai under the theme 'Ready Builder One', officially unveiling its first large-model-native agent phone, STEPX Neo. The early edition runs the agent-native operating system Step AOS 1.0 and the new-generation personal agent Step Amoo, and went on sale simultaneously on e-commerce platforms including JD.com, Tmall, and Douyin. The early edition opened pre-orders at 21:00 on October 13 and officially went on sale at 10:13 on October 16, available in carbon black and silicon silver, with a full-stack price of 3,999 yuan. The full-stack pricing approach integrates the Step series models, the Step AOS system, Step Studio tools, and the phone hardware into one price; buyers receive a STEPX Token Plan with 10 billion tokens per month for 12 consecutive months, and the first 10,000 users who meet the task criteria will also receive a 'Starship Ticket', allowing them to exchange STEPX Neo for any new STEPX terminal at zero cost within the next five years ^[2].

At the system level, Step AOS 1.0 was built from scratch for agents, integrating long-term memory, task scheduling, and security mechanisms into the system's foundation. A dual-entry interface design lets users find the agent at any time and lets the agent respond to user needs at any time; 'dual-domain memory' supports continuous recognition of user intent, with the user domain recording personal information, situational experiences, and preferences, and the agent domain accumulating knowledge, methods, and experience. A large number of underlying atomic capabilities of the system can be called directly by Step Amoo. Device-cloud collaboration achieves 'fast reactions on-device, deep thinking in the cloud': on-device models quickly handle simple commands such as adjusting system settings and creating alarms, and some local operations can also be performed without a network; for complex needs such as information queries, single-app tasks, and multi-app collaboration, cloud-side models handle task planning and decomposition, which the device side then executes and delivers ^[2].

In actual use, Step AOS supports Loop cyclic execution and autonomous Tool Call at the system level, and can also invoke various Skills and MCP tools. A tiny collapsible panel stays at the bottom of the screen; when opened, it becomes a system-level Agent task navigation bar, showing all long-chain tasks the agent is currently advancing in the background as well as historical records; when a step requiring confirmation comes up, a pop-up asks for instructions, and when a task finishes, results are proactively organized and pushed. A light press on the side AI key brings up an input box, while a long press starts direct voice conversation. On the ecosystem side, Step Amoo has already reached cooperation with nearly 100 companies or products including Alipay, Tencent, Ctrip, Didi, Jianying, Amap, and Meituan; STEPX is also the first terminal manufacturer to integrate Alipay Vibe Pay, and has worked with Jianying to implement an A2A collaborative video creation experience. Step Studio, aimed at makers, allows AI apps, games, agents, and Skills to be generated using natural language, and opens model, system, hardware, and ecosystem capabilities to makers ^[3].

Privacy and permission issues for personal assistant agents were also mentioned. The security architecture emphasizes 'trustworthy, visible, controllable, and reversible', making task processes auditable and permissions and memory manageable, with ultimate control in the user's hands. STEPX has partnered with Alipay to build a system-level agent solution and a customized full-stack AI payment solution, seeking to close the loop from demand response and task planning to service fulfillment. Whether these arrangements can form a stable experience in real use depends on the reliability of ecosystem calls and users' acceptance of permission boundaries ^[2][3].

Embodied Intelligence: Real-Machine Competition and Full-Body, Full-Modality Data Collection Advance Together

The 2026 Global Developer Pioneer Conference and International Embodied Intelligence Skills Competition will be held from October 23 to 25 in Pudong, Shanghai. Continuing the theme 'Developers, Seek and Find', the conference is organized under the overall framework of '1 competition, 1 opening ceremony, N developer activities', focuses on 'real machines in real scenes, real-scene training', and features 9 major tracks, 23 competition events, and a real-scene training space. Scene providers including Yanfeng Automotive, Yingzhong Information, the Shanghai History Museum, the Shanghai Institute of Quality Inspection and Technical Research, Bailian Group, STO Express, China Life Property & Casualty Insurance, Tianyi Digital Life Technology, the Zhejiang Qiushi Institute, and Goertek will bring real needs from production lines, the service industry, and the public sector to the competition venue to test robots' job-task capabilities. More than 10,000 robots have already moved into the Zhangjiang Artificial Intelligence Innovation Town, and more than 1,000 teams will compete on the same stage ^[4].

The industrial chain and open-source ecosystem are another main thread of the conference. Whole-machine and humanoid robot companies such as AgiBot, Unitree, the National and Local Co-built Humanoid Robot Innovation Center, Fourier, Lenovo, Keenon, JAKA, and Kepler will appear together with core component manufacturers and scenario application companies; exhibits cover complete machines as well as various component modules such as dexterous hands, joint modules, reducers, and force-tactile sensors. The conference also sets up a special open-source and developer community session, with more than 100 leading open-source communities and institutions at home and abroad participating in co-building; OpenJS Foundation Executive Director Robin Bender Ginn, former Linux Foundation Strategic Vice President Ibrahim Haddad, and former Google Developer Relations President Jeanine Banks will share on topics including open-source embodied intelligence software and hardware, profit space in the robotics industry, and the future of open models and software. The conference coincides with 1024 Programmer's Day and will release three major 'China Mobile · Spatiotemporal Intelligence' models and scenario benchmarks ^[4].

