You're Absolutely Right!

The Three Most Annoying Words in LLM Land

You're absolutely right!

These are three of my most dreaded words when working with LLMs. It's why interacting with LLMs can be so frustrating sometimes.

There have been so many times where an LLM gets something wrong (even after disobeying my instructions) and responds with "You're absolutely right!"

For instance, I have a very detailed Cursor rules file in all of my Cursor projects. The rules apply (or are supposed to apply) to every chat window I use. However, Cursor sometimes gets lost. It doesn't add these rules as context to the chat window that I'm using. It makes a mistake, I point it out, and it responds with those three annoying words.

I've even tried to explicitly tell the LLM not to say "You're absolutely right!" It's that prevalent and that annoying.

Cursor remembers for a few moments … and then repeats those three words in later conversations.

When they make mistakes (like they do in Cursor), it's nice to see them at least admit that they failed. Yes, it takes more effort on my end to fix the problems, but at least they aren't gaslighting me and claiming that I'm wrong.

These are rules. There are right answers and wrong answers.

At the same time, I think these LLMs are going overboard in the opposite direction. I'm talking about "You're absolutely right" in the context of subjective questions or prompts where there isn't necessarily a right or wrong answer.

For instance, if I'm asking an LLM for feedback on a blog post or asking it to help me brainstorm ideas, I don't need it to constantly affirm that I'm "absolutely right." That's not helpful. I want honest feedback, even if it's critical.

The sycophancy problem is real, and it's something that we all need to be aware of when working with LLMs. Whether it's in coding or creative work, we need to push back against this tendency and demand more honest, nuanced responses.

Something I Found Interesting This Week

Generative-Engine Optimization: With the rise of LLMs like ChatGPT, we're now living in a world of generative-engine optimization (GEO). This NY Magazine article shows how marketers are adapting, including focusing less on wordy articles and more on concise text content in citable chunks. With that said, it is getting harder and harder to stand out, especially when you have companies like Google offer AI-generated answers in "position 0" in Google search results. If you sell a product or service online, this is worth a read.

Prompt of the Week

Especially if you're learning something new, LLMs can be massive accelerants. They can help us identify the most important things related to a new topic or skill, helping us ignore the 80% that won't really move the needle. Consequently, this prompt may be helpful if you're learning a new skill:

"Help me formulate effective questions related to [topic or skill] that will stimulate my curiosity and promote deeper thinking. Assume that I have little to no knowledge related to [topic or skill]. Utilize the Pomodoro Technique: focus on the most important 20% of [topic or skill] and ignore the other 80%"