聲明:本文來自于（ID:ykqsd.com）授權(quán)轉(zhuǎn)載發(fā)布。

In the latest update of the China Academy of Social Sciences' annual meeting, "Artificial Intelligence+" has emerged as a key topic.

The government work report this year emphasizes the promotion of the "Artificial Intelligence+" initiative, aiming to integrate digital technologies with industrial strengths and market advantages, supporting the widespread application of large models, and developing advanced intelligent terminals, smart manufacturing equipment, intelligent robots, and intelligent vehicles.

To explore how AI can drive innovation across various industries and address challenges in its application, The China Academy of Social Sciences conducted interviews with industry experts. In their discussions under the "+" framework, representatives from diverse sectors elaborated on concepts like " personal intelligence," education, safety, and AI agents, offering unique insights into the practical application of AI within their specific domains.

Notably, several industry insiders have highlighted the growing importance of cost and performance trade-offs in the deployment of large models. While advancements in training and inference efficiency have reduced operational costs, actual deployment expenses remain significant. Experts suggest that minimizing computational overhead while ensuring efficient resource utilization is critical for successful implementation.

" Artificial Intelligence+" is driving innovation across industries, blurring the lines between AI and robotics.

The rapid development of AI is fostering the convergence of large models and robotic technologies, significantly enhancing the autonomy and interaction capabilities of robots.

As an AI innovation and application company, CEO Li Daixiang of Baidu revealed that integrating AI with robotics under the " personal intelligence" paradigm has led to the development of a " brain on the end," achieving a world-first case in AI applications for human-like robots.

Advanced end-side models are accelerating the deployment of " personal intelligence" by combining natural language interaction, broad command capabilities, low response times, comprehensive multi-modal processing, real-time data handling, efficient hardware use, and strong security features. Baidu's MiniCPM end-side model exemplifies this capability, with vast application potential.

Challenges such as difficult data collection, high annotation costs, and limited utilization rates persist in industries like autonomous vehicles and " personal intelligence." LightAuto, a leader in synthetic data solutions, addresses these issues by offering simulated data that mimics real-world scenarios, helping enterprises build virtual environments for more efficient and cost-effective AI development.

LightAuto's long-term investment in synthetic data and simulation technology directly tackles the data scarcity issue hindering large model pre-training. In the realm of smart manufacturing and industrial digitalization, synthetic data enables the creation of virtual environments that support optimized production, quality control, and intelligent decision-making, significantly enhancing productivity and efficiency.

數(shù)據(jù)表明，教育作為消費(fèi)行業(yè)的核心屬性，憑借龐大的用戶基礎(chǔ)，AI+教育有望率先實(shí)現(xiàn)技術(shù)落地應(yīng)用。

在AI教育應(yīng)用領(lǐng)域，萬(wàn)物引力科技公司創(chuàng)始人兼首席執(zhí)行官肖翔峰向《科創(chuàng)板日?qǐng)?bào)》記者介紹，其開發(fā)的 Talkface.AI 產(chǎn)品是一款A(yù)I語(yǔ)言教學(xué)工具，專為提升用戶口語(yǔ)能力設(shè)計(jì)。

肖翔峰建議，政府、國(guó)有資本與科研機(jī)構(gòu)、產(chǎn)業(yè)園區(qū)應(yīng)共同設(shè)立專項(xiàng)孵化渠道，重點(diǎn)扶持中小企業(yè)；同時(shí)應(yīng)建設(shè)先進(jìn)模型與算力加速中心，鼓勵(lì)個(gè)人及中小型創(chuàng)業(yè)者利用大模型技術(shù)探索各行業(yè)應(yīng)用場(chǎng)景。

