從深港居民使用頻次最高的場景做起Starting where two cities meet, every day. Starting where two cities meet, every day.
深港人工智能應用創新中心，專注邊緣 AI 與端側推理，切入點是過關通勤、兩地消費、就醫配藥、升學查詢這類每天都在發生的事 —— 高頻、能被真實用戶當場檢驗，而兩地規則不同，正是通用模型最容易失手的地方。 A Hong Kong–based centre specialising in edge AI and on-device inference. We start from what happens every day — crossing the border, spending on either side, filling a prescription, checking a school rule — because those moments are frequent, judged on the spot by the people living them, and governed by two different sets of rules, which is exactly where general-purpose models come apart.
深港人工智能應用創新中心，專注人工智能前沿技術的應用落地；切入點是深港兩地居民使用頻次最高的生活場景。The Centre works on putting frontier AI to use, and it starts where people on both sides of the border actually spend their days.
我們不從行業大詞出發，而從一件具體的事出發：一個人早上要過關、中午要點餐、下午要配藥、晚上替孩子查升學規則。這些事每天發生上百萬次，兩地的規則、語言與證件卻不一樣 —— 通用模型在這裡出的錯，普通人一眼就能看見。這既是難處，也是最好的產品起點。We do not begin with an industry category; we begin with something concrete. Someone crosses the border in the morning, orders lunch, fills a prescription in the afternoon, and looks up a school rule that evening. These things happen millions of times a day, and the rules, the language and the paperwork differ on either side — so when a general model gets it wrong, an ordinary person sees it immediately. That is the difficulty, and it is also the best place to start a product.
從研究、適配到部署，我們關注的始終是同一件事 —— 把前沿能力，變成一個人今天就能用上的東西。From research and adaptation to deployment, we focus on one thing throughout — turning frontier capability into something a person can use today.
深圳的工程與製造、香港的研究與國際化，在中心交匯成一條從前沿到落地的完整鏈路。Hong Kong's research and global reach, Shenzhen's engineering and manufacturing — converging into a single path from frontier research to deployment.
Where ideas become hardware
Where research meets the world
中心位於深港之間 —— 把香港的研究、資本與合規，接上深圳的算力、製造與場景。兩地之間每天往返的人流，本身就是最好的測試場：同一個需求，在兩套規則下要有兩種正確答案。Positioned between the two cities, the Centre connects Hong Kong's research, capital and compliance with Shenzhen's compute, manufacturing and scenarios. The daily flow of people between them is the test bed: one need, two sets of rules, two correct answers.
圍繞「端側智能」展開：模型跑在人手上的設備裡，離線可用、資料不上傳 —— 這在跨境、弱網、涉及個人資料的日常場景裡不是加分項，而是前提。Built around on-device intelligence: the model runs on the device in someone's hand, works offline, and keeps the data there. Across a border, on a weak signal, with personal information involved, that is not a bonus — it is the precondition.
在終端設備上完成 AI 推理：低延遲、數據不出本地、離線也能用，兼顧效能與隱私合規。AI inference that runs on the device itself: low latency, data that never leaves, and offline operation — balancing performance with privacy and compliance.
適配國產 AI 芯片與端側算力平台，讓模型跑在手機、車機、一體機這類人手可及的設備上，成本、功耗與時延都在可控範圍內。Adapted to domestic AI chips and edge compute platforms, so models run on the devices people actually hold — phones, in-car units, all-in-one terminals — with cost, power and latency kept in range.
連接兩地的研究、資本與真實用戶，把方案放進日常場景裡反覆試 —— 好不好用，由每天真的要用它的人來判斷。Connecting research, capital and real users across both cities, we put solutions into everyday use and iterate — judged by the people who have to rely on them each day.
中心不只做技術，更串起生態各方 —— 讓高校的前沿研究、深圳的端側算力、兩地的真實用戶與國際資本，在同一件具體的事情上對得上。Beyond technology, the Centre links the players — university research, Shenzhen's edge compute, real users on both sides and international capital — around one concrete thing at a time.
我們不按行業分類選題，而按一個人一天裡真的會遇到的事情選題。這些場景每天發生、能被使用者當場判斷好壞、而且兩邊的規則、語言與證件都不一樣 —— 通用方案往往正是在這裡失手。We do not pick problems by industry category; we pick them from what a person actually runs into in a day. These moments are frequent, judged good or bad on the spot, and governed by different rules, languages and documents on either side — which is precisely where generic solutions fail.
