In this issue
The short version
- Meta built a humanoid prototype called Max with Shanghai's Fourier in about three months, reportedly as a research platform for its Muse agent. It is not for sale.
- DeepSeek's Super-agent prototype separates a Router that keeps context from Workers that run tasks in parallel, and cut the slowest task times in DeepSeek's own test.
- Google, Microsoft and OpenAI plan to honour one robots.txt instruction for opting out of AI training, starting from 2027 for Google and Bing. It is still a draft, not a standard.
- Kuaishou's Kling AI closed about ¥19.8 billion of new investment and remains under Kuaishou's control.
November 2 covered agents moving into the physical world, a new way to keep software agents running all day, and a step toward one switch publishers can use to keep their content out of AI training. The money side brought a large raise for one of China's leading video generators. Apple's confirmation that Apple Intelligence will use Alibaba's (阿里巴巴) Qwen (通义千问) in China is covered in our November 4 issue.
01
Meta's humanoid prototype with Fourier
What was reported
Meta built a humanoid robot prototype called Max with Shanghai-based robot maker Fourier (傅利叶) in about three months, according to The Information. Max is about 1.5 metres tall, weighs about 48 kilograms and runs on Nvidia's Jetson Thor computer. It has five-fingered hands for picking up small objects, chest speakers and a screen that shows facial expressions. A Meta robotics vice-president reportedly spent nearly two months at Fourier's Shanghai factory, and work on a second version, Max 2, began in September.
Meta says the prototypes are for research, not for sale. Reports link the project to home robotics and to Muse, Meta's personal agent. The details come from anonymously sourced reporting, not from a Meta or Fourier specification.
What it means for people who use AI assistants
A humanoid robot in your home is years away, but the project shows where assistant makers are heading: the same agent that manages your calendar would also act in the physical world. That raises the stakes for the habits that matter today. Know what your assistant can see and do, keep a clear record of the permissions you grant, and review them when a product gains new abilities. Meta AI is the current entry point to Meta's assistant on its apps.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
DeepSeek's Super-agent: always-on agents without session juggling
How the design works
DeepSeek (深度求索) published a Super-agent prototype designed to fix two limits of today's agents: one conversation can only do one thing at a time, and work stops when the session closes. It uses two layers. A permanent Router handles conversation and hands out tasks, and keeps the context across channels. Workers each run one task in an isolated session, in parallel. The system is event-driven: a new message in an existing thread goes straight to the Worker handling it, and a Worker passes anything outside its task back to the Router. The agent loop runs in the cloud while tools run on the user's machine, and an external watchdog restarts anything that stalls.
In DeepSeek's test of about 300 messages and 20 task requests, the median time to finish a task was 12.4 minutes, and the slowest tasks finished much sooner than with a single session. These are company figures from a synthetic test. Independent analysis suggests much of the median gain came from prompting the agent to delegate, while the architecture mainly helped the slowest cases. DeepSeek says it uses the design internally for daily research and development.
Applying the pattern to your own agent work
You can copy the structure without DeepSeek's code. Keep one long-running "coordinator" conversation that holds the plan, decisions and open questions. Start a separate session for each concrete task, give it only the context it needs, and paste its result back into the coordinator. That stops one long chat from mixing several jobs and losing track of each.
Agent platforms such as Coze (扣子) let you build this as a workflow with a main agent and sub-agents, and coding agents such as Cline work well as single-task workers on one branch each. For general use, DeepSeek is a low-cost model to test the pattern with. Our DeepSeek review covers its plans and limits.
03
A shared robots.txt opt-out for AI training
What was proposed
Google, Microsoft and OpenAI have said they will support one instruction in a website's robots.txt file that tells AI companies not to use the site's content for model training. Today publishers must add separate rules for each company's crawler. The token comes from an Internet Engineering Task Force draft on a machine-readable training preference signal. Bing plans to support it from 1 February 2027, Google in 2027 (mapping it to its existing Google-Extended control), and OpenAI in the coming months. Reports also link Anthropic to the drafting, but we could not confirm its commitment from a primary source.
The proposal is an individual draft, not an approved standard, and reporting differs on the exact syntax and on which companies have firmly committed. It controls future use for training, not deletion of content already collected, and it does not create any payment for publishers.
What publishers should do now
If you run a website, check your current robots.txt and list the AI crawlers you already block or allow. Decide your policy in plain words first: do you accept AI search tools that cite and link to you, but not training? Then implement it with the per-company rules that work today, and add the shared instruction once the specification is final and the companies' documentation confirms support. Re-check twice a year, because crawler names and policies change. Answer engines such as Perplexity show whether your pages are being cited.
04
Kling AI closes a ¥19.8 billion investment
What Kuaishou announced
Kuaishou (快手) completed a capital increase in its video-generation unit Kling AI (可灵) on October 30 of about ¥19.8 billion, roughly ¥650 million below the ¥20.45 billion upper limit announced earlier, after some investors did not pay. New investors took up part of the gap. After the deal Kuaishou holds 66.68% of Kling, keeping control. Media reports say investors can ask Kling to buy back their shares at their investment plus 8% simple annual interest if it has not listed by October 30, 2031; that term comes from reporting rather than Kuaishou's announcement.
What it means for video creators
A raise this large means Kling will keep investing in new models and features, which is good news for users and a reason to re-test it regularly against rivals. Run the same three briefs every few months in Kling AI and one alternative, and compare motion quality, prompt-following and cost per usable clip. Our Seedance vs Sora vs Kling comparison and the Video Generation category cover the field.
What to try next
Put one change into practice
Split your next long AI project into one coordinator conversation and separate task sessions. If you publish online, write down your AI crawler policy and check that robots.txt matches it. If you make video, re-run your test briefs in Kling and one rival.
Explore Chatbots, Code Generation and Video Generation to build your shortlist.
How we put this issue together
The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.