Weekend Reflections #12 | The Speed Collapse

[Views are my own]

I have been testing local LLMs for my personal assistant project.

I started with a 20B model. Too slow on my hardware. Smaller models were faster, and for my use case the difference was acceptable. I am not trying to build a reasoning engine. I want something that can synthesize notes, cluster ideas, track my agenda, and run research tasks while I am busy.

Working at my speed, or slightly ahead of it. That would be more than fine.

And then I stopped and asked myself: "how much do I need it faster?"

At the slower pace, something felt different. I could follow what was happening. I could disagree with a step. I could question an output. The pace was human-sustainable, and the relationship between me and the work was still intact.

Speed did not only improve the experience. Past a certain point, it weakened my ability to stay meaningfully involved.

I started calling this Speed Collapse: the point where AI moves faster than humans can meaningfully verify, absorb, or question what it produces.

Speed Collapse happens when the production loop has higher throughput than the verification loop.

This is different from the Speed of Doubt I wrote about in an earlier reflection.

Speed of Doubt is a perception problem: the work feels untrustworthy because it arrived too fast. Speed Collapse is structural: the human system no longer has the throughput to verify it, regardless of how it feels.

Set aside the question of whether humans should stay in the loop. The harder question is whether, past a certain speed, meaningful presence is structurally possible at all.

If an AI produces what looks like six months of work in one week, how long does it take a human to verify that work? The check either becomes a rubber stamp – technically present, meaningfully absent – or gets delegated to another AI. Unless the review is grounded in independent evidence, tests, logs, or escalation rules, the system starts producing the work and certifying its own quality.

You can draw the governance diagram correctly and still produce a fiction, because the rhythm at which work arrives makes the review a rubber stamp before anyone has lifted a pen.

That is not just a governance problem. It is a human throughput problem.

AI agents can produce dozens of product requirement updates, customer insights, and roadmap recommendations overnight. The PM may still be in the loop, but the loop only matters if they can understand the assumptions, compare them with evidence, and stop the wrong ones before they become direction.

But there is a softer version that is more common and more dangerous. The human is still in the loop. The inspection just becomes superficial because the volume makes depth impossible. Review becomes a checkbox, not a judgment.

A useful test is not whether a human is still assigned to review the work. It is whether they can see what changed, inspect the assumptions, stop the flow, and reverse the decision. Without those four, the loop is still there on paper. It is latency disguised as control.

Speed Collapse does not require full delegation. It only requires that the human presence stops being meaningful.

The same dynamic works in both directions.

For the producer, the team shipping AI-assisted changes, the pace can exceed their own capacity to understand what they are shipping. For the consumer, the user absorbing those changes, the pace can exceed their capacity to internalize them. A change that arrives before the previous one was understood is not really adopted. It is just displaced by the next one.

At our current speed, that is basically every day. In many AI products, this is already the default rhythm.

I am not arguing that AI should be slower. The value is exactly that it can work while I am doing something else, faster, in parallel, handling what I would handle if I had more hours. That is the right relationship.

The design question is placement: where does human judgment need to enter, and in what form, so that presence is structural rather than ceremonial?

Speed Collapse measures not the rate of production, but whether the human on the other end remains capable of meaningful presence. Change without internalization does not compound. It resets.