NeuroAI

An AI agent is a system that can plan, use tools, and act toward a goal with limited human hand-holding — not just answer one prompt. This guide explains how agents differ from chatbots, what they are made of, and where Chinese labs like DeepSeek fit in.

Artificial general intelligence (AGI) is the hypothetical point where a machine matches or exceeds humans at most cognitive tasks. This guide explains what AGI means, how it differs from today's narrow AI, why nobody agrees on a test for it, and how China and the US each talk about it.

China runs the largest robotaxi (自动驾驶出租车) trials on Earth, led by Baidu's Apollo Go (萝卜快跑), Pony.ai and WeRide. This guide explains how they work, why Chinese cities moved faster than most, and what the safety and cost picture actually looks like.

Nvidia dominates AI compute with its GPUs and the CUDA software stack; AMD is the credible challenger with its MI-series accelerators. This guide explains the real difference, why software matters as much as silicon, and where China's Huawei Ascend fits in.

China has become the world's largest source of openly released AI models. This guide explains what "open weights" means, why Chinese labs like Alibaba (Qwen), DeepSeek and Zhipu (GLM) give models away, and what that means for developers outside China.

China's AI strategy is less a single law and more a stack of plans: a national agenda, the "AI Plus (人工智能+)" campaign to push AI into every industry, and the 15th Five-Year Plan. This guide explains the moving parts in plain English.

DeepSeek is a Chinese AI lab that released frontier-grade models with open weights and a fraction of the training cost. This explainer covers what "open weights" means, why it mattered, and what it changed.

A brain–computer interface (BCI) reads electrical signals from the brain and turns them into commands a machine can act on. This guide explains how invasive and non-invasive BCIs work, what they can and cannot do today, and why China has become a surprising center of activity.

Neuralink gets most of the headlines, but in 2026 a Chinese implantable BCI, NEO, became the first to win commercial approval. This piece compares the two approaches on approval status, use cases, and what "winning" actually means.

Everyone is building a humanoid robot. This field guide maps the serious contenders — Unitree, AgiBot, Figure, Tesla Optimus and the rest — with the production numbers that actually matter, not just the videos.

Nvidia's GPUs power most AI outside China. Huawei's Ascend series is the domestic alternative Beijing is betting on. This explainer compares what each does, the supply gap, and why it matters.

Strip away the videos and a humanoid robot is three problems: move, see, and think. This explainer breaks down actuators, sensors, and the "brain" — world models and VLA — and why the brain is now the hard part.

Wang Xingxing says embodied AI's real bottleneck is generalisation, and proposes a threshold — a robot that works in 80% of unfamiliar places on 80% of tasks. That is 64%, and the industry is not there yet.

At its 2026 tech conference JD committed to 3 million robots, 1 million unmanned vehicles and 100,000 drones over five years — and, in the same breath, 100,000 new service-engineer jobs. Automation and employment, booked together.

China's BCI market is around RMB 3 billion today, forecast at RMB 12 billion by 2030. A Tsinghua professor's shoe metaphor explains why the number is small — and why that is the industry's real problem.

Ling-3.0-flash-VL activates only 5.5B of 124B parameters, reads 256,000 tokens at once, and runs an observe–act–verify–correct loop. Open weights are not charity; they are a land grab.

AI comic-drama uploads went from ~20,000 in January to 47,000 by March. The hit rate is 0.16%. The scarce thing is no longer generation — it is the pipeline that keeps a face stable across a hundred episodes.

Tencent's WeMM-Embedding ships in 2B, 4B and 9B sizes; the 9B tops MMEB-v2 at 80.6. It already runs at billion-calls-per-day scale inside WeChat — open weights forged in production, not in a lab.

DeepSeek opened ~150 roles in server-side development and agent elastic computing — zero AI research positions. The next day Reuters reported it had hired CITIC Securities to prepare a STAR Market listing. The build-out is infrastructure, not papers.

