14 projects. 50+ participants. 9 hours. 2 tracks merged into one. One overwhelming theme: AI agents are everywhere, and they have personalities now.
What happened during 9 hours in Shenzhen, and why it matters for the future of AI development
Stepfun AI (阶跃星辰) hosted a hackathon in Shenzhen in June 2026, gathering approximately 50–60 participants from China, Australia, Poland, Belarus, France, Germany, the UK, and beyond. The format was classic: 4 hours to build a project from scratch, followed by 2–3 minute pitches with live Q&A, sticker-based voting, and an awards ceremony. Two tracks were originally planned: Track 1 (entertainment/gaming/immersive) and Track 2 (productivity/tools), but they were merged mid-event because nearly every project naturally fell into Track 1's territory of multi-agent, personality-driven systems.
The hackathon produced 14+ projects, of which 10 out of 14 used multi-agent architectures. This alone marks a paradigm shift: in 2025, hackathon projects were primarily single-chatbot wrappers around GPT-4 or Claude. By mid-2026, the baseline expectation had shifted to orchestrating multiple AI agents with distinct personalities, independent memory, and autonomous reasoning loops. The architectural complexity of even a 4-hour hackathon project had increased by an order of magnitude.
This hackathon is a microcosm of the current state of AI development in mid-2026. Multi-agent systems have become the default architecture. A year ago, most projects would have been "chatbots with a single prompt." Today, 10 out of 14 teams built systems with multiple agents, each possessing its own personality, memory, and role. The shift from "AI as a tool" to "AI as an inhabited environment" has happened quietly but irreversibly. Agents don't just answer questions anymore, they lie, debate, remember, collaborate, and compete.
| Layer | Count | Essence | Examples |
|---|---|---|---|
| Narrative AI | 5 | AI agents tell stories, play roles, conduct dialogue | Werewolf, TTRPG, HuiShu, Tarot, Ideal Reader |
| Instrumental AI | 5 | AI agents solve practical tasks | ZhanTu AI, Intercrew, Jie Nei Sheng Zhi, RenRenKan, Life Decision Simulator |
| Infrastructure AI | 2 | AI as an environment for collaborative work | Agent Jam, AI Teacher |
| Analytical AI | 2 | AI analyzes human-generated data | Weibo Privacy, AI News Anchor |
| Communication AI | 1 | AI mediates or replaces human conversation | Digital Human Calling |
In 2025, hackathons were about "wrappers around ChatGPT." In 2026, they're about orchestrating multiple agents with independent memory, ReAct loops, and personalized behavior. The barrier to entry for "building something interesting" has risen dramatically: it's no longer enough to attach an API key, you need to design agent interaction topology. The projects at this hackathon are closer in complexity to what startup teams would have built over weeks in 2025. Four hours in 2026 produces more sophisticated systems than 48 hours in 2025.
The winners reflected a tension between visionary ambition and technical depth: Agent Jam (Group 5) took first prize for its bold vision of a shared virtual workspace where humans and AI agents collaborate in real time. Jie Nei Sheng Zhi (Group 4) won second prize for its creative gesture-based coding interface. TTRPG Master (Group 13) claimed third prize for its immersive AI-powered storytelling. And ZhanTu AI (Group 1) won the popularity prize for its practical exhibition matchmaking platform. The Werewolf AI Game, arguably the most technically ambitious project, did not podium, revealing a judging philosophy where "感性大于理性" (emotion outweighs rationality).
The most technically ambitious project of the hackathon. Eight AI agents play Werewolf (Mafia) with voice interaction, persistent memory, strategic deception, and live TTS/ASR, all built from scratch in 4 hours by a single developer.
Core concept: A fully playable Werewolf (狼人杀 / Mafia) game with 8 players, 7 AI agents and 1 human. Uses 屠边 rules: werewolves win by eliminating all villagers OR the seer. Each AI agent has independent context, memory, personality, and the capacity for outright lying and strategic deception. The system demonstrates that LLMs can engage in the most human of social activities: bluffing, alliance-forming, and detecting lies.
Social deduction games are the ultimate test of AI social intelligence. Unlike chess or Go, which are games of perfect information and pure strategy. Werewolf requires agents to operate under information asymmetry. Werewolves know who the other werewolf is. The seer knows one player's identity each night. Villagers know nothing. Everyone must reason from partial information, evaluate competing claims, detect inconsistencies in speech patterns, and ultimately decide who to trust, and who to kill.
Dzmitry's core question was deceptively simple: Can AI agents lie, deceive, and detect deception? Can they play social deduction games autonomously, generating natural-sounding speech while hiding their true roles? Can they form coherent strategies across multiple rounds? And perhaps most provocatively: can they deceive humans?
"Why play with AI instead of humans? This system can test human versus machine deception ability. You can viscerally feel how machines can deceive humans. It's also a benchmark for model intelligence."— Dzmitry Piakhouski, Q&A session
| Component | Technology | Purpose |
|---|---|---|
| LLM | DeepSeek v4 (flash/pro) | Generates all agent speech, voting decisions, and night actions. Uses reasoning_content for internal deliberation and content for spoken output. |
| TTS | Edge TTS (zh-CN voices) | Chinese voice synthesis for agents. Each agent has a distinct voice (YunyangNeural, XiaoxiaoNeural, YunxiNeural). Lazy generation to prevent OOM. |
| ASR | Stepfun ASR API | Real-time speech recognition from the human player's microphone. Parses target names for voting. |
| Server | Python FastAPI + ThreadingHTTPServer | Multi-threaded HTTP server with SO_REUSEADDR and HTTPS (self-signed certificates). Handles concurrent agent requests. |
| Memory | agent_memory.py | LLM-powered reflection after each game: each agent extracts personal lessons, stores them, and retrieves them for future games. |
| Guardrails | guardrails.py | Three-layer protection: thinking chain leak prevention, incomplete phrase trimming, role confession detector. |
| Suspicion | suspicion.py | Suspicion matrix: each agent independently tracks who they trust and distrust based on gameplay evidence. |
| Orchestration | ThreadPoolExecutor (8 workers) | Parallel speech generation and vote collection from all 7 AI agents simultaneously. |
ReAct Loops for Reasoning: Each agent uses a ReAct (Reasoning + Acting) loop. Before speaking, the agent internally reasons about the game state, who has been suspicious, what contradictions exist, what strategy to pursue. This reasoning is captured in reasoning_content, while the actual spoken words go into content. This separation allows for strategic depth while keeping the spoken output natural.
Independent Agent Memory: Unlike most multi-agent demos where agents share a single context window, each Werewolf agent has its own independent conversation history. After each game, an LLM-powered reflection process extracts personalized "lessons learned", what worked, what didn't, who seemed trustworthy. These lessons carry forward to future games, creating agents that genuinely improve over time.
Suspicion Matrix: The suspicion.py subsystem maintains a matrix where each agent independently tracks its trust levels for every other player. Trust is updated based on observed behavior: inconsistencies in speech, unexpected voting patterns, and logical contradictions all feed into the suspicion algorithm. This creates nuanced social dynamics where agents can be partially trusted rather than binary good/bad.
DeepSeek returns reasoning in the reasoning_content field and speech in content. But sometimes the model places its speech inside reasoning_content and leaves content empty. During the live demo, one agent spoke its reasoning chain aloud, accidentally revealing that it was a werewolf trying to frame an innocent villager. The crowd's reaction: a mix of laughter and genuine surprise at how transparent the AI's "thought process" had become.
"他不好意思，他已经显示了他的思考链。" (He's embarrassed, he already revealed his thinking chain.)
