In my last post, I wrote about my complex feelings regarding opencode
and the general AI workflows in how I integrate it into my workflows.
To save you some time, my general conclusion is that it's currently
"not great, not terrible" and can be used for automating certain tedious
tasks as well as being great for code reviews. I generally find it to
be generally useless for building, well, anything that I'm interested
in building.
Over the past few weeks I have had a few really interesting use cases that lead me to feel that much of the focus on AI is backwards.
The focus of the AI tech and thought leaders in the industry is fixated on:
- Replace Software Engineers
- Replace Writers
- Replace Artists
- Replace Data Scientists
- Replace Doctors
But this idea of "How to replace X, Y, Z" is an absolutely terrible mentality. AI's most valuable contribution is not in replacing humanity (it does this poorly).
AI's greatest contribution is in being a way that we can augment outselves.
Learning Assistant
As I mentioned before, I fell into a trap over the last few years of replacing my Japanese Language studies with AI tooling instead of self study and reliance. In part, this was due to the nature of my work putting a heavy focus on immediate efficiency rather than long term benefits. It is also because of how tempting it is.
Humans are always tempted with the path of least resistance, especially an (at the time) untreated ADHD brain where running headlong into a roadblock requires inexorable will to defeat. Leaning on the machine is always a temptation.
However, I find myself coming back to lean on AI in my second attempt at learning. The big difference is that I've fundamentally shifted my mentality when approaching who to integrate AI into my learnings. Instead of approaching things from a "AI do Japanese for me" way, I've shifted it into being a strict, albeit helpful, teaching assistant. As always, it is no replacement for proper study materials, or instructors but they aren't always available at all times and AI can easily fill in that gap. (such was the motivation behind making my Joshu tool which helps me to get corrections to my writing)
This also goes further with other study. While a major copyright concern, it is a fact of the modern age that the best way to learn is learning in the modern age is by just trying something and searching the internet for solutions. As always, the difference in approach fundamentally comes down to whether you are just copy pasting the answer, or trying to learn why the answer fixes it.
This use-case has just been absolutely mindblowingly good at helping me noticeably level up my language skills, but it requires the discipline to never copy paste the output. Read it, distill it, and write it yourself. That's the only way that it contributes to yourself.
Brain Storming Assistant
My second place is how I tend to get lost in my own ideas and have difficulty starting. I recently did an experiment with de-rusting in Java (it's just really popular where I live) and one of the things that I was trying to figure out what just "Where do I start?".
Asking AI "I wanna study java by making a webserver, where should I start" gave me a great starting point where various search results just lead me to more questions.
It doesn't replace me actually doing the work (though again, it's a good learning tool when you don't copy-paste) but it does help me to get unstuck and break out of a writer's block.
This might just be me, but sometimes I need something that can just kind of give me a direction that I'm incapable of deciding on myself. This part is the fundamentally important use of AI in my workflow.
It can be the first place I stop prior to talking to a mentor or subject matter expert in unsticking myself.
AGAIN ingesting the information and considering it myself is the critical part. Copying and pasting is no good if you care about your own knowledge and skills in the "age of AI".
Philosophy AI Augmented Development
One of the things that I feel like I need to preface a lot of this with is that, fundamentally, my personal development and understanding is on me. I need to be at the core of any tools. It doesn't matter if it is Neovim, my cell phone, obsidian (my note taking app), git, or whatever. The tools are methods of augmenting myself and AI is no different.
While some might be fine with allowing the AI to write the code for them, the liability is always on me and my understanding of language/code. This mostly leads me into the scenario of using AI tooling to augment my various workflows (learning, studying, developing, reviewing). The big difference is that I am at the center of the AI workflows, and I cannot reasonably provide it to I need to
Usual Concerns
It still stinks that AI is kinda just, really bad for the planet in a era where sustainability should be a much more important topic. It's a shame that the great AI datacenter race is taking priority over attempting to determine how to better build sustainable infrastructure.
I still find that model vendor locking is a big source of anxiety with relying on these tools. Most of the personal AI frameworks that people build out heavily rely on various proprietary models such as Anthropic's Opus and given how early we find ourselves in the AI adoption and lifecycle, it's hard for me to embrace building more of my life around an AI infrastructure when we aren't past the AI bubble yet.
While I am absolutely sure that AI will play a role in the world after the bubble, there will be winners and losers of the bubble. We have no idea whether Anthropic, Google, Meta, or some other company will come out on top. So using proprietary models will always feel a bit uncomfortable. Though, lacking a spare $100k USD to build my own solution, the only option is to use online models.
Perhaps the best place to start is aiming for the opensource models that I have the possibility of running at home. I cannot afford to build a scaffolding for my life that vanishes with a company declaring bankruptcy