I was always excited about the possibilities recent developments in LLM-powered workflows provide. At the same time, I was also wary about the larger implications of these things, from the whole theft problem to wondering “I wonder when this thing is going to become self-aware”.
I was also skeptical about how so many people seemed to jump onto every new AI tool, before fully considering whether it’s a good use of their time or (if paid) money.
My first foray into generative AI was a few years ago, when I trained my own little model on my own books in a virtual ghostwriter experiment in Google Colab. It was a fun half-day project, but the results were unusable - mostly because I didn’t have enough material to give it. In the AI explosion of 2023, I took things slowly, trying out and keeping myself up to date on many things but not committing to anything. I wanted to take time to feel out each possible application in terms of its:
- Real-world usefulness
- Efficiency
- Security
- Privacy
For most of 2023, most of the tools I tried just weren’t good enough to work into my day to day life. But over time, that has changed, and I’ve slowly adopted a few different AI-powered “aids”.
I won’t go into that progression here. GPT 3 to 3.5 to 4 changed things drastically for me. Midjourney v5 changed things, too.
The earliest usable use cases for me included creative assistants, mostly brainstorming aids for my sci-fi stories. Then came art generation for book marketing materials.
Finally, programming aids came last. For a long time, AI tools just did not produce code that I found trustworthy enough to bother with. Of course you always have to cross-check everything it writes anyway, but there has to be some balance of quality expectation there. It took a while for AI to meet that threshold of “This output is useful enough for me to actually benefit from” rather than “I’m going to spend more time checking and editing this than I would’ve if I just wrote this code myself.”
With all that said, here are some tools I finally ended up paying for this year after this slow ramp-up.
What I used and don’t anymore
Midjourney - $60/month
The first AI service I started paying for was Midjourney, at $60 per month (for stealth mode, which was a requirement for me) when v5 came out. I used it for book marketing materials and it was really good for that. Then they introduced inpainting and it got even better.
For many months this was the only AI service I paid for. I no longer do, because that $60/month subscription can instead go to the others listed below. If it wasn’t so expensive just for stealth mode, I’d still be paying for it today.
What I’m using now
GitHub Copilot - $10/month
A month or so ago I tried GitHub Copilot and saw just how much more efficient AI workflows integrated into an IDE can be. I know Copilot had been out for a while, but I felt no urgency to dive straight into it until it saw more usage and testing. I was especially wary of using it due to possible security concerns early on.
So finally I tried it for a personal project. Copilot didn’t always get it right, but it usually gave me a good starting point. It’s also decently good at generating comments and boilerplate (function signatures etc). And being right in the IDE, there’s no need to toggle back and forth between windows all the time anymore. That quick iteration speed makes all the difference. Even if it produces incorrect syntax for a suggestion, it often gets the spirit of what I’m trying to do right.
GPT Plus - $20/month
GPT Plus registrations were closed for a while, but when they opened back up I signed up. This gave me access to both GPT 4 and DALL-E through ChatGPT.
DALL-E is limited compared to Midjourney for my marketing illustration material purposes. It is just not as good. But at the price, it’s good enough. Plus, I think there’s plenty of room for me to learn to improve my prompts here. I miss inpainting a lot. I did get some DALL-E credits to try out that feature in labs, but it was not good and I’m not willing to pay more for it. I think a proper workflow for this would include getting a Photoshop subscription to do manual edits, but I’m not invested enough for that just yet.
GPT 4 is great, especially for brainstorming and coming up with phrasing alternatives. I also tried it out to discuss some software architecture decisions I was thinking about for a personal project. Even though you have to view its suggestions with a critical eye, it can be a great tool to talk through a programming problem or idea with.
Cursor Pro - $20/month
Cursor is Visual Studio Code-esque IDE with AI operations integrated into it. Aside from a GitHub Copilot integration, it also provides a handy “AI” sidebar where you can easily ask it questions about the project (referencing the actual project files as needed). Cursor also has convenient debug and “Ask AI” buttons in various parts of the UI, but in reality I’m pretty sure those are just different entry points to the same chat functionality. At $20/month, you get 500 GPT 4 fast queries per month and unlimited slow queries (much cheaper than paying for OpenAI API credits in my case, but you can also provide your own OpenAI API key and use that).
I’ve used it to review my code, and it’s been useful. But I’ve had more than one time where Cursor very confidently told me why something wouldn’t work as I expected, to the point of making me doubt myself before realizing that it is indeed completely wrong. So again, you still need to know enough to use your own judgement with the output.
Elevenlabs (intermittent) - $22/month
I am also using Elevenlabs intermittently to produce AI-narrated audiobook content. This is not the highest priority for me, but so far Elevenlabs has had the highest-quality most human-like voice synthesis I could find. I also like the ability to generate different voice models for different characters easily.
There are problems with the workflow. They provide a long-form generation tooling, but if I want a high-quality result that sounds as human as possible, with appropriate intonation, pacing, mood, etc., I basically need to generate each line sentence by sentence (often regenerating many times). There is not yet enough granular control over these output properties. But when I have the time to narrate a chapter, it can come out sounding very good. The workflow takes a lot of time and effort. It’s definitely not a case of “just paste your chapter into this tool and it’ll narrate it for you” - at least not if you actually care about the quality of the result.
If I had all the money in the world I’d pay for human narration, but that’s not an option right now and this, so far, is the next best thing.
Conclusion
I generally err on the side of cutting tools rather than adding more of them. We’ll see how long these last and how useful they will continue to be over the next year.