The Curious Companion: Ep. 12 – Why ChatGPT Sometimes Sucks (and What to Do)​

Curious Reader!

Welcome to this week’s ChatGPT Curious companion newsletter.

What you came for is below, and you can CLICK HERE to listen to the episode if you decide you prefer earbuds to eyeballs.

Happy reading!

In this episode we dig into why ChatGPT sometimes gives you garbage outputs and what you can do about it. From hallucinations and drift to bad prompts and long chats, we’ll cover the common pitfalls, what to do when it sucks, and how the imperfections of AI might actually be keeping your own skills sharp.

When ChatGPT Sucks

We’ve all been there. You ask ChatGPT to do something for you, it gladly obliges and then spits out something completely unusable.
WTF?

I’d be lying if I said I’ve never lost my temper and demanded that it figure its shit out. Of course it always validates my frustration but not once has it ever then gone out to generate an improved output.

Why doesn’t berating it improve the outcome? That’s a rhetorical question. A question that is actually worth asking is: Why did it suck in the first place? To understand that it’s helpful to simply understand how ChatGPT works.

If you want a deeper dive into how ChatGPT works, head back to episode 1:

• ​Ep. 1: What You Actually Need to Know about ChatGPT​
• ​Curious Companion Ep. 1​

Quick Refresher: How ChatGPT Works

The key thing to remember: It’s math, not magic.

ChatGPT does not reason. It does not memorize. It does not “think” or understand. It takes your input and predicts the output that has the highest probability of being correct. Think of it as predictive text on steroids.

• Your input is broken into “tokens” (common sequences of characters)
  • GPT-4 had about 100,000 tokens in its vocabulary.
• ChatGPT processes your input as tokens and generates a response by predicting what token comes next, one by one, based on probability and patterns in its training data
• ChatGPT is by definition a probabilistic model:
  • It is based on likelihoods instead of certainties. Instead of always giving the same answer every time (deterministic), a probabilistic system uses probability to decide the outcome.
  • For ChatGPT, that means it’s predicting the most likely next word or token based on patterns in its training data, not pulling from a fixed memory bank. This is why asking it the same question twice, or asking it to perform the same task twice, can and will generate (slightly) different results.

Because ChatGPT is predicting, not recalling, it can give you something factually incorrect. That’s what we call hallucinations: Plausible sounding but factually incorrect outputs that are fabricated or unsupported by real data. They’re not gibberish, they’re just wrong.

This gap between probability and reasoning is one of the reasons we’re not anywhere near AGI (artificial general intelligence) yet, and also why sometimes ChatGPT just plain sucks.

Where the Suckiness Shows Up

Worth noting, what I’m really focusing on in this episode is when ChatGPT is generating something for you, as opposed to when it’s giving you an answer to a question. For example, when you’ve asked it to create an outline, a summary, or a draft, and it just completely drops the ball.

What You Can Do About It

Here’s the good news: you’ve got options.

1. Better input = better output.
​Make sure your prompt is clear and specific. For more help, check out episode 8 with my clarity checklist:

• ​Ep. 8: Getting ChatGPT to Do What You Want​
• ​Curious Companion Ep. 8​

2. Focus on what you want it to do.
​ChatGPT (like humans) struggles with negative rules. If you say “don’t do this,” expect it to do exactly that. I stay telling it that I want no em dashes, and guess what I keep getting? The gotdamn em dashes.

3. Check the memory.
​ Look at what it’s stored about you and make sure it’s still accurate and relevant. For a refresher on how to check stored memories, check out episode 9.

• ​Ep. 9: Helpful ChatGPT Features You Might Not Know About​
• ​Curious Companion Ep. 9​

4. Watch for drift.
Drift is the gradual or sudden change in how ChatGPT responds, caused by updates to the system rather than anything you did. The biggest thing I’ve noticed here is that drift may mean you need to update or modify any saved instructions you have for your Projects.

I recently had to do this for one of my Projects because what ChatGPT was producing for these Companions was requiring way too much changing and editing.