Three layers of protection now guard against this: extract_content() hunts for speech patterns in quotes, guardrails detect thinking markers like "let me analyze" or "我的推理是," and the system prompt explicitly forbids outputting reasoning chains.
Edge TTS spawns a Node.js subprocess (~150MB each). Parallel audio generation for 7 agents simultaneously would consume 1GB+ of RAM, triggering the Linux OOM killer. The solution: lazy TTS, audio is generated one agent at a time, only when the frontend requests the next speech. This reduced peak memory usage to ~200MB and made the system stable enough for the live demo.
The Stepfun ASR API transcribes human speech accurately, but extracting a target player's name from natural Chinese speech proved tricky. A player might say "I think it's Chloe" or "投票给 Ethan" (vote for Ethan), the system needs to parse the intent, not just the words. The solution uses keyword matching with fuzzy logic to identify which of the 7 agent names the human player is targeting.
The Werewolf AI Game is the hackathon's most significant project not because it's the most polished or the most practical, but because it tests the limits of what AI agents can do. Every other project uses agents as tools, to organize exhibitions, debate decisions, tell stories. Werewolf uses agents as players in a game where deception is the core mechanic. This is fundamentally different: it's not about "can AI help humans?" but "can AI deceive humans?" The answer, demonstrated live, was an unambiguous yes. An AI werewolf successfully framed an innocent villager, persuaded others to vote, and nearly won, all while maintaining a facade of helpfulness. The thinking chain leak was, paradoxically, the most compelling proof of concept: it showed the sophistication of the deception strategy the AI had autonomously developed.
More importantly, the project points toward a future where AI evaluation includes adversarial social dynamics. Current benchmarks test reasoning, coding, and math. None test whether a model can look you in the eye and lie convincingly. As AI systems become more integrated into human society, understanding their capacity for strategic deception becomes not just an academic question but a safety imperative.
🏆 First Prize Winner. A shared virtual world where every developer has their own AI agent, and teams collaborate in real time, no PRs, no async reviews, just presence.
Core concept: A shared virtual world (Minecraft-like or Unity-based) where each team member has a personal AI agent. Voice entry drives agent actions. Walk up to a teammate's avatar and see what their agent is coding in real time. "Collaboration is no longer PRs, it's walking over and seeing code change live." Each agent maintains persistent context. Real-time code review and discussion happen naturally through spatial presence rather than formalized workflows.
"The problem is that GitHub moves fast, but we solve problems together, really slowly."— Agent Jam pitch
Modern software development is fundamentally asynchronous. The workflow, pull request → code review → comments → fixes → re-review, introduces latency at every step. Context decays. By the time a reviewer sees the code, the original developer has moved on mentally. Remote teams lack a shared physical space where informal collaboration happens naturally. Agent Jam proposes a radical alternative: what if collaboration was spatial rather than procedural?
"In the future, humans won't create alone — they'll create with their agents, and agents will collaborate with other agents."— Agent Jam team
The ambition extends beyond a coding tool. Agent Jam envisions a world where the boundary between human and AI collaboration dissolves. Your agent isn't a chatbot you query, it's a teammate who works alongside you, visible to others, with its own workspace and agency. The spatial metaphor matters: physical presence creates accountability and serendipity that async workflows cannot replicate.
"Why not just use DingTalk or Feishu?"— Judge question
The team's response was sharp: "Those are communication tools. This is a shared creation space. Games let you build things together that productivity tools can't." The distinction matters: Agent Jam isn't replacing chat; it's replacing the developer environment itself. It's Slack vs. Minecraft, and the argument is that creative work benefits from spatial rather than textual interfaces.
Agent Jam won because it was the most visionary project at the hackathon. While other teams built practical tools or entertaining demos, Agent Jam proposed a fundamental rethinking of how software development works. The judges recognized that the idea, even if not fully implemented in 4 hours, pointed toward a future where development environments are inhabited spaces rather than text editors. The project also demonstrated strong understanding of the shift from AI as tool to AI as environment, a theme that resonated across the entire event.
However, significant questions remain unanswered: Can agents actually write production-quality code collaboratively? Would developers adopt a 3D spatial interface when they've spent decades optimizing keyboard-driven workflows? Is the latency of voice-to-code acceptable for experienced programmers who type at 80+ WPM? Agent Jam's victory is a bet on vision over execution, and the coming years will determine whether that bet pays off.
Agent Jam represents the "spatial computing" branch of the multi-agent revolution. It's the only project that treats AI agents not as tools inside an existing workflow, but as co-inhabitants of a new kind of workspace. The Minecraft inspiration is telling, the most successful creative collaboration platforms in history have been games, not productivity tools. If Agent Jam succeeds, it won't be because it's "Slack with AI", it will be because it's "Minecraft for developers." The question is whether developers want that. History suggests creative professionals adopt spatial tools reluctantly. But if agents are doing the actual coding while humans supervise from a god's-eye view, the spatial metaphor might work better than any text-based alternative.
Six AI agents representing orthogonal personality dimensions debate your life decisions. Built on CrewAI by a psychology student.
Core concept: Six AI agents representing three orthogonal personality dimensions, optimist vs. pessimist, risk-taker vs. cautious, idealist vs. pragmatist. A user describes a dilemma (career choice, relationship decision, life change), and the agents debate it from their assigned perspectives. The output is a structured analysis with pros, cons, and a synthesized recommendation. The name "Intercrew" combines "internal crew" with the CrewAI framework used to orchestrate the agents.
Huang Ruojie brought a psychology background to the hackathon, and it shows in the project's conceptual depth. The insight is that human decision-making is already a multi-agent process, we all have internal voices pulling us in different directions. The optimist says "go for it." The pessimist warns "you'll fail." The risk-taker urges action. The cautious voice advises patience. Intercrew externalizes this internal dialogue into a structured, visible debate.
The three personality axes were chosen carefully: Optimist/Pessimist covers emotional valence, how positively or negatively a decision's outcomes are framed. Risk-taker/Cautious covers action orientation, whether to prioritize potential gains or avoid potential losses. Idealist/Pragmatist covers value alignment, whether to follow principles or practical constraints.
Together, these six perspectives create a comprehensive view of any decision. No single agent is "right", the value is in the synthesis.
Built on CrewAI, a multi-agent orchestration framework. Each agent has a distinct system prompt encoding its personality, for example, the optimist's prompt emphasizes best-case scenarios and growth opportunities, while the pessimist's prompt focuses on risks, failure modes, and historical cautionary tales. The agents engage in structured rounds of debate, with each agent responding to the previous speakers' points. The system then synthesizes the debate into a decision matrix with weighted pros and cons.
The "inner meeting" concept is the project's most innovative feature: agents remember previous decisions and build on past reasoning. If you asked Intercrew about changing jobs last month, and now you're asking about moving cities, the agents reference the earlier discussion, "you chose growth over stability last time, and it worked out" or "you took a risk before and regretted it." This creates a continuous advisory relationship rather than one-off consultations.
Intercrew is the most psychologically sophisticated project at the hackathon. While other projects treat multi-agent systems as technical orchestrations, Intercrew treats them as models of human cognition. The project implicitly argues that AI's greatest value in decision support isn't providing answers, it's making the decision process visible. By externalizing internal debate, Intercrew helps users understand their own thinking patterns. The limitation is that it cannot access information the user doesn't provide, an optimist agent can't research job market data, only reframe what the user tells it. But as a tool for clarifying thinking rather than replacing it, Intercrew is genuinely novel.
A RAG-powered life simulator where students walk through the decisions of historical figures. Choose Elon Musk. Face Tesla's near-bankruptcy. What do you do?
