In this issue
The short version
- AgiBot's GO-3 Preview robot model is designed to keep improving after deployment. AgiBot reports benchmark wins and plans to open-source it by the end of 2026.
- Tencent's Buddy One links Tencent Docs, Meeting, Drive and ima under one account, one credit pool and one agent that answers with sources.
- Guoxing Aerospace, which builds satellites for in-orbit AI computing, passed its Hong Kong listing hearing on its third attempt.
- Marvell was reported to be in talks to buy Lumentum. The deal was announced on November 2.
The October 30 issue covers AI that learns on the job, from robots in a convenience store to an office assistant that reads across all your work apps. It also covers a Chinese space company that wants to put AI computing in orbit, and early reports of one of the year's largest chip deals. Each section explains what was announced and what to do with it.
01
AgiBot GO-3 Preview: a robot model that keeps learning
What AgiBot announced
Chinese robot maker AgiBot (智元机器人) released GO-3 Preview on October 29, a foundation model for robots designed to learn at every stage: pre-training, post-training and after deployment. Its architecture links task reasoning, prediction of what will happen next and action generation. It is trained on first-person video of people doing tasks together with robot movement data, so it can attempt unfamiliar household and kitchen tasks and learn longer, fine-grained jobs from a few demonstrations.
After training, a system called Genie Evolver improves the model using a fleet of about a hundred real robots, physics simulation and a world model. Once deployed, a framework called Genie Master lets robots spot their mistakes, adjust and keep useful experience; AgiBot demonstrated this with robots working in a real convenience store. AgiBot says GO-3 Preview tops three robot benchmarks: RoboDojo, RoboTwin and RoboColiseum. These are results AgiBot presented at its own event for a model labelled as a preview. The company plans to open-source GO-3 Preview and Genie Master by the end of 2026.
What learning after deployment means
The important idea is not the benchmark scores but continuous improvement in real conditions. The same idea already applies to software agents: they get better when you capture what went wrong and feed it back. Whatever agent you run, keep a short log of failures, write a one-line rule for each recurring one and add it to the agent's instructions. Review the log monthly. That small habit gives you the benefit of "learning after deployment" without any new technology.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
Tencent Buddy One: one AI across your office apps
What Tencent launched
Tencent (腾讯) launched Buddy One on October 29, an AI system for individuals that connects Tencent Docs (腾讯文档), Tencent Meeting, Tencent Drive and the ima knowledge-base app. All of them share the WorkBuddy agent, one account and one credit pool. After linking accounts in WorkBuddy for PC (version 5.9.0 or later), you can ask a question in one sentence; the agent searches the documents, meetings and knowledge bases you have authorised and answers with the source app marked for each item. Credits can be spent in any of the four products. Tencent also positioned Tencent Drive as the shared storage layer for its apps and agents.
Tencent says memory sharing across products and more apps will follow. For now the cross-product features are mainly the shared account, search over authorised content and the shared credits.
Getting value from a cross-app assistant
An assistant that reads across your documents and meetings is only as good as the material it can find. Before linking everything, tidy one project: give documents clear titles, put meeting notes in one folder and delete old drafts. Then ask practical questions that span apps, such as "what did we decide about the launch date, and where is it written?", and click through the sources to check each answer.
Link only the accounts you need, and review what the assistant can access every few months. The same approach works with other ecosystems, such as Claude or ChatGPT connected to your files. Our AI research workflow shows how to keep sources traceable.
03
Guoxing Aerospace clears its Hong Kong listing hearing
What happened
Guoxing Aerospace passed its listing hearing at the Hong Kong Stock Exchange on October 29, on its third attempt. Founded in 2018, it builds AI satellites, runs in-orbit operations and provides ground-station services, and was one of the first Chinese commercial space companies to work on AI computing in orbit. As of October 20 it held orders for 114 AI satellites worth ¥3.475 billion. Its "Star Computing" plan aims for a constellation of 2,800 satellites, with the first 100 launched by 2028 and the full network by 2035. The company has made losses in each reported period, and the final price and listing date were still to be set.
Putting space-computing stories in context
Space computing now has a commercial company heading to the public market, alongside research projects from Google and Chinese institutes. For people who use AI services, nothing changes soon. When you read these announcements, separate signed orders and launched satellites from long-term plans, and check whether the company is profitable. Company databases such as Crunchbase AI in the Finance AI category help track the businesses involved. This is general guidance, not investment advice.
04
Marvell in talks to buy Lumentum
On October 29, reports said chipmaker Marvell was in talks to buy Lumentum, which makes optical components that move data inside AI data centres. The deal was announced four days later, on November 2, and we cover the confirmed terms in our Special Edition for November 1 to 7. The episode is a reminder to treat "in talks" reports as unconfirmed until a company announces terms.
What to try next
Put one change into practice
Start a failure log for any agent you run and turn recurring problems into rules. Tidy one project's documents before you connect a cross-app assistant to it. Read space and chip headlines for what has been signed, not only for the plan.
Explore Data Analysis, Chatbots and Finance AI 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.