ASAdi Shikinopeningthehood.hashnode.dev·Sep 23 · 10 min readDesigning an Asymmetric SLM Pipeline: From Long-Context Reading to Efficient InferenceAs promised, this is a slightly longer post for those who enjoy going a bit deeper. Since the release of DeepSeek V4.1 Flash, three ideas have stayed with me, mainly because I thought they could have 00
ASAdi Shikinopeningthehood.hashnode.dev·Sep 2 · 11 min readYour Eval Set Is Your Product SpecUp until now, I wrote mostly about how AI models work. We talked about embeddings, MLPs, backpropagation, why raw MLPs were not enough, and transformers. This post is a bit different. It is not about 00
ASAdi Shikinopeningthehood.hashnode.dev·Aug 17 · 10 min readThe Inside of a Model - Part D: TransformersPreviously, we talked about MLPs, how models learn, and why raw MLPs, CNNs, and RNNs eventually led us here. This post is about the architecture that changed modern AI more than any other architecture00
ASAdi Shikinopeningthehood.hashnode.dev·Aug 6 · 13 min readThe Inside of a Model - Part C: Why MLPs Were Not EnoughUp until now, we talked about MLPs, how they process vectors, and how they learn using backpropagation. We saw that a neural network can take an input vector, pass it through layers, calculate a loss,00
ASAdi Shikinopeningthehood.hashnode.dev·Jul 29 · 11 min readThe Inside of a Model - Part B: How a Model LearnsIn this post, we finally get to the question that makes neural networks feel almost magical: how does a model actually “learn”? Small recap from the last post: we talked about MLPs and how they proces00
