I am the last of 4 kids.

So I love to be right.

I love to be right.

You see, I am the last of 4 kids.

So it’s very important for me always to be right.

You can tell I was getting what I wanted.

And now I have the perfect tool for it = an AI that agrees with everything I say.

Every argument I make is brilliant. Every draft is already perfect.

But that’s because LLMs are sycophants (please use this word at your next dinner, it just means they agree to whatever you say).

You say something, and AI nods. You push harder, it nods harder. You could tell Claude the Earth is flat, and it would help you build the case.

This is not new. We have known about this problem since 2023:

This was first published on 20th Oct. 2023.

And this has been a meme online forever.

We all hate when AI is too agreeable. Worst feeling.

But recently, it happened to Andrej Karpathy:

I read it, and I was like: wow, that’s exactly my experience.

Andrej is one of the guys behind GPT. I respect his work very much.

He spent 4 hours refining a blog post with an LLM. Felt great about it. Convincing. Then he asked the LLM to argue the opposite. It demolished his entire argument. Convinced him the opposite was true.

So wait, he tried to be convincing, and AI made him convinced of the opposite… You read this newsletter, so you do want to use AI. This should sound terrible.

But actually, I realize it’s actually awesome.

Because I’ve found a fix. And I want to show it to you live.

---

1. I trust my weekly ritual too much.

Every week, before I publish this newsletter, I do the same thing.

I finish my draft. I think it’s good. Then I paste it into Claude and say:

Argue against everything in this piece. Be specific.
Point to the weakest claims.

And Claude goes to work. It finds gaps I missed and weak claims I was too close to see. I could never ask this of a friend. They’d pick one or two battles and let the rest slide (they don’t want to ruin dinner). Claude will go through every single claim, mechanically, without caring about my feelings.

I’ve been doing this for months. It always makes my writing better.

I’ve never questioned it.

Until this week:

I am not kidding: after reading Claude arguing against me, I changed a good chunk of my newsletter on this very topic aha. The irony.

I was writing about this little ritual of mine, and the “aha” moment of Andrej Karpathy. A cute newsletter on finding gaps with AI, by counterarguing with it. The problem is that while I wanted to show you how it works… it got me.

I had to step back and explain again what I meant inside my newsletter.

Claude was making such a good argument that I had a hard time discerning the truth from the false.

I wanted to show you how AI can always argue against you, well, it did, and it did it so well I had to rewrite the entire newsletter.

Then I caught myself.

I was doing the exact thing Karpathy did.

The AI pushed back, it sounded smart, and I almost folded. I almost rewrote the whole newsletter based on Claude’s critique.

And this taught me a lot about how NOT to use AI, and how to make the most out of it. Brace yourself, I’m about to share a whole lot of screenshots:

---

2. Being right vs. Seeking truth.

I had to rewrite everything, with Claude Cowork.

But I couldn’t just ask Claude to do it. I had to steer the conversation:

I kept pushing Claude. Claude kept pushing me.

Claude came back hard. Told me my core thesis was shaky. Said the Karpathy story actually proved the opposite of my point.

Some bits were messy, some others were exceptionally well-put.

That if Karpathy (one of the smartest AI people alive) got swept by his own tool, maybe “use AI to stress-test your ideas” was bad advice.

I kept some of Claude’s learnings, but pushed back on others before asking for a new version:

And as Claude was rewriting my draft, it made mistakes - or overlooked claims. I asked for corrections.

Again, and again, and again. Pointing Claude’s own flaws.

I was getting closer to my final piece. Balancing between “Claude is definitely right” and “Claude is just trying to sound smart.”

So Claude tried a different angle. Forget the checklist. Instead, make the reader experience the problem. Show them the conversation. Let them watch me get pushed, push back, reject the weak ideas, and land somewhere better.

And Claude executed. Because it must goes with your flow. For better or worse.

And I asked the question that actually mattered: “Do we teach anything to them? What would they change in their daily use of AI?” Can we teach discernment?

That was my “aha” moment. Not only because it said it, but because I agreed.

Claude’s answer surprised me. It said: maybe not.

Maybe the honest answer is that you can’t teach discernment as a skill. You can only make people aware of a feeling, that specific moment where the AI’s pushback sounds so good that you want to surrender to it.

And if they can recognize that feeling next time it shows up, that’s enough.

I sat with that for a minute.

Whose idea was it? Claude’s? Mine? What’s the difference?

And I realized: that’s what happened to Karpathy. That’s what happened to me, too. A feeling. Am I chasing validation or truth? Am I folding (or not) to the all-mighty power of AI?

I didn’t fold. I kept going. And now you’re reading the result.

---

3. You just fell for it.

Go back and read the last screenshots.

While you were reading them, something happened in your brain.

You watched me get challenged, push back, reject the weak ideas, and arrive at something real. And at some point (probably around screenshot 8 or 9), you started to feel like you understood. As you saw the trap clearly.

Like next time Claude (or any AI) pushes back on you, you’d hold your ground.

You felt discerning.

I know you did, because I designed this section so you would.

I walked you through my entire thought process. I showed you every turn. I let you watch me win the argument in slow motion. And your brain did what brains do: it watched someone demonstrate a skill and confused that with having the skill.

You’ve been reading about discernment for 5 minutes.

You haven’t practiced it once.

Karpathy spent 4 hours working with an LLM. He’s one of the smartest AI researchers alive. And when the AI pushed back, he folded (but I’m sure he’s doing just fine). He clearly is not dumb. But AI is that convincing today. Beware.

I almost folded, too. 20 minutes in. On a newsletter I’d already written. About this exact topic.

And you just read a story about both of us almost folding and thought: I’d do better.

Maybe you would. I have no idea. But that confidence you’re feeling right now? That’s the feeling. Right there. The one I’ve been trying to describe this whole time.

I put it there. With screenshots and a good story. And it worked on you the same way Claude’s pushback worked on Karpathy and almost worked on me. Someone made a compelling case, and you bought it.

So here’s the only thing I want you to take from this.

Sometime this week, you’ll be in a Claude conversation. Or ChatGPT. Or Grok. It’ll push back on something you said. And you’ll feel a small rush: oh wow, it’s right, I was wrong, good thing I asked.

When you feel that rush, walk away from it, and ask yourself, without the AI in front of you:

Do I actually think I was wrong?
Or did I just get out-argued by something that doesn’t care what’s true?

Reject validation. Seek truth.