Greetings from a world where…
Jeopardy is appointment viewing during this streak
…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Feature Translation: When AI Enters the Harness Era: MiniMax as a Case Study for the New Cloud Infrastructure of AI Agents
Context: In February, a small OpenAI team disclosed the results of a five-month experiment: “building and shipping an internal beta of a software product with 0 lines of manually-written code.” Using OpenAI’s software engineering agent Codex, the three person-team used a “harness engineering” approach, summed up in four words: “Humans steer. Agents execute.” In this world, human AI engineering skills have shifted from merely designing context-rich or helpful prompts:
“Our most difficult challenges now center on designing environments, feedback loops, and control systems that help agents accomplish our goal…”1
In this week’s feature translation of a jiqizhixin (Synced) article, we dig into how Chinese companies are adapting to the Harness era. This article focuses on an interesting partnership between the Chinese AI start-up MiniMax’s AI agent (MaxClaw) and Alibaba Cloud.
Key Takeaways: MiniMax’s MaxClaw agent is in the first tier of OpenClaw-based adaptations, but how does it deal with fundamental engineering barriers? Remember, the Harness approach gives AI agents significant autonomy and access to high-risk privileges.
The article describes the core problem set: “When hundreds of thousands of agents concurrently execute these complex tasks with highly strict access privileges and multiple steps, traditional system architectures frequently struggle to cope…How does MiniMax manage to withstand the massive traffic surges generated by MaxClaw and MaxHermes while ensuring business stability?”
Synced highlights four “chasms” that limit the diffusion of AI agents into enterprise-grade production environments: 1) security boundaries — since agents are increasingly running directly on top of local operating systems, they have high-risk privileges (e.g., deleting key files); 2) state volatility involved in executing long-running tasks; 3) need for scheduling multiple agents; and 4) tension between cost consumption and drastic workload fluctuations.
Enter MiniMax’s partnership with Alibaba Cloud’s Container Service for Kubernetes (ACK) and Container Compute Service (ACS), which tries to address those four engineering challenges associated with large-scale harnessing of agents.
On the security challenges, Agent execution occurs in the ACS Agent Sandbox, which isolates and contains the risk of prompt injection or privilege escalation attacks. Per Synced, “the scope of the resulting risk is strictly contained within that specific instance, making lateral movement or container escape extremely difficult.”
To handle unpredictable computational demands — AI agents sometimes have relatively idle periods, punctuated by sudden spikes of workload — Alibaba Cloud supports elastic scheduling of compute.
MiniMax’s evolution points toward the importance of cloud computing platforms as “AI super computers.”
Citing an IDC survey that finds AI inference accounts for 47% of all token usage and API calls, the article notes that the AI industry is undergoing “transition into a phase of large-scale post-training and inference execution, as the focus of computing power rapidly shifts toward Agent-centric scenarios.”
If enterprises collectively operate hundreds of thousands of AI agents, round-the-clock seven days a week, then cloud computing platforms and modern container services like Alibaba Cloud’s ACK and ACS will serve as “cloud-native operating systems” for AI supercomputers.
FULL TRANSLATION: When AI Enters the Harness Era: MiniMax as a Case Study for the New Cloud Infrastructure of AI Agents
ChinAI Links (Four to Forward)
Must-read: 27 Years Later, Still Bombing the Wrong Targets
Jon Lindsay is an Associate Professor at Georgia Tech, where he researches the impact of information technology on military power. I’ve really enjoyed his Substack posts lately. This one analyzes the U.S.’s mistaken strike of an Iranian elementary school, drawing on his previous experience as an intelligence officer when the U.S. military accidentally attacked the Chinese Embassy in Belgrade:
I thought so because I had seen something like this before. Almost three decades ago, I served as an intelligence officer during Operation Allied Force, the Kosovo Air War of 1999. During that war, the US military accidentally blew up the Chinese Embassy in Belgrade. A mistake of that magnitude seemed unimaginable at the time. How could we possibly not know what we were hitting, and then hit that of all things? Anyone who walked by the street in Belgrade would plainly see it was an embassy, but US analysts only looked at overhead satellite imagery. The Chinese had told their counterparts in the US Department of State that they had relocated to that new building a year before, but that info had not been input into the military database.
Should-read: (Alibaba’s) DingTalk and (ByteDance’s) Feishu collectively pivot to CLI: Is MCP dead? Is GUI on its way out? (in Chinese)
How should AI agents interact with computers, data sources, and humans? Xiaochai Ye [叶小钗], a developer with previous experience at Tencent, Ctrip, and Baidu, shares thoughts on why some Chinese enterprise AI platforms prefer the Command-Line Interface over the Model Context Protocol.
Should-read: Xiaoice’s [小冰] “Unicorn” Illusion (in Chinese)
Back in December 2023 (ChinAI #248), we deep-dived into the history of Xiaoice, previously China’s most popular chatbot and first incubated in Microsoft before being spun off. Two and a half years later, Xiaoice has been shut down. What happened? This Leiphone article tries to explain.
Should-read: China’s robot champion has everything to lose
For Reuters Breakingviews, Robyn Mak profiles China’s Unitree, which recently applied to go public on China’s STAR market.
Thank you for reading and engaging.
These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99
I’ve gathered some additional notes from the blog. On the novel problems produced by full agent autonomy: “Initially, humans addressed this manually. Our team used to spend every Friday (20% of the week) cleaning up “AI slop.” Unsurprisingly, that didn’t scale.” The solution required some level of human taste: “Instead, we started encoding what we call “golden principles” directly into the repository and built a recurring cleanup process. These principles are opinionated, mechanical rules that keep the codebase legible and consistent for future agent runs.” Source: https://openai.com/index/harness-engineering/
"harness era" is a wild term for "everyone's building on someone else's infrastructure" — though honestly not wrong?