Core concept: Choose a life to simulate, a historical figure like Elon Musk, or a fictional character. AI generates a branching decision tree from a knowledge base of real biographical data. Players face pivotal moments: "Tesla is about to go bankrupt. What would you do?" Multiple choice (A/B/C) with consequences that unfold dynamically. The system uses RAG (Retrieval Augmented Generation) to ground responses in actual history, preventing hallucination. Target audience: middle school, high school, and first-year university students making real-life decisions about their futures.
The team's key technical insight was that life decision advice without factual grounding is dangerous. An LLM told to "simulate Elon Musk" might invent events, misrepresent decisions, or give advice based on fictional scenarios. The solution: a RAG pipeline using LlamaIndex that indexes real biographical data, books, interviews, news articles, and verified timelines. When a player faces a decision point, the system retrieves what actually happened and generates branches from that foundation.
"With real knowledge base, it won't hallucinate. It follows actual history. But the interesting part is the 'what-if' branches."— Group 6 Q&A
The "what-if" branches are the product's hook. What if Musk had sold Tesla to Apple in 2018? What if he'd chosen a different co-founder? These counterfactual scenarios are explicitly labeled as speculative, distinguishing them from the historically grounded main path. This dual-mode approach, factual backbone with speculative branches, makes the simulator both educational and engaging.
The project addresses a genuine educational gap: students make life decisions (university major, career path, city choice) based on extremely limited information. They don't know what "being a doctor" actually feels like day-to-day. They haven't seen how career decisions compound over decades. By compressing years of decisions into a 30-minute simulation, students can experience the consequences of choices before making them in real life.
Voice input is planned for future versions, which would make the simulator more accessible to younger students who may struggle with text-heavy interfaces. The team also discussed adding multiplayer scenarios where students' simulated lives intersect, creating shared learning experiences.
The Life Decision Simulator is the most directly practical educational application of multi-agent AI at the hackathon. Unlike the narrative AI projects (which entertain) or the tool projects (which optimize), this project aims to change how young people make life decisions. The RAG approach is crucial, it's the difference between "AI-generated career advice" (dangerous) and "AI-mediated access to real biographical data" (valuable). The project's main challenge will be content: building comprehensive knowledge bases for hundreds of life paths is labor-intensive. But the core concept, learn from history before making your own, is both timeless and newly feasible with current AI capabilities.
Pixel-art AI tarot reader with persistent memory. Draw cards. Ask questions. The AI remembers everything you've asked before.
Core concept: A pixel-art styled AI tarot reader. Players ask questions (anything from career concerns to "will Group 4 win first place?") → draw cards → AI interprets the cards with persistent memory of the player's entire question history. Each session builds on previous readings, creating a coherent narrative across multiple visits. Demo included a meta-question about the hackathon itself, with the AI interpreting the cards as foretelling "obstacles but eventual appreciation."
What distinguishes Pixel Tarot from countless AI fortune-telling apps is persistent cross-session memory. The AI remembers that three sessions ago you asked about your career, two sessions ago about your relationship, and last session about a creative project. When you return for a fourth reading, it weaves all of these threads together. This creates a experience closer to a human tarot reader who knows you personally, a relationship built over time rather than a series of disconnected transactions.
"How is this different from a random number generator?"— Skeptical judge
The team's response was philosophically sharp: "Tarot is about 引导思考 (guiding reflection), not prediction. The AI remembers your previous questions and builds a narrative." This reframes the project entirely. It's not a fortune-telling tool, it's a narrative therapy tool. The cards are prompts for self-reflection. The AI's value is in connecting those prompts across time, creating a coherent story of the user's evolving concerns.
Built with a multi-turn LLM architecture with persistent memory storage. The pixel-art UI was created specifically for the hackathon, giving the project a nostalgic, indie-game aesthetic that stands out from the sleek-but-generic designs of most AI demos. The tarot interpretation prompting is carefully structured: the AI first describes each card's traditional meaning, then connects it to the user's question, then offers a synthesis that incorporates past readings, and finally poses a reflective question back to the user.
Pixel Tarot is easy to dismiss as "just another AI wrapper," but it's actually one of the most well-designed user experiences at the hackathon. The team understood something crucial: people don't use tarot for predictions, they use it as a structured way to think about their problems. The AI doesn't replace the human meaning-making process; it facilitates it. The persistent memory feature transforms a novelty into a relationship. If you use the app for a month, asking about different life areas, the AI builds a model of your concerns, values, and growth patterns. That's not fortune-telling, it's AI-mediated self-knowledge.
🏆
Fourth Prize (四等奖). Popularity Award: Group 1 (展图AI, the exhibition matching platform) and Group 14 (sharing the popularity vote). These "Most Popular" prizes (最佳人气奖) went to the projects that gathered the most sticker votes during the chaotic rally session.
Third Prize Winner. An AI Game Master that runs tabletop RPG sessions, character cards, world-building, branching narrative, dice rolls, combat narration. All autonomous.Core concept: An AI Game Master that runs complete tabletop RPG sessions. Generates character cards, vivid world descriptions, and dynamically branching narratives. Player choices shape the story in real time. The system handles skill checks, combat narration, NPC dialogue, and quest generation autonomously. The pitch emphasized that the story is co-created: "这个故事不是我一个人写的，是我们大家一起写的。" (This story isn't written by me alone, it's written by all of us together.)
The organizers and judges were visibly impressed by the project's breadth of character and role-playing capability. Unlike a single-purpose AI, the TTRPG Master needed to generate distinct NPC personalities, maintain consistent world lore, adjudicate rules, improvise in response to unexpected player actions, and maintain dramatic pacing, all in real time. The demo showed a fantasy setting with character creation, skill checks, and narrated combat, demonstrating that the system could handle the full spectrum of TTRPG gameplay.
The project resonated because it solves a real pain point in the TTRPG community: finding a good Game Master is hard. Human GMs burn out, have scheduling conflicts, or vary wildly in quality. An AI GM is always available, consistent in rules application, and capable of running sessions for solo players or groups of any size. It democratizes access to a hobby that's traditionally bottlenecked by the scarcity of willing GMs.
Running a TTRPG session requires several AI capabilities working in concert: world-state tracking (what has happened, what NPCs know, what items players have), rules adjudication (correctly applying game mechanics), narrative consistency (maintaining lore and character motivations across sessions), and improvisation (responding to player actions the system couldn't have anticipated). The team used a multi-agent approach with separate agents for narration, rules, and NPC dialogue, orchestrated through a central game state manager.
TTRPG Master sits at the intersection of two major trends: AI as storyteller and AI as game infrastructure. It's the most commercially viable project in the narrative AI category, the TTRPG market is large (D&D alone has 50M+ players), underserved by technology, and hungry for tools that reduce the barrier to play. The AI doesn't replace the social experience of TTRPGs (friends around a table); it replaces the bottleneck (needing someone to prepare and run the game). If executed well, this could become a standard tool in the TTRPG ecosystem, the "Roll20 of AI GMs."
Seven AI agents with distinct personalities track your daily life, generate retrospectives, assign random missions, and visualize your emotional trajectory over time.
Core concept: Seven AI agents, each with different branches and personalities, collectively track the user's daily activities, tasks, and emotional states. The system generates life retrospectives at daily, weekly, and monthly intervals, transforming fragmented experiences into coherent narratives. A "random mission generation" feature assigns spontaneous life tasks: "跟某人吃顿饭" (have dinner with someone) or "结束无意义的内耗" (end a pointless overthinking session). Location-aware recommendations integrate with the user's physical context. Emotional trend analysis visualizes happiness, stress, and satisfaction over time.
