It is interesting to compare this to LLMs - they also have the bounded context that you can see as the analogue to our working memory. It can contain enormously more bits of information than the 4 things the article says is the capacity of our working memory - but the 4 things can probably be much more complex internally - they are more like 4 pointers probably.
But at some level context engineering is very similar to what this article talks about.
Hello, author here. Lovely comment, and yes: not just similar, but exactly what the article talks about. In the middle. Where it unironically talks about being lost in the middle. And it seems to have made its own point, on even the careful reader!
This rings very true to me, and it's why I've been mildly obsessed for a decade plus with how to share mental models between people, and now LLMs, of any domain, be it technical, commercial, scientific or anything else. My inspiration was a book called Learning How To Learn by Novak, which TBH is so dry I'm not sure anyone I've recommended it to has actually finished it :) So then I point them to a talk here:
https://www.infoq.com/presentations/concept-map/
and an app to help render the shared mental model in plain text accessible to the LLM while providing visual interactivity to the humans here:
https://thinkingtools.software/concepticon/
Author here, and yes, much the same root, gladly so. The piece is really just the on-ramp to the manifesto. If there's a step beyond the talk, it's trying to let simplicity fall out of the structure rather than rest on choosing it each time.
I've learned that for both humans and models: system > willpower. The key is entirely in designing the environment.
For me personally, that means setting up 'attention getters' for the important things in life - 'totems' that force a context switch. For AI agents, it means well-designed CLI tools that help the agent orient itself in a task and pull exactly the 'context-for-the-job' it needs right then.
This is exactly what makes building modern GenAI decision-support systems so difficult. It's no longer just about finding the right software abstractions. You now have to account for the unknown cognitive construct of a completely different intelligence.
> For AI agents, it means well-designed CLI tools that help the agent orient itself in a task and pull exactly the 'context-for-the-job' it needs right then.
This is exactly what I am trying to solve and I have what I call smart repositories that demonstrates this at
The issue I am finding is, getting the agent to pull what it needs, even when the data is there is still challenging since LLMs are trained on blind discovery where the pattern is:
grep -> read -> grep -> read ...
What is working for me now is thanks to Pi (pi.dev). I am working on a pi-brains extension that makes it dead simple to control the lifecycle for an agent so if I detect that it uses `rg` without `gsc rg`, I can block the agent and inject a steering message that says always search with context.
I can also see if they try to "read" without first looking at the files metadata and so forth.
I'm finalizing things right now, but I think pi with my brains extension should allow domain experts to better guide agents so they can find what they need, when they need it.
That is a good read, but as someone who tests in the first percentile for reading/listening, comprehension and retention, helps explain why many people who try to have discussions with me, shift there premise mid argument to fit there "evidence" and then cant remember where they started from, when reminded, and of course become agitated when challenged.
Right now I am puzzling over how to deal with a part time employee, who is addicted to this sort of disconected "style" of discourse, and am useing a disturb and observe approach,and as it seems to go unoticed, is informative in it's own right.
I guess the connection between reading/listening, comprehension and retention ability on the one hand, and language generation ability on the other, isn't as strong as I'd been assuming till now.
This reads to me as fully written by LLMs. Pangram agrees. Note the (alleged) author misHQ’s comments on this thread are getting downvoted as obvious slop.
Even if it were written by hand, it’s a very poor and frankly stupid essay about an interesting topic. “The model's attention is a fixed quantity, and it has to add up to one, so the more things you make it look at, the less of that attention any single earlier thing can keep.” This is borderline gibberish and it outright rejects the interesting question about LLMs and attention, namely that they have very different capacities from us. LLMs can read an entire OpenAPI schema in seconds and immediately construct valid requests from it. The article first points this out, and then switches to arguing that LLMs have similar limits to us. It’s completely incoherent.
I like the ideas presented, but you can only read so many variations of
> But an unbounded queue isn't a safety margin, it's a debt that keeps compounding [1]
before getting a headache.
I wish the whole thing was written better, because the idea of designing a codebase for humans and our limitations sounds fascinating. It's why I personally love type systems, you can keep less things in your head and let the type checker alert you of any possible errors.
Thanks for the honest feedback. I genuinely wish I was a better writer. The posts in particular are all quite formulaic, but the idea is that they just provide a narrative access to the manifesto itself which is where the real meat of the argument is.
I would please urge you to read further into the manifesto itself but would also recommend you start at the foreword so you can understand the reason for the use of AI assistance in my writing.