5. Start a new chat.
​This one is underrated but highly effective. Long chats tend to freeze, glitch, or just veer off course. Start a new chat if you’re working with significant amounts of text. Additionally, if you’re wanting multiple revisions, often times it’s best to just start a new chat and revise the original. This is because each time you ask for a revision, you’re not getting edits to a clean copy, you’re getting edits to edits of edits. Probability based on probability based on probability. At some point, it’s better to wipe the slate and start over.

Know Your Shit

Beyond the technical fixes, the biggest thing you can do to help ChatGPt suck less is simple: Know your shit.

You can’t blindly trust it. The best way to use ChatGPT is as an assistant, always keeping in mind that you’re in the driver’s seat.

Example: I had it create markdown files from my Intensive call docs and it mixed up the weeks. When you know the content, you catch the mistakes, and you can get ChatGPT back on track.

Why the Suckiness Might Be a Good Thing

Here’s the twist: I think that ChatGPT’s suckiness has some value.

Mistakes force you to practice. They help you sharpen your skills.

Yes, using AI and automating tasks can be helpful, but there’s an important distinction to be made between skills of operation and skills of understanding.

• Skills of operation = executing the task.
• Skills of understanding = knowing the principles behind the task.

Examples of skills of operation we’ve lost without collapse:

• Math by hand → calculators (but still understanding mathematical concepts)
• Memorizing phone numbers → smartphones
• Film photography → digital

In each of these, the specific style of operation became obsolete, but the understanding remained.

Where things get tricky is with comprehension and expression, because operation and understanding are fused. Typing instead of writing? Fine. Dictating instead of typing? Fine. Having ChatGPT both think and write for you? Not fine.

We all know this. I’m not trying to go down that path in this episode.

I will however say this: To me the actual solution to any concern about outsourcing one’s thinking and expression, isn’t banishing AI, the solution is changing society such that we actually truly value and champion thinking and individual expression

How I Used ChatGPT This Week

Each episode I include a section where I briefly discuss how I, or someone I know, has used ChatGPT that day/week.

This week, it’s my guy DocJoeO. Joe shared a handful of use cases with me, but the one that stood out (because I haven’t mentioned it in an episode yet) was using ChatGPT to generate meals for the week, along with the accompanying grocery list. Stellar use case.

My girlfriend, Lex, absolutely does this as well, and while I’ve yet to use it to generate full meals, I have used it for smaller cooking questions, like when I wanted to make a marinade but only had a few things in the cupboard and also didn’t want to spend 97 hours letting it sit.

The coolest part about using ChatGPT for recipes is how much you can customize things:

• Certain flavor profiles
• Fewer ingredients
• Shorter prep time

Whatever you want, ChatGPT can make it happen. Whether or not it tastes good…well, I guess you’ll just have to hope the recipe it generates isn’t a hallucination.

So yeah, ChatGPT sometimes sucks. I’ve given you a few strategies to help it suck less, but the most important thing to remember is to use it as an assistant, not a replacement, and let the imperfections sharpen your own thinking.

Alrighty, that’s it for today’s episode. As always, endlessly grateful for you and your curiosity.

Questions, comments, concerns, additions, subtractions, requests? Hit reply or head to the website (​chatgptcurious.com​) and use that contact form. I’d love to hear from you.

Catch you next Thursday.

Maestro out.

AI Disclaimer: In the spirit of transparency (if only we could get that from these tech companies), this email was generated with a very solid alley-oop from ChatGPT. I write super detailed outlines for every podcast episode (proof here), and then use ChatGPT to turn those into succinct, readable recaps that I lightly edit to produce these Curious Companions. Could I “write” it all by hand? Sure. Do I want to? Absolutely not. So instead, I let the robot do the work, so I can focus on the stuff that I actually enjoy doing and you get the content delivered to your digital doorstep, no AirPods required. High fives all around.

Did someone forward you this email?
​Click here to join the ChatGPT Curious newsletter family.​

Stay curious.​