Ayoub EL KHAIARI·Nov 4Does a Smaller Model Still Look at the Same Things?I distilled a vision transformer into two small models and compared what they look at. Four questions, one honest result.
Adnan Masood, PhD.·Nov 4When LLMs Edit Other LLMs: Weight-Space Programming and the Conditions for Recursive…The evidence behind GPT-6 Astra’s Qwen weight edits, their relationship to fine-tuning, and the conditions for continual learning and…A response icon1A response icon1
Sreekar Reddy Pindi·Oct 26How We Distilled GPT-4 into a 7B Edge Model for Empathetic AI MentorshipInside Baagupadu’s approach to synthetic conversations, QLoRA fine-tuning, and local deployment for more accessible AI mentorship.A response icon1A response icon1
Devansh pandey·Oct 23The Teacher Was Less Accurate Than the Student. Distillation Still Helped.Same recipe, two datasets, opposite results, and what the teacher’s confidence had to do with it
The Provenance AI Report·Oct 13Why I’m exploring model distillation for NusaThe first article in a series about teaching a small model to handle a specific application’s decisions.
Adnan Masood, PhD.·Oct 5Parameter-Efficient Discriminative Fine-Tuning of Qwen3.5–4BInside a zero-budget LLM fine-tuning lab: memory failures, resumable pipelines, and hard-won lessons from one DGX Spark.
InVirtualCrew.AIbySameer Nigade·Oct 3Knowledge Distillation for Home Price Prediction: A Hands-On Python GuideHow knowledge distillation turned a single decision tree into a much better home-price predictor, and ran 100× faster than the model it…
Mehmet Özgen ÖZDOĞAN·Oct 2Shrinking Big Models for Local AI: A Practical Roadmap to Teacher-Student DistillationAfter all the serving and GPU-sharing work from the last few posts, I kept running into the same underlying tension: the models that are…
Nexoraw·Oct 1Why Smaller AI Models Learn From Bigger OnesThe AI industry spends a lot of time talking about bigger models.
Chayan Mann·Sep 29I Distilled GPT-6 Luna into a 1.5B Model on My MacBook. Here’s What BrokeHow I replaced a big model and a 3,900-token prompt with a 35 MB adapter that routes support tickets in 0.3 seconds.A response icon2A response icon2
