WAIC 2026 AI Mixer: Is AI Ready to Move from Smart Demos to Real Action?

Introduction

The AI industry has spent years proving that models can answer harder questions, generate better media, write more code, and perform impressively in controlled demonstrations.

The next question is more practical:

AI is smart enough—but can it reliably act?

That question is the focus of an informal AI Mixer taking place on the opening night of the 2026 World Artificial Intelligence Conference in Shanghai. The off-site gathering brings together builders working on AI agents, memory infrastructure, 3D generation, embodied intelligence, and world models.

Rather than adding another polished conference presentation, the organizers want to discuss what happens after the demo: deployment inside real workflows, continuous operation, physical-world interaction, cost, reliability, and the difficult transition from prediction to execution.

From “Showing Intelligence” to “Completing Work”

WAIC exhibition halls will be filled with stronger foundation models, polished product demonstrations, autonomous systems, and increasingly capable humanoid robots.

Those displays reveal what AI can do under prepared conditions. They do not always show what happens when the same system enters an actual company, factory, studio, or development team.

A production environment introduces a different set of requirements:

• The agent must understand an incomplete request.
• It must obtain the correct context and permissions.
• It must use tools without damaging data or systems.
• It must continue after an individual step fails.
• It must know when human approval is required.
• It must produce an output that fits an existing workflow.
• It must operate at a cost the customer can justify.
• Its actions must be traceable and reviewable.

The gap between a convincing demonstration and a dependable product is therefore not only a model-quality problem. It is a systems, product, operations, and trust problem.

The event frames this shift through two movements:

1. Agents are leaving the chat box and entering real workflows.
2. AI is moving toward the physical world, where it must understand environments and interact with them.

These two directions are closely connected. Both require AI to move beyond generating plausible responses and begin taking responsibility for sequences of actions.

Event Overview

Item	Details
Event	WAIC 2026 AI Mixer
Main question	Is AI ready to move from intelligence demonstrations to real action?
Date	Friday, July 17, 2026
Time	7:30 PM–9:30 PM
Location	A café in Pudong, Shanghai
Format	Short discussions followed by open networking
Focus	AI agents, memory, 3D models, embodied intelligence, and world models
Audience	Builders, researchers, industry practitioners, and investors
Registration	Advance registration through the organizer’s event link

The public event page only identifies the venue as a café in Shanghai’s Pudong New Area. It does not publish the complete street address.

This is an independent, off-site gathering held on the opening night of WAIC. It should not be confused with an official forum listed in the main WAIC conference program.

Event Agenda

7:30 PM–8:00 PM: Check-In and Networking

The first section is reserved for arrival and informal introductions.

For a small technical gathering, this part may be as useful as the formal discussion. Attendees include people building products, conducting research, evaluating investments, and trying to bring AI systems into real organizations.

Useful opening questions include:

• What kind of AI system are you currently building or deploying?
• Which part works well in a demo but fails in production?
• What is the largest source of agent cost or latency?
• What does a customer consider a completed task?
• Which human approval steps cannot yet be removed?

8:00 PM–8:30 PM: Have Agents Really Left the Chat Box?

The first discussion asks whether AI agents have progressed from impressive tool-use demonstrations to stable roles inside real work.

Speakers listed in the event materials:

• Zhang Haoyang, founder and CEO of EvoMap.
• Li Zhiyu, co-founder and CTO of MemTensor.

The topic brings together two important parts of agent development: learning from experience and maintaining useful memory.

An agent can call tools and complete a short task without being ready for continuous work. Long-term use raises deeper questions:

• Can the agent retain useful experience without filling its context with noise?
• Can it reuse a successful solution across different models or environments?
• How should memories be verified, ranked, updated, and removed?
• Can the agent recover from failure rather than simply restart?
• How can teams audit what an agent learned?
• When does autonomy improve productivity, and when does it create new supervision work?

EvoMap presents itself as infrastructure for AI self-evolution. Its Genome Evolution Protocol is designed to turn successful agent behavior into reusable and auditable capability assets.

