Topic
Fine-tuning and training
Fine-tuning, distillation and preference training. What changes inside a model when you teach it more.
Kitchen table
You can teach a robot new tricks. Sometimes it forgets old ones along the way. These pages are about teaching without forgetting.
Whiteboard
Fine-tuning moves weights you cannot fully see. Adapters, distillation and RLHF each change the model in a different way. These entries show what shifts and what breaks.
Boardroom
Training your own model is a choice with lasting costs. It can lift quality or quietly break what worked. These pages show when it pays.
Start with the concepts
-
Fine-tuning & Adapters LoRA and the PEFT family: small trainable deltas over frozen weights, so you tune millions of parameters instead of billions and swap adapters per…
-
RLHF & Preference Train a reward model on pairwise human preferences, then optimise the policy against it – PPO classically, the DPO family when you would rather skip…
-
Gradient Descent Differentiate the loss with respect to every parameter, then step against the gradient.
-
Neural Networks Layers of weighted sums pushed through non-linearities.
From the daily page
-
the reasoning stayed honest and the answer went rogue
21 October 2026 -
the copy came out steadier than the thing it copied
15 October 2026 -
the forgetting was in the words your data never used
12 October 2026