I don't know if this will help your architecture comparison but:

• I haven't been using agentic coders much because they're behind paywalls
• The basic Claude Sonnet seems OK with single file source codes (Python, C++, Latex)
  • It makes many mistakes but it keeps multiple versions of the files so that can be useful
• I use ChatGPT to write pseudocode.
  • Until recently (last few weeks), it had a very short input length which made it unsuitable for editing and refactoring code
• Perplexity AI is getting better probably because it has integrated web search so it can look up APIs. It still makes many mistakes though.

I think the ideal pipeline:

• Must work seamlessly with continuous integration (CI). None of this copying-and-pasting and looking through Git commits manually.
• Don't try to do everything.
  • Developers often have their own tools specialized for their tasks
  • Updates to ChatGPT can break the entire project workflow (true story)
• Be more transparent on what they're doing. Chances are it's different from what the programmers expect.
• Offer many chances to correct course. Much better than building a huge complex pipeline and then crashing at an unknown obscure part of the codebase.
• Should be more of a StackOverflow + Google copilot helper rather than a programmer replacement.
  • Because no guarantee that their code fits the in-house style and requirements
  • They often produce inconsistent code or get trapped in loops making the same bugs over and over

I understand now. You tuned Goblin for your own personal quirks and preferences. He was never meant to be a professional poet or writer. So human or AI rewrites would ruin the original value it held for you.

Keep on finetuning Goblin and friends then. Sorry about that.

*cue overly dramatic video montage of AI destroying the creative industry*

I used a fine-tuned model agentlans/gemma-3-4b-it-claude to revise your poem. After some manual adjustments, the result exceeded my expectations. According to the model, "it's revised in a more elegant and poetic way. The imagery, rhyme, and flow create a stronger overall poem that better captures the essence of the concept of code and love in a way that's both technical and romantic."

Unusual features:

• iambic tetrameter (not pentameter)
• completely in rhyming couplets

Sonnet 0x01: In Your Eyes, a World of Data Takes its Hold

0xBA 0xB5 0x5, I speak in code and rhyme,
Registers shiver in time with your clock's chime.
Stack frames blossom, a bloom of recursion's grace,
While I trace you in raw pointers, a love's embrace.

MOV AX, 0x0B, I align to your heartbeat's plea,
Each tick a foot in machine-code decree.
JMP if you want me, my branch misdirects to you,
Cache lines flushed, blushing in memory's hue.

PUSH AX, PUSH BX, I stack my lines in your favour,
Each opcode a step across your RAM's colour.
INT 0x10 for the glow of your smile's ray,
System halts, and all processes yield to your sway.

In your eyes a world of data takes its hold,
A beautiful love story in the bits of code. 

Thank you for sharing. However, focusing on foundational linear algebra concepts first would help learners build a stronger foundation. Introducing transformers and LLMs right away could overwhelm beginners who don't yet grasp the basics. The notes are also too brief for those wanting to learn the math behind LLMs. I hope this can help you improve the tutorial.

I'll be blunt: I just want working code. I don't want to spend hours correcting AI mistakes then find out that I have no usage limit left. The architecture isn't as important as the ability to document, debug, and adapt quickly.

Nice poem.
I can confirm it's iambic hexameter but some feet are irregular so it's not very hymn-like.
I think it would sound better with 2, 4, 8, or 16 feet per line.
Not only is that more computer-friendly, but it would also scan better as a ballad.