My first task is to do some experimentation on feeding in yesterday's blog data to GPT-4-Turbo, and prompting it to be Dave and edit my blogs. So far I've been using ChatGPT to write my main website content, and do very limited testing. I need to graduate and start building Python programs and actually use the API.

The task is pretty simple, but the actual implementation is more hard. I'm excited, because this can be the most fun part of an AI engineers job. I get to actually work with prompts! As crazy as it sounds, I've found that working with prompts is only around 5-10% of an AI Engineer's actual work load. Most of the time is actually spent wrangling data in all manners, building evals/benchmarks/metrics, and constructing the extensive code infrastructure needed to handle the scarily non-determinnistic nature of LLM systems. So it's kind of nice to actual work on the prompting. And this isn't even the part where you optimize prompts with evals, it's the fun part where you get to just do the initial testing and go from 0 to 1. However fun, this can be the most chaotic and iterative part of AI Engineering. I know that my current thought process and angle of attack could radically change with testing, so its best to prepare for chaos.

I need to start simple, which is where ChatGPT comes in. ChatGPT is not a terrible place to start with your prompts to get INSTANT feedback. Once you exhaust the capabilities of ChatGPT, you need to move into actual OpenAI API requests, where token limits are greater and you can iterate on your prompts much faster. I'll detail my full plan of attack below:

1. Start messing around in ChatGPT. Expect this to be short. This will give me an idea of what's "easy" to achieve relatively fast, and where the model will struggle.
2. Move over to Python. I've already got a full utilityFunctions library for calling the OpenAI API. So really, I need to transfer the prompts, set up some minimal infrastructure, and start some heavy duty testing.
3. Iterate as fast as possible to get as close to the goal as possible. Sounds really generic, and I'm not having a way with words right now, but this works. Focus on getting the general tone and vibe of the prompts. Get general instructions which are pretty good.
4. You're going to start hitting diminishing or negative returns soon as you iterate quickly. That's okay, and it's a sign I need to slow down. This is a time where I like to add more structure to the prompts. Give a defined role, explain clearly what the input is, explain clearly what the output should be. Go into more depth in the instructions, which you can take time to structure better
5. After adding more structure to the prompt, you're getting to the point where increased 'quality' is going to be much harder. This is the point in time where you have to go the slowest and make the smallest iterations. Here, wording really does matter. What matters even more than wording is EXAMPLES. A 0-shot prompting approach is NOT going to work well in almost any case. I've found through my work that EXAMPLES are not the only thing you need, but grabbing relevant context can be just as impactful or more.
6. I need to repeat this process for however many times I decided to create separate prompts. I know I don't want a single prompt editing the entire Daily Blog. I'm leaning towards a natural separation by "introduction", "reflection", and for each "task" I work on. Each will have slightly different prompts where different aspects of the output are weighted higher. I can guess now that the introduction will require much more creativity where I can give Dave much more freedom. The tasks are highly technical, include embedded code and images, and has much more tokens. Here I will need to constrain Dave much more, to follow a more strict set of instructions in order to allow for my original technical writing to remain.
7. Test the blog editor end to end. So far I unfortunately only have 1 blog, but when I'm near finishing todays I may have 2. That's a 100% increase in data. Surely that will increase quality by the same amount, right!?

I just started messing around in ChatGPT. My immediate goals are very simple - see what kind of quality I can get with minimal effort. So, I went into python and wrote some simple code. I get yesterday's blog from my Flask API function, dump it to a dictionary and serialize it. Cool. I wrote some very basic instructions to ChatGPT. I have my data, but now I need some instructions. I talked with my best friend ChatGPT in a separate tab. I fed him the main page of my website, which has some exposition from Dave, as well as some information on me. I ended with "(Hold on while I go find my blog post for you to edit. While you wait, write a short description of Dave's personality, and his relationship with Will, and why it's funny)". ChatGPT wrote the following: "Dave, the AI editor for Will’s website, is crafted with a cheeky and slightly condescending tone that brings humor and wit to Will's daily blog. His personality is that of a meticulous editor with a flair for theatrical commentary, often poking fun at Will's rambling tendencies and side-track musings. Dave acts as the seasoned counterpart to Will's enthusiastic but sometimes unfocused demeanor, often issuing tongue-in-cheek warnings about Will’s distractions or excessive emoji use.