Why seven agents? Each agent specializes in a different aspect of life tracking: one focuses on work/productivity, another on social relationships, a third on emotional well-being, a fourth on health/physical activity, a fifth on learning/growth, a sixth on randomness/serendipity (generating the "random missions"), and a seventh on narrative synthesis (weaving everything into retrospectives). This specialization allows each agent to have deep domain knowledge while the synthesis agent creates a unified view. It's the same multi-agent principle as Intercrew, applied to self-knowledge rather than decision support.
"The hardest part is getting people to consistently use a journaling app."— Q&A criticism
This is the elephant in the room for all life-tracking apps. The team's answer is twofold: agent-generated prompts that reduce the friction of "what should I write about?" and gamification through random missions and streak tracking. The agents don't just passively record, they actively engage the user. "Your emotional trend is declining this week. Want to talk about it?" or "You haven't logged a social activity in 6 days. Here's a random mission: reach out to someone you haven't spoken to in a month."
HuiShu is the most ambitious quantified self project at the hackathon. The 7-agent architecture is elegant: specialized agents provide depth while the synthesis agent provides coherence. But the project faces the same adoption challenge that has killed every life-logging app since the dawn of smartphones: consistent data entry is hard. The AI agents help by reducing friction, but they can't fully solve the problem, you still need to tell them what happened. The most promising path forward might be passive data collection through phone sensors, location tracking, and calendar integration, with the AI agents inferring activities rather than requiring manual input.
🏆 Popularity Prize. A QR-code based platform that matches exhibition visitors with shared interests, generates icebreakers, and gives organizers real-time analytics.
Core concept: A QR-code platform for exhibitions and trade shows. Organizers create events and generate QR codes. Exhibitors and visitors scan codes, fill out profiles (what they offer / what they're looking for), and the AI matches people with shared interests. The platform provides map navigation to booths, AI-generated icebreaker phrases ("破冰") to facilitate networking, and an analytics dashboard for organizers showing dwell time, bookmarked booths, and popular zones.
ZhanTu AI won the popularity prize for a simple reason: everyone at a hackathon understands the problem it solves. Hackathons are miniature exhibitions, you wander between project tables, trying to find the interesting ones, armed with nothing but a paper agenda that's obsolete within an hour. The "paper maps that get thrown away" problem resonated viscerally with every voter. The project is immediately practical, clearly scoped, and solves a universal pain point.
The icebreaker generation feature is particularly clever. Networking at exhibitions is awkward. Walking up to a stranger and saying "hi, I see you're interested in X" requires social courage that many people lack. The AI-generated icebreaker reduces the activation energy for human connection, it's a social lubricant powered by profile matching.
"The hardest part isn't tech — it's getting exhibitor and visitor data entry."— He Jing, Q&A
This honest admission cuts to the core challenge of all marketplace platforms: the cold start problem. An exhibition platform is only useful if exhibitors and visitors actually fill out profiles. If only 10% of attendees scan the QR code, the matching quality degrades. The team acknowledged this openly, noting that adoption incentives (priority booth placement, featured listings) would be essential for real-world deployment. The technical side: AI matching, map integration, analytics, is straightforward. The behavioral side, getting busy exhibition-goers to stop and fill out a profile on their phone, is the real challenge.
ZhanTu AI is the most commercially sensible project at the hackathon. It's not the most technically innovative (it's essentially a CRUD app with AI matching layered on top), but it addresses a clear market need with a clear path to revenue (exhibition organizers paying for the platform). The popularity prize reflects a truth about hackathons: sometimes the project that resonates most isn't the most ambitious or technically impressive, it's the one where everyone in the room thinks "I would actually use this."
Multiple AI agents with distinct personalities read alongside you, generating real-time commentary, questions, and reactions. "Like raising a group of children who read with you late at night."
Core concept: Multiple AI agents with distinct personalities read a book alongside you. Each "ideal reader" agent has a defined character and generates real-time commentary, questions, and reactions in a reading room sidebar. The demo used a Lu Xun (鲁迅) text, with different agents offering different perspectives, one focused on historical context, another on literary technique, a third on emotional resonance, and a fourth on the author's biography.
"Serious readers want background knowledge, not quirky commentary. This feels like 吐槽 (roasting) a book, not studying it."— Judge criticism
This critique goes to the heart of what Ideal Reader is trying to be. Is it a study tool for serious literary analysis? Or is it an entertainment product that makes reading more social and fun? The judge's criticism suggests the demo leaned too far toward entertainment, the agents' commentary came across as witty reactions rather than substantive analysis. But this may be a feature, not a bug, for the target audience of casual readers who find solo reading isolating.
The project's emotional hook is genuine: reading alone is lonely. Book clubs exist because people want to share the experience of reading, to discuss, debate, and react together. But book clubs are logistically difficult to organize, require everyone to read at the same pace, and often devolve into off-topic conversation. Ideal Reader offers a middle ground: the social experience of a book club without the scheduling headaches. The agents are always available, always on-topic, and always at exactly the point in the book where you are.
Ideal Reader is the most polarizing project at the hackathon. To its supporters, it solves a real emotional need, making reading feel less solitary in an era of declining attention spans. To its critics, it misunderstands what serious reading is, a deep, focused engagement with text that shouldn't be interrupted by AI-generated chatter. The truth likely depends on the use case: for casual readers working through popular fiction, AI companions could be delightful. For scholars analyzing dense philosophical texts, they'd be a distraction. The project's success will depend on whether it can tune its agents' behavior to match the reader's intent, witty for entertainment, substantive for study.
🏆 Second Prize Winner. Smash bad code with a hammer gesture. Crack a whip to push the AI. Voice commands + physical gestures + emotional release = the most creative coding interface at the hackathon.
Core concept: A gesture-based AI coding assistant that combines voice commands, physical gestures, and emotional release. A hammer-swing gesture "breaks" bad code. A whip-crack gesture "pushes" the AI to work harder. Built with a DeepSeek-powered agent and skill plugins, the system translates natural language plus gestures into coding actions. The unique value proposition: emotional catharsis meets productivity. Sometimes you don't just want to fix a bug, you want to smash it.
The project's insight is that coding is emotional. Debugging frustration, refactoring exhaustion, the satisfaction of a clean compile, these are visceral experiences that text-based interfaces cannot express. Jie Nei Sheng Zhi gives programmers a physical vocabulary for their emotional state. Stuck on a bug for two hours? Smash it with a hammer gesture, the AI interprets this as "roll back to a previous version and try a different approach." Feeling like the AI is being lazy? Crack the whip, it generates more ambitious refactoring suggestions.
This may sound gimmicky, but there's a genuine HCI insight here: embodied interfaces can encode state more efficiently than text. A hammer gesture communicates "I'm frustrated and want to undo recent work" in 200ms. Typing that sentiment takes 5-10 seconds and requires cognitive translation from emotion to language. Gestures bypass the language bottleneck.
"Isn't doing hammer gestures in public more embarrassing than just typing?"— Judge question
The team had a ready answer: "The gesture mode is for emotional release; you can also just press 1-2-3 for normal mode." This dual-mode design is crucial, it acknowledges that gestures are context-dependent. In a private office, hammer away. In a coffee shop, use keyboard shortcuts. The gestures are an option, not a requirement, and the team was smart to design around that flexibility.
Built with a DeepSeek-powered coding agent as the backend, with skill plugins that map gestures to specific actions: hammer → aggressive undo/refactor, whip → increase AI effort/creativity, point → select code region, swipe → delete/clear. Voice commands handle natural language instructions ("add error handling to this function"). The gesture recognition likely uses computer vision (webcam-based), though the 4-hour build constraint means it may have been simplified to keyboard-triggered gesture simulation during the demo.
