Written by Tim
on 
7 minutes to read
Agents Platforms Harness Attention

Cheap Words, Expensive Attention

When the writer stops doing the work of compression, that work does not disappear.

I love a cryptic crossword. The challenge of deciphering what the setter is hiding in plain sight. A simple clue with a literal and a cryptic part hidden within. The joy of breaking apart words with their many meanings and the rush of solving a tough one can keep me engaged for hours

I also spend a bunch of time on X, maybe too much time. I do tend to find this time valuable in as much as there are often some nuggets of gold buried within my feed. I collect the bits that provoke thought or offer some window into a novel idea. There is a beauty in the medium of twitter that feels incredibly prescient, and increasingly so in tha agentic era.

The beauty, to me, comes in the form of compression. An idea has to fit into a limited character format. The Setter has to squeeze layers of misdirection into a handful of words. The Poster has to remove the cruft and get to the point quickly. The point then stands on it own, it may accumulate likes and comments and follow ups, but the point is exactly that. The point.

Communication

Shannon gave us a way to think about information in a signal and how efficiently it could be transmitted. At Bell Labs, bandwidth was scarce and expensive. Compression mattered. Twitter introduced a bandwidth constraint and made this compression mandatory. Posters were forced to encode the thought into 140 characters. The encoding took time and effort from the poster.

Writing

Writers have always had to take time and effort over their output. There has always been constraint on a writer to distil their thoughts cleanly, clearly and concisely.

Even non writers. Office folk, you and I, have needed to write documents to share ideas. You’ll head to an empty doc, be greeted by a blank screen and then set to work filling the screen with words. If you are lucky, you will start typing and every word will come out in the right order to say exactly what you want to say and no more. If you are like me, this process is a lot more iterative. You’ll write some bullet points, try and flesh them out with a bit of narrative, replace some bullet points, re-write a clunky phrasing. After some time it will broadly say what you want. if you care enough, you start removing things. You notice that two paragraphs are really making the same point. A sentence exists only to get you to the next sentence. A whole section turns out not to matter. The document gets shorter, but the idea gets clearer.

The act of writing forced the Author to confront their own mess. Writing well is a form of compression over thought

Working

In days gone by, if you received a long form document from a colleague that spanned many pages you would have thought that they had laboured long and hard. A ten page document implied that somebody had spent time thinking, drafting, restructuring and editing. It wasn’t proof of quality, but it was at least a reasonable proxy. The writer had paid something to produce it. To a manager, people who could consistently produce hefty documents were regarded as thoughtful, diligent, even as leaders. The manager is satisfied at the evidence of output even if they haven’t got the time to read it

Goodhart says hello

As AI entered the workforce, early adopters could quickly match the output of their more verbose colleagues. An LLM was a hidden superpower that could fill an empty page with confident and approximate verbiage, indistinguishable at a quick glance from the product of hours of work.

What does this do to our word-o-meter when suddenly the proxy of length no longer correlates with effort. Worse even, the proxy of length no longer correlates with thought. Surely we re-calibrate?

Tokens. Tokens Everywhere

We would expect that as the measure becomes detached from the thing it’s measuring that people would consider the measure worthless. The assumption that we are entirely rational is a bold one. The Emperor’s clothes do look lovely don’t they.

We aren’t through it, not yet anyway. We are awash with tokens. I prefer to use the term token here instead of words because whilst they do look like words, the traditional fruit of thought and effort, they no longer carry that guarantee. Tokens spread abundantly over a page to give the impression of thought while separating it from effort.

Effort

Back to Shannon briefly. If the sender does less work to compress the information, then surely the receiver has more signal to process to recover the same information?

Exactly. Effort has been moved in the system from the writer to the reader. A writer’s thought has gone from their head, into an LLM that will happily construct a 10 pager around it. The thought - as valid as it may be gets hidden. The task of recovery now falls on every reader. The tokens may be nearly free for the writer, but they are not free for the reader. A human still encounters them as words. They still consume attention one word and sentence at a time.

I saw this on X recently, it looks cool, and I appreciate the thought that has gone into it.

But it goes to highlight the point - If the summary contains everything most readers needed, perhaps the summary was the thing worth publishing.

Prose is not only the pro’s output

This pattern is not unique to prose. Many of the words that appear on my screen come in the form of code. Code by its very nature tends to be terse. The main consumer of code is a computer that can turn the words into magic. Coding languages have always tended to prefer brevity, initially as a space saving endeavour, but over time as an understandability, safety and explicitness endeavour. Good code should be simple enough for a human reader to read and reason over too.

Writing code has for the longest time felt like only 10% of the job. Good software engineering has always involved compressing a large problem space into a small implementation. Requirements, edge cases, architecture, trade-offs, constraints, conversations: all of that thought gets compressed into relatively little code. In that sense, code is another residue of thought.

Perhaps this is why Twitter feels so noisy at the moment. When code is cheap, activity is cheap too. Repositories, frameworks and abstractions balloon before we have established whether they should even exist. So many agent harnesses, eval maxxing hoping for a problem to solve..

Another harness
We’ve enough bloody shovels
Go looking for gold