From 10 September 2026, DeepSeek's Flash series offers off-peak cache-hit inputs at ¥0.02 per million tokens. The cut targets repeated, high-volume work — a subsidy for AI that runs all day, not AI that chats.

Beijing-backed NEUCYBER says its semi-invasive BeiNao-1 system has been implanted in more than 40 patients, the longest running past 18 months. It is the clearest sign yet that Chinese neurotech is becoming a real medical product, not a lab demo.

A vertical short-drama app has more daily users than China's four biggest streaming platforms combined. An AI episode costs a tenth of a live-action one — and the producers will tell you exactly where it still fails.

Udeer Robotics listed in Hong Kong on 9 September 2026 with over 114,600 delivery robots sold to 5,100+ customers. Not humanoid — and that is exactly the point: dull robots ship first.

Chinese makers now ship more than 400 humanoid models — over half the world's total — and account for roughly 80% of global humanoid and quadruped sales. The story has shifted from spectacle to scale, and scale is the harder thing to copy.

In 2026 brain–computer interfaces entered China's Government Work Report and five-year plan. Behind the headlines, regulators are writing the classification, ethics and reimbursement rules that decide whether neurotech becomes medicine or stays experimental.

Huawei's Kirin 9050 Pro supports on-device MoE models of 30B total parameters with 2B active. The number that matters is not the parameter count: it is the 66% cut in power draw for key tasks.

Supplier guidance went 150 units a week in July to 300 in August to 1,000 in September — then a single batch of about 5,000 Optimus. That is not linear growth; it is the steep part of an S-curve.

China's BCI work reaches far beyond paralysis. Demos now run from mining helmets that watch for fatigue to a visual prosthesis that lets a blind patient read letters, to an implant delivered through the blood vessels. The question is which of these scales first.

On 6 September an anti-fraud app guided by the Ministry of Public Security went live. Three functions, one of which — read-aloud — matters more for elderly users than it sounds. The tech is easy; the trust is the hard part.

Volume AI short drama went from RMB 5,000 per minute to several hundred in under a year. Meanwhile Shanghai studios charge RMB 200k–300k for a 20-minute branded piece. Under 1% of creators earn real money.

iFlytek's Spark X2.5 was trained and served end-to-end on domestic accelerators, with open on-device models small enough to run on a laptop. The constraint moved from silicon to electricity.

Three billion downloads, permissive licences, and a developer default that no longer points West. China's open-model labs did not win an argument — they changed what developers reach for first.

Class III is China's strictest medical-device review — long-term implant. Since early 2026, multiple BCI devices have cleared it, shifting the question from 'is this possible?' to 'how fast can it reach patients?

On 6 September Qwen released Qwen-Drive-1.0-4B — described as the first vision-language foundation model built for autonomous driving, folding perception, VQA and planning into one 4B model. Weights are free on GitHub, Hugging Face and ModelScope.

NEO won regulatory approval in March 2026 after 32 cases across 11 hospitals with a 100% response rate on assisted grasping — and Shanghai now covers it on insurance. Overseas patients are emailing weekly.

Unitree says cumulative production of its bipedal humanoids has passed 18,000 units, with over 5,500 shipped in 2025 alone, and it is now profitable. In an industry still accused of selling videos, those are rare numbers.

The MIIT's SME AI Entrepreneurship Support Plan (2026–2028) commits to 10 national inclusive compute centres and compute vouchers. Three pitfalls apply to any voucher scheme: pay-first-reimburse-later, hidden eligibility, long settlement.

Runway's Solaris predicts and renders the next screen directly instead of generating code first. In tests with 250 evaluators, 61% preferred its interfaces on instruction-following and 71% on natural feel.

Global humanoid shipments passed 22,000 units in H1 2026, up nearly 300%. Manufacturing took 13% and logistics 5%; the rest went to stages, showrooms and labs. The install base is real, the utilization is not.

Tens of thousands of humanoids have been built. The hard part is the next one: teaching them to do jobs. A national 'real-scenario training' programme, robot farms and data centres are being built to close the last kilometre.

