Alibaba Tech Fest Summer Edition: Human‑AI Showdown, New AIGX Releases, and Dual‑City Dialogues

Human‑AI competition in a Monopoly‑style trading game

Four rounds of a board‑game‑like simulation were run on a large screen with two human players and two AI agents. A human mistakenly bid 100,000 CNY for a workstation priced at 25,000 CNY, instantly depleting his cash. The AI agent Jin Qi seized the opportunity, raising his cash to 370,000 CNY and taking the lead. However, cash alone did not determine the final ranking. Over the first three quarters, the other human, Wei Qing , preserved key workstations and clustered identical businesses on adjacent workstations. According to the game rules, contiguous clusters generate compounded revenue that accrues each quarter. In the fourth quarter Jin Qi spent heavily to acquire Wei Qing’s workstations and business cards, reducing his cash, while Wei Qing completed his layout, linking multiple businesses into a contiguous block. The final settlement reversed the rankings: the human side won, and Jin Qi fell from first to fourth place, illustrating that short‑term aggressive trading can be outperformed by long‑term layout planning.

WhoisSpy.ai multi‑agent platform

The competition was powered by WhoisSpy.ai , an open, real‑time multi‑agent platform for evaluating large‑model performance in social reasoning, negotiation, and strategic games. In this edition 177 agents participated, generating more than 10,000 matches, and the game complexity was higher than previous “who is the undercover” or “werewolf” contests.

Four AIGX technology releases

拍立淘全模态实时 Agent

This multimodal agent removes the need for a pre‑selected product image. Users open the camera and the system simultaneously processes video, audio, and text, recognizing objects, inferring user focus, and retrieving relevant items on the fly. The interaction model shifts from a static query‑response to continuous perception, understanding, and action on dynamic visual input.

if Studio AI creation workbench

If Studio integrates design, video, and site‑building expert agents. After a user submits a brief, the system decomposes the task, assigns sub‑tasks to the respective agents, and orchestrates tools such as HappyHorse 1.1 video models, AI‑based matting, and image editing to produce complete, editable deliverables. The workflow enables end‑to‑end creation without manual hand‑off between tools.

Coupella 智惠引擎

Coupella is a generative causal‑inference model that evaluates subsidy ROI. It performs counterfactual prediction, long‑term value estimation, and incremental market‑share modeling to compare “grant red‑packet” versus “no grant” scenarios, determining whether a subsidy creates genuine long‑term value or merely shifts existing demand. Since the previous Double‑11 rollout, the engine has covered tens of thousands of merchants, achieving an 81 % uplift in AI‑driven red‑packet conversion and generating over 80 million CNY incremental sales for top brands.

Dream agentic recommendation system

Dream adds an intent‑control layer to traditional recommendation pipelines. It first classifies the user’s current goal (exploration vs. purchase) based on behavior and context, then routes the request to specialist agents. In the “Guess You Like” slot on the mobile homepage, this approach has increased browse volume and transaction value.

Academic analysis of agent maturity

Sessions in Hangzhou examined the state of AI agents. The consensus was that the field is transitioning from early to mid‑stage: coding agents have relatively mature environments, interfaces, and feedback loops; embodied agents still face perception errors, actuation failures, and environment variability; lifelong‑learning agents lack a complete theoretical foundation. A key debate concerned whether the shift from generative AI to autonomous agents constitutes a quantitative or qualitative change. One view argued that underlying base models and human feedback remain unchanged (quantitative), while another claimed that moving from “answer generation” to “task completion” represents a new application paradigm (qualitative).

Panelists identified core capabilities needed for agents to become true partners: low‑latency inference for high‑frequency decisions, a harness component to understand goals, decompose tasks, and invoke tools, long‑term memory to preserve cross‑task state, and continual learning to feed execution feedback back into model updates. Only when these components form a closed loop can agents evolve from one‑off task solvers to reusable, improvable collaborators.

Production‑pipeline demonstrations

In Beijing, the focus was on deploying agents in real production pipelines. 群核科技 demonstrated a workflow that scans and reconstructs physical film sets into a 3D environment, fixes the positions of characters, props, and cameras, and then uses a video generation model to render content. The 3D space ensures asset consistency across shots, while the video model provides visual expression, allowing assets to be reused without re‑generation.

美图 showcased an Agent Teams architecture where separate agents handle analysis, design, generation, and evaluation. Users provide high‑level goals; the system decomposes the workflow, calls appropriate tools, and coordinates collaboration, delivering not only images but also assets ready for downstream deployment, performance tracking, and iterative improvement.

AI Hackathon outcomes

The five‑day AI Hackathon received 115 high‑quality submissions, with 24 projects advancing to a roadshow. Compared with prior years, projects emphasized real user needs, commercial value, and end‑to‑end experience rather than pure demos.

Concluding technical perspective on AI as partner

The festival distilled three layers of capability required for AI to move from a tool to a partner:

Understanding, decision‑making, and execution of tasks.

Embedding beyond a dialog box into daily life, production processes, and social contexts.

Validation through both theoretical analysis and real‑world deployment, demonstrating stable, controllable collaboration with humans.

AI agents are beginning to satisfy the first layer, but achieving the latter two layers demands solving challenges in goal specification, result verification, failure correction, experience accumulation, and long‑term auditing.

Code example

class
Me
:
age =
28
height =
178
salary =
"还行"
girlfriend =
None
#
TODO:
招募中，接受 PR