ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News3 min read·2026-10-19

Former ML Engineer Samuel Verboomen: Show Friends What AI Can Do First, Then Talk About Risks

Samuel Verboomen spent nearly a decade in machine learning before transitioning to AI safety. He found that the core problem for non-technical people is not a lack of information but a mental picture of AI capabilities still stuck in 2022. He designed a two-hour interactive presentation that uses a capability showcase as bait, progressively introducing risk discussion across five parts, and validated a key design choice at his first session with six friends: letting attendees first operate a nearly useless base model, then switch to an assistant model — the contrast made 'what predicting the next word can achieve' very concrete.

ImportanceMaterialEvidenceE3 inspectableWrite-upStandard

2026-10-19 08:53 GMT+8

Samuel Verboomen worked as a machine learning engineer for almost ten years, holding a CTO role at startups. Six months ago he quit to transition into AI safety and launched his own non-profit. But the LessWrong post he wrote is not really about his own journey; it is about a problem he has repeatedly observed over the past half year: people outside the AI field are not indifferent to AI — they simply do not know what AI can already do.

Chatting with his stepmother, he noticed that for her AI is still basically a chatbot: you ask it a question, copy-paste the answer, and use it to draft an email. He cites US polls showing many people say they are worried about AI, but in his experience few understand why it could actually be dangerous. That cognitive gap led him to conclude that talking about risks directly is ineffective: it is hard to make someone fear something they do not yet grasp.

So he designed a two-hour interactive presentation structured in five parts: first show what AI can already do today, then cover AI history and how it works, then introduce who is building these systems and why nobody is willing to slow down, then move into risks — starting with the easiest to understand, misuse, through systemic effects, and ending with what sounds most like science fiction: loss of control. His own words were 'I wanted to trick my non-tech friends into learning about AI safety' — inviting them to learn 'how to use AI better,' but the real goal was to progressively bring them face to face with the risks.

Sign in / Sign up
ReadingResearchRadarInvestment framework

At the first session he invited six friends. These were engineers, people in sales and finance — not exactly disconnected from technology. He says it went well, but the most interesting part was what genuinely surprised them: first operating a base model that is mostly unhelpful, then switching to an assistant model that can genuinely change your daily work. That jump from 'useless' to 'useful' made 'what predicting the next word at scale can produce' very concrete.

Verboomen is upfront that this is one person, one session, six friends' worth of experience — he calls it a 'personal impression.' The presentation is open-source on GitHub, with a storytelling script and presenter-mode notes, so anyone can deliver it, take only the parts they find useful, or fork it and make their own version. He puts the underlying logic simply: six months ago, despite nearly a decade in ML, he had no idea about most of these risks. All it took was a friend who spent time explaining them.

He leaves readers with open questions at the end: Should AI safety be everyone's concern, or is it too technical and better left to experts and governance professionals? Could mass involvement in the conversation itself be dangerous? Or is it the only way to pressure AI companies into slowing down?

Sources

  • 原始出处lesswrong.com↗
Continue researching →