Field notes 006: Becoming really good at using AI
Ask questions, stop and read, keep writing
I’m a software engineer. I’ve noticed that I become a better engineer when I’m writing my learnings down. In this section of my newsletter I post daily learnings as I work to become a fantastic software engineer.
I wrote my way out
Yesterday I had one of my favorite workdays in the last few weeks. I overcame a lot of inertia in the morning to ship like, three net new features. I did it by writing down my questions and breaking through the AI-enabled paralysis I was experiencing the day before.
I have to basically eat the discomfort of asking, “what parts of this view.py file do I have to make changes to? how do I create a new route in this application? how do I create a dropdown in this application? ok, looks like I have to update the data model, how do I do that? when the form is submitted, how is the request populated and where does the request hit?”
Today I actually did this! I really did. Here’s a list of the questions I asked that helped to unblock me:
I kept actually asking those questions and it really helped. AI was helpful once I targeted my questions more specifically. There are some changes that AI is really really bad at making, and they are often really simple ones. Then you just have to roll up your sleeves and get your hands dirty.
AI = Sandevistan
What I’m learning is how to leverage AI really effectively. To build and ship quickly, build and ship quickly.
AI currently does not meet the threshold of job risk for engineers. I basically think it’s nontrivial to skillfully build with Cursor/Composer in production environments with large codebases.
When should you use AI when coding? What’s the right amount of oversight? Some folks (including me, previously) have said you should never use it to write code you couldn’t have written yourself. After seeing some colleagues move incredibly fast in new domains thanks to AI, I don’t believe this anymore.
In unfamiliar domains, AI is like a risky cybernetic implant - if you’re using it to build faster with new tech, you need to be skilled with talking to AIs, with problem-solving, with knowing when to stop and read the docs. Can you learn the domain quickly? Do you understand some shared fundamentals about the tech you’re working with? Can you ask the right questions and ask them a lot? Do you have experience debugging?
Like when my request was failing with Django, a framework I’m new to, I know to debug I can:
look at the POST request
comment out the change I made
try adding a style with the style utility in Chrome devtools
try adding a style as an inline style (ignore tailwind)
If you are skilled in using AI like this, you can build incredibly fast in new domains. If you are not skilled, you will waste your time, energy, and sanity. You will go cyberpsycho.
AI is an excellent servant, terrible master. “Please do this specific task” >> “Please do this complex task.” Probably it will continue to get better at this. Maybe becoming a founder is a better hedge - building companies and products quickly. Who knows. I should write that essay where I work through my fears about AGI.
Call your LLM by their name in new messages, whatever you’ve named it. It’ll summon their identity up again if they’ve forgotten.
Learnings from prod breaking again
Yesterday, prod broke again. This time I was the one with the most knowledge on how to fix. what a rush.
Working on a live platform with real users is wild, we immediately got people demanding refunds and calling us frauds. It was really helpful to just focus on stopping the bleeding, and save the longer-term investigation for later.
This was one of those black magic WTF-why-is-this-happening bugs. Once again, really helpful to ask lots of questions out loud, testing hypotheses and just running with it. And really good to have our new hire Daniel next to me to pair and think out loud.
There is always a way. You always have options. Never kill yourself (figuratively).
Roundabouts > traffic lights
Roundabouts >> traffic lights, thank you to Ben Orenstein for his incredible talk. A simple but powerful operating principle useful is to act and notify your team instead of waiting on them for approval. It helps speed along the little things.
Roundabout: “Hey [designer], I updated the gap to be 20px to match the rest of the site instead of the 10px in the design. Lmk if I should revert.”
Traffic light: “Hey [designer], the gap in the design is 10px but the gap on the rest of the site is 20px. Which should it be? Pls lmk so we can push this out.”
Other team learnings
Take the quick, easy wins to delight your CEO. A feature she anticipated would take real engineering time ended up taking like, 10 minutes to implement because I had already built an extensible system. I made the choice to step away from my current task to show that off, and it was worth it because it took so little time.
The written specs in Figma will diverge from the actual live application, but it’s about adhering to the spirit of the Figma and how it looks, until those discrepancies are fixed (could be never).
I want to be able to understand AI systems and build the coolest shit with them
I’m so impressed by my CEO’s ability to build HDR in such a short time. Git for compute?? Branches?? Holy shit. I want to be good enough to build those. I asked him how he comes up with such novel shit for our company and he said he just reads the papers. “I just read the Deepseek paper. Also, I have like 8 years of experience building these ML systems.”
OK fair. I wanna read more, but I don’t make the time for it. I want to stay close to the infra side, and build the coolest shit with AI. To that end I want to read more papers and books. I want to follow my curiosity. I open up these research papers and I notice that I’m actually interested in reading them, I just click away.
You should really ask your friends how to debug your psycho-emotional-spiritual problems. I talked to Andrew about debugging “why don’t I read more?” and we came up with:
start Soylent (it’s coder food so that coders can spend time reading)
stop scrolling Twitter, start scrolling Arxiv. I installed the blocker Undistracted.
Read about Linux and containers (the zines, and some docs)
Read ML papers that come out, just scroll Arxiv. Yeah inevitably I won’t be able to understand some of it but I want to start to immerse myself. I also want to grok more about what reinforcement learning actually is.
Read our repos
Ask questions
It’s so exciting to live in a world where training ML models is about to become so much more accessible. I want to understand this deeply, because I’m actually interested - and because I want to leverage its full power in what I build.
Reminders to myself
My goal is to become more and more connected to the work that I do, to follow my curiosity and desire for skill, and to work with people I can learn from and help. I want to get closer and closer in connection with myself, and know that this supports powerful work - they’re not in conflict. Be unafraid. Don’t make myself less than. Take care of future Parth.
Start Soylent seems extreme lol