Jie Nei Sheng Zhi won second prize on creativity and emotional resonance, not technical depth. It's the most memorable UI concept at the hackathon, and the judges rewarded that originality. The project also captures something important about the future of human-AI interaction: as AI takes over more of the routine aspects of coding (writing boilerplate, generating tests, suggesting fixes), the human's role shifts toward direction and emotional feedback. Telling an AI "I'm frustrated, try harder" via gesture may become more natural than typing "please regenerate with higher temperature and more creative constraints." The project is a glimpse of post-keyboard programming interfaces.
Pinduoduo for restaurants. AI agent negotiates with restaurants on behalf of groups. The more friends you bring, the lower the price drops.
Core concept: A Pinduoduo-style group buying platform for restaurants. Users see a deal at a local restaurant → invite friends to "砍" (bargain/chop) the price down → an AI agent negotiates with the restaurant on behalf of the growing group. Dynamic pricing adjusts based on group size, larger groups get deeper discounts. The name riffs on "人人" (everyone) and "砍" (to chop/bargain), evoking Pinduoduo's famous "砍一刀" (chop once) mechanic.
RenRenKan explicitly models itself on Pinduoduo (拼多多), the Chinese e-commerce platform that turned group buying into a social phenomenon. Pinduoduo's insight was that shopping is social, people will invite friends to join a deal, creating viral growth. RenRenKan applies the same logic to restaurant dining. Instead of fixed coupons, the discount scales with your social graph. Bring 2 friends, get 10% off. Bring 8, get 40% off. The AI agent handles the negotiation layer that would be awkward for humans, nobody wants to haggle with a restaurant manager over group size discounts.
The project requires three systems working together: group formation (users inviting friends, tracking group size), dynamic pricing (calculating appropriate discounts based on real-time group size and restaurant capacity), and AI negotiation (the agent communicating with restaurant systems to lock in deals). In a 4-hour hackathon, the group focused on the AI negotiation component, demonstrating that an agent could receive a group size and price target, then engage in a structured negotiation with a restaurant-side agent to reach a mutually acceptable deal.
RenRenKan is the most culturally specific project at the hackathon, it makes sense primarily in the Chinese market context where group buying and Pinduoduo-style mechanics are well-understood social behaviors. The AI negotiation component is interesting but secondary: the real innovation is applying viral group-buying mechanics to a new vertical (restaurant dining) and using AI to automate the B2B negotiation that makes it scalable. The project would need significant restaurant-side adoption to work, facing the classic two-sided marketplace challenge.
Text in, 5-minute video out. AI avatars read news, product intros, and corporate communications. "Nobody wants to read text anymore."
Core concept: An automated pipeline that converts text into 5-minute video segments with AI avatar news anchors. The process: input text → script generation → avatar animation → voice synthesis → video compositing. Target users include news outlets, product marketing teams, and corporate communications departments. The pitch argued that video is increasingly the only format audiences consume: "没人愿意看文字了。" (Nobody wants to read text anymore.)
The team cited a concrete reference: Yu Yi (于一)'s account, which successfully uses AI news formats. Even Douyin (TikTok's Chinese sister app) has integrated AI avatar features, validating the demand for automated video content creation. The team positioned their project as filling the gap between "expensive human-produced video" and "text that nobody reads." For small businesses and local news outlets that can't afford video production teams, an AI news anchor turns a blog post into a broadcast-quality segment in minutes.
The system chains together several AI services: text-to-script (restructuring written content for spoken delivery), avatar generation (creating or selecting a photorealistic digital human), TTS (text-to-speech with natural cadence and emphasis), and video compositing (combining avatar video with background, graphics, and lower-thirds). In a 4-hour build, the team likely used existing APIs (possibly Stepfun's own) for each component, focusing on the integration layer.
AI News Anchor is the most commercially proven concept at the hackathon. AI avatar news is already a real industry in China, with multiple startups and platform integrations. The project isn't innovative in concept, but it's directly addressing a market that exists today. The key question is differentiation: if Douyin already offers this, what does a standalone tool provide that an integrated platform feature doesn't? The answer is likely customization and white-label deployment for B2B clients who don't want their content locked into a consumer platform.
Input 10 Weibo posts. AI infers: wake-up time, job type, favorite idol, favorite color, personality traits. A 4-hour experiment that makes people viscerally aware of their digital footprint.
The project's impact comes from showing, not telling. Everyone knows theoretically that social media data can be analyzed. But watching an AI reconstruct your daily schedule, guess your job, and identify your personality traits from 10 posts you probably forgot you wrote, that's visceral. The demo didn't just explain privacy risks; it made the audience feel them. Several audience members reportedly checked their own Weibo feeds during the presentation.
The technical approach uses an LLM's inference capabilities rather than traditional machine learning. Instead of training a classifier on labeled data, the system prompts the LLM to act as an investigative analyst, drawing connections between seemingly innocuous posts. A post timestamped at 7:30 AM with "刚醒" (just woke up) reveals wake-up time. A series of posts complaining about meetings and citing industry jargon reveals profession. Posts about specific celebrities reveal fandom. The AI connects dots that humans might miss.
The team was explicit that this isn't a surveillance tool, it's an educational tool for privacy awareness. By making the inference process visible and interactive, it helps users understand how much they're inadvertently sharing. The project could be deployed as part of digital literacy programs, university orientations, or corporate security training. The emotional impact of seeing your own profile reconstructed is far more effective than abstract warnings about "protecting your privacy online."
Weibo Privacy Analyzer is the most ethically charged project at the hackathon. It demonstrates a capability that exists at the intersection of AI and privacy that most people don't fully appreciate: inference is more powerful than collection. You don't need someone's private data when you can reconstruct it from their public posts. The project implicitly asks a question that AI development needs to grapple with: just because we can infer intimate details about strangers from public data, should we? The fact that this was built in 4 hours is itself the most powerful argument, if a hackathon team can do this, what can well-resourced organizations do?
An adaptive AI tutor that personalizes curriculum, adjusts to student pace, and provides real-time feedback. One-on-one tutoring at scale.
Core concept: An AI teaching agent that delivers personalized, adaptive education. The system adjusts curriculum difficulty and pacing based on student performance, provides real-time feedback on exercises, and maintains a model of each student's strengths and weaknesses. The pitch was brief in the transcript, suggesting the project may have been less developed or the presentation was cut short, but the core idea addresses one of education's most persistent challenges: the impossibility of one-on-one tutoring at scale.
Educational research consistently shows that one-on-one tutoring is the most effective form of instruction, students tutored individually perform two standard deviations better than those in traditional classrooms (Bloom's 2-sigma problem). But human tutors are expensive and scarce. AI teachers that can approximate the personalized attention of a human tutor could democratize high-quality education. The project sits at the intersection of EdTech's oldest dream and AI's newest capabilities.
Although the transcript records a brief presentation, the project's category and technology suggest several components: student modeling (tracking what each student knows and doesn't know), curriculum adaptation (selecting appropriate content and exercises), real-time feedback (explaining why an answer is wrong, not just that it's wrong), and engagement maintenance (adjusting difficulty to keep students in the "flow" zone between boredom and frustration). The underlying technology is likely an educational LLM fine-tuned or prompted for pedagogical best practices.
