Rest of World
· 1 min · Original

The AI talent war in Asia is about regulators and distribution, not model researchers

OpenAI and Anthropic are poaching senior Asia executives from Meta, Google, Microsoft and Amazon. Rest of World counted at least nine such moves to US AI labs in the past year. Meta’s former India and Southeast Asia VP Sandhya Devanathan, for example, has joined OpenAI to run Southeast Asia and Australia. What is scarce here is not ML knowledge. It is executives who have run a country-level P&L, dealt with Indian or Southeast Asian regulators, built government relationships and persuaded large enterprises to adopt unfamiliar technology. One recruiter describes them effectively as “mini-CEOs” for each market. Once frontier models become sufficiently comparable, an important competitive advantage shifts from research talent toward distribution, policy access and trusted local relationships.

Rest of World
· 1 min · Original

Google Earth’s 24-hour AI experiment damaged something more valuable than a feature

For roughly one day in July, Google let users generate photorealistic AI alterations directly inside Google Earth at real geographic coordinates. Researchers promptly generated fake nuclear infrastructure in Iran, a bomb crater next to a Gaza hospital and other fabricated scenes. Google withdrew the feature within 24 hours and promised stronger safeguards. Mahsa Alimardani’s deeper point is about the “liar’s dividend.” Google Earth has long been used by OSINT investigators as a baseline against which war footage, atrocities and satellite imagery can be verified. Once a trusted evidentiary tool can itself manufacture plausible scenes, genuine evidence becomes easier to dismiss: an actor does not need to prove an image is fake, merely make “maybe it’s AI” credible. Generative AI’s worst misinformation effect may ultimately be reducing confidence in authentic evidence rather than convincing everyone to believe individual fakes.

Rest of World
· 1 min · Original

China’s lawyers, architects and engineers are becoming gig workers training AI

This piece captures the AI labour transition unusually well. Rest of World found underemployed Chinese professionals doing evening annotation work that requires genuine domain expertise. One Shenzhen architect whose income had halved teaches models how to produce building proposals and profitability analyses. Lawyers generate and grade legal reasoning; experienced software engineers design agentic coding tasks. The jobs can pay only hundreds of yuan for several hours, and rejected work may earn nothing. Yet China actively wants this: its National Data Administration is encouraging “high-quality datasets” with higher “knowledge density.” ByteDance’s Xpert says it has recruited more than 50,000 experts. China’s training-data market is forecast at 7.8bn yuan this year, up 25%. The paradox is almost perfect: deteriorating white-collar employment makes expert knowledge cheaper precisely when AI companies need that knowledge to automate more white-collar work. The scarce input for the next model generation may increasingly be tacit professional judgment rather than another trillion tokens of internet text.

Rest of World
· 2 min · Original

Taiwan’s six-year hunt for China’s undercover chip labs

Rest of World obtained previously unpublished Taiwanese government data showing how extensive that human-capital struggle already is. Taiwan has investigated 166 technology cases involving alleged concealed Chinese operations or talent poaching over six years, plus 67 China-linked trade-secret investigations. The clever part is corporate disguise. One engineer thought he had joined an American AI-chip company on a $120,000 salary — twice the local norm — but discovered there were effectively no American colleagues and he was sending designs to coworkers in Nanjing. Prosecutors say Chinese companies have repeatedly created Taiwanese shells, registered through foreign nationals or used entities in jurisdictions such as Samoa to disguise their origin. Rest of World examined court records from 36 investigations; all 36 produced convictions for operating without the required approval, and 33 involved R&D or chip design. But actual trade-secret theft is much harder to prove: a Hsinchu prosecutor says about 70% of illegal-operation cases lead to prosecution versus roughly 20% of trade-secret cases. Stopping technology transfer becomes dramatically harder when the relevant “technology” is an engineer’s experience rather than a lithography machine. ### The chip war is becoming a labour-policy war Export controls work best on things that cross customs borders. Korea and Taiwan increasingly need corporate-ownership checks, employment restrictions, IP enforcement and espionage law because expertise crosses borders inside people. The danger is obvious too: legitimate international recruiting and R&D can increasingly be interpreted through a national-security lens.