If you took philosophy in college or knew someone who did, you probably remember having conversations along the lines of “what if, you know, everything I observe is just an evil god tricking me?” This is the Cartesian demon1 that has haunted more undergraduates than pop quizzes. But after a few late-night talks, the demon becomes a memory when we, as they say, touch grass.

Rolling hay in Minnesota.

Of course these days we have a Cartesian demon at our beck and call. Mine is called Claude and I’ve been interacting with him fairly regularly for the last few weeks. This makes me a late adopter and all the hot takes have been taken. But that’s ok because we’re going to have a lot of time with our Cartesian demons and their impact on our culture. In spite of it all, time is on our side, if we let it.

Playing Cribbage with my demons

The other day I asked Claude to help me with a Cribbage game I was playing online. Despite not being specifically trained to play the game, he did seem to know the rules well enough and gave competent advice. Or at least he did until my opponent led a king and Claude warned against playing my 9:

Right now the count sits at 10 (from the King). Playing a 9 would bring the count to 19, but that puts a 10 and a 9 back-to-back in the sequence. If the opponent has an 8, they’d play it and score a 3-point run (10-9-8). That’s a real risk you’d rather avoid.2

If you don’t know the rules of Cribbage or only vaguely recall them, that certainly seems like a reasonable precaution. But it’s not and I replied that there was no risk of a run when playing a 9. Claude answered:

You’re right, and thanks for catching that — I mixed up two different things cribbage uses for a card. For pairs and 15s/31, face cards count as 10. But for runs, what matters is rank order (A-2-3-4-5-6-7-8-9-10-J-Q-K), and there a King is rank 13, way off on its own at the top of the sequence. A King and a 9 are four ranks apart, so there’s no way a single card from your opponent could stitch them into a run. I shouldn’t have flagged that as a risk — that reasoning was wrong.3

Nice that it agreed I was right and the updated explanation is mostly accurate.4 So it sure seems like Claude has learned and will be better next time. The promise of a machine learning system is that it can be trained.

But as I played through the game, I kept finding other little misunderstandings confidently proclaimed by my demon. For instance, it told me that it didn’t matter if I played a Jack or not in a situation where it was clearly better to play the Jack. Only by linking to a tip on the American Cribbage Congress website was I able to convince Claude it was mistaken.

Training requires me to already have a solid grasp of Cribbage—its rules and strategy—along with confidence that my demon is wrong. No doubt I could train my demon to be better at Cribbage than me, but it would be a long, frustrating road with no particular destination in the end. I already regularly lose to better human players on the internet.

Avoiding learning JavaScript by letting my demons do the work

Now Claude knows5 plenty of things better than I do. For instance, it’s really good at downloading a GitHub project and getting all the prerequisites in place to product running code. Not long ago I’d find an interesting project and spend half a day discovering it wasn’t useful for me. With Claude’s command line tool, I can cut that to half an hour. Even better, I don’t need to acquire even one more JavaScript factoid.

When I was building my consultant business,6 I built a Discourse plugin that allows anonymous contact forms. I’m sorta proud of how I hijacked Discourse’s incoming email API. Other than that, it’s a plugin written by someone who doesn’t want to learn more about JavaScript than he already does. Unfortunately the JavaScript framework that Discourse uses got upgraded and my plugin stopped working. The official suggestion for plugin authors in my situation is to ask the Discourse AI. It helped, I suppose. Sadly I learned a bit more JavaScript and didn’t help future developers.

I have sneaking suspicion that developers who actually know JavaScript will be horrified by my code just as I’m horrified by Claude’s flailing about when it comes to Cribbage. So I decided to make it a little better by using Claude to add a few tests. That worked great and I didn’t need to add to my meager supply of JavaScript arcana!

Of course, I can’t be sure it’s really testing anything, so asked a different Claude7 to tell me what the first Claude’s tests did. It verified the first Claude wrote good tests and suggested a couple more things to test. Sadly those tests failed in CI and I had to ask yet another Claude to clean up the mess. Having learned my lesson, I’m ignoring further advice from my evil demon to avoid learning JavaScript.

Wasn’t this supposed to have career advice of some kind?

It turns out LLM technology writes decent code. Logically that makes coding a less valued skill in the modern job market. During my forced sabbatical a friend asked for some help building a website. One option was to ask someone on Upwork or Fiverr to build something. I volunteered to build a minimum viable product (MVP), which turned out to be a lot harder than planed. If he were to ask today, I’d recommend he employ an evil demon.

If an eager student asked me about joining the ranks of software developer, I’d start by warning the job will soon be maintaining code written by LLMs for people who don’t know much about coding and don’t care to know more. People, in other words, like me. That might sound ok, but remember the people who use evil demons to code also probably trust them to some degree. The job, even more than it is now, will be to a rock-solid understanding of the craft and to demonstrate you can solve problems effectively.

Since LLM demons know all sorts of arbitrary things and since it’s in their nature to please their humans, virtually every career choice will require demonstrating superior understanding of the craft. Programmers might have it easy compared to people who must employ “soft skills”. Everyone, to a degree, can be a community manager with the help of an AI demon. There is no easy test to prove my approach is more effective than one suggested by a model powered by all the words on the internet.

So my career suggestion for the new world we face is the same as Descartes’. We must expose our opinions to doubt and tear down our own assumptions. And then, with greater understanding of the fundamental principles of the field, rebuild our mental models. And we must be prepared to do that over and over, refining our craft and responding to new input. Then, and only then, will our demons be time-saving devices rather than distractions.

  1. Turns out it’s not called that , but eventually everything René Descartes wrote about acquires the name Cartesian X so I might as well get on the train early.↩︎

  2. Light editing to remove pointless natter.↩︎

  3. Natter left in this time.↩︎

  4. Face cards certainly don’t count as 10 for pairs and the “15s/31” phrasing is screwy.↩︎

  5. I’m not actually taking a stance on whether, philosophically speaking, AIs have the ability to “know” things the way we do. My guess is that machines can get pretty close to replicating our brain function, but don’t have minds. Whether this is relevant to you or not depends on whether you subscribe to Cartesian dualism.↩︎

  6. In other words, I was unemployed.↩︎

  7. This is another difficulty with training my demon. He keeps dying and getting replaced with another demon that forgot everything I told it. Still, it’s handy to get a second opinion sometimes.↩︎