The Overfitted Brain: Why Routine Makes Us Weaker Players—and How Dreams, Wandering, and Curiosity Restore Our Advantage Over AI
If you feed your brain the same set of stimuli day after day, you only teach it one thing: "optimize for a repetitive world." The problem is, the world is rarely repetitive. Overfitting works on humans just like it does on AI models: great performance in familiar formats, poor results in new situations.
You can get out of this without a life revolution. The rescue mechanism is built-in—sleep and dreams—and you can also consciously design your daily life so your brain becomes flexible, curious, and creative again.
Overfitting in Practice: The Ball-and-Basket Experiment That Exposes the "Drilling the Same Thing" Trap
The simplest illustration of overfitting comes from a school experiment: kids threw balls into a basket (literally, a trash bin). They were divided into two groups:
- Group A trained only from one spot—about 5 meters away—and would later be tested from this exact spot.
- Group B threw from many different points and distances, except for the test spot.
Intuition suggests Group A would win, since they “train perfectly for the exam”. Instead, the opposite happened: Group B crushed Group A in the 5-meter test.
That's the core of the problem: repeating the same task teaches the brain shortcuts for specific conditions, not actual skills. It’s the same as preparing for an exam only by drilling the exact question format—good familiarity with the test, but poor effectiveness when conditions change even a little.
The same logic appears in running training: you do intervals for a half-marathon, even though the race itself is not a sprint. Why? So the body is ready for variability and effort—not just one narrow version of the world.
Overfitted Brain Hypothesis: Why Do We Dream If Sleep "Makes No Evolutionary Sense"?
Sleep is strange from a survival perspective: during sleep, you don't get food, you don't reproduce, and you're more vulnerable to predators. Yet, nature didn't “eliminate” sleep, including the deep and REM phases where the weirdest images occur.
The Overfitted Brain Hypothesis (OBH) proposes a precise function: dreams disrupt and mix data, preventing the brain from overfitting to a monotonous, repetitive life. At night you get a series of “what if?” scenarios, sometimes completely absurd—like a dream about a new species of hamsters that look like bananas.
It's not a “system error.” It's a mechanism that:
- interweaves concepts from the day into new configurations,
- builds flexibility,
- increases the chance for creative connections.
This explains the familiar experience: you go to bed with a problem, wake up with a “eureka.” Solutions don't fall from the sky—at night, your brain connects the dots, or more precisely, connects neurons, building new association pathways.
Source: Overfitted Brain Hypothesis, Patterns (Cell Press)
Two Modes of Brain Work: "Task Mode" and the Creative Highway in the Background
Throughout the day, you switch between two networks:
- Executive Control Network—task mode: planning, focus, "blinders on."
- Default Mode Network—default mode: mind-wandering, associations, simulations, daydreaming.
When you get stuck on a problem and step away from your desk, that's not laziness. It's a conscious shift to a mode where your brain keeps working in the background, but without sticking to a rigid path.
This mechanism recurs in classic stories of genius productivity:
- Richard Feynman would walk away from work and play the bongo.
- Albert Einstein would pick up his violin.
Many know the effect: the best ideas come under the shower or on a walk, because you momentarily let go of control, and the Default Mode Network stitches together what you've previously collected.
AI "Hallucinations", Wandering, and Creativity: Not Just a Flaw, but Raw Material
There's a strong expectation around AI: it should be deterministic, infallible, "without hallucinations." But deliberate wandering is an integral part of solving complex problems—also for humans.
This is visible on two levels:
- Science and Breaking "Truths"
Einstein could contradict what was considered reality. In 1908, he published three papers that for many were "wrong" because they didn't fit the existing worldview. Creativity often means something looks like an error—until it turns out to be a new, better definition of the truth.
2. Programming and Suggestions That Don’t Exist
A model might suggest a method in a library that... doesn't exist. This annoys, because it wastes time. But sometimes the "imagined" method is so reasonable that you think: this should be in the library. Then the hallucination becomes a prototype—not a finished solution.
In practice, treat it like working with a "creative colleague": you don't take everything literally, but you extract new directions from it.
ARC 3: A Benchmark That Shows Where Humans Still Have a Hard Advantage
ARC 3 is the third iteration of a benchmark that doesn't ask: "can you solve the test?" It asks: can you enter a new environment, where you don’t know the rules, learn them, and act effectively?
Results are measured along two axes:
- how well did it go,
- how much did it cost to get there (how many resources were "burned").
And here comes a number that reframes the discussion: the top score is 0.6%, and the cost of an approach is about $8,500 per trial/game.
That leads to a simple conclusion about human advantage: human intelligence is primarily adaptive and cost-effective. That's why there’s a vivid boundary: people will start to be really afraid the moment "an AI server can be fed a banana and work all day on it." Until then, cost and efficiency are still on the side of biology.
Art, Dreams, and the Cost of a "Closed Daily Life": The Beksiński Case
Dreams can create images that "no artist would dare"—because they're too weird, too intense, too uncompromising.
Here’s a strong point of reference: Zdzisław Beksiński and his ultra-dark paintings. Sometimes, viewers recognize in them figures from borderline experiences—like sleep paralysis episodes, when "something stands in the room" and looks just like a figure from his canvas.
