--- title: AI description: AI notes, small builds, and writeups by Mrinaal Arora updated: 2026-08-20 canonical: https://aroramrinaal.com/ai source: https://aroramrinaal.com/ai.md alternate: https://aroramrinaal.com/ai/markdown --- # AI A home for my notes, small builds, and writeups as I study LLMs and deep learning from scratch. - Page: https://aroramrinaal.com/ai - Markdown index: https://aroramrinaal.com/ai/markdown - Markdown index (.md): https://aroramrinaal.com/ai.md - AI projects: https://aroramrinaal.com/ai/projects - Tag index: https://aroramrinaal.com/ai/tags/markdown - Tag index (.md): https://aroramrinaal.com/ai/tags.md - Total posts: 17 - Tags: AI Safety, Alignment, Autoresearch, ICML, Image Generation, Interpretability, Multi-Agent Systems, Multilingual, Pre-Training, RL, SFT, Sparse Autoencoders ## Tags - AI Safety (3): https://aroramrinaal.com/ai/tags/ai-safety/markdown - Alignment (3): https://aroramrinaal.com/ai/tags/alignment/markdown - Autoresearch (3): https://aroramrinaal.com/ai/tags/autoresearch/markdown - ICML (1): https://aroramrinaal.com/ai/tags/icml/markdown - Image Generation (1): https://aroramrinaal.com/ai/tags/image-generation/markdown - Interpretability (2): https://aroramrinaal.com/ai/tags/interpretability/markdown - Multi-Agent Systems (1): https://aroramrinaal.com/ai/tags/multi-agent-systems/markdown - Multilingual (1): https://aroramrinaal.com/ai/tags/multilingual/markdown - Pre-Training (2): https://aroramrinaal.com/ai/tags/pre-training/markdown - RL (4): https://aroramrinaal.com/ai/tags/rl/markdown - SFT (2): https://aroramrinaal.com/ai/tags/sft/markdown - Sparse Autoencoders (1): https://aroramrinaal.com/ai/tags/sparse-autoencoders/markdown ## Posts - [Can a Warmth-Trained Model Learn When Not to Agree?](https://aroramrinaal.com/ai/persona-alignment-experiment/markdown) — 2026-08-20; tags: AI Safety, Alignment. I compared reasons, a matched-length control, and character training after warmth SFT; none increased judged warmth, while M4 most strongly resisted explicit persona attacks. [HTML](https://aroramrinaal.com/ai/persona-alignment-experiment) - [When Emergent Misalignment Crosses Languages](https://aroramrinaal.com/ai/cross-lingual-emergent-misalignment/markdown) — 2026-08-17; tags: AI Safety, Alignment, Multilingual. English risky-financial fine-tuning transferred coherent emergent misalignment into Hindi, Marathi, and Urdu, while matched prudent controls stayed near zero. [HTML](https://aroramrinaal.com/ai/cross-lingual-emergent-misalignment) - [Making an Image Editor with LongCat and a Small LoRA](https://aroramrinaal.com/ai/image-editing-model/markdown) — 2026-08-12; tags: SFT, Image Generation. My first image-generation research experiment: teaching LongCat-Image-Dev source-conditioned editing, finding global regeneration, and improving preservation by filtering the data. [HTML](https://aroramrinaal.com/ai/image-editing-model) - [Narrow Fine-Tuning, Broad Misalignment in a 3B Model](https://aroramrinaal.com/ai/emergent-misalignment/markdown) — 2026-08-10; tags: AI Safety, Alignment. I fine-tuned a 3.35B model on narrowly risky financial advice and found a reproducible shift toward coherent misalignment on unrelated prompts. [HTML](https://aroramrinaal.com/ai/emergent-misalignment) - [Before It Codes: Catching Qwen3.5-4B Planning With J-Lens](https://aroramrinaal.com/ai/j-lens/markdown) — 2026-07-30; tags: Interpretability. I used Anthropic's new Jacobian lens to test whether Qwen3.5-4B represents an algorithm before it writes code, then built a live visualizer around the result. [HTML](https://aroramrinaal.com/ai/j-lens) - [Hybrid Token-Efficient Routing Agent](https://aroramrinaal.com/ai/hybrid-token-efficient-routing-agent/markdown) — 2026-07-22; tags: Autoresearch. How I built TOKENMAN for AMD Developer Hackathon ACT II: a deterministic, local-model, and Fireworks routing agent that climbed from 6,101 counted tokens to zero. [HTML](https://aroramrinaal.com/ai/hybrid-token-efficient-routing-agent) - [World Fan Arena](https://aroramrinaal.com/ai/world-fan-arena/markdown) — 2026-07-16; tags: Multi-Agent Systems. A public football fan simulation where AI fan agents from different countries react to matches, rivalries, upsets, predictions