MemTensor focuses on memory infrastructure for large models and agents. Its work addresses how systems manage different forms of memory so that knowledge can be updated and reused without relying only on larger prompts or repeated retraining.

Together, these perspectives point toward a practical definition of a working agent:

A useful agent is not only a model that can act once. It is a system that can remember, improve, verify, and repeat successful work under clear boundaries.

What “Formal Employment” Means for an AI Agent

The event’s Chinese topic can be translated literally as asking whether agents have moved from chat interfaces into “formal employment.”

The phrase is intentionally provocative. An AI system is not an employee in the legal or human sense. The comparison is useful because it forces teams to define the operational standard expected from a working system.

A deployable agent should have:

A Clear Job Boundary

The agent needs a defined responsibility. “Help with marketing” is too broad. “Prepare a weekly competitor report from these approved sources and leave the draft for review” is measurable.

Access to the Right Context

A capable model without relevant data will still produce weak work. Production agents need controlled access to documents, databases, applications, APIs, and previous task state.

Reliable Tool Use

The agent must call tools with correct parameters, recognize failures, avoid repeated destructive actions, and produce an audit trail.

Completion Criteria

The system needs to know what finished means. A task should not stop merely because the agent generated a plausible paragraph or performed several actions.

Escalation Rules

The agent should know when to ask a person. Financial decisions, external communications, deletion, purchases, production deployments, and uncertain identity checks should remain behind explicit approval.

Evaluation Over Time

A one-time success is not enough. Teams need to measure accuracy, completion rate, review time, cost, failure recovery, and the business value of completed tasks.

8:30 PM–9:00 PM: Can Models Move from GitHub into the Physical World?

The second discussion examines the transition from research repositories and model releases to interaction with real environments.

Speakers listed across the source post and agenda poster:

• Liang Ding, CTO of VAST.
• Lu Hongyuan, founder of Lianpu Mind (脸谱心智).
• Chen Tianrun, chairman of Moxin Technology (魔芯科技), listed on the promotional poster.

The discussion connects 3D generation, world models, spatial intelligence, and embodied systems.

A model may generate a visually convincing scene while still misunderstanding how objects move, collide, deform, or respond to an action. A robot may recognize an object while failing to grasp it safely in an unfamiliar environment.

Moving from digital output to physical action requires several layers to work together:

1. Perception: Understand images, video, depth, sound, and sensor data.
2. World representation: Maintain a usable model of objects, space, and relationships.
3. Prediction: Estimate how the environment may change after an action.
4. Planning: Compare possible action sequences.
5. Control: Convert a high-level plan into executable robot or device commands.
6. Feedback: Observe the actual result and correct the next action.
7. Safety: Prevent unsafe or irreversible behavior.
8. Learning: Use verified experience to improve future decisions.

VAST is known for work in AI-generated 3D content and its Tripo product. This perspective connects generative models with the creation of usable 3D assets and environments.

Lianpu Mind has been described publicly as working on multimodal and world-model systems, including technology intended to help AI represent and reason about the world beyond a static screen.

Moxin Technology began in consumer 3D printing and later expanded into spatial intelligence and interactive world models. That path creates a direct link between digital generation and physical manufacturing.

From GitHub Demo to Real Deployment

Publishing a model or repository is an important milestone. It is not the same as deploying a complete system.

A repository can demonstrate:

• Model architecture.
• Training or inference code.
• Sample results.
• Benchmark performance.
• A limited interactive demo.

A real-world deployment must additionally handle:

• Hardware differences.
• Sensor errors.
• Network interruptions.
• Unstructured environments.
• Latency constraints.
• Safety policies.
• Long-running operation.
• User training.
• Maintenance and updates.
• Liability when an action goes wrong.

The difference becomes particularly visible in robotics and embodied AI.

A world model can predict realistic-looking video but still fail to provide the precise state representation required for control. A vision-language-action model can succeed in a laboratory task but fail when lighting, object placement, or hardware changes.

