How I'm Using AI

AI is a divisive technology, and it has a lot of flaws. I’m not an AI evangelist, but I don’t think it is a useless technology. I think we need more people engaging in realistic conversations about the technology and about how they can, do, and don’t use it. This post is me trying to be part of the change I’d like to see by writing realistically about how I use and approach AI, including what doesn’t work.

Timeline

As AI is evolving, so is my usage of it. This is my attempt to keep a personal timeline of experimentation, usage, and adoption. The AI space is a fast-moving beast, and I think it interesting to keep tabs on how my usage changes. This post isn’t entirely comprehensive but is fairly representative.

2022

I referenced the output of GPT-3 via OpenAI’s Playground prior to ChatGPT’s release when writing assignments and spent a lot of time experimenting with the technology. I was conscious that GPT-3 was very limited in capabilities but that it was exciting and good for writing generic boilerplate.

I switched to using GPT-3.5 via ChatGPT shortly after its launch. Fairly capable, but still extremely flawed.

I gave Craiyon a test, but its image output was a garbled mess and interesting only as a proof of concept.

2023

I experimented with DALL·E 2 for generating concepts and inspiration. It gave me a vague source of inspiration for some specific concepts that hadn’t been explored before but was overall extremely poor in quality and not particularly useful.

I created Jarg at this time to interact and experiment with models not yet available in ChatGPT for free users, such as GPT-4, though I kept using ChatGPT as my primary interface.

My main usage throughout 2023 remained testing, experimenting, and extraordinarily menial tasks rather than actual usage, though I could see genuine applications of the technology beginning to form and approach viability. I did manage to find bona fide, applicable usage as a really good context-regarding thesaurus and rhyme machine.

2024

In 2024 I began seriously using LLMs for code generation, although under strict supervision. I wasn’t confident enough to permit them to work unsupervised, but I was happy to use them for generating one-off, non-mission-critical tools or for doing menial but verifiable tasks such as writing regular expressions. They do well with languages such as BASH, JavaScript, PHP, and Python, but I find them exceedingly poor for HTML and CSS.

I tried briefly to use LLM editor integrations such as GitHub Copilot but disliked the user experience. I felt I lost granular control over their actions and that they were sloppy in pulling in useful context when attempting to solve problems – which was the main benefit as far as I was concerned. I continued to use web interfaces and copy in any necessary code. I tried LLM line completions but found them slow and inferior to non-AI implementations.

Mid-year, I moved over to Anthropic’s Claude 3.5 Sonnet. Like many others, I found its coding abilities incredibly impressive, though was disappointed by its continuous bias towards more complex technologies when I just wanted it to use HTML/CSS/JS. I remained using ChatGPT as a fallback when I ran out of messages thanks to its strict usage limitations.

I locally hosted several models, but my computer wasn’t capable of running anything reasonable with any degree of speed.

I experimented with DeepSeek V3 and found it capable but inferior to Claude 3.5 Sonnet. I also really disliked the censorship.

I tested Perplexity and found it interesting but prone to hallucinations and therefore completely unfit for use as a general search engine. ChatGPT Search was equally poor as a general search engine, but I found utility in ChatGPT pulling in web content to avoid the limitations of its knowledge cutoff.

Late in the year I began using LLMs for help with my Japanese learning. Mainly for the purpose of quizzing me on words and phrases or to converse with me and offer corrections as necessary. Language is the bread and butter of a Large Language Model, and it did a very good job. I used DeepL on occasion to verify it wasn’t just spouting rubbish, and it wasn’t any of the times I verified.

2025

I switched to Claude 3.7 Sonnet when it released in February, though didn’t find it a particularly impressive upgrade and noted it seemed a tad overzealous – introducing unnecessary complexity and being too verbose. The remaining lack of web search capabilities left it feeling behind the curve.

LLMs continue to be poor in HTML and CSS, with a huge amount of unnecessary complexity and massive accessibility flaws in their implementations. Claude in particular goes way overboard with ARIA.

In March I transitioned over to using Google’s Gemini 2.5 models for everything besides coding, and they proved hugely superior in most regards except for ‘personality’. Their swift response times and fantastic use of web content make them incredibly useful. Shortly after my switch to Gemini, I began informing it (and ChatGPT, when I occasionally use it) who I am, which prompts it to pull in useful context.

I tested ChatGPT’s image generation (gpt-image-1) via ChatGPT and found it very impressive, but I don’t have a need for image generation and found that its images tended to have a certain undesirable property regardless of style and are tinted yellow despite my best attempts to avoid it.

I gave Cursor a go around this time, but it failed in all the same ways GitHub Copilot did and continues to.

In April I began using Gemini’s writing editor Gem for getting feedback on my grammar and flaws. I disagree with many of its suggestions and find it often tries to neuter any character, but it does act like a fresh pair of eyes and points out structural/grammatical/spelling issues.

Towards the conclusion of April, I began experimenting with the freshly released Qwen3. It is the first model I’ve found that runs reasonably locally on my laptop and has output worth using compared to current flagships.

With the introduction of Claude Sonnet 4 in May, I moved from Claude 3.7 Sonnet. Sonnet 4 It is far more capable than previous iterations and doesn’t fall victim to the same needless introduction of unnecessary complexity that 3.7 did. I was also pleased with Opus 4, but haven’t found Sonnet lacking enough to justify paying for it.

Specific Usage

The two tasks I use AI models for the most are writing and coding, but I’m conscious I use them for these tasks differently from most people I see discussing them.

Writing

I love writing, and I don’t want an LLM to do my writing for me. However, I do appreciate it providing feedback like a human reviewer would. Things such as picking up on wordiness or less-than-ideal grammar are fantastic use cases.

I often write alone for long periods and get caught up in my own head, so having something to riff off is beneficial. Like most writers, I sometimes hit a bit of a block, so being able to get an LLM to give me a nudge to push me along is great for my productivity. It is accessible and will provide feedback that, while not necessarily always good, does usually inspire or trigger some thought.

For instance, when struggling to rephrase a clunky sentence, asking an LLM for a few alternative phrasings often sparks the exact direction I need, even if I don’t use any of its suggestions directly. A debug duck, if you will.

Coding

I will happily get AI models to generate large portions of code, but any non-vibe-coded AI-generated code I commit must be at the quality of code I would write myself and is checked along the way. Code being AI-generated is not itself an excuse for it being bad, and vigilant oversight is required.

When I use AI to generate code, I do it separate from my editor. Every experience I have had with in-editor AI has been poor unless I give it complete control, which I dislike doing for obvious reasons. AI is a tool, not a replacement, so leaving it to do things on its own usually goes awry.

Inline suggestions are the most egregious of all the ways AI tries to help me in-editor. It gets directly underfoot and in my way to the extent it can. Suggestions are worse than what can be sourced from my LSP or elsewhere in my buffer and are frequently so wrong they’re distracting.

Instead, I’ll always opt for an external interface where I supply the relevant code and allow it to be handled, and this works extremely smoothly. This is also often more cost-effective given available free plans for flagship models.

It is also incredibly useful for creating small, throwaway tools as part of the development process, as I’ve discussed before.