⚖️ 𝐀𝐜𝐜𝐨𝐫𝐝𝐢𝐧𝐠 𝐭𝐨 𝐭𝐡𝐢𝐬 𝐒𝐭𝐚𝐧𝐟𝐨𝐫𝐝 𝐜𝐥𝐚𝐬𝐬, 𝘸𝘦𝘪𝘨𝘩𝘵 𝘥𝘦𝘤𝘢𝘺 𝘥𝘰𝘦𝘴 𝘯𝘰𝘵 𝘩𝘦𝘭𝘱 𝘵𝘰 𝘱𝘳𝘦𝘷𝘦𝘯𝘵 𝘰𝘷𝘦𝘳𝘧𝘪𝘵𝘵𝘪𝘯𝘨. ✨ 𝐀𝐜𝐭𝐮𝐚𝐥𝐥𝐲, 𝐰𝐞𝐢𝐠𝐡𝐭 𝐝𝐞𝐜𝐚𝐲 𝐝𝐨𝐞𝐬 𝐬𝐮𝐩𝐩𝐫𝐞𝐬𝐬 𝐨𝐯𝐞𝐫𝐟𝐢𝐭𝐭𝐢𝐧𝐠. 𝘛𝘩𝘢𝘵 𝘥𝘰𝘦𝘴𝘯’𝘵

When fine-tuning a model, you want to make sure that the layer quality metrics for the fine-tuned model are at least as good or better than the base model. Here, I am comparing the instruction fine-tuned Qwen2.5-14B-Instruct to its base model, Qwen2.5-14B Notice that the base

😂

2. is exactly the problem we are solving with WeightWatcher.ai

One of the hallmarks of complexity theory in statistical mechanics is the amazing fact that seemingly very different mechanistic models can lead to the same long-range critical behavior. This can be confusing to the uninitiated when comparing models that look nothing alike at

Some of Nassim Nicholas Taleb followers have decided to troll me--It would not be the internet without trolls. His minions block me so I will clarify here First, uh, yeah, I don't anything about heavy tails 🤣 jmlr.org/papers/v22/20-… nature.com/articles/s4146… arxiv.org/abs/2507.17912 On

統計力学講義開始：ハミルトニアンで特徴づけられる「構成要素の時間発展の法則（力学）」とエントロピーで特徴づけられる「現象の整理から得られる普遍的体系（熱力学）」をつなぐのが統計力学と位置づけ、「要素の力学」「現象論」「その両者をつなぐ体系」というのが他にも色々あることを紹介。

Gemini with Google Colab is cool

For the Nassim Nicholas Taleb trolls following me, I rederived your guru's "masterful contribution" as an elementary application of 1-D RG theory github.com/charlesmartin1… yawn