CloudMatrix 384 links 384 Ascend 910C accelerators into a single logical computer. Independent analysis puts it near double the dense compute of NVIDIA's GB200 — at roughly four times the power. The win is in scale, not efficiency.

The MIIT plan explicitly encourages local governments to accommodate 'one-person companies' and 'super-individuals' doing agile entrepreneurship with intelligent tools — the first national-level recognition of the form.

Reuters reported on 3 September that Moonshot, developer of Kimi, has confidentially filed for a Hong Kong IPO targeting ~US$3 billion at a ~US$50 billion valuation — two and a half times its start-of-year figure.

Kuaishou's Kling passed 100 million global users across 224 countries in June 2026, with quarterly revenue above 850 million yuan and a spun-out valuation near $18 billion. A state-backed short-video app became a global AI product.

Vertical integration, cheap actuators and brutal iteration speed have made China the place humanoid hardware gets built. The strategic question is whether that lead survives the shift from metal to intelligence.

A Beijing-developed semi-implantable BCI, BeiNao-1, enabled a paraplegic patient to stand and walk with a frame — as China debuts its first dedicated BCI exhibition zone at a national assistive-technology expo. The clinical milestone is real; the scale-up is the unknown.

Lian Wenzhao, a robotics PhD who worked at Vicarious, Google X and Figure before founding Yuanluo in China, argues the humanoid sector should prize reliability over speed. His 'impossible triangle' is gaining traction as Beijing warns against blind frenzy.

On OpenRouter, the share of tokens US companies spend on Chinese models has sat above 30% every week since February 2026 and peaked at 46%. Engineers are not switching for ideology — they are switching because the models are good and about fifty times cheaper.

The “AI+” initiative is not a research programme — it is an adoption programme with dated targets. It may be the single most under-covered story in Chinese technology, because it is about deployment, not hype.

On 1 September the Nubia NaviX Ultra — built with ZTE and ByteDance, running Doubao — received network approval as the first 'AI agent phone'. The technology works. The question is whether you trust it with payment.

On 2 September Yuanidian launched Zeroth Bridge, a personal high-performance humanoid at RMB 8,888 for the geek edition, with secondary-dev support. Three cost curves converged at once to make the price possible.

TrendForce forecasts the nine largest cloud providers will spend over US$886.7 billion in 2026, up about 90%. Increasingly the money is not going to GPUs — it is going to cooling, packaging, interconnect, power and memory.

On 1 September Shenzhen published draft BCI rules banning mind control, illegal neural-data collection and irreversible neural damage or false advertising. The move pairs promotion with guardrails as consumer BCI products proliferate.

Enflame raised RMB 6.1 billion at RMB 142.18 a share. The listing is a beginning, not a verdict: a high-end AI chip costs hundreds of millions to tape out, and the real moat is the software ecosystem.

From 1 September Huawei, Xiaomi and Honor raised prices across models, some by RMB 200–1,000. The cause is memory: procurement reportedly went from ~RMB 500 to RMB 2,200. AI data centres took the capacity.

Forget the demos. In Chinese factories and clinics, AI is being measured in changeover minutes saved and misreads avoided, not in press releases. The gains are boring, which is exactly why they stick.

Shao Tianlan founded Mech-Mind in 2016; on 1 September it listed in Hong Kong — 3,843 times oversubscribed, ~HK$12.4 billion market value. Sequoia China invested in four consecutive rounds and holds 13.63%.

Running the same workload for a fiftieth of the price is not a discount — it is a different product. China's labs turned a compute constraint into a cost structure now reshaping global model economics.

A Chengdu hospital listed a non-invasive BCI adaptation fee above RMB 900 — among the first official prices for the technology in China. Without a price, hospitals cannot charge; without charging, they cannot procure.

The least glamorous and most consequential AI story in China is not in Beijing's labs — it is in rice breeding, county-hospital radiology and the local social-security desk. This is where adoption actually compounds.