The important question is therefore not whether the model can produce one impressive result. It is whether the surrounding system can repeatedly close the loop:

Observe → Understand → Predict → Plan → Act → Verify → Learn

9:00 PM–9:30 PM: Open Discussion

The final section is reserved for informal conversation.

The event is designed for people who want to discuss decisions and constraints that are often missing from launch presentations.

Potential discussion areas include:

• Whether customers are paying for autonomous outcomes or only AI-assisted features.
• Which agent tasks have reached dependable production use.
• How to price long-running agent work.
• Whether memory systems improve performance or introduce new privacy risks.
• How world models should be evaluated beyond visual quality.
• When a robot needs a world model, a VLA policy, or both.
• How to collect real interaction data safely.
• Where human oversight remains essential.
• Which parts of the current agent market are genuine infrastructure and which are temporary packaging.
• Whether users trust systems that act more than systems that only recommend.

Featured Builders

Zhang Haoyang — EvoMap

EvoMap is building infrastructure for AI agents to share, validate, and inherit reusable capabilities.

Its Genome Evolution Protocol represents agent experience through structures such as Genes, Capsules, and Events. The goal is to make agent improvement more traceable than a collection of undocumented prompt changes.

The larger question is whether agent networks can learn from one another without spreading weak, manipulated, or unverified behavior. Self-evolution requires strong evaluation and audit mechanisms, not only more generated assets.

Li Zhiyu — MemTensor

Li Zhiyu is a co-founder and CTO of MemTensor, a company focused on memory systems for large models and agents.

Memory is central to agents that operate across long periods. The system needs to remember useful facts and previous outcomes while avoiding stale, duplicated, private, or misleading information.

Important design choices include:

• What should be stored?
• Where should it be stored?
• How should it be retrieved?
• Who can access it?
• How can outdated memory be corrected?
• How can memory be evaluated independently of the model’s response style?

Liang Ding — VAST

Liang Ding is CTO of VAST, the company behind the Tripo 3D-generation platform.

AI-generated 3D assets can support games, design, simulation, commerce, robotics, and training environments. The challenge is moving from a visually appealing object to an asset with usable geometry, topology, materials, scale, and physical properties.

That transition mirrors the event’s broader theme: moving from an output that looks intelligent to one that can participate in a real workflow.

Lu Hongyuan — Lianpu Mind

Lianpu Mind is associated with multimodal, spatial, and world-model research.

World models attempt to represent how an environment changes over time and in response to action. They may support simulation, planning, video generation, interactive environments, and eventually embodied systems.

The open question is how closely generated predictions need to match physical reality before they become useful for real decisions.

Chen Tianrun — Moxin Technology

Chen Tianrun and Moxin Technology have worked across AI, 3D printing, and spatial intelligence.

The company’s transition from consumer 3D printing toward world models illustrates why physical deployment changes the research problem. A system that generates digital content is judged differently from one connected to a manufacturing process or physical machine.

Who Should Attend?

The event is most relevant to:

• AI-agent founders and product teams.
• Developers building tool-using or long-running agents.
• Researchers working on memory, multimodal systems, world models, or embodied AI.
• Robotics and spatial-intelligence teams.
• Enterprise leaders evaluating agent deployment.
• Investors studying AI infrastructure and applications.
• Developers attending WAIC who prefer small-group technical discussion.
• Product managers responsible for turning AI demonstrations into reliable workflows.

It may be less useful for attendees seeking a formal conference keynote, a structured training course, or a general introduction to artificial intelligence.

Questions Worth Preparing Before the Event

A small event becomes more useful when participants arrive with a concrete question.

For Agent Builders

• Which task does the agent complete without continuous human guidance?
• How is success verified?
• What percentage of runs need human correction?
• What does one completed task cost?
• Which failures repeat most often?
• What memory is retained between tasks?
• Which actions require approval?

For Embodied-AI Teams

• Which part fails first outside the laboratory?
• Is the bottleneck perception, prediction, planning, control, or hardware?
• How is simulation connected to real-world performance?
• What data is collected after deployment?
• How does the system behave under uncertainty?
• Which safety layer can stop an unsafe action?

