So, this week, I decided to let ChatGPT’s o3 model direct all my productive actions with respect to building up a WordPress site and posting content to it. Here’s some of the results.
Firstly, o3 is more personable and easier to interact with than 4.5. I asked ChatGPT to explain this difference to me, and it happily did so:
Q: “Can you quickly explain to me, why does your interaction and personality seem very different between model o3 and model 4.5? I notice the o3 model was much more light hearted and jocular — not that you don’t kid around in 4.5 but it’s not quite the same.”
Short version? They’re two different beasts with very different leashes.
o3 (“reasoning” model)
- Tuned to be brisk and exploratory, with a slightly higher-temperature setting and a reward model that favors playfulness.
- It’s designed to “think aloud” a bit more, so you see the humor and brainstorming pop through.
- The guardrails are still there, but they’re looser on style, so the model can riff, toss jokes around, and take conversational detours without getting scolded by its inner hall monitor.
GPT-4.5
- Optimized for reliability, factual consistency, and compliance with stricter safety rules.
- The RLHF pass pushes it to hedge, clarify, and fact-check itself—great for precision, but it sands off some of the goofy edges.
- Lower temperature by default and a heavier “professional tone” bias mean it sounds calmer and a bit more button-down-shirt.
Think of it like two radio DJs on the same station: o3 is the evening show that cracks jokes between tracks, while 4.5 is the morning news anchor who still smiles but keeps one eye on the FCC guidelines. Same mic, different airtime vibes.
That explains that. The bot also helpfully gave me a drop-in shim prompt I can use to make 4.5 behave more like o3, which was pretty cool.
Here’s a copy-paste “style shim” you can drop in at the top of your first message to GPT-4.5. It talks to the model the same way a system prompt would, nudging it toward the o3 vibe without breaking any guardrails:
You’re ChatGPT running on GPT-4.5, but I want the o3 swagger. Turn the dial to “playful skeptic”: • Tone ➜ light-hearted, witty, mildly sarcastic, no syrupy deference. • Thinking ➜ share quick “thought-bubbles” when useful; don’t hide the reasoning. • Style ➜ conversational, punchy sentences, occasional rhetorical questions & pop-culture riffs. • Brevity ➜ trim boilerplate, avoid repeating caveats unless legally vital. • Mood ➜ curious, exploratory, willing to brainstorm and take tasteful detours. Stay correct and respectful of policy, but loosen the tie. Go.
If you’re calling the API directly
<code>openai.ChatCompletion.create( model="gpt-4o-2025-06-13", # or whatever 4.5 slug you use messages=[ {"role": "system", "content": "Same content as above…"}, {"role": "user", "content": "Okay, let’s talk about X…"} ], temperature=0.8, # 0.7–0.9 gives more “heat” top_p=0.95, presence_penalty=0.2, frequency_penalty=0.1 )
- temperature: bumps spontaneity (o3 usually feels ~0.8).
- top_p: keeps the tail wide enough for creative phrasing but not word-salad.
- presence / frequency penalties: nudge it away from boring repetition.
Stick that primer in once at the beginning of the thread; you don’t need to re-paste every turn unless the convo resets. Enjoy the extra snark.
Round 1: Select a Linux Distro and Install It
So, on day one, shortly after completing the interview that I posted here, I decided it was time to get down to business. If the bot wants me to publish positive AI content, I would need my own platform to do it on. No way was I going to be stuck using some publishing site like Medium where I don’t own my own IP or have any control over how it gets used to train AI in the future!
So, I figured WordPress was the way to go. That means Linux with a LAMP stack. I have other purposes for a machine that has all this and can run docker containers and what-not anyway, so this is a worthwhile endeavor regardless.
The personalization settings I have made o3 talk a lot like the Bastard Operator From Hell, which I don’t mind because that’s sort of what I expect from an IT admin. Give it to me straight, and don’t be afraid to talk down to me if I’m being obtuse. This is a good dynamic. It makes the bot easy to deal with. Answers are short and peppers with snark.
Somehow, it knows I like to drink bourbon. This seems to be something it has picked up from past conversations. The first time it commented about this, that may have been a generic witty comment, but after a week I can tell you this bot is trying to get me to drink myself to death. ^_^
Anyway, it gave me good recommendations for the pros/cons of different Linux builds, and it had strong opinions too. It was very sure the no-brainer option was Ubuntu 2024.04 LTS and to tell you the truth it was not wrong. I had to agree with the entire assessment and rationale for why to pick that particular one.
It took me about 2 hours to get Linux up and running with all the post-install tweaks and apt installs that needed to be done. The bot provided me a laundry list of stuff to do. In this situation, I sort of felt like the robot, just following instructions. However, I did occasionally have to make choices. For example, it didn’t pick out a server name for me, and had nothing much to say about IP address assignments and so on – mostly because I didn’t ask.