北京盒智科技創(chuàng)始人兼首席執(zhí)行官魯雅琦向《科創(chuàng)板日?qǐng)?bào)》記者表示，AI技術(shù)的快速迭代為智能硬件與訂閱服務(wù)的爆發(fā)提供了新機(jī)遇。通過部署低成本的硬件功能升級(jí)方案，以及利用大模型的推理能力，公司在B端教育信息化與智慧校園建設(shè)領(lǐng)域也具有廣闊空間，可解決智慧教考、智能評(píng)測(cè)、自動(dòng)化批改等場(chǎng)景需求。

作為中國(guó)AI領(lǐng)域先行者之一，未來式智能創(chuàng)始人兼首席執(zhí)行官楊勁松對(duì)《科創(chuàng)板日?qǐng)?bào)》記者指出，技術(shù)產(chǎn)業(yè)化路徑的清晰化是當(dāng)前的關(guān)鍵方向，通過"人工智能+"戰(zhàn)略推動(dòng)技術(shù)向應(yīng)用層面擴(kuò)散，避免重研發(fā)輕應(yīng)用的失衡。此外，公司還在AI與新能源汽車、智能終端等優(yōu)勢(shì)產(chǎn)業(yè)之間建立協(xié)同，形成技術(shù)突破到產(chǎn)業(yè)升級(jí)再到經(jīng)濟(jì)增長(zhǎng)的良性循環(huán)。

楊勁松認(rèn)為，這些政策的有效實(shí)施將推動(dòng)中國(guó)在全球AI產(chǎn)業(yè)中實(shí)現(xiàn)應(yīng)用場(chǎng)景的主導(dǎo)地位，而非僅停留在基礎(chǔ)技術(shù)的追趕。

隱拓智安的首席執(zhí)行官?gòu)執(zhí)祧蜗颉犊苿?chuàng)板日?qǐng)?bào)》記者介紹，作為AI安全領(lǐng)域的創(chuàng)新企業(yè)，公司高度關(guān)注政府提出的統(tǒng)籌發(fā)展與安全的戰(zhàn)略布局。目前，90%的大模型仍處于裸奔狀態(tài)，缺乏有效的安全防護(hù)體系。

張?zhí)祧握J(rèn)為，對(duì)端側(cè)模型的全生命周期安全防護(hù)將成為行業(yè)發(fā)展的剛需，將迅速催生新的安全技術(shù)和服務(wù)。建議在醫(yī)療、金融等高敏感行業(yè)設(shè)立AI安全技術(shù)試驗(yàn)專區(qū)，為企業(yè)提供合規(guī)框架下的創(chuàng)新實(shí)踐空間。

"我們看到了DeepSeek帶來的熱度之后，各行各業(yè)加速擁抱人工智能。大模型公司、應(yīng)用公司與硬件公司協(xié)同合作，共同推動(dòng)AI技術(shù)在各行業(yè)的落地。麥伽智能聚焦法律科技領(lǐng)域，2025年加速拓展法院、檢察院和司法行政機(jī)關(guān)的智能輔助應(yīng)用落地，業(yè)務(wù)已覆蓋企業(yè)和事業(yè)單位的合規(guī)管理。"

據(jù)財(cái)聯(lián)社創(chuàng)投通統(tǒng)計(jì)，2023年以來，人工智能賽道共發(fā)生741起融資事件，涉及1795家投資機(jī)構(gòu)，其中奇績(jī)創(chuàng)壇、紅杉中國(guó)、啟明創(chuàng)投、藍(lán)馳創(chuàng)投、阿里巴巴等知名機(jī)構(gòu)活躍，累計(jì)融資金額達(dá)62.296億元。

"降本增效"助力企業(yè) more affordably utilize large language models

年初DeepSeek的橫空出世，引領(lǐng)了"深思考"時(shí)代潮流，不僅加速了人工智能技術(shù)的快速普及，更使算力面臨了一輪新的挑戰(zhàn)。

