Prateek | AI • Data • Tech

Didascalia

Most people think learning LLMs is about finding the “best course.” It rarely is. It’s more about quietly building the right mental layers over time. These repos felt different when I came across them. Less noise, more signal. — 📚 The reading layer Normcore LLM Reads isn’t trying to impress you. It just sits there with papers, systems thinking, and real deployment lessons. The kind of stuff you usually only notice after something breaks in production. Felt like understanding LLMs instead of just using them. — 🧪 The evaluation layer Awesome-LLM-Judges shifted something subtle. You realize generating answers is easy. Knowing if they’re good, safe, or even correct is the harder problem. Kind of like shipping features without ever measuring if they work. — 🧠 The control layer Prompt Engineering repos look basic at first. But over time, you start noticing how small phrasing changes completely alter outputs. Same model, different behavior. It’s less about prompts, more about learning how the model “listens.” — ⚙️ The systems layer Awesome-LLM-Inference is where things get real. Latency, batching, KV cache… all the parts you ignore until your system slows down and users start feeling it. This is where AI stops being a demo. — Putting these together changed how I think about LLMs. Not as a tool, but as a system with layers you can actually reason about. And once you see that, most “AI learning content” starts feeling incomplete. — 💾 Save this when you want a clearer path beyond surface-level AI 💬 Comment “LLM” and I’ll share more like this 🔁 Follow for more grounded AI engineering notes