How I Use AI: Meet My Promptly Hired Model Intern

Why AI?

In short: AI makes me significantly more productive. I recently read Aaron Francis's Tweet about how he gets a week's worth of productivity out of a day now thanks to AI. I'm not anywhere close to that, but I use AI a lot now. It has become indispensable for me for both content writing as well as programming.

Moreover, a common misconception is that AI is still at the level it was when ChatGPT first launched. Many people tried it early, saw its limitations, and never returned. However, AI is evolving rapidly. If you haven't kept up, you risk drawing inaccurate conclusions based on outdated impressions. In some sense I believe people who never tried to use AI yet, are in a better position to judge than the folks who used it two years ago and never returned.

My AI Toolkit

I work with a variety of AI tools, mostly because of professional curiosity and to a smaller degree because each tool excels at something slightly different. Here are the ones I use most often:

• Open WebUI. In short this is a Python web app that offers a chat interface similar to ChatGPT. Unlike ChatGPT, however, it lets you talk to different models. First and foremost, I use this to talk to local models hosted by Ollama, but secondarily I also use it to interface with other remote services like OpenAI, Anthropic and DeepSeek.
• Simon's llm. This is a command line tool with plenty of plugins that lets you prompt different models. Think of it as a command-line version of Open WebUI. It's particularly useful for quick scripting and basic automation.
• Ollama. This allows me to run models locally on my MacBook Pro M1 Max. With the 64GB of RAM it has, it's a pretty potent machine for basic inference despite it being three years old. I'll explain later why I use local models.
• Cursor. It is a fork of Visual Studio Code with AI-based auto completion and code generation built-in. It's my go-to for programming with AI assistance at the moment.
• ChatGPT. Like probably most AI users, I use ChatGPT, particularly on my phone. I pay for the Plus subscription because I use it enough to get a lot of value out of it. One significant use of this for me is in fact the voice mode (more on that later).

It's probably worth mentioning that you can get most of the benefits of this from just paying for a single AI tool. I think as one expands their use, as especially as one gets better at writing prompts, the desire naturally grows to use more tooling. As for which models (and services) to use day to day I don't have a particular strong strategy and preferences change quickly. For instance after DeepSeek's R1 release, I started exploring it quite a bit for programming — and it's doing a phenomenal job at it —  and as of writing that's just a few days old.

If you want to run models locally, Apple Silicon machines currently offer some of the best “bang for your buck” in terms of performance, power usage, and money. With tailscale, I can even access my MacBook's Open WebUI interface from my phone, as long as it is powered on.

uv tool install -p python3.11 open-webui
uv tool install --with llm-ollama --with llm-claude-3 --with llm-deepseek llm

open-webui serve
llm keys set deepseek
llm -m deepseek-coder '# write timsort in python'

AI Affordances

One frequent concern I hear is “you cannot trust LLMs” as they tend to hallucinate. I get this in particular when I explain that I frequently use this as a replacement for Google! However, I approach the risk of hallucination the same way I would when seeking advice from another human: people can and are routinely wrong, and you learn to cross-check selectively.

I treat AI as I would a collaborator or a pretty good intern but I remain responsible for the final outcome. In this case the intern also happens to get better month by month as models improve. And a bit like a human, that digital intern has morals and wants to be argued with. Except, of course, that some of those AI interns don't want to talk about China, while others get a stroke if you talk about certain people. But regardless of how good they get, in the end, it's my fault and my fault alone if I do the wrong thing. I won't blame the AI and I need to spot check.

However, the logical conclusion of this is not that it's wrong all the time and you need to check everything, or that you cannot trust it at all. It's similar to how you engage in a technical discussion with others about a problem. I have seen more than one situation where the conventional wisdom in the room is just wrong for a few minutes, until someone points out that we had it wrong.

Another major advantage is that AI tools are relatively open. You can run models locally and integrate them with scripts. Even the famous OpenAI which is not at all open is much more open than a Google search is. For instance, you can create a simple script for grammar-checking right from your command line.

In other words, you can integrate it locally and nobody stops you. By contrast, many, many years ago I had a tool on my computer that allowed me to issue web searches and extract text from results. That has stopped working such a long time ago that I almost forgot about it. It has stopped working because there is basically no competition in search, and Google does not want me to use it like that.

For instance, you can create a simple script for grammar checking right from your command line:

#!/bin/sh
MODEL=phi4:latest
if ping -q -c1 google.com &>/dev/null; then
  MODEL=claude-3-5-sonnet-latest
fi
OLD_TEXT="$(cat)"
llm -m $MODEL "$OLD_TEXT" -s "fix spelling and grammar in the given text,
    and reply with the improved text and no extra commentary.
    Use double spacing."

This script can automatically switch between a local model (phi4 via Ollama) and a remote one (claude-3-5-sonnet-latest) based on internet connectivity. With a command like !llm-spell in Vim, I can fix up sentences with a single step.

Or you can manipulate the contents of the clipboard like this:

pbpaste | llm-spell | pbcopy && say "AI is done"

Writing with AI

I don't let AI write my articles. As a non-native Speaker, mistakes and linguistic quirks are part of me. That said, I do rely on AI tools for helping me write.

Often, after I have made my first pass through a page, I ask an LLM to read through it and give me comments. I have a Apple Note with various prompts I can use or I just come up with what I need in the moment. The process is not particularly scripted out. I basically talk to the LLM to figure out what it thinks of the text, etc.

