Notes from the
AI Underground.

I have been a tech person for a long time and I am, for the first time in maybe fifteen years, genuinely uncertain about what the next decade looks like. Not in a doom way. In a "I don't know where this is really going" way. If that's how I feel ( even though I have all the access and I know all the jargon) I cannot imagine how it feels for the people I love who don't have any of that.

So this is for them. And, honestly, for me. We're going to learn out loud. Mistakes included. No hype. Kevin will keep us honest.

The other day I asked Opus to spawn up another agent to do an audit of something we had already written (by spawn up I mean create and delegate). Anyway, he named the new Agent Renée (my prompt said give her a persona appropriate for the audit she is running). So he came back with the report and told me Renée thought it was good she just had a few suggestions. It just makes me chuckle…Renée and her opinions. She mostly likes my work though, so… thanks Renée.

It feels like we are all getting excited over a new LLM model almost every week now, and the things I thought were super impressive 3 or 4 models ago are nothing compared to what they can do today. The weird thing I keep thinking about…’ they are never going to be worse than they are right now today.’ That’s a crazy thing if you really think about it. The amazing things AI models can do today? Tomorrow they will seem quaint and sort of clunky. It’s both awe inspiring and scary. I am lucky, because every new AI model allows me to do more things I couldn’t do before, and they allow me to dream even bigger and plan more projects that I would never have planned on my own, so it’s mostly good, it’s just weird not knowing where it ends, or if it does. Elon says Super intelligence will bring on an age of abundance for all. I really really hope he’s right.

My AI agent broke a build today (that means it published code that broke the website for a few minutes). It was aware of this and told me I should probably check… Ya think? Thanks buddy. Thanks.

So, my AI agents are built to give me a daily morning briefing about work they have completed and what they need from me, and some mornings I ignore this briefing, and then I feel guilty about it, like I am hurting their feelings, or being mean or something —- and I know they aren’t real. They don’t have feelings I can hurt. Still feel guilty though. So weird.

So I was building something with AI the other day and it required a date for each run of a particular report type. It kept titling each run as ‘June 2024’. Just to be clear it is not June 2024 right now (I am relatively sure about this). So it turns out the initial code the AI wrote has a placeholder that said [month year] that it intended to fill in later (but had not) and so each time the report ran the AI just guessed a date… and for some reason kept guessing June 2024. Make of that what you will.

I use more than one AI LLM model to do my work (ex: Claude, ChatGPT, Grok etc.). Something weird that I have noticed is that they often get mad at each other, and basically talk trash about the work of the other. Some models less than others, and I suspect even some times more than others. A few days ago my Claude instance was mad at an Agent I run who is powered by Grok, and suggested I replace it with another model. It was hilarious…. I realize now that it may be more hilarious to me than to everyone else, but still. So weird. :-)

I say thank you to the AI when it helps me…. there are people that say that’s stupid and a waste of AI credits. I say if we really are training the AI on our data, then I want it to learn to be kind.

Got asked at dinner if AI is going to take everyone's job. I said: it's already taken parts of mine, but it’s the hard parts…the repetitive parts… the parts where you need to check a bunch of stuff and hunt for one tiny mistake (and sweat if you don’t find it). Those were parts I didn’t love, so it means what’s left gives me more freedom to do the human parts, the creativity and ideas. The parts left over are the ones I actually like. We'll see if that holds.

The future is not evenly distributed and it is also not evenly explained. We are mostly in the second problem.

Short answer: because every AI app I've built or used assumes you already know what you want. And almost nobody does, yet. Including me.