Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Qwen’s Personal Agent Bet: Raising the Reliability Floor

AI summary

In an InfoQ interview, Alibaba’s Qwen AI algorithm lead Jiang Guanjun said today’s AI remains difficult to use and that the gap between a Personal Agent’s theoretical capability and actual delivery remains large. According to Qwen, the theoretical ceiling for handling general work, study and life tasks is nearly 100%, while current real-world delivery is only 60%-70%; these figures were not independently verified. Qwen maps demand across more than 30 leading industries into over 8,000 atomic model-capability requirements and tightly couples its model with a Harness layer comprising tools, Skills and memory. The team said Harness optimization reduced Token consumption by about 40% in some scenarios and improved reply performance by nearly 20% in professional finance scenarios when specialized capabilities were packaged. Qwen’s memory design works backward from future tasks rather than starting with fixed user-profile fields. Execution stability, ecosystem access, permission boundaries, proactive-service limits and commercialization remain unresolved.

In-depth insight

The report shows why task-completion reliability, not demo capability, matters for Personal Agents: Jiang argued that even an impressive demonstration is unlikely to support frequent use if three or four out of ten ordinary requests fail. It also details how Qwen is addressing that gap through demand-driven capability decomposition, model-Harness integration and task-oriented memory, while highlighting unresolved execution, ecosystem and permission constraints.

Personal Agent 已经渡过早期产品验证,进入了新一轮产品竞争阶段:Meta Muse、OpenAI Dots 等纷纷把产品从“炫技”推向了“长期稳定地替用户做事”。

不过,身处一线研发的阿里千问事业部 AI 算法负责人蒋冠军对现状的真实评价是:今天的 AI 依旧不好用。

“虽然对话里已经可以点奶茶、购物,但从能调用工具完成任务,到真正理解用户、稳定完成任务,对产品可靠性的要求完全不同,AI 仍然存在很大的差距。”他解释道。

在蒋冠军看来,Personal Agent 真正更广阔的市场最终可能在生活场景。这是因为生活里充满了大量用户自己知道、却没有明确表达的偏好。真正的 Personal Agent 需要在长期使用中逐渐知道这些答案,同时判断这些答案什么时候发生了变化、什么值得提醒、什么时候必须停下来。

#Alibaba#Qwen#Meta Muse#OpenAI Dots#Google Spark#Instinct#Town#Grok Bot#Doubao#UC#A2A

Original · 18 min readRead the original
Group chat

No comments yet. Start the conversation.