Ramji

Builder at heart, Director at work, Applied ML by trade

[Ramji @maran_ramg]

Synthesis is the ultimate form of understanding!

[Ramji @maran_ramg]

Resumed my game development after a hiatus but this time with Antigravity and its browser subagent. It’s nothing short of magical to see it use the browser to run the test ms I had, and making code changes to the backend. Never a better time to build!

[Ramji @maran_ramg]

Curiosity is the learning rate for the gradient ascent of life!

[Ramji @maran_ramg]

@leerob More recently there is another emotion that I am having hard time grappling with: Joy. Building good software gives me joy and when that creation process is delegated, the joy is gone with it. Maybe we will recalibrate but for now I like to still build stuff with AI boost.

[Ramji @maran_ramg]

@_mohansolo I loved the IDE so much that I upgraded my pro subscription to ultra. But never have I before but quote limits so fast, which negatively affects the user experience. Hope the increased limits are sufficient

[Ramji @maran_ramg]

AI assisted coding is only pleasurable if you understand what it does, verify what’s generated, and at some capacity, partake in the creation process. Co-creation is more satisfying than cracking the whip on AI to build something that you care about. But does it really matter?

[Ramji @maran_ramg]

After a month’s break, I resumed my game development. Because I used AI to write almost all of the code, I felt disconnect rather reconnect with the project. The process of generation imbues a sense of ownership, which was completely missing as I didn’t “generate” anything other than the idea! I am gonna write some code from scratch myself to reestablish the lost connection. Maybe I will only use AI for the things I can’t do and the things I don’t want to do. What’s the point of using it if there is no joy in creation!?

[Ramji @maran_ramg]

@deedydas Can you elaborate on the last bit of using smaller models as tools for larger models? You mean larger models could use these smaller models to accomplish a task well instead of wasting tokens? Combine instruction following goodness with task specific precision? @deedydas

[Ramji @maran_ramg]

Continued implementation of Phase 1, which is game dynamics, using Codex. It is a somewhat disjointed experience to work in the console, then copy over error logs from browser when something fails. Need to find a better way to do this - Day 12!

[Ramji @maran_ramg]

I wanted to find out how codex would handle this task and to my surprise it did fantastically well. It recognized responses API and custom prompts being a feature of OpenAI. Gonna use codex for few weeks.

[Ramji @maran_ramg]

After finishing up the dialogue generation piece, I wanted to refactor and cleanup the code. Cursor AI, in all its wisdom, didn't recognize the new Responses API and custom prompt I have been using, and decided to delete everything I added yesterday. Thank Linus for git - Day 11!

[Ramji @maran_ramg]

In Phase3 of my game dev, I realized that each character in game could be modelled as an AI Agent, but for simplicity, I left them all as a single LLM call with structured output. All tests are passing now, but I should verify context spill over from one piece to another - Day10!

[Ramji @maran_ramg]

Phase 3 of my game development almost nearing completion. Got the responses api working and generated situational dialogues for characters. Along the way, realized how useful OpenAI playground has become. Lot of nifty features to test the models and apis - Day9!

[Ramji @maran_ramg]

Interestingly, Claude4 in Cursor has a hard time to follow the documentation link. Even though it reads the page and understands how responses api works, it generates code with older completions api. And also chooses older models even though documentation has examples of gpt5

[Ramji @maran_ramg]

Refactored llm interactions in my game to use the responses api. But while running tests I had to refactor all the code to ES6 standards. Claude made all of it work. Interestingly, some of the new code didn’t work but Claude4 cleverly identified those as a separate issue - Day8!

[Ramji @maran_ramg]

Claude implemented the phase3 of my game development, which is situational dialogue generation, but it used the chat completions API. Spent rest of the time reading about Responses API and looks like the code needs a refactoring - Day7!

[Ramji @maran_ramg]

After a week of illness, got back to game development with phase-3 which is integrating LLM with character traits to generate dialogues and actions. Brainstormed the implementation plan with Gemini and then Claude coded it up. Testing tomorrow - Day6!

[Ramji @maran_ramg]

My lack of familiarity with JS was nagging at me. So I got Claude to explain the codebase to me and how my game’s frontend and backend work. It was a pretty good learning for me and I am keeping the code as-is for now. Maybe I will learn JS by the time I finish my game - Day 5!

[Ramji @maran_ramg]

Made good progress with my game by implementing personalities of game characters. Claude4 on Cursor was too eager and implemented a UI for testing. Claude4 chose JS for implementation but I don’t know the language. Does it matter if the app works?! - Day 4

[Ramji @maran_ramg]

Took my project brief & development plan for phase 2 of the game to Cursor. GPT-5 came up with a plan of action but couldn't execute a single step. Switched to Claude-4, same execution plan but spotless execution. Insane that I can just switch models to get results - Day 3!