• The Beijing training base in Shijingshan district will generate over 6 million data points annually across 16 specific scenarios including manufacturing facilities, retail outlets, elderly care centers, and smart homes.
• The boot camps are part of China's broader push into "embodied intelligence" - AI integrated into physical machinery. In March, the government highlighted this as a key future industry for the first time. Beijing rolled out guidelines in late August targeting over 90% adoption of AI-powered devices across the economy by 2030.
• Multiple cities are racing to build their own facilities. Shanghai announced its robot training base would begin field tests by month's end, while Beijing's facility will collaborate with counterparts in Zhejiang, Shandong, Anhui, and Henan provinces.

The "robot boot camp" strategy reveals how China is planning to compete in the humanoid robot race, by creating infrastructure that generates the training data needed to rapidly improve robot capabilities. Whether this centralized approach delivers better results than Silicon Valley's competitive model remains to be seen, but the sheer scale of investment signals China's serious commitment to becoming a leader in the humanoid robot industry.

• Trained on over 400,000 comprehensive UK Biobank medical records, Delphi treats each disease's diagnostic code as a "token" (similar to how ChatGPT processes language) and learns statistical patterns of how conditions connect and follow each other. The AI incorporates lifestyle factors like body mass, smoking, and drinking habits to generate personalized predictions.
• The Delphi system is also easily transferable. Without any modifications, it successfully mapped health trajectories for nearly two million people in Denmark's National Patient Registry spanning half a century. Its prediction accuracy barely dropped despite moving to an entirely different population and dataset.
• As impressive as it's been thus far, it's not without limitations: the training data has an overrepresentation of people who are white, middle-aged, and educated. It also has very little information on people over 80. The AI also struggles with conditions like Type 2 diabetes that have highly variable trajectories depending on lifestyle changes.

Nonetheless, the Delphi system represents a shift from treating diseases to predicting and preventing them. The ability to forecast health trajectories across multiple conditions could help healthcare workers prepare plans for aging populations while giving individuals unprecedented insight into their future health risks.

• Using ChatGPT on the Azure-based Nebula One platform, Connock simply prompted the AI to research and impersonate him, then build personalized material. The agent designed an impressive six-module course that was "intellectually challenging" and pulled from his entire multimedia output - books, speeches, articles, even university lectures he didn't know had been recorded.
• China's DeepSeek was far less impressive and didn't include a lot of the content that GPT did to design its training corpus. Connock humorously stated that: "there is no greater diss in the age of AI than a leading LLM deeming your book about AI irrelevant."
• Google's Gemini and X's Grok also didn't do as well as ChatGPT. They both struggled with basic facts: Gemini hallucinated biographical details about Connock and Grok attributed a quote from DeepMind's Demis Hassabis to Connock instead.

In his reflection about the experiments, Connock argued that the overall approach represents education's future: AI tutors offering hyper-personalized courses for each student, but with the caveat that human teachers will need to guide the learning experience based on individual students' learning needs.