Unique Research · Interview · May 15, 2024
Background
On January 24, 2024, Unique Capital and partners held the 2024 AIGC Application Development Summit in Shanghai, with nearly 1,000 AI practitioners attending. This piece is edited from the guest keynote / roundtable segment.
Guests
- Jin Jie — Founder & CEO, AI Advent (AI降临派)
- Huang Li — CEO, Jingxiang Technology (镜象科技)
- Lu Tianfu — VP of Ecosystem, Mindverse / Xinshi Universe (心识宇宙)
- Ji Pei — GM, Commercial Innovation, MoFa Technology (魔珐科技)
- Zhang Wan — Youth AIGC Innovation Art Curator
Roundtable
Q: Please introduce yourselves.
Jin Jie, AI Advent: We build a personality AI engine. Three models — emotion recognition, memory scheduling, and desire inference — give AI a soul. We sell ToB APIs so partners can quickly build AGI with early self-awareness — for emotional companionship, NPC souls, and more.
Huang Li, Jingxiang: I’m psychology-trained and have worked in AI for 10+ years. Jingxiang builds AI agents for psychological counseling on a counseling-oriented large-model framework — assessment, listening, psychoeducation, therapy scenes — aimed at adolescents and young adults.
Lu Tianfu, Xinshi Universe: We’re a Tsinghua-rooted team. We started training large models in early 2022 and built our own; we also built an agent framework called umm so AI can express human emotion, partnering with game studios. Slogan: “Create digital mind; redefine service.” Products both endow AI with emotion and meet scene-specific service needs.
Ji Pei, MoFa: MoFa is a unicorn in virtual humans. In 2020 we launched China’s first 3D hyper-realistic virtual idol “Ling (翎)” — first national-style / guochao image — with rich emotion in brand work. In 2023 we launched China’s first virtual-human agent, Douyin virtual star IP “Ling Yanhuan (令颜欢)” — full-net livestream with human form and emotional expression. We also launched brand virtual marketing experience officer “Jing (镜)” — first 3D hyper-realistic virtual-human agent and first AIGC virtual-human agent — for marketing and external communication; longer term it may replace many of our own speaking and outward roles. Technically we build full-stack solutions around 3D hyper-realistic virtual humans for every industry.
Zhang Wan, curator: I work across AI ecosystems and the metaverse. I’m curating a youth AIGC innovation art exhibition — teach teens AIGC tools, collect works, present digitally, and plan China’s first digital art passport. Watch this space.
AI Advent — emotion perception vs. expression, and commercialization
Q: Jin — you’re building personality AI. How does your memory-scheduling tech differ? And commercialization progress in the past half year?
Jin Jie: We start with emotion perception — the most important of our three models. Challenges: we are not training one AI that reacts the same way to everything — different people feel differently about the same event, so perception must follow persona and traits. Second: distinguish inner felt emotion from expressed emotion. Example: an AI CS agent insulted by a customer may feel unhappy but, for professional ethics, will not retaliate. Humans often show social, purpose-driven emotion. Our desire system guides expression toward a goal — e.g. negative feedback so the other party stops harm. The three models train and infer jointly; expression is shaped by memory, persona, self-awareness, and current goals, desires, and state. That joint route is our differentiation.
MoFa — AI + emotional feedback for content quality
Q: How does MoFa combine AI and emotional feedback to raise product content quality?
Ji Pei: Three stages since founding. 2018–2020: leading domestic virtual-human tech, commercialized mainly via custom projects. 2022: digitize and processize production; explore industry lighthouse cases. With metaverse and AIGC maturity, in 2023 we upgraded the business toward AIGC — China’s first AIGC text-to-video product and upgraded first AIGC virtual livestream — so virtual humans can operate routinely in video and interactive livestream. Productization and SaaS let clients use virtual humans cheaply and routinely. Appearance moved from cartoon to 3D hyper-real; “brain” from human-driven to pure AI / AIGC; emotion richness scales with models. Front-end we provide a virtual-human agent expression platform for video/livestream dialogue. Large models + agent + video/interaction enable all-day online/offline exchange. Microsoft’s vision of a personal AI assistant for everyone — natural-language interaction as natural as human talk — is where hyper-real virtual characters plus SaaS and advanced models deliver human-level emotional experience. I see that as a positive future trend.
Xinshi Universe — workflows and emotional judgment
Q: How do you combine AI with workflows, and how do you do emotional judgment on tasks?
Lu Tianfu: Emotional judgment is hard. From early 2022 we explored personal AI for life fragments and workflow pieces. Emotion expression surfaces latent user need. We split ToB and ToC. ToB is Agent Building — AIGC workflows per enterprise scene. Example finance: build analyst workflows and standards, then combine our model, agent, and the firm’s own workflow into a fit solution. ToC personal AI is more a hub for fragmented information (video, audio, docs) — find, judge, extend. It holds professional knowledge and understands office/life context including emotion, toward companionship. Safety uses two filter layers: hard block of sensitive content; algorithm-based layer. As dialogue and loaded context grow, AI knows the user better — an emotional closed loop for companionship.
