Disclaimer: this is a bunch of toots on a thread I published on Mastodon that I put together as an article, in case it isn’t obvious.

I’ve got an hypothesis of AI and cognitive effects.

They aren’t as clear cut as one rule for all.

With the vast majority of people, I think it’ll dumb them down. It’ll cause a cognitive decline.

But for a peculiar circle the opposite might be very true.

The dumbest example I can put to simplify this is: do you belong to the set of people who ask it something and accept the output or to the set of people who questions, processes, learns and can keep up paying attention above all to the the thinking blocks while forever never trusting the LLM and avoid it when they know they won’t be able to read due to tiredness or exhaustion those blocks ? 🤔

I’m not talking about laziness, I’m talking about innate brain bandwidth, something that can’t be trained but born with. If you naturally process information quickly then an LLM is like getting rid of a lifelong straight jacket. Something that can produce information (right or wrong) at finally the same pace that you do.

Let me give a better working example, since I’m a heavy local AI user: choosing deliberately a dumber model.

We all here know the term rubber ducking, don’t we? If not, well, it is more or less when you talk to yourself aloud step by step about how to solve a problem. The exercise makes it easier to think and faster to reach the solution, because it is an active cognitive exercise.

Enter the dumb but fast model, for example LFM2.5-8B. It is dumb as fuck, so you assume already ahead of prompting everything it says will be either wrong, inaccurate, a lie or stupid.

Which means every problem, you try to solve with it would require far far more YOUR own knowledge than the model. The model just becomes a thing to keep your focus because it replies back very quick, more than 400 tokens per second on my PC (token generation).

If you’re the kind of person who reads the thinking blocks, with something that fast and yet that dumb, you’ll think very carefully what you prompt.

Every iteration of it giving you the wrong feedback, makes you refine your prompting, add more detail, add more steps, YOU are building the abstraction in your mind with the help of a thing so dumb that forces you to go to the very last detail. Add QDrant to have some long-term memory past the 128 thousand tokens max, so it can keep up with YOUR reasoning and the result is:

The LLM is actually training you, the roles have reversed.

A smart LLM can do things on its own, a dumb LLM is just an inquisitive toddler.

That is how I think it’s possible to use them for a cognitive not even enhancement but improvement. Dumb models are making me think deeper and by doing so I’m connecting more information, learning quicker, using better my focus and attention, and ultimately hacking the shortcomings of having a brain that never stops and it’s ever hungry for data.

The End

Let me give you another example: real life.

My partner and I have a notoriously different working-memory size.

He thinks I’m more intelligent than him, I am very aware I am not. Whatever he has his focus onto he can solve that problem as fast as I can.

The difference is angles. I can see many angles in parallel, he cannot. So when a problem requires many levels of detail I’m faster reaching a solution BUT when a problem is straightforward to solve, he’s usually faster, my tree-like way of thinking gets in the way. I’m not overthinking, I just physically can’t avoid branching into angles.

So what’s my solution to simpler problems: choose the first apparently viable solution and stop there. The same rule to avoid analysis paralysis works to avoid letting my brain branch out. Hence why “less is more” is in my right forearm tattooed as a reminder.

Now with this background, what happens if I just take the first option: errors. Thousands of them per day.

So what do I do when I fall? I stand up quickly and try again. Brain has its own life, the moment it sees the error it’ll branch out in multiple other “apparently” viable options, solving problems become a refinement of the previous iteration, what worked before is preserved, what caused the error is re-analysed in parallel, the apparently best solution chosen. My working memory size allows for this, solving daily life as the Travelling Salesman problem.

His working memory size, as it is more linear, requires priorities. And when things require priority, there isn’t enough bandwidth for errors so he tries to avoid them. So my partner, like most people, is error-adverse, while I love errors because I learn from them much faster than from being right.

See again with this example, how a dumb LLM amplifies the natural way my brain works? And this tree-way of thinking is common to most people with high cognitive abilities so I’m not alone.

Now let’s go back to the blanket statement at the beginning “I think for the vast majority of people it’ll cause a cognitive decline” I wrote.

An excellent question would be? How the fuck would I know when I can’t think the way they do?

So what I know for certain is only that for a subset of people it can definitely cause a cognitive improvement. While for the vast majority, it’s at best: guessing.

And that guessing is not far from the “videogames will dumb people down”. It never happened, the opposite was true.

The one thing I know is the human brain demands using itself, and when an activity becomes dull, it finds another one more stimulating or complex.
And that’s the same for everyone.

People who used AI and noticed a cognitive decline, ends up STOPPING to use it. Boredom stopped the decline.

So now we know

• For tree-way-like thinking people causes an improvement.
• For a subset of the vast majority it bores them.
• Teenagers seem to hate it 👇

https://fortune.com/2026/05/20/why-do-kids-hate-ai-gen-z-backlash/

So you’re seeing live on this thread a small example of how my branching brain works, I started from a question, an ignorant assumption and real empirical data based on self-experimentation and awareness of my own differences.

From there I work my way back the branch of the tree to realise how my assumption (“vast majority of people…”) is an error, and branch from the new knowledge if “wait, I can’t actually know” how a subset of that other group who happens to be the one noticing cognitive decline, also ends up getting bored about it & quitting (so, no decline, just wisdom).

And then the branch of teenagers hating it which is quite a big chunk of the world population. So…I don’t really know the answer, of what it’ll do for certain to the vast amount of people, what I can maybe deduct (new branch of a branch) is that not only the hype may end soon but rejection will be heavy and once that rejection becomes mainstream the big 2 (Anthropic, OpenAI) will cease activities.

And for me that means simply, no new models.

But wait, I’m already using dumb models anyway and thriving with them. e.g.: Obsidian never worked for me, a vector database and a simple RAG works flawlessly.

Why would I need any new ones, why would I stop using the ones that work fine for me, why would others in the same situation? We don’t need fatter GPUs, we don’t care about 5 trillion parameter models of frozen statistical knowledge. What for? Wikipedia is fine, googling is fine. Small dumb models can search info online fine instead of making shit up.

Now let me drop one big controversial question and go back to watching a “Puked by my hockey rival” derivative again (they are all Pride and Prejudice varieties):

Did we dispose of all the controversially acquired medical knowledge all the way to 2026 and start afresh ever OR did we judge and jail those who acquired it controversially and yet kept the knowledge?

Why is me running a local LLM any different?

Good night.