Size:

Human Taste Became Invisible.

We are in the age of slop. Generating sites is basically free but the generated ones drift toward lookalikes. Making a site no longer conveys your skills.

Craftsmanship has been covered under loads of stochastic rumbling. Even the truly skilled sacrifice their skills to fancy autocompletion. They direct the model, they know what to cut out. They allow the eloquent electricity to save them time. The problem is when we want to distinguish the deliberate from random: Did the author mean it, or was it luck?

We sold skills to LLMs. And we lost taste.

Before, developers bragged about scale and beauty, and they had a reason to. You couldn't just stick together stack overflow posts and hope they worked well without a second thought. You needed the glue that was years of experience. Skill was the most visible trait.

Taste lurked along with it. And they were trained together, when you practice your vocabulary, you learn what sounds good. You start to get that feeling. Which is exactly taste. Before, you could expect taste from someone with skills, because both were the outcome of practice.

Then the probabilistic casino came in, and the skills are no longer necessary, you can just prompt claude, and then iterate until satisfaction. This created a world where we have beautiful sites, with human and machine choices blended together. When looking at the web, you just see yet another site that looks like another, and you can only wonder: "Was this deliberate, or slop?".

Previously, this was easy to answer, you just looked at the site, you saw it was clean, loaded fast, had some interesting quirks and clever mechanisms. Now? Now everyone can add them and the decision "Is it slop" can only be made retroactively after reading the whole thing.

Constraints surface taste.

Models "think" by writing, so their output drifts verbose. They are trained on the internet, so they are a distilled version of it. Distilled out of probabilities. If something occurred often, it will be generated often.

By pushing this to the extreme, if we find a field where everything is tailored, and the iterations are rarely public so the training data is small. And the data it is left with is only the end product. It does not have access to given up features, thinking process of the reducer and the tricks he used to reach that size. We can have a pretty hard time using our thinking sand to get meaningful answers.

If we add limited size on top, we are faced with a really hard problem for the probabilistic casino. Because now, not only does it work in unfamiliar territory, but also it must do the exact opposite of its nature - meaningfully reduce.

Meaningful reduction requires taste, and the taste of the machine drifts towards typicality. Giving only 1024 bytes over the wire to create a website, and it faces a really hard challenge, you can try it right now. Tell it to create a minimal website that is under 1024 bytes, has actual content, and then try tailoring it to your needs.

A 3MB site can be spewed out at once, with a single prompt. A 1kB one has to be carefully prepared, because you can only fit so much. Say, if we add this feature, the other won't fit, which one do we value more? This is even more exaggerated when we measure compressed size. With brotli for example. Now we have a real challenge, because we can add a feature that makes other features basically free. That feature gets backreferenced and in reality costs only 10B while being over 50 chars in length.

Example being: My card tilts towards your cursor, the cat on the photo tilts away from it. The code looks long, but adding the second tilt only cost a few extra bytes, because it backreferenced. Making the choices interact in non-obvious ways.

Iteration is cheap, Choice is yours.

The limitations make this a hard problem, because it is branching. And each branch choice has cost in possibilities. Skills are exposed here, because that branching forces intuition, and sentient stones will either drift toward the median, or ask the operator for choice.

And wrong choice forces reversion and restart, because each branch is unique, and going into one costs the other. Constrained size makes the amount of paths finite, making the search meaningful and choices unforgiving.

The choice between paths distills taste. Talking rocks can generate you paths. But removal of what is un-important for the person is the depiction of their taste. And the limits force choice at every stage.

So LLMs bury skill and taste. Limits dig them up again. And the site is signed by the builder.

To see other cool sites, check 1kb.club or js1k.com