With a title like “Good Vibrations”, you’re probably thinking I’ll talk about the Beach Boys, but no.

Instead, I want to go back a bit further, and talk about jazz legend Lionel Hampton.

Hampton started his career as a drummer, and combined brilliant technique with wonderful showmanship. An excellent example can be found here, where he demonstrates his ability, while also juggling his drumsticks and showing his simple joy in performing.

As a performer and band-leader, Hampton worked with many of the most famous musicians of the twentieth century, including Louis Armstrong, Benny Goodman, Buddy Rich, Charlie Parker, Quincy Jones, and dozens of others.

Hampton was probably best known as a vibraphonist. Also known as the vibraharp, and with apologies to musical pedants who will note that it’s a metallophone that resembles the steel marimba, it’s probably easiest to describe it as an electric xylophone with metal keys.

There’s an interesting video describing the vibraphone and showing how it works, but to summarize, there are three main features which differentiate the vibraphone from similar instruments. The metal keys can “ring” (like a bell) for quite a long time, the “sustain” pedal (similar to a piano) controls a bar which allows the performer to control whether the keys are allowed to “ring” or not, and then there’s the “resonator”.

Under the vibraphone’s keys, there are “resonator tubes” which amplify the sound, and a resonator, which controls the speed at which a set of small, metal discs rotate to cover and uncover these cylinders. This creates a tremolo effect (ie, a “trembling”, caused by the variation in volume), which can be controlled by the performer.

And that is how we get back to Kanban.

I recently mentioned that I was interested in learning more about Rust, and decided that I wanted to create a new version of my prior Kanban board.

In the course of all of this, I’ve been using AI, both to try and accelerate my Rust learning-process, and also to learn more about how AI will affect programming in future.

In contrast to “vibe coding”, however, I’ve been building the code myself, and using AI for troubleshooting, helping with things I don’t yet understand, and trying to get an idea of current best-practices.

It’s been quite interesting, and I’ve learned a lot about both the strengths and weaknesses of AI. It seems to depend highly on what is being asked, and on how much you want to share with the AI service. If you are trying to troubleshoot a block of code and don’t understand the compiler error, AI can easily identify the stupid typo you keep missing – this should not be trivialized, though, as it can be an enormous time-saver.

In contrast, if you are trying to resolve a complex set of web callbacks, AI will happily lead you down a deep rabbit-hole that leaves you oscillating between two wildly different design approaches – The first time this happened, I wasted a significant amount of time going in circles, before I realized what was happening. The next time, I caught it early and avoided the trap.

For a beginning programmer, expect to waste a lot of time, treat every response with distrust, and ask a lot of questions. And be careful – an AI won’t hesitate to confidently tell you to do something that any experienced programmer will tell you is stupid.

If you have prior programming experience, I think that AI can be a very useful tool for learning a new language. Always ask for clarifications, ask it to compare its approach with one that you would use in another language, and ask follow-up questions. You’ll still waste time, but progressively less as you learn more about the language.

For experts, I think AI can be extremely helpful, but be careful what you ask it. Ask it to run tests, ask it to adapt existing code to a new situation, ask it for documentation and test cases, but if you’re looking for it to design large, complex systems, don’t expect elegance or efficiency. In fact, expect it to take longer to get the AI to refine its design to something useful than it would for you to design it yourself.

And if the AI keeps cycling back to one point, even after multiple attempts to redirect it, one option is to simply reset the session.

An AI will generally address the surface issue – give it an error message, and it will prevent the error. But will it actually fix the problem? Steve Gibson described the adventures of Microsoft developer Stephen Toub in episodes 1027 and 1028 of Security Now, and how Copilot, even with detailed prompts pointing in the right direction, “solved” an issue by essentially sweeping it under the rug.

Still, I’m learning a lot about Rust, and spending a lot of time going down (interesting) rabbit-holes. While this new version may look similar to Kanban 1.0, I’ve implemented some fairly sophisticated error-handling, visual themes (eg, colours, spacing, postit sizing), a settings page, session variables, the ability to easily show/hide columns, and a bunch of other stuff that was not even on the radar in V1.

With regard to AI, take it for what it is – a great way to search for options and ideas, a great way to help troubleshoot well-defined issues, and a great way to generate a lot of code.

... So, a big improvement?

Well, yes, but I didn’t say “a lot of GOOD code”.

That said, and setting aside power consumption and other infrastructure questions, I see two key drivers for future AI LLMs (Large Language Models). The first is the quality of the training data – Garbage In, Garbage Out, and the balance between the quality and the quantity. The second is some method of defining “confidence” – some sort of Bayesian evaluation that will test the consistency of an item and compare it with reality.

And in this current world, we really need reality!

Cheers!