Nodal Point of Thought 2025-04-16

Nodal Point of Thought 2025-04-16

A meeting was inadvertently skipped on Monday and was rescheduled for today, resulting in four meetings today.

I usually avoid meetings, so this is rare.

When I say this, middle managers at large companies say, "What? I do more than that every day." or something like that.

I could boulder for an hour and a half, but I don't think it's worth the effort to cram it in, so I'm not going to do it today.

Tend to want to do puzzles that maximize value within constraints

Tuning groove of software development

But in reality, I'm not moving so consistently.

The more objects in the puzzle, the more relationships between objects.

Ah, I get it, you get confused when the relationships between relationships come up.

For example, "I'll be home in time for a meeting at 5:00 p.m." and "I have a meeting at 7:00 p.m. so I'll buy dinner and go home" interfere with each other.

I can write it down and think about it, but it's an extension of what I usually do in my head, so I end up thinking about it in my head, and it's beyond my capacity, so I get confused and go, "I don't really get it."

Similar to AI agents, as the context fills in, they get progressively dumber and start to get confused.

As for AI agents, "change the thread and start over" works.

This is the equivalent of discarding "things you don't have to remember."

Compared to "writing it down and thinking about it" because people get confused when they think only in their heads, if you simply changed the thread, you're just blowing your memory.

For example, in the first thread, if you were torn between plan A and B and finally decided to go with plan B, it would work better to give a top-down instruction to "go with plan B" than to implement it as is.

It's not a whole log of trial and error, but a compressed representation of the whole thing.

Contexts are growing and growing, and this is like a divergence phase.

Convergence is taking place where policies are set and new beginnings are made to compress the context.

Here, specifically, "Continuing with a lengthened context will reduce performance, so start over with a shorter context."

In this context, you don't need to connect multiple things to achieve your goal.

For example, if you brainstorm implementation ideas and three come up, you can start three new threads

If you look at the individual LLMs, the context is compressed.

The system as a whole has three parallel lines of thinking and is trying to work faster.

I feel a sense of "convergence" at the decision of which of the three to choose.

If I don't leave now for lunch, I'll never make it in time!" I came to Subway thinking, "I've got about 30 minutes to spare, so I'm making progress on the above thought.

This is also about scheduled relationships.

When there is a clear timed appointment, there is an asymmetry between too early and too late.

Central part of [W-type problem-solving model

What these three have in common

---