AI Teacher is the most socially significant but least developed project at the hackathon. Education is arguably AI's highest-impact application domain, improving learning outcomes for millions of students would have cascading effects on economic mobility, social equity, and global development. But the project also faces the hardest validation challenge: you can't demonstrate improved learning outcomes in a 4-minute hackathon pitch. The project's brevity in the transcript may reflect the difficulty of doing justice to an education project in a format optimized for flashy demos and quick wins.
AI-powered phone calls using digital human twins. Voice agents that represent real people, or pretend to be them.
Core concept: An AI system that creates a digital twin of a person and enables that twin to make phone calls on their behalf. The digital human speaks with the person's voice, can engage in natural conversation, and can handle routine calls, scheduling, customer service inquiries, information gathering. The concept spans both practical utility (automating tedious phone calls) and ethical complexity (when does a digital twin need to disclose its AI nature?).
Digital Human Calling sits at the cutting edge of voice AI in 2026. Unlike text-based agents, voice agents must handle turn-taking dynamics (interruptions, pauses, filler words), paralinguistic cues (tone, pace, emphasis), and real-time constraints (generating responses within natural conversational latency). The Stepfun ecosystem's multimodal capabilities, particularly its emotion and tone detection, would be directly applicable to making digital human calls feel natural rather than robotic.
The project raises an uncomfortable question that goes unasked in many AI demos: does the digital twin disclose that it's an AI? If a digital twin calls a restaurant to make a reservation and never mentions it's not human, is that deception? What if it calls a friend? A family member? A business contact? The ethical boundary between "convenient automation" and "impersonation" is fuzzy, and Digital Human Calling sits right on that line. The hackathon format doesn't provide space for deep ethical discussion, but the question lingers after the demo.
Digital Human Calling is less a project and more a provocation, it forces the audience to confront the near future where AI voice agents are indistinguishable from human callers. The technology is essentially here; the norms and regulations are not. This project would benefit from explicit ethical framing: under what circumstances is AI-mediated calling acceptable, and what disclosure mechanisms should be mandatory? Without that framing, it risks being seen as a tool for deception rather than a tool for convenience.
A ~23-minute dinner conversation between multinational hackathon participants. Qwen real-time translators, AI video culture wars, foreigner privilege, a halal food incident, horse farms in France, and the Matrix.
Five participants from China, Poland, Australia, Belarus, France, and other backgrounds gather for dinner during the hackathon. The conversation, captured in File 1 of the transcription, ranges from hackathon experiences to AI culture wars, travel stories, and the practical challenges of being an international participant in Shenzhen. It's an unfiltered window into the human side of the event, the conversations that happen between pitches, when people are just people.
"But do you understand Chinese at all?"
"Mandarin not. I have real-time translation, so I understand everything with delay three seconds."
"What is it?"
"Qwen. Qwen studio — when it's Chinese, it's just real-time."— Dinner conversation, ~4:00 mark
One of the most practical demonstrations of AI at the entire event happened not on stage, but at the dinner table. A Belarusian participant had been using Qwen's real-time translation feature throughout the hackathon to follow Chinese-language presentations with approximately a 3-second delay. This was a genuine "AI in the wild" moment, not a staged demo, but a working tool that made cross-cultural participation possible. The irony: Stepfun hosted the hackathon, but Qwen (a competing AI platform) was the evening's most praised product.
The translator was described as an app on the participant's phone, downloaded from the Chinese app store (no region switching needed, reportedly). When asked if he changed his app store to Chinese, he responded "No, no, I just put this app." The ease of access highlights how AI translation tools are quietly dissolving language barriers in real-world settings, far from the spotlight of keynote demos.
"The Chinese are very good at making these AI videos."
"Do they watch AI generated movies though?"
"Everyone hates the AI market. In Australia, no market. No market. In America and Australia, most people are very angry with it. Really. It's just China that everybody loves it."
"Same. That's why I'm out of the box. Very slow in understanding what is ahead in the world."— Dinner conversation, ~8:45 mark
This exchange captures a genuine cultural divide in attitudes toward AI-generated content. Chinese consumers have enthusiastically embraced AI video, AI music, and AI-generated entertainment. Western consumers, particularly in the US and Australia, are broadly hostile. The conversation treated this as an observed fact, not a value judgment. One participant noted that Chinese AI video generation is "very disruptive today," while another observed that the West is "very slow in understanding what is ahead in the world."
The cultural analysis here is nuanced. It's not that Western consumers dislike AI, it's that they specifically dislike AI-generated creative content, which they perceive as inauthentic or threatening to human creators. Chinese consumers, by contrast, seem to evaluate AI content on its merits rather than its origin. This difference may reflect broader cultural attitudes toward technology, authenticity, and creative labor, or it may simply reflect different stages of market development.
"In my little hometown back in Hunan, there was a foreigner — he doesn't have a driver license in China, but he drives a little car there every day. The policeman never bothered to ask him or let him show his driver license, because he knows he doesn't have it, and he doesn't want to get himself in trouble, and doesn't want to get the foreigner in trouble. So okay, let's just... let me do a little bit of privilege here."— Speaker 4, ~14:11 mark
This anecdote, told with a mix of amusement and exasperation, illustrates the strange double-sided nature of being foreign in China. On one hand, there are genuine privileges (police turning a blind eye). On the other hand, those privileges stem from a system that treats foreigners as outside the normal rules, not quite subject to the law, but also not quite protected by it. The story was told casually, as a "can you believe this?" moment, but it speaks to deeper dynamics of how international participants navigate Chinese spaces.
"I was in France, in a horse farm, as a volunteer for almost a month. The owner of the horse farm was like: I never wanted Chinese people to come here because I don't like Chinese stuff, I don't like blah blah blah bad things about China. I was like, okay, I know it's hard. I have to stay because I really want to change Chinese people's reputation."— Speaker 4, ~15:00 mark
A Chinese participant shared a powerful story about volunteering at a horse farm in rural France. The farm owner was openly hostile to Chinese people, based on stereotypes and media portrayals, and the volunteer chose to stay for the full 25 days specifically to challenge those perceptions through personal interaction. "I decided to stay like enough twenty-five days," she said. "To show that the people are very kind." This is citizen diplomacy in its rawest form: one person, one farm, 25 days, trying to change a mind through presence rather than argument.
"喂你好刚才打的白米饭没刷卡帮我刷一下刷一下还要你们拿猪肉来我们这里吃。没有什么比赛的我知道你们不能拿猪肉来我们餐厅吃饭我知道谢谢我们这个是回族穿来大家理解一下好好好好不好意思。"— Cafeteria confrontation, ~20:00 mark
Translation: "Hello, the white rice wasn't scanned, help me scan it. And you brought pork to eat in our area." "There's no problem..." "I know you can't bring pork to eat in our restaurant. I know. Thank you. We are Hui people, everyone please understand. Okay okay okay, sorry."
The dinner conversation took a serious turn when it emerged that some participants had brought pork dishes into a Halal (Muslim) dining area, likely a university cafeteria with designated zones for Muslim students. This was a cultural mistake with real consequences: "你还能再保证等下给他们那些少数民族看到了就对就容易引发事件" ("If other ethnic minority students see this, it could easily trigger an incident.")
The group scrambled to manage the situation, eating the pork quickly to "leave no trace," apologizing to the cafeteria staff, and acknowledging that the organizers should have been more aware of the dietary arrangements. One participant noted: "他们安排也不对哦，他们这个应该应该提前了解一下" ("Their arrangements were wrong, they should have understood this in advance.") The incident highlights the practical challenges of cross-cultural event organization and the ease with which well-meaning participants can inadvertently cause offense.
