#llms

We Don't Code Anymore

Since late 2025, many developers around the world have discovered that the quality of the code generated by AI models has increased significantly. They have started getting comfortable giving instructions to the coding agents & letting them do the actual coding part. Many people I've talked to (friends, colleagues, and others I meet at tech events) have told me the same, that they haven't written code by hand since they've discovered how good these models have become.

But somewhere in that hype, not writing code by hand got mixed up with not looking at code.

Personal Projects

It is true for me when it comes to my personal projects. I don't write code by hand, nor do I look at the code. I focus mainly on two things:

1. UI/UX - because the default designs by the models have that strong AI smell that makes me dismiss the product right away.
2. Cloud setup - I have a lot of experience with AWS, so during the planning stage, I tell the agent which services to use (mostly serverless) and for what purpose.

Production Code

But when I work with production code, I make that distinction explicit. Here's my usual approach:

1. I gather every minute detail of the requirement. I try to fully understand exactly what the requirement is, what the scope of it is, and why it is being done. This is crucial because I want to make sure the agent gets all the context it needs to do the changes. I think, telling the agent why a change is being made will result in better output than just telling it what to do.
2. I make sure I tell it to ask me questions if anything is unclear or ambiguous. The newer models tend to be more proactive and wanting to just get the task done. I find that adding this instruction makes it pause to clarify rather than assume incorrectly.
3. I read its commentary of what it is doing to catch the incorrect assumptions it is making and steer it regularly. The model can still make a few incorrect assumptions despite being instructed to clarify, so I make sure to read the commentary and when I catch one, I send another message to steer it in the right direction.
4. I spend a lot of time questioning the diff. I usually want the cleanest and smallest possible change that works for any requirement. And a lot of times, questioning a change will result in the model realizing that it "overengineered" and edits it to make a more minimal change. I have also been using the /show-me skill more recently to understand larger diffs.

Claude Code has a mascot named Claw'd which features in all the feature announcement posts. I find these animations extremely cute and wondered if it is possible to create such animations for myself.

I showcased before how coding agents can actually be used as general agents to do things outside of coding - like infographic video generations. I also recently subscribed to Google AI Pro plan and have been using the Gemini CLI with Gemini 3.1 Pro to do some frontend design changes for FanMeter.

So, I thought why not test both these agents to see which one can generate the better animation. I launched both of these agents in their respective YOLO modes (--dangerously-skip-permissions for Claude Code and --yolo for Gemini CLI) and gave this prompt:

Create a pixel art animation GIF of a 26 year old guy who spends his week teaching AWS classes (offline), working fulltime (WFH) as a DevOps engineer and is also addicted to AI software development (Claudoholic). High FPS, high definition. Not less than 10 seconds.

Gemini was the first one to generate it and it gave me this. I'd give it 4/10 at best:

Claude took a while and gave me the much better output. An impressive 7/10:

Interestingly, because I use Claude as my daily driver, it knew about how I got the idea of Fan Meter and added the part where I wake up in the middle of the night to build something

OpenAI announced the release of their new coding model GPT-5.3-Codex-Spark today, only a week after the release of GPT-5.3-Codex. They say that it has been designed for real-time coding capable of serving more than 1,000 tokens per second. Real-time coding here means to see the results of your requested changes immediately by getting near-instant responses. It runs on Cerebras for high-speed inference.

When I read 'ultra-fast model', I first thought of Fast mode for Opus 4.6 in Claude Code. But the primary difference is that Fast mode is the same model with different API configuration that prioritizes speed over cost. Codex-Spark is a different model with a drop in quality and capabilities.

Also interesting to note that the reduced latency is not just due to the improved model speed, but also because of improvements made to the harness itself:

"As we trained Codex-Spark, it became apparent that model speed was just part of the equation for real-time collaboration—we also needed to reduce latency across the full request-response pipeline. We implemented end-to-end latency improvements in our harness that will benefit all models [...] Through the introduction of a persistent WebSocket connection and targeted optimizations inside of Responses API, we reduced overhead per client/server roundtrip by 80%, per-token overhead by 30%, and time-to-first-token by 50%. The WebSocket path is enabled for Codex-Spark by default and will become the default for all models soon."

I wonder if all other harnesses (Claude Code, OpenCode, Cursor etc.,) can make similar improvements to reduce latency. I've been vibe coding (or doing agentic engineering) with Claude Code a lot for the last few days and I've had some tasks take as long as 30 minutes.

I am undecided on how I feel about LLMs (especially reasoning models). I have always been careful about my thoughts and decision making. I like to do things most people label as "boring" work, like DYOR (Doing Your Own Research) and RTFM (Reading the Fucking Manual).

My personal experience has been that doing the "boring" work is essential to think clearly. It is what solidifies the concepts & strengthens the fundamentals. Good decision making requires clear thoughts & strong fundamentals.

But given that now LLMs have done the boring work (pretraining) and can also do reasoning, anyone using LLMs is no longer thinking. And because everyone is using LLMs, everyone is basically thinking the same. The lack of diversity in thinking bothers me a lot.

When I look at a PR (pull request) with full of AI generated code, I don't know how to feel about it. Is it frustrating that the PR author has not done the thinking or does it really matter if the code works?

LLM thinking comes at a price, and it can think deeply if you pay more. If you do the "boring" work yourself, you fall behind. Does money matter now more than ever? Food for thought (no pun intended)