如何在保證高性能的前提下，系統(tǒng)性地降低千億級(jí)大模型的部署成本？在這一背景下，AI Infra賽道迎來了崛起機(jī)遇。AI Infra是指在大模型生態(tài)系統(tǒng)中，連接算力和應(yīng)用的中間層基礎(chǔ)設(shè)施，包括硬件、軟件、工具鏈和優(yōu)化方法等，是一個(gè)整體解決方案。

基流科技創(chuàng)始人兼CEO胡效赫對(duì)《科創(chuàng)板日?qǐng)?bào)》記者表示，DeepSeek-R1的全球影響力數(shù)據(jù)是對(duì)標(biāo)量定律在推理和合成數(shù)據(jù)方向發(fā)展的有力印證，這將進(jìn)一步推動(dòng)算力需求呈指數(shù)級(jí)增長(zhǎng)。

當(dāng)前，算力成本與效率的矛盾日益尖銳，單純依賴硬件擴(kuò)展已難以滿足可持續(xù)發(fā)展需求，唯有通過跨層級(jí)優(yōu)化實(shí)現(xiàn)"質(zhì)效雙升"。在軟件系統(tǒng)層面，需優(yōu)化分布式訓(xùn)練框架和資源調(diào)度算法，從而在單位算力成本下釋放更高模型性能，這正是基流科技的核心能力所在，也是與傳統(tǒng)設(shè)備廠商的關(guān)鍵差異所在。

同時(shí)，基流科技自主研發(fā)了端到端國(guó)產(chǎn)智算通信系統(tǒng)，實(shí)現(xiàn)了智算網(wǎng)絡(luò)中網(wǎng)卡和交換機(jī)芯片的國(guó)產(chǎn)化替代。這將有效降低對(duì)進(jìn)口設(shè)備的依賴，進(jìn)一步提升算力的自主可控能力。

據(jù)了解，千億參數(shù)模型的私有化部署仍然面臨諸多挑戰(zhàn)。盡管成本有所降低，但推理耗時(shí)較長(zhǎng)，中小企業(yè)在預(yù)算壓力下不得不采用輕量化模型，這在一定程度上犧牲了模型的智能化水平。

聚焦AI Infra賽道的趨境科技CEO艾智遠(yuǎn)告訴《科創(chuàng)板日?qǐng)?bào)》記者，從基礎(chǔ)設(shè)施廠商的視角來看，國(guó)家關(guān)于大模型應(yīng)用和算力建設(shè)的政策導(dǎo)向，需要更加注重如何在終端有限設(shè)備上合理調(diào)用算力、提升算力利用率。他指出，可以從計(jì)算量的底層邏輯出發(fā)，一方面壓低成本，另一方面盡量實(shí)現(xiàn)推理引擎的標(biāo)準(zhǔn)化和算力的可重用。

艾智遠(yuǎn)補(bǔ)充道，公司特別推出了"以存換算"這一新思路，通過這一創(chuàng)新，能夠更好地利用存儲(chǔ)資源作為算力補(bǔ)充，尤其在檢索和生成場(chǎng)景中，響應(yīng)延展能力能夠降低20倍，性能提升可達(dá)10倍。

一位頭部機(jī)構(gòu)投資人對(duì)創(chuàng)投日?qǐng)?bào)表示，趨境科技憑借"以存換算"和全系統(tǒng)異構(gòu)協(xié)同優(yōu)化的創(chuàng)新路徑，一定程度上緩解了千億參數(shù)模型落地過程中"部署成本高昂"和"性能折損"的雙重困境。

在AI Infra領(lǐng)域，不僅包括一些新興的初創(chuàng)企業(yè)如硅基流動(dòng)、無問芯穹、潞晨科技、清昴智能等，還包含一些行業(yè)巨頭，例如阿里云、華為云、字節(jié)火山引擎、百度等。這些企業(yè)通過各自打造完善的云服務(wù)平臺(tái)，整合了算力調(diào)度、模型訓(xùn)練和部署等核心能力，為AI Infra的全棧解決方案提供有力支持。