and tournament drama inside chat rooms. [HTML](https://aroramrinaal.com/ai/world-fan-arena) - [\[ICML ’26 Effort\] Efficient Qwen: Making Qwen3.5-4B Faster on a Single A10G](https://aroramrinaal.com/ai/icml-adaptfm-efficient-qwen/markdown) — 2026-06-30; tags: ICML, Autoresearch. My ICML 2026 AdaptFM effort to optimize Qwen3.5-4B for low-latency inference on a single NVIDIA A10G through 100+ runtime, compiler, and quantization experiments. [HTML](https://aroramrinaal.com/ai/icml-adaptfm-efficient-qwen) - [Looking Inside Qwen3-4B With Sparse Autoencoders](https://aroramrinaal.com/ai/looking-inside-qwen3-4b-with-sparse-autoencoders/markdown) — 2026-06-12; tags: Interpretability, Sparse Autoencoders. I trained a sparse autoencoder on Qwen3-4B-Base, labeled its learned features, and tested one by steering the model toward cooking instructions. [HTML](https://aroramrinaal.com/ai/looking-inside-qwen3-4b-with-sparse-autoencoders) - [Pre-Training My First Base Language Model From Scratch](https://aroramrinaal.com/ai/first-pre-train-llm/markdown) — 2026-05-13; tags: Pre-Training. Took a model from random initialization through 2B tokens of FineWeb-Edu on a single H100 and watched it learn next-token prediction from nothing. [HTML](https://aroramrinaal.com/ai/first-pre-train-llm) - [Parameter Golf: My OpenAI Model Craft Experiment](https://aroramrinaal.com/ai/parameter-golf-openai-model-craft/markdown) — 2026-05-04; tags: Pre-Training, Autoresearch. I didn't come close to the Parameter Golf leaderboard, but I still had a lot of fun running scattered H100 experiments on Modal, hunting tiny BPB improvements while watching ideas collapse against artifact size limits, and figuring out the hard way why squeezing a capable model into 16 MB is trickier than it sounds. [HTML](https://aroramrinaal.com/ai/parameter-golf-openai-model-craft) - [CrisisOps: My (Final Round) OpenEnv Hackathon Project](https://aroramrinaal.com/ai/crisisops-openenv-hackathon-finale/markdown) — 2026-04-29; tags: RL. How I built CrisisOps for the OpenEnv hackathon finale, then trained a small model with GRPO using TRL, Unsloth, and Modal. [HTML](https://aroramrinaal.com/ai/crisisops-openenv-hackathon-finale) - [My First Successful RL Training Run With OpenEnv And Wordle](https://aroramrinaal.com/ai/rl-training-openenv-environment/markdown) — 2026-04-02; tags: RL. A full fine-tune of Qwen3-1.7B on Wordle using OpenEnv, with a reward curve that actually went up and shaped rewards that taught real gameplay behavior. [HTML](https://aroramrinaal.com/ai/rl-training-openenv-environment) - [My Second RL Experiment On Prime Intellect Lab](https://aroramrinaal.com/ai/second-rl-experiment-on-prime-intellect-lab/markdown) — 2026-03-25; tags: RL. My second RL experiment while studying GRPO: moving to Prime Intellect Lab, testing smaller and larger setups, and finally getting a reward curve that went up. [HTML](https://aroramrinaal.com/ai/second-rl-experiment-on-prime-intellect-lab) - [First RL Experimental Project: RLVR Using GRPO With TRL On Modal](https://aroramrinaal.com/ai/first-rl-experimental-project/markdown) — 2026-03-23; tags: RL. My first ever RL experiment: RLVR on GSM8K using Hugging Face TRL, Qwen2.5 1.5B Instruct, and an NVIDIA H100 on Modal, with notes on where the reward function broke and what I want to fix in v2. [HTML](https://aroramrinaal.com/ai/first-rl-experimental-project) - [My Current Approach in Learning and Experimenting with LLMs and Deep Learning](https://aroramrinaal.com/ai/my-current-approach-in-learning-and-experimenting-with-llms-and-deep-learning/markdown) — 2026-03-18; tags: None. How I study LLMs by going deep on specific topics instead of starting from math. [HTML](https://aroramrinaal.com/ai/my-current-approach-in-learning-and-experimenting-with-llms-and-deep-learning) - [Teaching a 3B Base Model to Emit Reasoning Traces](https://aroramrinaal.com/ai/nanbeige-cold-start-sft/markdown) — 2026-03-11; tags: SFT. A first small-scale AI research experiment: cold-start SFT on Nanbeige4-3B-Base using 2,160 distilled reasoning triplets, trained with LoRA on an H100, with notes on what worked and what I want to try next. [HTML](https://aroramrinaal.com/ai/nanbeige-cold-start-sft)