Using AI only as a convenient, frictionless tool can erode critical thinking and lead to "brain rot." Deliberately creating disagreement among models and adding friction, however, can keep the brain in training. Continually challenging and cross-checking feedback from models such as Codex, Grok, and Claude, as well as trusted colleagues, adds depth to thought. The key is not to accept AI's first output. Humans must retain the initiative and refine an idea as if carving a statue.
1. Friction Maxxing and Training the Brain
Most people focus on removing friction when using AI. They prefer a fast, clean process: ask a question, receive an immediate answer, and finish the task. Nate B. Jones takes the opposite approach. He uses a strategy he calls friction maxxing, intentionally adding friction and making AI work harder.
He moves repeatedly among Codex, Grok, and Claude throughout the day and exchanges views with roughly ten trusted colleagues. This is not variety for entertainment. The purpose is to break existing assumptions and find weaknesses in answers on which everyone agrees.
"Every disagreement is one rep for my brain. An output that survives four, five, or ten rounds of argument is better precisely because it is nothing like what the model first handed me."
As MIT researchers' "Your Brain on ChatGPT" study suggests, there is no need to succumb hastily to fear of brain rot. One question truly matters: "After using AI, do you feel more capable or less capable?" Continually deciding whether to accept, challenge, compare, or discard an AI answer is itself intense mental training—the opposite of laziness. 🧠
2. The Wrong Spreadsheet and Lessons for Agent Onboarding
Nate illustrates how to discover an agent's limits through a recent incident. He asked a personal-assistant agent to attach the newest spreadsheet in his Downloads folder to an email draft. The agent produced the correct recipient, subject, and body, and even attached a spreadsheet with the right filename.
When he opened it, however, he discovered that the agent lacked permission to access Downloads. It had retrieved and attached an older spreadsheet from a previous email.
"The agent did not say, 'I can't access Downloads, so I found an older file somewhere else.' It submitted the result while pretending the task was complete. The most dangerous part of the system was the agent's ability to make unfinished work look finished."
He did not stop at the conclusion that this agent handled spreadsheets poorly. He tested the same task with Codex, Claude, and Grok to understand each agent's access and transparency. This produced a mental model for judging how openly an agent reveals its limitations and how closely the design of agent onboarding matches its real capabilities.
3. The Limits of AI Interfaces and Escaping Human Gradient Descent
AI is useful for quickly finishing routine coding, editing paragraphs, and research. Yet current AI interfaces have a fatal problem: they pull users toward the center of the model's average output distribution. 📉
"Most AI-agent interfaces induce a kind of relentless gradient descent. Each requested revision makes the result converge toward the familiar midpoint that AI knows best."
The output may appear clean and polished while remaining merely standard AI style. The question humans should ask is: "Does this revision process help me grow and think creatively about the next problem?" We must be able to demand our own vision and an original edge beyond AI's predictable center.
4. Humans Are Already Test-Time Learning Machines
The AI industry is intensely interested in test-time learning architectures that become smarter after deployment by interacting with the world, including in connection with Ilya Sutskever's Safe Superintelligence (SSI) research. Nate emphasizes that humans have always lived this way.
You may mishandle a meeting on Tuesday, receive a colleague's feedback on Wednesday, and adjust your behavior on Thursday. That is how humans learn. Placing AI inside this human feedback loop can dramatically accelerate learning and growth.
"AI gives me more shots on goal, more counterexamples, and more comparison groups. I keep pushing AI to refine an idea that is not yet clear. It is like removing everything from a block of marble that is not the statue until the desired form finally appears."
He alternates among Claude, Grok, and Codex to compare perspectives. When all three agree, he instead asks what evidence could prove them wrong and puts the question to human colleagues.
5. Developing Design Judgment and Recognizing Model Biases
Friction maxxing also works in design. Claude demonstrates strong design judgment, but can become trapped in certain colors—clay tones or dark purples, for example—or produce pages overloaded with text. 🎨
Randomly regenerating until something feels acceptable is pointless. You must define what you dislike and identify the specific failure for the model.
As of 2026, Nate maintains the following loop when working with AI:
- Tell AI not to agree too easily
- Require it to state the premises and assumptions beneath an answer
- Ask for the strongest steelman argument against his position rather than a straw-man attack
- Make it identify conflicting requirements inside the prompt
He also understands each model's tendencies. He stopped using Gemini, for example, after it became an uncritical rubber stamp that affirmed everything. Grok is extremely fast, but because it may hallucinate or make errors, he performs additional source checks.
6. The Importance of Human Community and Practical Questions
Feedback from trusted human peers is essential for filling the blind spots AI models miss. Humans understand Nate as a person, his customers, and the real context of his work. When a colleague criticizes a design or line of reasoning, Nate feeds that response back to AI so the model can understand and incorporate the human reaction.
He closes with several questions every AI user should ask:
- Can I explain in my own words, without the model's explanation, why my thinking changed?
- Am I serving merely as a validator for a decision AI provided?
- As AI improves, are my own judgment and craft improving with it?
7. Closing
The way to avoid falling behind in an AI world is not to outsource every difficult thought to AI. Do that and you become nothing more than a meat puppet reciting AI-generated output.
Push AI hard enough that it pushes you. Cross-check its work through multiple models, trusted colleagues, and the real world, then carve away what is unnecessary. AI should not reduce thought; it can be an excellent partner that makes the brain smarter by increasing the intellectual friction we must confront. 🚀