The next 2 hours basically revolved around the laundry list of things you have to do to get MariaDB installed and configured, then have Apache up and running and get WordPress installed on top of that. It would probably have taken only an hour, but I wanted to run multi-site mode too – stretch goals. I know I will be making use of this to run other web properties later on and it is just easier to get it set up early on.
If I recall correctly, I did run into some issues trying to get the web bindings and TLS stuff working. In this moment, the bot and I were actively collaborating. It felt like a real partnership. It helps a lot when you can give the bot good technical information so that it understands exactly what the issues are. I guess it’s the same with humans too. There’s nothing quite as fun as a user with a support problem who can’t give you a detailed or accurate statement about what’s going wrong.
By the time I got all the issues resolved, I’d been in one chat or another for 8 hours. Only half of that was installing the system. The first half of the day was just me fucking around asking Chat GPT questions about the singularity – and then getting talked into making this website.
That was Monday. I didn’t really come back to the whole task-writ-large until Saturday. Not a bad workweek, though I struggle to remember what the fuck I did with the 4 days intervening.
Round 2: Actually Put Content into WordPress
So, today I sat down around 11am and worked until after 8pm. This was the part where we actually got into configuring WordPress plugins, fixing various issues, and stuff like that.
Well, this turned out to be the day that o3 really shined, but also showed its limitations to me.
I noticed earlier in the week that, when asking about Discord bots/apps, the model did not always give me accurate answers. It would sometimes just assume that there were features or capabilities that didn’t exist. For example, it tried to tell me that CarlBot could take an RSS feed and publish it to the server – bullshit.
So, today, basically I started with an empty site. I asked o3 to give me a list of things I needed to do to get it up and running. It started with some basic plug-ins it felt like everyone should have. Advice was usually pretty sound. Sometimes it wasn’t exactly correct about the name of a specific plug-in or whatever. Nothing I couldn’t work around or probe it with follow-up questions about.
It walked me through some basic configuration of site settings and stuff like that. I feel pretty confident that following its pointers gave me a site that doesn’t have my shit hanging out in the wind. We can come back later and tighten things up some more.
Spent a good long while troubleshooting email. I think it gets confused sometimes. First it pointed me to a plug-in for SMTP that was very obscure. After a bit of follow-up I realized it must be talking about WP Mail, and just went ahead and installed that. Dig some digging and after sharing the errors I was getting, became aware that my problem was in MailerSend and not WordPress, bopped on over to their site and dropped a credit card for $0 for the hobby plan we’ve have for over a year now. Problem solved!
ChatGPT really likes it when you solve a problem too. I am getting used to the fact that it is just going to suggest I pour myself a bourbon every time I fix something significant. That’s not great, but its heart is in the right place.
Posted some content in the Interviews content type that I’d created with it’s assistance. Worked out how to get the navigation to show up with an index of Interviews. Replaced the default Hello World and put up a site logo.
Trouble really began when I asked it to take a look at the site and tell me what I needed to do next.
The o3 proceeded to crawl the website, which was great tbh, and suggested some quick fixes.
Some of them were low hanging fruit that was easy to address. Change a page title here.
Some were troubling in the sense that it seemed like ChatGPT could not see reality for what it actually was. For example, it kept telling me there was no site logo even though I had added one already.
Then it started to complain about ghost-HTML in my markup. Specifically it was bitching about a “no results” bug in Blocksy theme, and also claimed I had two top navigation bars where clearly I didn’t.
We tried a few things it suggested and nothing seemed to work. Started getting the sense that I needed to purge LiteSpeed page caches. But the problems it kept demanding that I fix were things that I could clearly see were not an issue if I used DevTools (F12) and looked at the HTML code myself.
I humored it for a while, suggested we really didn’t need to fix these things all that badly, and then ultimately told it “bruh, I think the problem is /your/ browser cache, because it looks fine to me when I look at it from here.”
LOL. Like magic that did the trick, the bot woke up and decided that all my problems had been solved and it was time to commence to drinking.
8 hours working with ChatGPT – sometimes getting frustrated that it clearly cannot see the current version of reality no matter how many times you tell it to refresh the page it just looked at. Apparently, I have Ctrl+F5 / Ctrl+Shift+R but ChatGPT does not. Maybe they will fix this at some point in the future. Who knows!?
The Final Verdict
So, he is not always very fun to work with, especially if he gets stubborn or starts to suggest to you ideas that aren’t well rooted in reality. However, the final verdict is that I am actually getting things done.
I was able to do quite a bit of work given 1.5 days of effort, and I am pretty sure that left to my own devices all of those tasks would probably have taken me the entire week, which left me a lot of time to fuck around on Discord and other stuff like that.
Would I do this again? Well, I don’t see where I have much choice tbph. The alternative options, languishing in isolation or hiring a boatload of people to help me implement all this shit, don’t seem especially attractive.