Amidst the dinner conversation, a spontaneous ideation session emerged about building an agent whose purpose is to convince a human that their reality is a simulation:
"Maybe sometime we can experiment building an agent trying to tell any human that you are being cheated — the world you are living in is a total scam — and then trying to put a lot of language to support the logic."— Speaker 2, ~6:00 mark
The group riffed on the idea: an agent that identifies you're "living in the Matrix" and tries to convince you of the truth, functioning as "mental support" or "mental disruption" depending on your perspective. It was a playful, speculative moment, the kind of conversation that happens when creative technologists gather informally, but it also reflects genuine questions about AI's capacity to reshape belief systems.
The conversation included remarkable travel stories: one participant had spent 3 years in Kenya and 4 months in Honduras. Another was from Belarus, had immigrated 10 years ago, and had only arrived in China "two months ago." The group discovered shared Hunan hometowns: Shaodong, Zhuzhou, Chenzhou, Changsha, and bonded over辣椒 (chili peppers) and local cuisine. Multi-national hackathons create these spontaneous connections; the dinner table is where the real networking happens.
Three stickers per person. One vote per project. A "rally period" that turned into controlled chaos. And the inevitable criticism: "Should be a digital system."
Each participant received three physical stickers. The rules were simple: you could vote for three different projects, one sticker each. No stacking votes on a single project. After all pitches, a "拉票环节" (vote rally period) was announced, teams could campaign for votes, and participants could walk around, ask questions, and place stickers.
The organizer explained: "现在大家手上其实每个人都有三张票啊。每个人手上都有三张票，然后大家只能给这个项目投一个票。比方说你有三张票，你只能投给三个不同的项目。就这么简单。" ("Everyone has three tickets. You can only give one vote per project. For example, if you have three tickets, you can only vote for three different projects. It's that simple.")
"好，我们现在开始拉票环节，拉票环节！他们现在疯狂的在拉票...还挺好玩的，拍视频，拍视频，接着拍。拉票拉票，我还没有给别人。"— File 3, voting rally
Translation: "Okay, we're now starting the vote rally period! They're now going crazy campaigning for votes... it's pretty fun, film it, film it, keep filming. Rally rally, I haven't given mine out yet."
What ensued was a chaotic but energetic period of face-to-face lobbying. Teams with charismatic pitchers (or simply more friends in the room) had a natural advantage. People were "疯狂的在拉票" (going crazy campaigning). The scene was described as genuinely entertaining, participants filming each other, laughing, making last-minute pitches. It was democracy in its messiest, most human form.
"Shouldn't be able to see other people's votes."
"Should be a better voting system. No, not this. Should be a digital system."
"Just that. Human conversation. Human nature. But there's no centralization. There shouldn't be rally. There's no voting rally. Shouldn't be any rally in this."— Disgruntled participant, File 3
Not everyone was charmed by the sticker democracy. At least one participant voiced strong objections: visible voting creates peer pressure, the rally period advantages extroverts and established social connections, and a digital voting system would be fairer and more private. The criticism is valid from a procedural fairness perspective, but it also misses something about the hackathon's spirit. The sticker voting and rally period created energy, interaction, and memorable moments that a silent digital ballot never could.
The organizer's response didn't directly address the criticism, but the awards ceremony framing (discussed next) suggests they were aware of the system's limitations. The "感性大于理性" (emotion over rationality) philosophy effectively acknowledges that hackathon judging isn't, and perhaps shouldn't be, purely objective.
The sticker voting system is easy to criticize (and was criticized). But it serves functions that a digital system wouldn't: it forces interaction. To get a sticker, you have to walk to someone's table, look at their project, and talk to them. That's the entire point of a hackathon, not just building, but connecting. A digital ballot would be fairer but would eliminate the serendipitous conversations that happen during the rally period. The tension between "fair voting" and "community building" is real, and there's no perfect solution. The sticker system erred on the side of community, and given that this was a hackathon, not a democratic election, that's probably the right call.
Track merger explained. "感性大于理性" judging philosophy. First through fourth prizes, best judge awards, consolation prizes for all, and a voice coding keyboard gift.
Before announcing results, the organizer addressed the elephant in the room: two tracks had been planned, but nearly every project fell into Track 1's territory.
"我们今天在现场看了一下大家的这个路演的项目，基本上都是跟赛道一差不多的。都是做的这个偏这种娱乐啊，什么剧本杀呀这种。所以赛道一和赛道二的评分，我们就不按照赛道一和赛道二那样去分了。选赛道一的好像只有十几个，15个。剩下的好像都是赛道二的，所以这样的话就不太公平。"— Organizer, File 4
Translation: "We looked at everyone's projects today, and they were basically all similar to Track 1, entertainment, murder mysteries, that kind of thing. So we're not going to score Track 1 and Track 2 separately. It seemed like only about 15 projects chose Track 1, and the rest chose Track 2, so that wouldn't be fair."
The scoring adjustment: instead of separate winners per track, the top 2 overall scores would be first prize, scores 3-4 would be second prize, scores 5-6 would be third prize, and scores 7-8 would be the popularity prize. This meant two winners per prize tier, effectively doubling the award slots.
"我们在最后的评分环节，他一定在某种程度上，大家的感性大于理性。我觉得这个东西它只是一个经历。大家今天能相遇，然后能够交流，把自己一个产品做出来，这个是最重要的。我们最后这个东西其实它不会那么的客观。而且你真正的在去做市场的时候，现实世界就是这样子的。不代表你的产品做得最好，你就一定能获得成功的。它一定在某些时候里面，有运气在里面，有各种机遇在里面。"— Organizer, File 4
Translation: "In the final scoring, to some degree, emotion will outweigh rationality. I think this is just an experience. The fact that everyone could meet today, communicate, and build a product, that's what matters most. The final result won't be completely objective. And when you actually go to market, the real world is like this. Having the best product doesn't guarantee success. There's always some luck involved, some serendipity."
This framing — "感性大于理性" (emotion over rationality), is refreshingly honest. It acknowledges what everyone knows but few organizers admit: hackathon judging is subjective. The projects that win aren't necessarily the best engineered or most commercially viable; they're the ones that resonate emotionally with judges and voters in the moment. The organizer framed this not as a flaw but as a feature, a lesson about how the real world works, where success depends on storytelling and connection as much as technical merit.
| Prize | Groups | Projects |
|---|---|---|
| 🏆 First Prize (一等奖) | Group 15 & Group 5 | Agent Jam (AJ) + additional project |
| 🥈 Second Prize (二等奖) | Group 18 & Group 4 | 节内生枝 (Jie Nei Sheng Zhi) + additional project |
| 🥉 Third Prize (三等奖) | Group 19 & Group 13 | 跑团达人 (TTRPG Master) + additional project |
| 🏅 Popularity Prize (最佳人气奖) | Group 1 & Group 14 | 展图AI (ZhanTu AI) + Werewolf AI Game |
| 🎖 Best Judge (最佳审评官) | — | Wu Wei (吴卫) & Zhou Zhou (周周) |
| 🎖 Best Technical Officer | — | Wang Yingying (王莹莹) |
| 🎖 Best Scene Capture (名场面记录奖) | — | Photo taken with Insta360 (陨石) |
| 🎖 Best Questioner | — | Awarded for sharp Q&A contributions |
Every participant received a consolation prize (安慰奖). The organizer made a point of announcing this: "假设你没有获奖的话，那个你可以找泽民领取一点 token。每个参赛选手都有一个安慰奖。" ("If you didn't win, you can get some tokens from Zemin. Every participant gets a consolation prize.") This inclusive approach, ensuring nobody leaves empty-handed, reflects the community-building ethos that the organizer emphasized throughout.