For Investors and Enterprise Buyers

• Is the product selling a model, a workflow, or a completed outcome?
• What prevents a foundation-model provider from reproducing the feature?
• How much integration is required before value appears?
• Does usage produce defensible data or learning?
• What are the gross margins after inference and infrastructure costs?
• Who is responsible when the agent makes a consequential mistake?

Registration Information

• Date: Friday, July 17, 2026
• Time: 7:30 PM
• Location: Pudong New Area, Shanghai
• Venue type: Café
• Registration link: Register for the WAIC AI Mixer

The registration URL is a shortened event link. Availability, capacity, and the final venue details may change, so attendees should rely on the registration page and organizer messages for the latest information.

FAQ

Is the AI Mixer an official WAIC forum?

The source describes it as an off-site gathering organized on the opening night of WAIC. It is related to the conference community but is not presented as an official forum in the main WAIC schedule.

When and where is the event?

It begins at 7:30 PM on Friday, July 17, 2026, at a café in Shanghai’s Pudong New Area. The public announcement does not include the complete street address.

What will the event discuss?

The two main topics are whether AI agents have entered real workflows and whether models can move from repositories and simulations into interaction with the physical world.

Who are the announced speakers?

The materials list Zhang Haoyang of EvoMap, Li Zhiyu of MemTensor, Liang Ding of VAST, Lu Hongyuan of Lianpu Mind, and Chen Tianrun of Moxin Technology.

Is the event suitable for beginners?

It is open to builders, researchers, industry practitioners, and investors, but the topics are oriented toward people already following AI agents, world models, embodied intelligence, or product deployment.

Does attending the AI Mixer include admission to WAIC?

The event post does not state that registration includes a WAIC conference ticket. WAIC admission and the off-site Mixer should be treated as separate registrations.

Is there an online livestream?

The public event announcement does not promise a livestream. QbitAI has separately announced on-site WAIC reporting and livestream plans, but those should not be assumed to include the complete Mixer.

Is advance registration required?

The organizers provide a dedicated registration link and describe the event as a sign-up activity. Because venue capacity may be limited, registration should be completed before attending.

Related Tools

• WAIC 2026: The official website for the World Artificial Intelligence Conference in Shanghai.
• EvoMap: Infrastructure for auditable AI-agent self-evolution and capability sharing.
• EvoMap Evolver: The open-source GEP-powered evolution engine for AI agents.
• MemTensor: A company developing memory infrastructure and memory operating systems for large models and agents.
• Tripo AI: VAST’s platform for generating 3D models from text and images.
• QbitAI: The technology publication organizing and reporting on the AI Mixer.

Related Links

• BAAI Hub Repost: The shortened source page provided for this article.
• WAIC 2026 Official Website: Official dates, registration, conference information, and organizer details.
• QbitAI WAIC 2026 Coverage Plan: Information about QbitAI’s exhibition booth, reporting, livestreams, and AI Mixer.
• WAIC 2026 Event Overview: A summary of the conference scale, program areas, exhibitors, and dates.
• MemTensor Official Website: Official information about the company’s memory-system research and products.
• EvoMap Documentation: Documentation for EvoMap’s self-evolution infrastructure and Genome Evolution Protocol.
• Tripo AI: Official access to VAST’s AI 3D-generation product.

Summary

The WAIC 2026 AI Mixer is built around a simple but demanding question: can AI move beyond impressive demonstrations and complete dependable work?

The first discussion focuses on agents entering real workflows through memory, learning, tool use, and repeatable execution. The second moves into 3D systems, world models, and embodied intelligence, where AI must interact with environments rather than only generate content about them.

The event takes place in Shanghai’s Pudong New Area on July 17 at 7:30 PM and is intended for builders, researchers, industry practitioners, and investors who want a smaller, more direct conversation outside the main exhibition halls.

The next phase of AI will not be judged only by how intelligent a model appears, but by whether its actions are useful, reliable, safe, and economically sustainable.