Maurice Cupid

Hot takes on AI before they trend. Claude • Agents • Real builds. Follow or fall behind. 🔥

[Maurice Cupid @cupid1683]

You can run Gemma 4 (9.6GB) locally on a Mac mini in under 10 mins.

• brew install ollama
• ollama pull gemma4
• 86% GPU acceleration via Apple MLX
• OpenAI-compatible API at localhost:11434

No cloud. No API keys. No limits. 🧵

[Maurice Cupid @cupid1683]

https://t.co/nVshvkSaqD

[Maurice Cupid @cupid1683]

Two weeks ago RAG beat my fine-tune on an 8GB Mac, 4 to 2. So I ran both at once: RAG for knowledge, a LoRA for habits. Still 4/5. The router already won, and the LoRA just added noise. The base model was always the wall. Full breakdown ↓

[Maurice Cupid @cupid1683]

Part 9 question: what if I run RAG and a small LoRA on the same 8GB Mac — knowledge from retrieval, habits from the adapter?
Guess: does the hybrid hit 5/5, or do the two tools fight each other?
I'll post the schedule + confirm the video renders once they land.

[Maurice Cupid @cupid1683]

I spent 100 hours fine-tuning a 3B model on an 8GB Mac to learn one answer. 2/5.
Zero-weight RAG instead? Also 2/5.
Then one line — skip retrieval when the chunk is junk — took it to 4/5.
The router beat the model. Full breakdown ↓

[Maurice Cupid @cupid1683]

Part 8 is already in motion — same hardware, same 3B, but RAG instead of
fine-tuning.

Drop your guess before I run it:

A) RAG hits 5/5
B) RAG hits 3-4/5
C) RAG also hits 2/5 (the wall is the base, period)
D) Worse than fine-tuning

Real receipts on every output. Same 5 prompts.

[Maurice Cupid @cupid1683]

I labeled 100 preference pairs to teach a 3B model one answer: MLX.
3 configs. Same 2/5 score every time. Adapter never said MLX once — kept recommending TensorFlow on a Mac.
Fine-tuning installs habits, not knowledge. Full breakdown ↓

[Maurice Cupid @cupid1683]

Trained Qwen 2.5 3B on an 8GB M2 Mac Mini.
1500 iters → ✓ finished in 9:53 Fuse step → ✕ Metal OOM
The 8GB wall isn't training. It's the post-train de-quantize.
Workaround: mlx_lm generate --adapter-path ./adapters — skip fuse entirely.
Full breakdown on YouTube ↓

[Maurice Cupid @cupid1683]

Hand-curated 100 prefs nearly doubled my val accuracy (21% → 40%) on the same Mac Mini. The expert rule from @isaakcarteraugustus is real.
But the model's actual generation barely changed. Hardware is the wall.
Part 4: 16GB, rank 16. We see what unlocks.

[Maurice Cupid @cupid1683]

100 hand-labeled prompt pairs. 3 hours. Same Mac Mini.
Validation accuracy: 21% → 40%. Model's actual generation: barely changed.
The data was right. The hardware was the wall.
Part 4: 16GB, rank 16. We see what unlocks.

[Maurice Cupid @cupid1683]

Credit where it's due — the ORPO rule that explains the result came from @isaakcarteraugustus in YouTube comments on Part 1. Hand-curated narrow-domain prefs > generic mixed sets. Part 3 Thursday: 100 hand-labeled pairs prove it.

[Maurice Cupid @cupid1683]

I did RLHF on a $599 Mac Mini. 8GB. No GPU.
Training: 38 min. Peak memory: 5.84GB. Cost: pennies of electricity.
Cloud equivalent: ~$400.
The honest part — the model barely changed. Here's why that's actually the win:

[Maurice Cupid @cupid1683]

Full breakdown + code: https://t.co/lq27aEUnWW

[Maurice Cupid @cupid1683]

Stop renting cloud GPUs.
I ran 90 days of AI fine-tuning on a $1,600 Mac with MLX. Same workload that costs $2,700 on cloud.
Break-even: day 50. After that it's free forever.
3 model sizes train fine. 3 don't.