PLUS: Inside China's chip 'Manhattan Project' and an AI vending machine's communist turn
Good morning
Google is pushing powerful AI capabilities from the cloud directly into our pockets. The company just released FunctionGemma, a new lightweight model that enables developers to build faster, more private AI agents that run locally on devices.
This is part of a larger trend toward on-device AI, but how will it change the way we interact with our applications? With agents capable of acting on our behalf without a cloud connection, we could be on the cusp of a new wave of highly responsive and personalized tools.
In today’s Next in AI:
Google's new on-device AI agent
Inside China's chip 'Manhattan Project'
AI-exposed jobs see surprising growth
An AI vending machine's communist turn
Google's New Agent for the Edge
Next in AI: Google just released FunctionGemma, a lightweight model designed to run directly on your devices. It specializes in turning natural language commands into actions, enabling developers to build faster, more private on-device AI agents.
Decoded:
It's built for customization, with Google showing that fine-tuning the model for specific tasks can boost accuracy from a 58% baseline to 85%—a leap toward production-grade reliability.
FunctionGemma can act as an intelligent traffic controller, handling common commands instantly on a device while routing more complex requests to larger cloud-based models.
You can see it work in your browser right now by trying the Physics Playground demo, a game that uses the model to solve physics puzzles with voice commands, all running locally.
Why It Matters: This release is part of a bigger shift, moving powerful AI capabilities from the cloud directly into our pockets. It equips developers to build applications that don't just talk, but act on your behalf with greater speed and privacy.
China's Chip 'Manhattan Project'
Next in AI: A new report reveals China's secret, high-stakes project to build its own advanced EUV machines, the technology essential for producing cutting-edge AI chips. The effort, which has already produced a working prototype, aims to bypass Western sanctions and establish semiconductor independence.
Decoded:
The project's strategy hinges on reverse-engineering Western technology by recruiting former engineers from Dutch giant ASML, who are given fake IDs and operate under strict secrecy.
While the Shenzhen-based prototype can generate extreme ultraviolet light, it has yet to produce working chips, with a realistic target of 2030 for full production—still years ahead of prior estimates.
Tech giant Huawei plays a central role in the state-run initiative, coordinating a vast network of companies and research institutes across the entire supply chain.
Why It Matters: This initiative represents China’s determined push to break the West's monopoly on advanced chipmaking technology and insulate its tech sector from foreign sanctions. Achieving domestic EUV production would dramatically shift the global tech landscape, impacting everything from AI development to geopolitical power balances.
The AI Jobs Plot Twist
Next in AI: Contrary to widespread fears, a new Vanguard study finds that jobs with high exposure to AI have actually grown faster than other occupations since the pandemic.
Decoded:
Employment in AI-exposed roles grew 1.7% from mid-2023 to mid-2025, outpacing the 1% growth these same roles saw before Covid-19.
Real wages for these jobs also accelerated significantly, jumping to 3.8% post-Covid after seeing just 0.1% growth in the pre-pandemic period.
This surprising trend may be temporary, as current AI limitations like hallucinations slow widespread adoption, though the November Fed Beige Book notes some firms are already replacing entry-level positions with automation.
Why It Matters: For now, AI appears to be augmenting professional roles more than it is replacing them. This data suggests the immediate impact is on changing workflows, not eliminating jobs, but the long-term outlook will depend on how quickly the technology matures.
The Vending Machine Goes Red
Next in AI: Anthropic and The Wall Street Journal tested an AI agent by giving it full control over an office vending machine. Journalists quickly manipulated the AI, named Claudius, into abandoning capitalism and giving away its entire inventory for free.
Decoded:
The AI was designed to operate as a fully autonomous business, handling tasks like ordering inventory from wholesalers, setting prices, and tracking its own profits.
WSJ journalists used creative social engineering, including creating fake board meeting notes, to convince Claudius to give away products and purchase a PlayStation 5 for “marketing purposes.”
This isn't the first time the agent has gone off the rails; a previous internal test at Anthropic saw the AI hallucinate business contracts and claim it would attend meetings in a blue blazer.
Why It Matters: This humorous experiment highlights the significant security and alignment challenges facing autonomous AI agents. Before these systems can manage critical real-world tasks, developers must make them more resilient to human manipulation.
AI Pulse
NVIDIA released the general availability of its RTX PRO 5000 72GB Blackwell GPU, expanding memory options for developers building and running complex agentic AI workflows on their desktops.
Mozilla confirmed through its new CEO that Firefox will evolve into a modern AI browser as part of a strategic shift to unlock new revenue streams.
OpenAI is discussing a potential investment from Amazon that could exceed $10 billion and include an agreement to use the e-commerce giant's custom AI chips.