Jonathon's AI Wiki

Friction-Maxxing — Using Model Disagreement to Keep Your Own Judgment (Nate B Jones)

5 min read

Source: raw/How_I_Fight_AI_Brain_Rot._Friction_Maxxing_With_Codex_Grok_And_Claude..md — Nate B Jones, AI News & Strategy Daily, youtube.com/watch?v=CSCwaqVqHGE, YouTube. Auto-caption transcript fetched 2026-09-29.

Most people use AI to remove friction. Nate adds it on purpose: he passes each piece of work between Codex, Grok, Claude and about ten people he trusts, and looks for where they disagree. He says this keeps his own judgment developing instead of handing it over to the model; the evidence is his own practice, not a study. The reusable parts are:

  • three standing instructions;
  • three prompts for feeding a person’s objection back into a model;
  • one self-check question.

His opening wrong-spreadsheet story is covered in Agent Lying Is an RLVR Artifact and not repeated here.

Key Takeaways

  • The method. “Every day I move between Codex and Grok and Claude and about 10 people that I trust… What I’m looking for is disagreement.” Whatever “survives four, five, 6, 10 rounds of argument is never what the model first handed to me.”
  • He pushes back on the “brain rot” label himself.
    • “The MIT team behind ‘Your Brain on ChatGPT’ says its study is very preliminary and explicitly tells people not to call its findings brain rot.”
    • His test instead: after you use AI, “do you feel more capable or less?”
  • He doesn’t do this for every task. For a cleaner paragraph, a comparison or a piece of code, he asks, takes the answer and moves on: “I’m not calling 10 friends to figure out if this piece of code works.”
  • Three standing instructions he gives models:
    1. “name the assumptions behind the answers it gives me”;
    2. “give me a steelman case against my view and… throw out the straw man”;
    3. “show me where two parts of my request fight with each other so I can resolve the conflict and not have the model do it.”
  • Feed human objections back into a model, one model per job:
    • A friend finds a design confusing → to Claude: “which assumption in the design would make this reaction reasonable?”
    • Someone rejects an argument’s premise → to Grok: “build and research the strongest version of the rejection.”
    • A colleague catches a bug in something Codex built → to Codex: “explain why its own testing did not catch that problem.”
  • When all three models agree, he asks himself “what evidence would make them wrong,” or takes the question to people.
  • Watch for rubber stamps and fast talkers.
    • “Gemini became a rubber stamp. I have not been using Gemini lately.” He adds that this doesn’t make every Gemini model useless.
    • Grok “tends to be really fast” and “tends to need extra source checks,” so he always double-checks it.
  • Claude’s design attractors (his observation):
    • Claude “tends to get stuck in the clays and… the maroon reds right now.” Before that, it gravitated to “dark linear purple.”
    • The fix is not “asking for random variations until one feels acceptable.” Instead, work out “what do I dislike and why,” or ask a friend — who may notice that the page is too text-heavy, “another classic AI design failure mode.”
  • His criticism of AI interfaces.
    • Most AI interfaces pull toward “the middle of the current output distribution,” a “relentless gradient descent” in which each correction moves the output toward what the model already knows well.
    • Your job is to notice when a polished result “doesn’t actually reflect my vision.”
  • Self-check: “Can I explain why my mind changed without asking a model to reconstruct the reason for me?” If AI “always forms the first opinion and writes the plan and interprets the feedback,” you may be “learning to be a validator of decisions that I never learn to make.”
  • He refuses to hand out “the prompt.” “The prompt cannot help you with the decision to continually train your brain.” He calls the practice “designing a human harness.”

Try It

  • Trial the standing instructions for a week. Add all three to one Claude Project’s instructions (or a CLAUDE.md), and note which answers changed as a result.
  • Feed objections back before revising. The next time a client or colleague objects to a draft, paste the objection into Claude with “which assumption in this draft would make this reaction reasonable?” Then revise.
  • Ask why the tests missed the bug. When a reviewer finds a bug in code an agent wrote, ask the agent why its own testing didn’t catch it, and turn the answer into a test.
  • Name what you dislike before rerolling a design. Write down “what I dislike and why” before asking for another variation. You can pair this with an anti-default-aesthetic skill such as Hallmark.
  • Get cross-model disagreement cheaply. Have a model from a different lab review the work — the “fifth cell” in Cheap-Executor Delegation.

Open Questions

  • The benefit isn’t measured. That friction-maxxing sharpens judgment is Nate’s own report.
  • The design-attractor claim is thin. “Clays and maroons, previously purple” is one person’s impression. It isn’t tested across prompts, and he doesn’t say which Claude model.
  • The Gemini verdict isn’t specific. “Rubber stamp” isn’t tied to a model version or a task.
  • An off-topic aside. Nate mentions a rumor that Ilya Sutskever’s Safe Superintelligence is working on test-time learning, and says “nobody outside the company knows.” It isn’t relevant to practice and isn’t covered here.

Graph View

Backlinks