多數產學研合作是企業把題目寫好、交給實驗室去做 —— 那樣得到的通常是一份外包，不是研究。我們反過來：把深港兩地高頻生活場景的現場開放出來，包括端側設備、真實用戶與兩地不同的規則，題目請老師和學生自己提。中心這一側的主線只有一條 —— FDE（前向部署工程師）：有人坐在使用者旁邊，把一句「不好用」翻譯成可度量的問題，再把方法送到日常能用的狀態。這條線上沒被回答的問題非常多，而且它們只有在現場才問得出來。Most industry–academia arrangements have the company write the question and hand it to a lab — which usually produces outsourcing rather than research. We work the other way round: we open the site — edge devices, real users on both sides of the border, and two different sets of rules — and the question comes from you. On our side there is one through-line: the forward-deployed engineer. Someone sits next to the user, turns “it doesn’t work well” into a measurable problem, and carries the method to the point where it holds up in daily use. There are a great many unanswered questions along that line, and they can only be asked on site.
中心不授學位、不佔招生指標、不參與錄取。學籍、導師關係、學術評價與畢業標準，全部留在高校一側。題目由高校導師與學生提出，中心不干預學術結論，只負責把現場、設備與使用者準備好。The Centre does not award degrees, does not hold admission quotas, and takes no part in admissions decisions. Enrolment, supervision, academic assessment and graduation standards all remain with the university. The question comes from the supervisor and the student; we do not interfere with academic conclusions — we prepare the site, the devices and the users.
目前正在與內地及香港的高校實驗室推進首批共建，以單實驗室、單課題、單學生為起步單元；一份一頁的提案就可以開始談。We are establishing our first partnerships with labs in the mainland and Hong Kong, starting with one lab, one topic, one student. A one-page proposal is enough to start the conversation.
FDE，前向部署工程師 —— 坐在使用者那一側，把問題定義出來，再把方法送到能被使用的狀態。它不替代學術訓練，而是補上論文之後、產品之前的那一段；這一段目前既缺人，也缺研究。The forward-deployed engineer sits on the user's side of the problem, frames it, and carries a method all the way to the point where it can be used. It does not replace academic training; it covers the stretch after the paper and before the product — a stretch short on both people and research.
使用者說出口的是「不好用」。第一項工作是把它翻譯成可度量的東西：哪一類輸入、多少時延、錯到什麼程度算錯、誰來判定。這一步做錯，後面所有優化都在解一道假題。What the user says is “it doesn’t work well.” The first job is to translate that into something measurable: which inputs, what latency, how wrong counts as wrong, and who decides. Get this wrong and every optimisation afterwards solves a fake problem.
算力、時延、內存、電量、私隱、離線可用性，每一項都是預算。方法的優劣要在這幾條線同時成立的位置上比較，而不是在資源無限的假設下比較 —— 這正是公開榜單給不出答案的地方。Compute, latency, memory, battery, privacy, offline availability — each is a budget. Methods have to be compared where all of them hold at once, not under an assumption of unlimited resources. This is precisely where public leaderboards stop being informative.
能裝到設備上、有權限與日誌、錯了有兜底、結果可復現。這是一項研究成果在真實使用者手上活過三個月的最低條件，也是學生履歷上最難偽造的一段。It installs on the device, has permissions and logs, falls back gracefully when it fails, and reproduces. That is the minimum for a research result to survive three months in real users' hands — and the hardest line on a student's CV to fake.
以下不是課題，是現場。每一條後面都躺著一堆還沒人回答的問題 —— 哪一個值得做成課題，我們想聽你的判斷。What follows are not topics; they are sites. Behind each sits a pile of unanswered questions. Which of them is worth turning into research — that judgement is what we want from you.
每天數十萬人往返兩地。最需要判斷的那幾分鐘，往往正好在弱網、換網、換卡的當口；口岸、接駁、票務的信息分散在互不相通的系統裡。一個只能在有網時工作的助理，在這裡等於沒有。Hundreds of thousands cross daily, and the minutes when you most need an answer are exactly the minutes of weak signal, network handover and a switched SIM. Checkpoint, transit and ticketing information sits in systems that do not talk to each other. An assistant that only works online is no assistant here.
同一件事，在兩邊有兩套正確答案：支付方式、找換、單據、報銷、退稅。模型不但要知道規則，還要知道自己此刻站在哪一邊 —— 而它常常不知道。The same task has two correct answers depending on which side you are standing on: payment method, exchange, receipts, reimbursement, tax refund. The model has to know the rules and also know which side it is currently on — and it usually does not.