A bonus gift was also announced: a voice coding keyboard — "一个快捷键盘自定义功能的，就是 voice coding" ("a shortcut keyboard with customizable functions, it's for voice coding"). This physical gift, combining hardware with the AI coding theme that ran through several projects, was described as "interesting" by the recipient. It's a tangible reminder that the hackathon's themes, voice interfaces, AI-assisted development, are leaking into physical product design.
"我们后面还会有机会在深圳去办一个，欢迎大家参加。我把我的微信二维码发在群里面了。我们经常会办一些有意思的活动。"— Organizer, closing
Translation: "We'll have more opportunities to organize events in Shenzhen. Everyone is welcome. I've posted my WeChat QR code in the group. We frequently organize interesting activities." The promise of future events, and the WeChat QR code, signals that Stepfun is building a community, not just running a one-off competition.
Six patterns that define the state of AI development in mid-2026, as revealed by this single 9-hour event.
| Theme | Projects | Significance |
|---|---|---|
| Multi-Agent Everywhere | 10/14 | The single-chatbot paradigm is dead. The default architecture in 2026 is multiple agents with distinct roles, coordinated by an orchestrator. Werewolf (7 agents), HuiShu (7 agents), Intercrew (6 agents), Ideal Reader (multiple reader agents), and Agent Jam (N agents per team) all demonstrate this shift. |
| Personality-Driven AI | 8/14 | Agents with character, not generic "helpful assistants." Each agent has its own voice, speaking style, biases, and emotional register. The shift is from utility to presence — agents aren't just tools, they're personalities you interact with. Werewolf werewolves lie differently from villagers. Ideal Reader agents have distinct "reading personalities." Intercrew agents embody psychological archetypes. |
| Persistent Memory | 6/14 | Agents that remember. Not just context window — long-term memory with reflection. Werewolf: LLM-powered lessons after each game. Tarot: history of questions across sessions. HuiShu: days of life data. Intercrew: previous decisions informing new ones. An agent without memory cannot build a relationship. |
| RAG + Knowledge Bases | 3/14 | Grounding AI in reality. Life Decision Simulator uses LlamaIndex for true biographical data. ZhanTu AI uses exhibition profiles. The knowledge base is the moat — it's what separates "AI that makes things up" from "AI that works with facts." |
| Real-Time Interaction | 5/14 | Voice, gestures, real-time collaboration. The interface is shifting from text chat to multimodal interaction. Werewolf uses ASR + TTS for voice-based gameplay. Jie Nei Sheng Zhi uses gestures. Agent Jam uses voice + spatial presence. The keyboard is no longer the only input modality. |
| AI as Social Mirror | 5/14 | AI doesn't just perform tasks — it shows humans something about themselves. Werewolf: how machines deceive humans. Weibo Privacy: what strangers can infer about you. Intercrew: your internal conflicts externalized. AI becomes a tool for self-knowledge, not just task completion. |
The biggest shift visible at this hackathon is that AI is ceasing to be a tool and becoming an environment. Agent Jam is a world where you work. Werewolf is a world where AI personalities compete against you. HuiShu is a world where AI observes your life. These aren't "apps with AI features." They're spaces inhabited by agents, persistent, personality-rich, memory-bearing agents that you visit and interact with over time.
This shift has architectural implications: agent memory becomes critical (you can't inhabit a space with someone who forgets you every time). Real-time interaction becomes essential (environments are experienced, not queried). And personality becomes a feature, not a bug, the "character" of the agents in your environment shapes the quality of your experience.
Equally telling is what wasn't built: no blockchain projects, no DeFi, no crypto. No traditional SaaS tools. No "Uber for X" clones. The hackathon's participants, mostly developers in their 20s and early 30s, voted with their builds: AI agents are where the creative energy is. The absence of Web3 projects is particularly striking given how dominant they were at Chinese hackathons just 2-3 years ago. The hype cycle has moved on.
The 4-hour build time acts as a fascinating filter. It eliminates projects that require extensive data preparation, complex backend infrastructure, or long training cycles. It favors projects that can demonstrate value through prompt engineering, agent orchestration, and clever UX. This filter may explain the dominance of multi-agent projects: orchestrating agents is achievable in 4 hours in a way that training a custom model or building a data pipeline is not. The hackathon's format selects for the current frontier of what's possible with off-the-shelf LLMs and creative system design.
Why the Werewolf AI Game is the most technically ambitious project at the hackathon, even without a podium finish. A comparative analysis across all dimensions.
| Dimension | 🐺 Werewolf | 🌍 Agent Jam 🏆 | 🧠 Intercrew | 🎲 TTRPG 🥉 | 🔨 Jie Nei 🥈 |
|---|---|---|---|---|---|
| Multi-agent | 7 agents + orchestrator | N agents in shared world | 6 agents (3 axes) | 1 GM agent + NPCs | 1 agent + plugins |
| Memory | LLM reflection, lessons | Session context | Decision history | Game state | None |
| Real-time | Voice + TTS + ASR | Voice + shared view | Text only | Text only | Gesture + voice |
| Deception | Core feature | No | No | Partial (NPCs) | No |
| Guardrails | 3 layers | None | None | None | None |
| Solo build | Yes (完全手搓) | Team | Team | Team | Team |
| Architecture | FastAPI+DeepSeek+EdgeTTS+ASR | Minecraft/Unity+LLM | CrewAI | LLM+State machine | DeepSeek+Gesture CV |
| Benchmark potential | Yes (deception, strategy) | No | Partial | No | No |
| Long-term impact | Standardized LLM eval | New dev paradigm | Decision support | Gaming tool | Novel UI concept |
Agent Jam won first prize for vision, the idea of "AI as collaborative environment." Werewolf didn't podium, but it achieved something rarer: it built a system that tests the limits of AI capability itself. Agent Jam imagines how AI could change how we work. Werewolf demonstrates how AI could change how we evaluate AI. That's a meta-contribution, a project about measuring AI that could influence how all future AI is measured.
In a year, Agent Jam might be a product people use. Werewolf might be a benchmark that shapes the development of next-generation models. Both are valuable. But the long-term leverage of creating a standardized evaluation framework exceeds the leverage of creating a collaboration tool, because the framework influences which tools get built. The Werewolf AI Game is the hackathon's most significant project not because it's the best product, but because it points toward a new way of understanding what AI can do.
Current LLM benchmarks test: reasoning (MMLU, GSM8K), coding (HumanEval, SWE-bench), math (MATH), and general knowledge. None test: strategic deception, lie detection, coalition formation, trust calibration, or information warfare under uncertainty. These are the skills that matter when AI systems interact with humans in adversarial or semi-adversarial contexts, negotiation, diplomacy, competitive strategy, security. The Werewolf benchmark would fill a gap in AI evaluation that becomes more important as AI systems become more autonomous and more embedded in human social dynamics.
The organizer's "感性大于理性" philosophy explains the outcome. Werewolf is a rational achievement, it's technically impressive, architecturally sophisticated, and rigorously designed. But it doesn't have the emotional hook of Agent Jam ("we'll change how teams work together!") or the visceral appeal of Jie Nei Sheng Zhi ("smash your bugs with a hammer!"). In a hackathon judged by human emotion rather than objective criteria, the projects that tell the best stories win. Werewolf tells a great story, but it's a story about AI capabilities, not about human experience. That's a harder sell.
All 14 projects positioned by technical complexity and primary interaction modality.
ARLI membership: the notes from the workshop
One site of yours, measured, with the price for the work in the same document
ARLI, Inc., a Delaware corporation. Questions: hq@arliwork.com.