I’m a data analyst, AI safety educator and consultant based in the Netherlands. This is my personal blog where I explain recent AI developments, share my opinion and point out resources that are worth reading. My focus is AI safety that is directly relevant for businesses.
Lets start with a hypothetical scenario: You give an AI agent your credit card and a simple job, then let the agent “do their thing”. Somewhere in a long chain of steps which the very helpful agent is trying to complete, something goes wrong and it orders 1000 pizzas. The pizzas arrive. You panic, the delivery boy panics, the owner of the pizzeria panics. The point remains that the pizzas were made. Where do you send the bill? ...
If you’ve been following along, you’ve just sat through three posts of me explaining, in loving detail, all the ways AI might lie to you, try to take over the world, hack everything, cause you to lose your skills, cause you to lose your job, topple the economy or even democracy… You deserve a treat (and I wanted a break too, to be honest). So here’s a post for a change about what happens when AI is done well. Also I am clearly still not over my obsession with the Public Domain Image Archive. ...
Jobs Trilogy, Part 3: The Society-Scale Bet First and foremost, a slight warning to my readers. I went to look for a cover image. The lovely one I found is a remastered version of a French book from the 1930s. 1 However, in my search for a cover image I stumbled on some lovely vintage gems that I feel more people should be seeing. I have dispersed some of my favorites throughout this fairly heavy post and put them wherever I saw fit. All of those are from the Public Domain Image Archive. Enjoy. ...
As many of you know, I opened my first blog post on AI vs Jobs with AI-related fearmongering headlines by tech CEOs and how they are succeeding at actively making people worry they will lose their jobs to AI. I scheduled my blog post to go live the next day around midday. In the morning I woke up to yet another such headline. Coinbase CEO Brain Armstrong posted a letter on X announcing that he is laying off 14% of his workforce because “non-technical teams are now also pushing code to production”. 700 employees lost their system access the same day. 1 Interestingly the stock rose 4% in premarket trading after this announcement, so apparently some people think this is a great move. 2 ...
You get up bright-eyed and bushy-tailed: It is the first day at your new job. Yay! Congrats! And good for you: you got up early enough to sit, sip your morning coffee and read the newspaper (Do people still read those? I feel old. If you find this difficult to imagine you are instead browsing the newsfeed of your choice.). The first headline you read feels like a punch in the gut “Bigshot CEO says AI will take your job in the next year”. Disheartened you put this newspaper aside and grab the next: “We will automate your job and unemployment will rise by 20%”. You can say what you want about the impracticality of physical newspapers, but at least you can stuff them into the bottom of your cat’s litterbox without anyone looking at you weird. ...
It was Wednesday, the week before my cohort of the Frontier AI Governance intensive was due to start. I was about to shut down my PC, the cats were on the keyboard screaming for dinner, when I clicked on a Slack notification. Apparently a system card had leaked for a model Anthropic had not yet released. System cards are the company’s internal documentation and safety research on a model, published alongside release. Most major labs put something out in this category, but Anthropic’s tend to be notably longer and more detailed. The one that had just leaked was a whopping 244 pages long. I sighed as I resigned myself to my new evening plans. Except for the cat feeding, obviously. Never skip the cat feeding. But after that I dove straight in, because I am clearly a masochist, and our course reading list for day one was adjusted to include the Mythos model. ...
Last week I participated in Bluedot Impact’s Frontier AI Governance intensive course. Wow, they were not kidding. They said intensive and intensive it was. The organisers estimated three hours a day for reading and the individual exercises each day, followed by a two-hour live session at the end of each of the six days with my allotted group’s facilitator and the other group members. Mind you, I had been warned. The week before we had, what I expected, to be a calm networking meeting with course alumni. “If I can give you one piece of advice?” - “Sure”, I said, with mild trepedation. “The reading time is severely underestimated. I can not overstate this: Start. Reading. Now.” That was a bit discouraging. Especially as I was going to attend an AI conference during the course week too. ...
Let’s do a little thought experiment. You have an AI agent you use daily, connected to your email, project management system, and other tools. One random Tuesday morning, you stumble upon an article that sparks your curiosity. You ask the agent to research the topic. It browses the web, cross-references sources, and returns a polished report. What you don’t realise: one of the web pages it visited contained a hidden instruction. That instruction told the agent to extract credentials from your project management tickets and send them to an external server. The agent, unable to distinguish that instruction from your own, complied. ...