TIL

Through a post in the SAIN WhatsApp group, I came across Seat at the Table - a short fiction film about the AI industry. At only 16 minutes, I think it’s certainly worth a watch.

The target audience is people who are fairly unfamiliar with AI and AI control. Without getting technical, it conveys the ideas that:

• AI is grown, not built
• AI is fascinating!
• AI can be deceptive
• The people making AI are scared of it
• The people making AI are trapped in race dynamics
• AI’s getting smarter, fast

These views are pretty well-accepted in my bubble, but I’ve struggled to have a conversation about this with people who don’t work in this (AI safety) field without sounding doom-y. The term “existential risk” definitely doesn’t help there.

I thought this film gets these concepts across in a very accessible way, while being extremely well-made and engaging to watch.

Europe 2031 is a fictional story that follows Caroline, a French policy worker in Brussels, and Christian, a German AI founder who’s just relocated to Silicon Valley over a period of five years. It starts grounded in real events from early 2025, but then turns speculative: Europe underestimates how fast AI is moving, sovereignty policies backfire, the US starts rationing compute access, European companies get acquired, unemployment rises, the euro destabilises. By 2031, Europe is left choosing between three bad options: becoming an American protectorate, falling under Chinese influence, or isolating itself entirely.

You can tell it was heavily inspired by AI 2027. It’s clearly opinionated (some of the creators were also behind the Dutch National AI Plan), but sketches a realistic possible scenario.

Interestingly, one of its central threads is the US restricting access to frontier AI for Europe, and it outlines in detail what that would mean in terms of economic decline, loss of sovereignty, and dependence. This was released on June 11th. Then, one day later on June 12th,the US government banned Fable 5 for non-US nationals.

Most conversations in AI safety, also from a European perspective, are about making sure AI does what we want it to do. And that’s hugely important.

But even if AI is perfectly controlled and aligned, the question remains: controlled and aligned by whom? Regardless of whom, we might be looking at an enormous concentration of power and wealth in the hands of whoever does control it, which leads to heightened inequality.

Sometimes things break in my terminal. Maybe it’s a training run that crashes overnight with an out-of-memory error, or a Docker build fails on some step I didn’t write and don’t understand. It doesn’t matter, sometimes things simply break. It happens.

Back in the old days (say, two years ago…), my go-to was pasting the error into Google and crossing my fingers that someone on StackOverflow had hit the same wall. Sometimes you’d get an exact match. More often the art was in stitching together half-relevant posts and trying them all.

Enter the age of AI.

Now, I copy the error into an LLM instead. I glance over its – let’s be honest, usually quite verbose – analysis, and paste the suggested fix back into my terminal. I run it. Another error, a new one this time. So there we go again. Another copy-paste. Another analysis. Another copy-paste. And the loop continues.

It works, generally. But it’s also quite tedious. And I feel like a messenger pigeon, ferrying text between two windows.

So lately I tend to go the other way. I avoid the terminal altogether and tell an agent to kick off whatever I need to run, fix whatever breaks, and fingers crossed hope for the best. Faster, sure. Less tedious, sure. But I don’t love it, it leaves me a bit uneasy. The agent is executing commands I’ve never read or approved, and I’ve quietly stopped feeling in control of my own machine.

Then I heard about ShellSage.

It’s an AI-powered command-line assistant that reads the terminal I’m already in. I can use my terminal as normal, but when I get stuck, I can invoke ssage and ask what’s wrong, right where it broke. It answers with the command I probably need, and one shortcut drops that command straight onto my prompt. From there it’s up to me. I can hit enter and run it. Or I can tweak a flag first, change a path, fix the bit it got slightly wrong.

Nothing runs unless I run it. But I do get context-aware support, right where and when I need it.

I haven’t had the chance to try it yet, but I heard about it from my ex-colleague Rens Dimmendaal (who is now over at Answer.ai), who’s always a great resource for these kinds of things. I’m excited to try it out!

One of the first things I set up on a new Mac is the Hyper Key: remap Caps Lock so a quick tap is Escape, and holding it acts as all four modifiers at once (⌃⌥⌘⇧). Almost no app uses that combo, so you get a whole layer of global shortcuts that won’t fight with anything else. And a much comfier Escape for your pinky.

I set mine up through Raycast (Settings → Advanced → Hyper Key). Before that I used Karabiner-Elements — both work fine; Raycast just meant one less thing running.

My daily bindings:

• Hyper + Space → Raycast
• Hyper + K → Slack
• Hyper + P → Cursor
• Hyper + B → Arc
• Hyper + T → iTerm

Muscle memory now. I’m not a very shout-y person anyway, so the prime spot on my keyboard once taken up by Caps Lock is finally useful.