AI Enters Its Working Era: Three Big Moves from Chinese Large Models in One Week

For the past year we were used to treating AI as a chat toy. This week it suddenly became a digital employee that can do work for you, guard network security and write tens of thousands of lines of code. The change came faster than expected, and the people who read it early have already started saving time, saving money and finding opportunities.

Start with the conclusion: the performing era of AI is over

On 13 August the official release of DeepSeek V4 Pro shipped. On Terminal Bench, which measures whether AI can really finish complex tasks on a computer, it scored 87.9 — only 0.1 point away from the global number one, and a big 15.8-point jump over the preview version. That means AI no longer just understands what you say; it can actually finish the job for you.

Three big moves

For the first time, open-source models arm-wrestle the closed ones

Alibaba open-sourced the flagship 2.4-trillion-parameter weights of Qwen3.8-Max; Zhipu GLM-5.3 broke through on both code generation and network security; and DeepSeek V4 Pro work capability is close to the global number one. DeepSeek output pricing is only 0.87 US dollars per million tokens — programming ability within a few percentage points, at a fraction of the price of closed models.

Network security became the AI entrance exam

When AI starts doing work for you and managing data, it first has to be trustworthy. Zhipu launched the Open Source Shield programme, and Chinese large models are moving from being able to write code towards being able to guard digital infrastructure. Security is the first threshold an agent must cross from demo to production.

Agents are becoming your digital employees

On GitHub, AI agents generate more than 17 million pull requests a month; Vercel open-sourced fx, a 6.4 MB lightweight coding agent with a 10-microsecond cold start. Agents are not demos any more; they are really working.

What this wave means for ordinary people

First, stop treating it as a chat toy. The real way to use it is to assign tasks — hand over the whole piece of work and let it deliver a finished result. Second, open-source models let a small budget lever a large capability, and this is the best window individual founders have ever had. Third, your knowledge base should be made ready to be used — building a knowledge base an AI can call is worth more than bookmarking ten thousand good articles.

Action suggestions

1. Change AI from something you ask to something you assign work to, and let it deliver finished results.
2. Judge tools by two measures: can it run cheaply (open source), and can you safely hand work to it (security).
3. Start tidying your knowledge base now, before your competitors put AI to work.

FAQ

Q: Can AI really replace human labour?

A: In the short term it is not replacement but a dividing line between people who can use AI and people who cannot. AI is good at repetitive work; judgement and building good workflows remain human value.

Q: How do ordinary people get to use open-source models?

A: DeepSeek, Qwen and GLM all offer APIs and clients, many of them free or cheap. Start by assigning tasks — let it build tables, organise material and write first drafts.

Q: How do you judge whether an AI tool is worth using?

A: Two things — can it run cheaply, and can you safely hand work to it (data security). If it meets both, it is real capability you can put into production.

Original work by SavantCat, first published at savantcat.cn. Any reprint must credit the source. Pricing is quoted per project for enterprise knowledge base and AI customer service engagements.