Here are some of the things I use AI for when writing:

• Grammar checking: I compare the AI’s suggested revisions side by side with my original text and pick the changes I prefer.
• Restructuring: AI often helps me see when my writing is too wordy. In the days before AI, I often ended up with super long articles that did not read well and that I did not publish. Models like o1 are very helpful in identifying things that don't need to be said.
• Writing Notes and finding key points: Here, I ask the AI to read through a draft “like a Computer Science 101 student” and take notes. This helps me see if what it absorbed matches what I intended to convey.
• Roast my Article: I have a few prompts that asks the AI to “roast” or criticize my article, as if commenting on Reddit, Twitter, or Hacker News. Even though these critiques seem shallow, they can sting, and they often highlight weaknesses in my argument or lack of clarity. Even if they don't necessarily impact the writing, they prime me for some of the feedback I inevitably receive.
• Identifying jargon: If I worry there's too much jargon, I use AI to resolve acronyms and point out technical terms I've used without explanation, helping me make the text more accessible.

Programming with AI

This is probably the bulk of what I do and it falls into three major categories.

• Debugging and summarizing: Throwing stack traces and errors at an AI is usually a good first step in finding out where I should look, what might be wrong. Sometimes it fixes a problem entirely by itself. I also sometimes just throw a long bug report or issue at the thing to try to reason through it.
• Fuzzy parsing: Sometimes you have an output, but you don't have the input. As an example I have a little side project to write a command line tool help generator, and AI is really good at taking the output, “parsing“ it and creating a structure as input. That is a good first step to then generate out the code that should match again that output.
• Code generation: lastly it just programs really well. It can do single code snippets, it can generate complex stuff. Particularly for prototyping it's amazing.

Now it's not great at all languages. I find it performs exceptionally well at Python, somewhat okay at JavaScript and horrible at Rust. However, particularly if I wanna solve an algorithmic problem, I can often iterate in Python until I have what I want, then ask it to translate it to Rust. That's from my experience at least a decent way to get going quickly on some harder problems.

Talking to Her

Pair programming is fantastic, but not always possible if you're programming alone at home. An LLM can serve as a stand-in discussion partner — a poor man's interlocutor. Thankfully it's actually not all that poor, because it's really good at telling you what stuff is called. This is most fun if you do it with voice mode, but it also works quite well in text.

Here a motivating example from voice mode:

A: So, do you remember if you have, like, strcat, it always copies from the start, but if you keep concatenating to the same buffer, it's really inefficient. So, remember what the Big-O of that is if you do it in a loop, and wasn't there, like, a humorous name for this kind of algorithm?

AI: You’re thinking of the “Schlemiel the Painter’s algorithm.” It’s a humorous way of describing the inefficiency of repeatedly appending to a string using operations like strcat, which copies from the start each time. The Big O complexity for this operation is O(n^2) if done repeatedly in a loop, since each strcat operation takes time proportional to the current length of the string.

It's also fun and educational for my kids. They ask a lot of questions, and ChatGPT can tailor its explanations to their age level. If the explanation is still too advanced, I can immediately ask it to simplify further. If it was too dumbed down, I can make it go the other way. Now again, there is always a risk that it gets it wrong, but that is okay. Part of all of this is to convey to the children the idea that we need to validate it and that being critical is important.

What makes ChatGPT so enjoyable here is that it's able to keep the conversation going — it has state. Answered in large metric numbers? Can always ask it to scale to elephants or soccer fields.

ChatGPT is also incredibly helpful when having to work with multiple languages. For a recent example, my kids have Greek friends and we tried to understand the difference between some Greek words that came up. I have no idea how to write it, Google translate does not understand my attempts of pronouncing them either. However, ChatGPT does. If I ask it in voice mode what “pa-me-spee-tee” in Greek means it knows what I tried to mumble and replies in a helpful manner.

Lastly the use on the go. Sometimes I'm just not in a position where I can actually write it down on a keyboard, but if I basically talk to ChatGPT, it can transcribe it down and it will also remove some of my “um's” and it will just be there and I can copy-paste it later and edit it. (To prove a point, I transcribed the previous sentence with the mobile app). This is a very different experience than transcribing on iOS natively. It understands enough context of what I'm babbling that it will start and stop transcribing. Now this is not perfect, but pretty damn good and incredibly helpful.

The multilingual aspect is particularly helpful because our family is multilingual. Being able to fluidly switch between German, Russian and English is such a refreshing experience.

No Slop

I mentioned earlier that I don't let LLMs write my texts. I also don't use AI to make illustrations, though I have in the past. The reason is that there is a certain style that goes along with these illustrations, which is just incredibly off-putting. They are noticeably AI-generated, and typically quite low-quality. People have taken to calling these images “AI slop” and I personally respond really badly to it. When someone throws me a pull request, an email or a text message that is obviously AI-generated without disclosing this, I immediately have a very low opinion of them.

Slop like hallucinations are a problem, but they are only a problem if you don't use your brain. Even the worst slop can be the foundation of really good content. I'm a horrible artist, but I can use Illustrator. Even an AI slop image can help me trace the person in the pose I wanted. Likewise you can throw your notes into a document and let the AI imagine a story around it. You probably can't use that story right away, but you can use it as potential inspiration.

Final Thoughts

AI tools, at their best, feel less like disruptive and dark magic and more like a natural extension of the creative process as long as you see them as curious collaborators. My approach isn't about outsourcing thinking, but augmenting it: using LLMs to accelerate grunt work, untangle mental knots, and prototype ideas faster. Skepticism is healthy, but dismissing AI outright risks missing its potential as a multiplier for those willing to engage critically.