---
id: "lib-2026-10-09-silicon-valley-101-alibaba-real-tasks"
type: "library"
title: "Why users don't feel stronger models: Zhang Kuo, Alibaba.com, revisited"
title_original: "模型越来越强，为什么用户没感觉？再访阿里国际站总裁张阔"
lang: "ZH"
org: "硅谷101播客 Silicon Valley 101"
source_type: "Podcast"
date: "2026-10-09"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-09-silicon-valley-101-alibaba-real-tasks/"
original_url: "https://www.youtube.com/watch?v=Fhx_qNS2P8g"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Podcast interview with the president of Alibaba.com giving specific figures from an open 107-task commercial benchmark, but the cost and pass-rate numbers are company claims not independently checked."
summary_basis: "transcript"
original_minutes: "46"
tags: ["ai technology", "business transformation", "real task benchmarks", "small model routing"]
frames: ["verification-capacity", "specification"]
---

# Why users don't feel stronger models: Zhang Kuo, Alibaba.com, revisited

Original title: 模型越来越强，为什么用户没感觉？再访阿里国际站总裁张阔

Tier B (Useful, limited): Podcast interview with the president of Alibaba.com giving specific figures from an open 107-task commercial benchmark, but the cost and pass-rate numbers are company claims not independently checked.

Summary based on the full transcript.

## In short

Episode E255 of the Silicon Valley 101 podcast (硅谷101), published 9 October 2026, with Zhang Kuo, president of Alibaba.com, the cross-border B2B marketplace, on a return visit six months after his first appearance. He explains why users of commercial agents do not feel the gains that public benchmarks report. His team built 107 tasks from real merchant work and released them as open source; the best frontier model passes about 61% without human help. He argues that non-coding work lags coding because commercial outcomes are heterogeneous, firm-specific and slow to verify.

## Key numbers

- 61%: pass rate of the best frontier model on Alibaba's 107 real e-commerce tasks, with no human intervention
- $3.69 vs $9+: cost to run all 107 tasks with Accio Work vs Codex and Claude Code, company figures
- 130%: growth in multimodal search on Alibaba.com, against 20% for search overall
- ~99%: his estimate of coding-agent penetration among developers

## What it reports

- He estimates coding-agent penetration among developers at about 99%; every other vertical is far behind.
- On Alibaba.com's AI search, overall search grew 20% while multimodal search grew 130%.
- A supplier-comparison spreadsheet a buyer used to maintain for a month can now be produced in about 24 hours.
- His formula: Agent = Model × Harness × Context, multiplicative, so a zero in any term gives zero.
- The 107 tasks cover seven categories, including sourcing and trade disputes, graded L1-L4 by horizon; the set keeps growing to prevent regression.
- Frontier labs report public benchmark scores near 99 out of 100; on the 107 tasks the best model passes "a little over 61%".
- Running all 107 tasks cost $3.69 with Accio Work against more than $9 for Codex and Claude Code, about one third.
- A post-trained small Qwen model comes close to top frontier models on simple, high-volume work.
- Payment and purchase approval stay with the human, as do budget allocation and product judgement.

## How they got it

A 46-minute conversational interview. Figures come from Alibaba's internal product data and its own benchmark built from merchant operations and Accio question data. The digest uses cleaned auto-captions of the Mandarin audio.

## Limits

- The benchmark, pass rate and cost comparison all come from the company selling the agent; the scoring method is not explained.
- An August 2026 press feature reported a top score of 56.1 on what appears to be the same benchmark, which conflicts with the 61% figure.
- No independent data on merchant revenue or time saved.

## Why it matters for companies

Public leaderboard scores say little about performance on a company's own processes; a set of real internal tasks is a better test before buying or scaling an agent. Routing routine steps to small, cheap models and keeping expensive models for hard cases is where much of the cost difference comes from. His line on approvals is a practical design rule: agents prepare, people press the pay or purchase button. Firms with years of proprietary transaction data hold the "context" term of the formula.

## Key passages in translation

> 我们现在认为Coding的市场渗透率 我估计有99% 因为很难看到身边有一个研发人员 不用任何AI去写代码

We now think coding's market penetration - I estimate it at 99% - because it is hard to find a developer around you who does not use any AI to write code.

> Agent = Model × Harness × Context ... 为什么是乘法 哪一项是0 整个这个公式结果就是0

Agent = Model × Harness × Context... Why multiplication? If any one term is zero, the result of the whole formula is zero.

> 最frontier的模型 在我们这个基准测试下面 能够通过率就是61%多一点点 它想要达到及格水平 现在都有点费劲

On our benchmark the most frontier model has a pass rate of just a little over 61%; right now it struggles even to reach a passing grade.

> 用Accio Work的估算成本 是3.69美元 我看到是你们给到的一个数字 Codex和Claude Code 都是在9美元以上

The estimated cost with Accio Work is 3.69 dollars - that is a figure I saw you provided - while Codex and Claude Code are both above 9 dollars. (The host, citing the company's figure.)

> 你用这个Agent的产品经营 当然给定一定的边界 就是你的预算 你的供应链是一样的情况下 这个Agent一定比一个人 去做日常的这些操作和判断 能赚到更多的钱

You run the business with this agent product, within given limits - your budget and your supply chain being the same - and the agent earns more money than a person doing these daily operations and judgements. (His definition of "commercial AGI".)

> 比如你要去支付 要采购这批货这个按钮 一定是交给人去判断的 因为你要对于你整个的经营结果负责

For example, paying, or the button to purchase this batch of goods, must be left to a person's judgement, because you are responsible for the results of your whole business.

Original: https://www.youtube.com/watch?v=Fhx_qNS2P8g