繁簡混排的菜單、粵語與普通話夾雜的口語、手寫價目與拍糊的照片。這是一個對端側小模型極不友好、對使用者卻極其日常的多語種理解問題。Menus mixing traditional and simplified characters, speech switching between Cantonese and Mandarin, handwritten prices and blurry photographs. A multilingual understanding problem that is brutal for a small on-device model and utterly ordinary for the person holding the phone.
兩地的藥名、劑量單位與處方規則不同，而使用者要的往往只是一次可靠的對照與提醒。這裡沒有任何討價還價的空間：資料只能留在端側，錯了要能被發現，講不清楚就必須說「去問醫生」。Drug names, dosage units and prescription rules differ across the border, and what the user usually needs is one reliable cross-check and a reminder. There is no room for negotiation here: the data stays on the device, errors have to be detectable, and when the answer is not clear the honest output is “ask a doctor.”
兩地學制、插班與報考規則變動頻繁，家長面對的是散落各處、時效性極強、且錯一步代價很大的信息。這是一個檢索與核實遠比生成困難的問題。School systems, mid-year transfer and application rules change often; parents face information that is scattered, time-sensitive, and costly to get wrong by one step. Here retrieval and verification are far harder than generation.
表格、證件、住址、公文用語，在兩地有各自的寫法與慣例。翻譯不是難點，「在這一邊該怎麼填才會被受理」才是。Forms, identity documents, addresses and official phrasing follow their own conventions on each side. Translation is not the hard part; knowing how it has to be filled in on this side to be accepted is.
不需要立項書格式，不需要先有結果。老師、學生都可以直接來信；同一個現場我們歡迎不同的提法。No formal template, no preliminary results needed. Supervisors and students are equally welcome to write; different takes on the same site are welcome too.
具體到一個人、一個動作、一次失敗。「跨境場景的體驗優化」不算，「在口岸弱網下，離線助理答不出換乘」算。Specific to one person, one action, one failure. “Improving the cross-border experience” does not count; “on a weak signal at the checkpoint, the offline assistant cannot answer the transfer question” does.
什麼指標、在什麼設備上、多少樣本、跟誰比。允許粗糙，但要能被別人重做一遍。Which metric, on which device, how many samples, compared against what. Rough is fine; reproducible by someone else is not optional.
設備、場景、使用者、工位、算力，還是一位在現場待過的人。說清楚了，我們才知道自己接不接得住。Devices, the setting, users, a desk, compute — or a person who has spent time on site. Say it plainly, so we can tell you honestly whether we can provide it.
兩週內回覆。合適的話，下一步是一起去一次現場，而不是先簽一份協議。We reply within two weeks. If it fits, the next step is a visit to the site together — not a signed agreement.
生活場景涉及個人資料，因此一律最小化收集、優先端側處理、參與者可隨時撤回。研究開始前先過高校倫理審查與中心的私隱評估；涉及跨境的資料處理，按兩地各自的規則分別執行。Everyday scenarios involve personal data, so collection is minimised, processing happens on the device wherever possible, and participants can withdraw at any time. Work begins only after university ethics review and the Centre's own privacy assessment; anything crossing the border follows each jurisdiction's rules separately.
方法與結論經脫敏審查後可正常發表。審查範圍僅限參與者身份與個人資料，窗口 15 個工作日，無異議即視為通過；中心不以審查為由干預學術結論。Methods and conclusions are publishable once desensitisation review is complete. Review covers participant identity and personal data only, with a 15-working-day window; no response means cleared. Review is never a lever over academic conclusions.
學生的學術成果與學位論文權利歸學生與高校。中心不主張對通用方法學的獨佔；僅當技術與中心既有產品或特定場景直接結合、形成可專利成果時，雙方就共有與實施許可另行約定。Academic outputs and thesis rights belong to the student and the university. The Centre claims no exclusivity over general methodology; only where the work combines directly with an existing product or a specific setting to form a patentable result do we agree separate terms on joint ownership and licensing.
學生畢業與發表，不受商業條款掣肘。這是前提，不是讓步。A student's graduation and publication are not subject to commercial terms. That is the premise, not a concession.
合作、洽談或諮詢，歡迎隨時來信，我們會盡快回覆。For partnerships, projects or enquiries, write to us anytime — we'll reply promptly.
lxy@shengangai.com