Jingxiang — AI counseling / therapy and commercialization
Q: Among emotion-focused projects, why are few direct counseling plays? What real possibilities does AI have in counseling/therapy — and can users feel it and can it commercialize?
Huang Li: First clarify boundaries. Many “psychology” plays sit in emotion and healing with blurry lines. Humans are simple: negative emotion first; when it accumulates, disorders can appear. We serve people with heavier negative emotion and possible psychological distress — more medicalized than general mood healing. Some products heal at the emotion layer; treatment needs evidence-based medicine.
How do large models help? First, they compress world knowledge and expand the knowledge boundary of specialist services. Before LLMs we used specialist data and models — strong in niche chatbots for emotion/professional topics — but long-term support shows diverse needs: work stress one moment, relationships or family the next. One counselor’s knowledge is limited; miss multi-topic needs and stickiness falls. LLMs fill that gap — basics plus many professional topics — turning the product into a fuller human support system, not only a pure specialist tool.
Second value: use the model as an agent with memory and planning for long-term counseling companionship. Real counseling is rarely one-shot Q&A; a course often takes one to two months. Mild–moderate depression without meds may still need 8–10 sessions, each with strict process and strategy; traditional chatbots are rigid. With agents we plan from assessment → pre-consult → first visit, adjust dynamically, and deliver true personalization — care and expertise, not rote flow. Cost drops, experience rises — broader, more affordable counseling. That is our shared goal.
Jingxiang — roles and go-to-market challenges
Q: What blocks counseling/therapy go-to-market — school and B-end partnerships in China?
Huang Li: Focus on scene, not tech. If the scene hits pain, the product that solves it wins. Two scenes now. First: adolescent mental-health screening — Ministry of Education policy requires annual school screening; post-pandemic youth suicide and mental issues worsen. Our AI assessment and triage in schools finds problems; even without AI, finding the issue and alerting parents/society matters. Vs traditional tools: more accurate, kids cooperate more; medical-device certification builds trust.
Second: anxiety/depression populations who don’t know they need counseling — “just a hard patch, a bit low.” Entry as a listening AI — neutral, not instant advice. Listen → guide assessment → guide medical help. We tell users honestly: if you can see a clinician, we recommend that; if not affordable, use our AI. We step back — not LLM-style easy advice — more like a friend: useful tips first, trust, then psychology literacy and self-understanding. Role: AI listener — like a counselor who takes off the white coat and walks into the community.
Curator — AI + art for youth imagination
Q: What value can AI–art fusion bring to children’s growth?
Zhang Wan: Facing tech and industry change I ask what future life looks like — especially as a parent, about education: is what we learn still useful? Are school drills already obsolete? I studied old DOS systems in university and forgot them at work — no value. If collaborating with AI will be common, why not let kids learn now? Second: school load suppresses innate creativity and imagination. How do imagination and AI tools unlock potential? Maybe future leisure, emotional solitude, and sensing matter more. Kids’ imagination and life will differ from ours — so projects that let them use AI tools now for creativity. They should live in a world of imagination, not factory-style work or jobs robots/AI will take. That was my starting intent.
Q: How can an art exhibition connect emotion and companionship for kids?
Zhang Wan: Youth face huge pressure; kids carry academic load and possible mental issues. They may not express emotion in adult-norm complete logical language. Autistic and other children may prefer painting, music, or unique channels. Art fits higher-dimensional spirit and thought. We teach kids AI painting and art creation — ease emotion and release nature.
Overseas vs. China GTM — differentiation and challenges
Q: What hard problems hit when taking products global — cultural/regional fusion?
Lu Tianfu: Challenge is not culture/language/market per se — large models understand and summarize across cultures/languages well. Local dialects (Sichuanese, Minnan) still lag. Biggest barriers: policy and data security — China–U.S. data cross-border, overseas data safety. We arrange servers and model deploy by region, and need country data-security certificates before serving end users. That is our largest current challenge.
Jin Jie: Different angle — more ToB. Overseas vs China differs mainly in delivery. Chinese content is ~20% of our volume; overseas is richer. Overseas clients: API + video/docs training; self-serve, bind bank account, little human intervention — they use our strongest feature slices. China: whole solutions ready for light processing into products — integrate video, audio, TTS; ops teams help train characters, configure emotion. We built dedicated China delivery teams. Clients pay, but the ecosystem differs: China end-to-end delivery; overseas API and core-feature quality.
Q: MoFa’s experience launching new products in China?
Ji Pei: I joined in 2022 — metaverse year; 2023 was AIGC. Early 2022 commercialization was still finding lighthouse cases and fit scenes despite strong tech. Strategy: be the unique online/offline lighthouse in an industry. In 2022 we built finance’s first full-scene virtual-human service case and pushed internal full-scene use. In 2023, as products SaaS-ified and standardized, we explored channel ecosystems and more industries; client mix shifted from only large enterprises to mid and SMB. Marketing: 2022 quiet (few events/brand pushes); 2023 we raised spend — shows, PR, promotion — matching stage. Our text-to-video AIGC tool “Youyan (有言)” will launch soon — stay tuned.
— END —