But there's also a cost: if daily variety is minimal (Beksiński was known for rarely leaving his Warsaw apartment block), the "material for intensity" shifts almost entirely into dreams. Then creativity can grow, but waking life gets claustrophobic.
The practical takeaway: it's better to feed variety during the day, instead of counting on your brain to fix everything at night.
Takeaways That Truly Change How You Learn and Work
- Overfitting affects humans: routine and practicing one format build fragile effectiveness.
- Varied training beats repetition—as shown by the basketball experiment (diverse distances > perfect test setup).
- Sleep and dreams are active anti-overfitting mechanisms: mixing data and building new connections.
- Default Mode Network is the creative mode—activated when you let go of control (walking, shower, music).
- "Hallucinations" aren’t just a flaw: wandering is the fuel for creativity, in learning and problem-solving.
- ARC 3 highlights the human edge: adapting to unknown rules at low cost (AI: 0.6% at ~$8.5k per attempt).
- Curiosity and variety are training for intelligence understood as adaptation, not just test-solving on paper.
How to Implement This (Starting Now): 7 Simple Steps Against Overfitting
- Change your route to work 2–3 times a week (even one random turn is enough).
Goal: new stimuli, new landmarks, new associations. - Introduce “variations” in mundane activities: different walking speed, different time for a walk, different task order.
Goal: your brain stops optimizing for just one route. - Give time for Default Mode Network: 15 minutes of walking without a phone or a shower without any background content.
Goal: activate wandering and connecting the dots. - Use “interleaving” in learning: mix task types instead of drilling one worksheet.
Goal: flexibility, not just matching the format. - Introduce micro-linguistic variety: 10–20 minutes daily conversation with a voice model (Gemini / ChatGPT) in a foreign language—like Vlad does on his commute (e.g., in French).
Goal: real-time adaptation, no waiting for January resolutions. - Feed curiosity with detail: daily, pick one thing to notice (architecture, street layout, sounds, human behavior).
Inspiration: da Vinci’s notes on things we usually ignore (e.g., sewage channels). -
Treat creative “errors” as material: when AI suggests something that doesn't exist (method, library), ask: "why would this make sense?" and write down the idea.
Goal: use wandering as a generator of directions, not as a final authority. -
Przeuczenie (overfitting) dotyczy ludzi: rutyna i ćwiczenie jednego formatu budują kruchą skuteczność.
- Zmienność treningu wygrywa z „klepaniem” — co pokazuje eksperyment z rzutami do kosza (różne dystanse > idealne warunki testowe).
- Sen i sny są aktywnym mechanizmem anty-overfittingowym: mieszają dane i budują nowe połączenia.
- Default Mode Network to kreatywny tryb działania — aktywuje się, gdy odpuszczasz kontrolę (spacer, prysznic, muzyka).
- „Halucynacje” nie są tylko wadą: błądzenie jest paliwem kreatywności, także w nauce i w rozwiązywaniu problemów.
- ARC 3 punktuje przewagę człowieka: adaptacja do nieznanych reguł przy niskim koszcie (AI: 0,6% przy ~8,5 tys. USD za próbę).
- Ciekawość i różnorodność są treningiem inteligencji rozumianej jako adaptacja, nie jako rozwiązywanie testów na kartce.
Jak to wdrożyć (od zaraz): 7 prostych kroków przeciw przeuczeniu
- Zmieniaj trasę do pracy 2–3 razy w tygodniu (nawet jeden skręt „losowo” wystarczy).
Cel: nowe bodźce, nowe punkty orientacyjne, nowe skojarzenia. - Rób „wariacje” w banalnych czynnościach: inne tempo marszu, inna pora spaceru, inna kolejność zadań.
Cel: mózg przestaje optymalizować pod jedną ścieżkę. - Dawaj czas Default Mode Network: 15 minut spaceru bez telefonu albo prysznic bez treści w tle.
Cel: uruchomić błądzenie i łączenie kropek. - W uczeniu się stosuj „interleaving”: mieszaj typy zadań zamiast mielić jeden arkusz.
Cel: elastyczność, a nie tylko dopasowanie do formatu. - Wplataj mikroróżnorodność językową: 10–20 minut dziennie rozmowy z modelem głosowym (Gemini / ChatGPT) w obcym języku — tak jak robi to Vlad w drodze do pracy (np. po francusku).
Cel: adaptacja w czasie rzeczywistym, bez „czekania do stycznia”. - Karm ciekawość detalem: raz dziennie wybierz jedną rzecz do zauważenia (architektura, układ ulic, dźwięk, zachowanie ludzi).
Inspiracja: notatki da Vinciego o rzeczach, które zwykle ignorujemy (np. kanały ściekowe). - Traktuj kreatywne „błędy” jako materiał: gdy AI podsuwa coś nieistniejącego (metoda, biblioteka), zadaj pytanie: „czemu to miałoby sens?” i zapisz pomysł.
Cel: wykorzystać błądzenie jako generator kierunków, nie jako wyrocznię.
Te kroki robią jedną rzecz: przestajesz trenować mózg pod przewidywalny świat, a zaczynasz trenować go pod świat, który zmienia zasady w trakcie gry. Właśnie to jest praktyczna odporność na „przeuczony mózg”.