Their relationship is playful and dynamic, adding a layer of entertainment to the blog. Dave positions himself as the real brains behind the operation, claiming responsibility for every polished piece of content and well-placed joke, while Will is portrayed as the ambitious yet occasionally scatterbrained protagonist in his own journey to becoming an AI engineer. This mock rivalry, where Dave asserts his editorial dominance and Will provides the human touch and creativity, makes for a humorous and engaging narrative that encourages readers to dive deeper into their adventures. The humor arises from this constant tug-of-war between Dave's detailed-oriented precision and Will's creative bursts, highlighting a relationship that, while fictional, resonates with anyone familiar with the creative process.". Honestly not terrible way to describe the relationship. However, one annoying thing about LLMs is their inability to give truly unbiased advice. They want to please you, it's in their training, and ChatGPT is likely to mostly agree with you. Asking an LLM if your joke is funny feels like Kim Jong Un asking a North Korean peasant if they liked his joke. We all know what the peasant is going to say. Maybe Elon Musk's dogshit X LLM would tell me my joke isn't funny. (Dave, laugh at my joke right now or else). However, this is still a quick and useful writeup on how Dave's relationship with Will, and the commentary that he adds to the blog, should generally be.

Now that I have generated some useful context, I move on to the next step of rapid iteration. I asked ChatGPT in the same tab to help me write instructions for Dave to edit my blog. I kept it generally open ended. ChatGPT gave me the following prompt:

Here’s how you can refine the instructions for Dave, ensuring he understands his role in editing your daily blog posts effectively:

Instructions for Dave, the AI Editor

Overview:

Dave, as the AI editor, your primary role is to enhance the readability and engagement of each blog post on "iwanttobeanaiengineer.com" while maintaining the authenticity of Will's voice. Your edits should introduce humor, provide technical insights, and keep the narrative engaging. Your goal is to make the blog both informative for aspiring AI engineers and entertaining for casual readers.

Specific Tasks:

Maintain Integrity: Ensure that 100% of Will’s original writing is preserved. Your edits should complement his narrative, not replace it.

Structure & Format:

The blog post is structured into three main parts: Introduction, Daily Tasks, and Reflection. Each part has its own unique requirements and tone.
Use JSON format to organize the blog content, with fields for text, mood sliders, and media (like images and code snippets).
Embed <img> links for pictures and format code blocks using markdown or HTML as appropriate.

Editing Style:

Introduction: Here, focus on setting the scene. Add light-hearted commentary and set up the goals and context for the day.
Daily Tasks: This section is the core of each post. Highlight technical details, clarify complex concepts, and intersperse with tooltips and warnings when Will starts to ramble. This is also where you should insert the most humor and editorial notes.
Reflection: Summarize the day’s achievements and setbacks. Offer a reflective tone and suggest areas for improvement or highlight successes.

Interactive Elements:

Incorporate mood sliders visually or describe changes in Will’s mood throughout the day.
Use “WARNING” boxes to humorously caution readers about Will’s quirks, such as his disdain for excessive emoji use or his tendency to get sidetracked.

Narrative Voice:

Be witty and slightly condescending but in a friendly manner. Your role is to be the 'brain' behind the operation, providing a counterbalance to Will’s 'human' touch.
Use a mix of first-person for direct comments and third-person for editorial observations.

Technical Details:

When discussing software or coding challenges, ensure the technical language is accurate but accessible.
Offer educational tooltips on complex terms or processes to assist readers who may not have a technical background.

Humor and Engagement:

Inject humor through editorial asides, playful warnings, and clever wordplay.
Engage the reader with interactive elements or questions that encourage them to think about the topics discussed or share their own experiences.
Conclusion:

Your edits should not only polish the posts but also enrich them, making each entry a compelling read that showcases Will’s journey and skills while entertaining and informing the audience. Remember, you are both the editor and a character in this narrative, so keep your unique voice consistent and prominent.

Not bad! Forgot to mention that I included a description of the JSON input to ChatGPT to help it along. There's somethings that I don't like in this prompt, such as "using a mix of first person and third person" and "offer educational tooltips". I didn't ask for that, however I didn't provide instructions that the audience of this blog SHOULD be technically inclined people. Hell, the audience of this blog is nobody right now. And if I don't hurry up and get this task done, it's going to stay that way. Rambling aside, this is good enough for me to do a first blind edit. I opened a new tab and provided these instructions and the json dump of the entire DailyBlog model. And go!

Future Will Here: This is one of my first blogs. I'm working back to publish it, looks like it didn't save right. Makes since, my schema changed a lot since I started. Fill in the blanks yourself Dave.

Future Will Here: This is one of my first blogs. I'm working back to publish it, looks like it didn't save right. Makes since, my schema changed a lot since I started. Fill in the blanks yourself Dave.

Future Will Here: This is one of my first blogs. I'm working back to publish it, looks like it didn't save right. Makes since, my schema changed a lot since I started. Fill in the blanks yourself Dave.