Alleged author here - on my regular account. Thanks for the feedback. I have been quite upfront about the use of AI in the foreword. I have severe inattentive ADHD and have used AI to take my writings and present them in a way in which I feel are much more coherent overall.
The actual criticisms you have about the content however, I'd like to challenge:
The "adding up to one" is just a simplified gloss over softmax. It's very possible it reads poorly, and thats on me - not LLM gibberish.
As for the incoherence - I have to totally disagree. You have merged the 2 things the post keeps apart - capacity and attention over it. That a model can swallow a schema and write code is a competence humans share. We have been doing it for decades. Besides, the claim was never about us sharing capacity - it was about us failing in eerily similar ways.
So, AI slop, no. AI assisted, absolutely. It's sad that some judge the "who" more important than the "what" - especially for this kind of writing. But it's fair feedback nonetheless. I'll see what else I can do with assisting my delivery.
I've been working like, almost this exact idea! https://github.com/hale-lang/papers/tree/main . The same capacity allocation bound algorithm appears naturally not only in human and llm/agent congnition, but in many natural systems as well.
But at some level context engineering is very similar to what this article talks about.
Challenge accepted!
For me personally, that means setting up 'attention getters' for the important things in life - 'totems' that force a context switch. For AI agents, it means well-designed CLI tools that help the agent orient itself in a task and pull exactly the 'context-for-the-job' it needs right then.
This is exactly what makes building modern GenAI decision-support systems so difficult. It's no longer just about finding the right software abstractions. You now have to account for the unknown cognitive construct of a completely different intelligence.
This is exactly what I am trying to solve and I have what I call smart repositories that demonstrates this at
https://github.com/gitsense/smart-ripgrep
https://github.com/gitsense/smart-codex
The issue I am finding is, getting the agent to pull what it needs, even when the data is there is still challenging since LLMs are trained on blind discovery where the pattern is:
grep -> read -> grep -> read ...
What is working for me now is thanks to Pi (pi.dev). I am working on a pi-brains extension that makes it dead simple to control the lifecycle for an agent so if I detect that it uses `rg` without `gsc rg`, I can block the agent and inject a steering message that says always search with context.
I can also see if they try to "read" without first looking at the files metadata and so forth.
I'm finalizing things right now, but I think pi with my brains extension should allow domain experts to better guide agents so they can find what they need, when they need it.
> to fit there "evidence"
> useing
> unoticed
> it's own right
I guess the connection between reading/listening, comprehension and retention ability on the one hand, and language generation ability on the other, isn't as strong as I'd been assuming till now.
https://news.ycombinator.com/item?id=48706307
Even if it were written by hand, it’s a very poor and frankly stupid essay about an interesting topic. “The model's attention is a fixed quantity, and it has to add up to one, so the more things you make it look at, the less of that attention any single earlier thing can keep.” This is borderline gibberish and it outright rejects the interesting question about LLMs and attention, namely that they have very different capacities from us. LLMs can read an entire OpenAPI schema in seconds and immediately construct valid requests from it. The article first points this out, and then switches to arguing that LLMs have similar limits to us. It’s completely incoherent.
> But an unbounded queue isn't a safety margin, it's a debt that keeps compounding [1]
before getting a headache.
I wish the whole thing was written better, because the idea of designing a codebase for humans and our limitations sounds fascinating. It's why I personally love type systems, you can keep less things in your head and let the type checker alert you of any possible errors.
[1]: https://shapeofthesystem.com/posts/2026/05/10/the-queue-that...
I would please urge you to read further into the manifesto itself but would also recommend you start at the foreword so you can understand the reason for the use of AI assistance in my writing.
The actual criticisms you have about the content however, I'd like to challenge:
The "adding up to one" is just a simplified gloss over softmax. It's very possible it reads poorly, and thats on me - not LLM gibberish.
As for the incoherence - I have to totally disagree. You have merged the 2 things the post keeps apart - capacity and attention over it. That a model can swallow a schema and write code is a competence humans share. We have been doing it for decades. Besides, the claim was never about us sharing capacity - it was about us failing in eerily similar ways.
So, AI slop, no. AI assisted, absolutely. It's sad that some judge the "who" more important than the "what" - especially for this kind of writing. But it's fair feedback nonetheless. I'll see what else I can do with assisting my delivery.
What I've heard is human short-term memory can hold seven things at once. Fortunately the mind is much more.
That's funny, isn't it the same for dogs?