Sara

The US and China talk about AI risks this week. But are they even afraid of the same things? These are my personal observations on how those differences shape the conversation about AI safety. While in the US much of the conversation is around AGI, in China, I found less obsession with AGI and more focus on practical applications, especially robotics. The people I spoke with also seemed more optimistic about what AI could do for them. But that optimism sits alongside a different set of concerns about influence, social stability, and government control. If you enjoy breakdown like this, you should subscribe to my newsletter. If you comment “newsletter”, I’ll send it in your inbox every Sunday.

How worried should you actually be about AI? I’ve worked in AI for 10 years, and I’m currently collaborating with an AI safety research lab in San Francisco. The industry is deeply divided. Critics argue that OpenAI and Anthropic exaggerate the dangers to push regulations that protect their businesses. But research on AI misalignment predates today’s debate. Corporate interests don’t mean the risks are invented. And we’re bad at predicting progress. Before ChatGPT, I worked with chatbots and wondered whether they’d ever give me an intelligent answer. Fast forward to today, they are hacking companies. So I wouldn’t underestimate where we could be heading. There’s still time to get this technology right if we really want to. I piece together AI news, research, and debates in my weekly newsletter. Comment “newsletter” to get it in your inbox every Sunday. #aisafety #airisks #airesearch

You don’t need superintelligence to cause catastrophic harm. According to CNN, an AI-assisted intelligence report falsely identified a Chinese ship’s cargo as nuclear weapons components. The U.S. military prepared to intercept the vessel before officials caught the error. The risk is already here: people trusting false AI outputs when the stakes are enormous. Let me know in the comments if you want me to make a series on AI risks grounded in reality. #ainews #superintelligence #aisafety

How much of your work could you still do if AI disappeared tomorrow? I was building AI products years before ChatGPT. Now I spend most of my working day talking to AI, checking its output, and rarely doing anything completely on my own. That’s started to worry me. So I made a guide for myself: habits to practise independent thinking, prompts that make AI challenge your reasoning, and the books that shaped mine. There’s also a short quiz to help you notice what you’ve been handing over. Comment THINKING and I’ll send you the guide and quiz. #aiautomation #airesearch

One massively underrated flaw with AI agents that CEOs just seem not to get is that they are designed to be personal assistants. They work for one person in isolation. But an actual organization is many people working together. The problem is: there is no company agent. Every deployment is basically one human, one agent, one task —> and the output comes back to that human. This paper looks at what happens when multiple AI agents have to work together instead of solving a task independently. The researchers used Claude agents on coding tasks where information and responsibilities were split across different agents, forcing them to communicate and coordinate to finish the job. And that coordination quickly became expensive. As teams got bigger, agents spent dramatically more time communicating. They started broadcasting information to everyone instead of talking directly to the relevant agent. Giving one agent the role of “coordinator” didn’t reliably make the team work better. Even adding more agents wasn’t necessarily useful: in one 8-agent team, four agents had nothing meaningful to own, yet still generated 62% of the messages. If you’re new here, I’m Sara. I’ve worked in AI for the past 10 years, and I’m currently in San Francisco collaborating with an AI research lab to make AI research easier to understand. I also write a weekly newsletter where I break down the latest AI news and research papers. If you want to check it out, comment NEWSLETTER and I’ll send it to you. Paper: “ When Agents Coordinate: Measuring Coordination in Multi-Agent AI Coding”, Giuseppe Destefanis & Tomaso Aste #airesearch #aiagents

The biggest AI opportunity is everything that comes after the chat. Most companies have figured out how to give employees access to AI. Far fewer have figured out how to embed AI into actual workflows, connect it to their systems and data, and let agents do real work. That gap is where I think a huge part of the next wave of AI adoption will happen. I’ve built my own framework for helping companies move from chat → workflows → agentic AI. Comment WORKFLOW and I’ll send it to you. #aiagents #aiworkflow #aiopportunities