Lessoned Learned: Using a LLM to Write a Blog Post Worth Reading.

90% of the content on the internet will be generated by LLMs, or so they say.

I wanted to put the narrative to the test. I decieded to use a LLM to write blog post. The goal was to write something worth reading. I wanted to be able to pass it off as myself.

To challenge the LLM, the topic was a personal experience. While I coud iterate on the prompt infinitely, I didn't allow myself to touch the output.

TL;DR -- It was A LOT harder than I expected.

My initial prompt was a breif and vague. Even still, I expected the LLM to craft a seemless narrative, fill in the story with vivid details. BOY, WAS I WRONG. The output was reminencent of a high-schooler barely passing 9th grade english.

But more surprising, it didn't intuit details from my experience. In hindsight, DUH. It's an LLM! Still, part of me just expected the LLM to get it.

As I iterated, 2 productive strategies emerged. First, tell it "how". Second, tell it "what". While the latter was by far the majority of the prompt, the former had a larger impact on the outcome (see below).

After a few hours of iterating, I realized I would have been better off writing the post myself. Using a LLM didn't save me any time. The output was worse. And the prompt ended up being longer than the post.

Did I reach my goal? I don't think so. I wouldn't publish as-is. But, I enjoyed the exercize. After the experience, I do think much of the internet's content will be LLMs gernerated. I'm just not convinced I'll have the patience to read it, at least as it is now.

UPDATE (3/16/23): Checkout the follow-up post. I run the same exercise with GPT-4. The results ARE NOT the same.

If you're going to do this yourself, here are some tips...

• Use ChatGPT, not the OpenAI playground. This actually ended up making the MOST difference.
• Put the "how" near the beginning of the prompt.
• Be make sure the prompt has structure.
• If you want the model to include a detail, tell it that detail.
• Use direct and clear adjective in your prompt.

The Prompt

Write an college-level english-class story in the style of the pulitzer prize winning author and journalist, Michael Lewis. The story is of the viewpoint of a startup founder. The main focus of the article is the importance of launching quickly and continuously.  The story should be in the first person. Use college-level words. Use clear and vivid imagery.  Do not be redundant.  Be concise.

To illustrate the point, tell the story of 2 products the founder built.  

The first product was called Functionary.  It is a marketing software.  Functionary was a failure.  Functionary launched three times once every two month.  Because he was building solo, it took him too long to launch each time.  When he did launch he didn't have enough bandwidth to both grow and improve the product quickly enough to make it a success.  When he did launch he'd talk to users, which taught him the main use-cases for the product.  When he learned the main use-cases for the problem he was solving, he realized even though his product was different than customer.io, the use cases were the same.  It was clear that for the main use-case customer.io solved the problem better than Functionary.

The second product he built with a friend he met back in college at Carnegie Mellon. They launched the product within a week.  They got their first user by the end of the second week. The second product, Farseer, is an GPT powered webpage scrapper, that lets users get contents from webpages by just explaining the content you want from the page.  The second pr approach was much better, because talking to users allows them to incorporate their feedback and iterate to something that people want more quickly. Working with another person meant they could divide the responsibility and iterate quickly enough to learn what customers wanted and iterate fast enough to keep them happy.

Start the story describing a moment the founder experienced right after launching Functionary for the third time. He is sitting at his desk overwhelmed and unclear what he was going to do next, just then he gets a call from his friend.  This call is about an idea he wants to hack-on together.

End the story by explaining how easy it is to make excuses to yourself about why you haven't launched a product. While not the easy way out, the Y-Combinator advice is what I should have followed from the beginning. Do not say, In conclusion.

The Result - GPT-3.5

I sat at my desk, staring blankly at my computer screen, feeling a mix of frustration and exhaustion. Each time I launched Functionary, my marketing software, users weren't interested. I had launched three times, once every two months.  I wanted to go faster, but working solo meant that I didn't have the bandwidth to both grow and improve the product quickly enough to make it a success.

I knew that talking to users was essential to learning the main use-cases for the product. Through these conversations, I realized that although Functionary was different from customer.io, the use cases were the same. It became clear that customer.io solved the problem better than Functionary.

As I sat there feeling defeated, my phone rang. It was a friend from college, someone I had met at Carnegie Mellon. He had an idea he wanted to hack-on together, a new product we could build. We talked about it for a few minutes, and I felt a spark of excitement.

That was the beginning of Farseer, an GPT powered webpage scraper that let users get content from webpages by just explaining the content they wanted from the page. We launched within a week and got our first user by the end of the second. The approach we took with Farseer was much better. Working with another person meant we could divide the responsibility. We spoke with customers, learned what they wanted and iterated quickly enough to keep them happy.

Through my experience with Functionary and Farseer, I learned the importance of launching quickly and continuously. It's easy to make excuses for why you haven't launched a product, but the Y-Combinator advice is what I should have followed from the beginning: launch early and iterate often. While it's not the easy way out, it's the best way to build a successful product.

w/o the How

As a startup founder, I've learned that launching quickly and continuously is crucial for success. I learned this the hard way with my first product, Functionary. Despite launching it three times, it failed because it took too long to launch each time and I didn't have enough bandwidth to both grow and improve the product quickly enough.

However, I didn't let the failure of Functionary stop me. I teamed up with a friend I met in college and launched our second product, Farseer, within a week. We were able to get our first user by the end of the second week. By talking to users and incorporating their feedback, we were able to iterate quickly and create a product that people actually wanted.

It's easy to make excuses for why you haven't launched a product. You might feel like it's not ready yet or you don't have enough resources to do it properly. But the truth is, launching quickly and continuously is the only way to truly test your product and see if there's a market for it.

In conclusion, as a startup founder, it's important to remember that launching quickly and continuously is crucial for success. Don't let fear or excuses hold you back from getting your product out there and testing it with real users. Follow the Y-Combinator advice and launch early and often to increase your chances of success.