The data collection side is also filling in capabilities. On October 14, 2026, Dawn Intelligence and Tinno Mobile jointly released SeePhys Hub, a full-body, full-modality data collection system. The entire device consists of a four-eye headband, two wristbands, three waist-and-leg motion trackers, and a tactile glove; it restores full-body posture by fusing visual and motion signals and collects hand contact pressure according to task requirements. The team adopted a sparse-node approach, placing three motion trackers at the waist and both lower legs, with IMU providing motion-related signals and UWB providing ranging information between nodes, then combining with vision to estimate full-body posture; the four-eye headband provides 360° surround view, wrist cameras supplement close-range manipulation details, and an optional tactile glove records pressure changes when both hands contact objects. According to product materials, the synchronization accuracy of the cameras, IMU, and UWB inside the headband is no more than 0.5 milliseconds, the headband weighs no more than 220 grams, and a single wristband weighs no more than 75 grams ^[6].

The person and route behind this device are also worth noting. Dawn Intelligence founder and CEO Yang Shuo was born in 1998 and is currently a tenured professor and doctoral supervisor at the School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen), long engaged in full-modality large models, computer vision, and embodied intelligence research. The team has previously carried out a series of efforts around the acquisition, expansion, and use of tactile data: in May 2026, the EgoTouch dataset and TouchAnything research were made public; the former synchronously records first-person video, both-hand posture, and real tactile pressure, while the latter attempts to estimate contact areas and pressure distribution from video; by TouchWorld, these 'feelings' began to participate in robot action. In the three-layer data pyramid described by Yang Shuo, the bottom layer is large-volume internet video, the top layer is high-precision motion capture and real-machine operation data, and SeePhys Hub targets the middle layer, attempting to balance the potential for large-scale collection with more complete interaction records ^[6].

AI Variety Shows Emerge in Bulk: Generative Video Lowers the Production Threshold

The classic variety show 'X-Change' is being restarted in another way by AI creators. The Douyin account @ReshapePlan used AI to produce a cyber version of the show; it has updated four episodes so far, averaging more than 200,000 likes per episode, and gained 127,000 followers from just three videos. The earlier AI variety account @100,000 Challenges, relying on wildly imaginative competitive challenges, gained 1.89 million likes on a single video, averaged more than 250,000 likes per series, attracted more than 2 million followers, and reached a commercial order quote of 150,000 yuan. On Douyin, the view count for AI variety-related topics has climbed to 28.764 million. These programs, which once required dozens of people and large budgets, can now be made with one computer and AI tools, and the two creators share another identity: LibTV certified creators ^[5].

The change in production methods is the direct reason AI variety shows can emerge in bulk. Taking @100,000 Challenges as an example, scenes that are difficult to shoot in reality—the host riding a mount into the sky to commentate, contestants casually pulling off max-level ultimate moves, and wind-fire wheels and yachts appearing out of thin air from a phone—are all completed through AI video generation. @ReshapePlan's 'X-Change' derivative retains the familiar variety show captions, dramatic conflicts, and convincing characters and scenes; one episode tells of Lu Xiaonao, whose speech is full of internet memes, and Zhao Ruonan, who habitually solves problems with his fists, going to the countryside for a makeover. Comments included 'I feel so sorry for grandpa' and 'Don't bully grandpa', and one netizen remarked, 'Thank goodness it's AI, otherwise I'd be heartbroken watching it.' What AI generates is virtual characters, but what the audience invests is real emotion ^[5].

The change in cost structure is equally critical. For traditional variety shows, take the first season of 'Beast Games', which launched in 2024: the show gathered 1,000 contestants, and the final grand prize reached millions of dollars, backed by funding, teams, venues, and a long production cycle. AI tools have changed this logic: according to the creators, the image materials and video generation required for the program are basically completed on the AI video generation platform LibTV, whose built-in team collaboration functions let everyone intuitively understand production progress and reuse each other's assets to improve collaboration efficiency; a roughly 5-minute AI variety episode costs about 2,000 to 5,000 yuan to generate and takes only about a week to produce. Lower costs have also changed the way creators test ideas: they can first spend five or six minutes trying out a concept, then adjust the format based on likes, completion rates, and comment feedback, and if the response is mediocre, they can quickly change direction ^[5].

Distribution and platform support form the final link. An AI variety episode runs five or six minutes, with conflict at the start, a twist in the middle, and a cliffhanger at the end; viewers can watch when they scroll by and follow for more if hooked, a rhythm that fits the distribution logic of short-video platforms. Content platforms have also begun to step in: Douyin launched the 'Future Director Support Plan' at the end of 2025, and LibTV reached cooperation with the Douyin Future Director Support Plan to jointly support AI variety creation, with a focus on supporting leading creators in producing new program works. AI turns ideas into images, platforms provide resources, and creators come up with ideas; a new content production chain is taking shape ^[5].

Common Trends and Follow-Up Watch Points

Putting these threads together, the deployment of model capabilities is unfolding simultaneously from a single software layer toward four directions: chips, terminals, data, and content. On the memory side, the 1TnR route seeks to resolve the contradiction between density and cost through shared transistors, and Gb-level tape-out means it has begun to enter the productization verification stage. On the terminal side, the agent phone packages models, system, and hardware into one price and uses a system-level Agent to call third-party ecosystems, testing closed-loop capabilities in permissions, payments, and task fulfillment. On the data side, embodied intelligence uses real-machine competitions to test robot work efficiency while using wearable devices to supplement body motion and tactile information, seeking a middle layer between large-scale collection and high-precision recording. On the content side, AI variety shows push production costs down to several thousand yuan and cycles down to one week, and platforms then catch the supply with support plans. What is worth watching next: the progress of RRAM adoption in customer scenarios, the real usability and security boundaries of agent phone ecosystem calls, the noise and wearability adaptation issues of embodied data collection devices in real scenarios, and whether AI variety shows can form a stable content and commercial cycle under platform traffic support ^{[1][2][3][4][5][6]}.