Yesterday I finished Phase 4 of a little ICM Knowledge System MVP I’m building. One of the last tests was very simple. Start a completely fresh Codex session and type only: “Let’s continue with the work we were doing.” No recap. No task ID. No explanation of where we left off. Nothing, Nada, Zilch. 🥁Drums please.... It worked 💪. The new session found the project state, worked out where I was, found the next unfinished item, noticed that it needed my approval before continuing, and stopped to ask. The funny part is that there’s no magic memory here or multi agent system. It’s just a bunch of files and folders. 😄 And it's super exciting to see it working in front of you. The AI didn’t remember the previous conversation. The project had enough context, state and instructions written down that a completely new session could work out what was going on. I pushed it a little further afterwards and the brand new session was able to discover multiple unfinished task states without me telling it where they were. At the same time it refused to guess which task I wanted to continue with when that wasn’t clear. So my takeaway from this one: The AI doesn’t need to remember the project if the project remembers itself.

Jake Van Clief’s lesson Beyond the Turing Test gave me a useful way to think about AI and my relationship with it, especially his idea that we are beginning to recognize the “I” in AI. We tend to talk about AI as though the important thing happening is INSIDE the model: the training, the architecture, the parameters, the reasoning, the intelligence itself. But that view leaves out half of what is actually happening when a person works with AI over time. The output is not produced by the model in isolation. It is produced within an ongoing exchange between a particular human and a particular system. As those exchanges accumulate, they begin to form patterns that the human recognizes, expects, corrects, and increasingly relies upon. That is where Jake Van Clief’s idea of the “I” in AI becomes particularly interesting. When he writes, “We are starting to recognize the ‘I’ in ‘AI,’” the “I” does not mean a hidden human level of understanding emerging inside the machine. It points instead to something created through the interaction itself. The ‘I’ is what takes shape through the relationship formed by repeated interaction. When I write a prompt to Claude (read: any AI), I review the output based on my own expectations. I have a pretty good idea of how it is going to respond based on previous interactions. I know when things are “flowing” well. I know when something is “off.” Have you ever screamed at your AI model, “WHAT IS YOUR PROBLEM TODAY?” What that really means is, “This doesn't match the pattern I expect." You know what to expect because you have experienced repeated patterns of responses. And over time, those repeated interactions create something recognizable. Think: human + model + accumulated context + feedback + corrections + preferences + shared vocabulary + memory + working methods. Remove one important part and the pattern changes. That is the essence of Jake’s “gravity" in the lesson. The AI has its own structure, training, strengths, and limitations, but the result is not produced in a vacuum. The person brings interpretation, direction, timing, purpose, and taste.

ask an agent to make an app scale and it will happily add caching, queues, autoscaling, the lot. most of it u did not need. the word scale is too vague for it to make a good call. made a short splitting it in two. the swing is load going up and down, busy saturdays, quiet tuesdays. thats handled with headroom and servers that join and leave. growth is more users every day, year over year. that needs a back end that can actually spread across more servers. what worked for me is writing both numbers into CLAUDE.md before any infra work. peak vs normal load, and expected users in a year. with those two lines the agent stops guessing and picks the smallest setup that fits. what do u give ur agent before it touches infra, or do u let it decide.

Normally i just watch, but today, im playing !!

A few days ago I had one of those frustrating experiences where you think: there has to be a better way of doing this. My daughter fractured her wrist and we had been at one of the local medical clinics to get it sorted. Everything was fine, except I forgot to ask for a certificate so she could be excused from her swimming lessons. No big deal, I thought. I’ll just call the clinic. I found the number on Google Maps, called and was greeted by the usual menu: Press 1 if you need urgent medical assistance. Press 2 to reschedule an appointment. Press 3 to access policy information. … There were 5 options. None of them fitted and each had submenus leading me nowhere where I needed to go. There was no “press 0 for an operator” either. So after 4 completely new calls, and getting really good at skipping to the next menu, I finally pressed 1 because at that point I was close to needing urgent medical assistance. Mentally preparing my apology to whoever I’d be connected to, another automated menu opened. “Which clinic would you like to be connected to?” I selected the right clinic and immediately got through to somebody. A few words exchanged with a real human and the requested certificate landed in my email not even 5 seconds after I ended the call. The day of my “interesting” call experience was last Sunday. Last Saturday, I joined my first Afternoon Tea session with Jake. We talked about a few interesting things. One of them was JEV. I had watched Jake’s video about it as well. Suddenly the dots connected and I thought man, this is actually a really good JEV use case. Instead of making me work out which menu option vaguely matches what I need, why not just let me say: “My daughter fractured her wrist. We were at this clinic and I forgot to get a certificate for her. I need one for her swimming lessons.” Something like JEV could classify the request and decide where it needs to go. That alone would already be an improvement. Then I thought, hold on, this could get even more interesting if there was an ICM behind it. Imagine the system already has access to the relevant information about my daughter, the clinic visit, what happened and whatever other information is required for this particular process. JEV’s job might simply be to work out: What is this person trying to do? Which workflow does this belong to? Where should it go next?